xref: /aosp_15_r20/external/mesa3d/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /**************************************************************************
2  *
3  * Copyright 2009 VMware, Inc.
4  * All Rights Reserved.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the
8  * "Software"), to deal in the Software without restriction, including
9  * without limitation the rights to use, copy, modify, merge, publish,
10  * distribute, sub license, and/or sell copies of the Software, and to
11  * permit persons to whom the Software is furnished to do so, subject to
12  * the following conditions:
13  *
14  * The above copyright notice and this permission notice (including the
15  * next paragraph) shall be included in all copies or substantial portions
16  * of the Software.
17  *
18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25  *
26  **************************************************************************/
27 
28 /**
29  * @file
30  * Texture sampling -- SoA.
31  *
32  * @author Jose Fonseca <[email protected]>
33  * @author Brian Paul <[email protected]>
34  */
35 
36 #include "pipe/p_defines.h"
37 #include "pipe/p_state.h"
38 #include "pipe/p_shader_tokens.h"
39 #include "util/bitset.h"
40 #include "util/compiler.h"
41 #include "util/u_debug.h"
42 #include "util/u_dump.h"
43 #include "util/u_memory.h"
44 #include "util/u_math.h"
45 #include "util/format/u_format.h"
46 #include "util/u_cpu_detect.h"
47 #include "util/format_rgb9e5.h"
48 #include "lp_bld_debug.h"
49 #include "lp_bld_type.h"
50 #include "lp_bld_const.h"
51 #include "lp_bld_conv.h"
52 #include "lp_bld_arit.h"
53 #include "lp_bld_bitarit.h"
54 #include "lp_bld_logic.h"
55 #include "lp_bld_printf.h"
56 #include "lp_bld_swizzle.h"
57 #include "lp_bld_flow.h"
58 #include "lp_bld_gather.h"
59 #include "lp_bld_format.h"
60 #include "lp_bld_sample.h"
61 #include "lp_bld_sample_aos.h"
62 #include "lp_bld_struct.h"
63 #include "lp_bld_quad.h"
64 #include "lp_bld_pack.h"
65 #include "lp_bld_intr.h"
66 #include "lp_bld_misc.h"
67 #include "lp_bld_jit_types.h"
68 
69 static void
lp_build_gather_resident(struct lp_build_context * bld,struct lp_sampler_dynamic_state * dynamic_state,LLVMTypeRef resources_type,LLVMValueRef resources_ptr,LLVMValueRef offset,LLVMValueRef * out_resident)70 lp_build_gather_resident(struct lp_build_context *bld,
71                          struct lp_sampler_dynamic_state *dynamic_state,
72                          LLVMTypeRef resources_type,
73                          LLVMValueRef resources_ptr,
74                          LLVMValueRef offset,
75                          LLVMValueRef *out_resident)
76 {
77    struct lp_type type = lp_int_type(bld->type);
78 
79    struct gallivm_state *gallivm = bld->gallivm;
80    LLVMBuilderRef builder = gallivm->builder;
81 
82    static_assert(sizeof(BITSET_WORD) == 4, "Unexpected BITSET_WORD size");
83 
84    LLVMValueRef residency =
85       dynamic_state->residency(gallivm, resources_type, resources_ptr, 0, NULL);
86 
87    LLVMValueRef tile_size_log2 =
88       lp_build_const_int_vec(gallivm, type, util_logbase2(64 * 1024));
89    LLVMValueRef tile_index = LLVMBuildLShr(builder, offset, tile_size_log2, "");
90 
91    LLVMValueRef dword_bitsize_log2 =
92       lp_build_const_int_vec(gallivm, type, util_logbase2(32));
93    LLVMValueRef dword_index = LLVMBuildLShr(builder, tile_index, dword_bitsize_log2, "");
94 
95    LLVMValueRef dword_size_log2 =
96       lp_build_const_int_vec(gallivm, type, util_logbase2(4));
97    LLVMValueRef dword_offset = LLVMBuildShl(builder, dword_index, dword_size_log2, "");
98 
99    residency = lp_build_gather(gallivm, type.length, type.width, lp_elem_type(type),
100                                      true, residency, dword_offset, true);
101 
102    LLVMValueRef dword_bit_mask =
103       lp_build_const_int_vec(gallivm, type, 31);
104    LLVMValueRef bit_index = LLVMBuildAnd(builder, tile_index, dword_bit_mask, "");
105    LLVMValueRef bit_mask = LLVMBuildShl(builder, lp_build_one(gallivm, type), bit_index, "");
106 
107    LLVMValueRef resident = LLVMBuildAnd(builder, residency, bit_mask, "");
108    resident = LLVMBuildICmp(builder, LLVMIntNE, resident, lp_build_zero(gallivm, type), "");
109 
110    if (*out_resident)
111       *out_resident = LLVMBuildAnd(builder, *out_resident, resident, "");
112    else
113       *out_resident = resident;
114 }
115 
116 /**
117  * Generate code to fetch a texel from a texture at int coords (x, y, z).
118  * The computation depends on whether the texture is 1D, 2D or 3D.
119  * The result, texel, will be float vectors:
120  *   texel[0] = red values
121  *   texel[1] = green values
122  *   texel[2] = blue values
123  *   texel[3] = alpha values
124  */
125 static void
lp_build_sample_texel_soa(struct lp_build_sample_context * bld,LLVMValueRef width,LLVMValueRef height,LLVMValueRef depth,LLVMValueRef x,LLVMValueRef y,LLVMValueRef z,LLVMValueRef y_stride,LLVMValueRef z_stride,LLVMValueRef data_ptr,LLVMValueRef mipoffsets,LLVMValueRef ilevel,LLVMValueRef texel_out[4])126 lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
127                           LLVMValueRef width,
128                           LLVMValueRef height,
129                           LLVMValueRef depth,
130                           LLVMValueRef x,
131                           LLVMValueRef y,
132                           LLVMValueRef z,
133                           LLVMValueRef y_stride,
134                           LLVMValueRef z_stride,
135                           LLVMValueRef data_ptr,
136                           LLVMValueRef mipoffsets,
137                           LLVMValueRef ilevel,
138                           LLVMValueRef texel_out[4])
139 {
140    const struct lp_static_sampler_state *static_state = bld->static_sampler_state;
141    const unsigned dims = bld->dims;
142    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
143    LLVMBuilderRef builder = bld->gallivm->builder;
144    LLVMValueRef offset;
145    LLVMValueRef i, j;
146    LLVMValueRef use_border = NULL;
147 
148    /* use_border = x < 0 || x >= width || y < 0 || y >= height */
149    if (lp_sampler_wrap_mode_uses_border_color(static_state->wrap_s,
150                                               static_state->min_img_filter,
151                                               static_state->mag_img_filter)) {
152       LLVMValueRef b1, b2;
153       b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
154       b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
155       use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
156    }
157 
158    if (dims >= 2 &&
159        lp_sampler_wrap_mode_uses_border_color(static_state->wrap_t,
160                                               static_state->min_img_filter,
161                                               static_state->mag_img_filter)) {
162       LLVMValueRef b1, b2;
163       b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
164       b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
165       if (use_border) {
166          use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1");
167          use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2");
168       } else {
169          use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
170       }
171    }
172 
173    if (dims == 3 &&
174        lp_sampler_wrap_mode_uses_border_color(static_state->wrap_r,
175                                               static_state->min_img_filter,
176                                               static_state->mag_img_filter)) {
177       LLVMValueRef b1, b2;
178       b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
179       b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
180       if (use_border) {
181          use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1");
182          use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2");
183       } else {
184          use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
185       }
186    }
187 
188    /* convert x,y,z coords to linear offset from start of texture, in bytes */
189    if (bld->static_texture_state->tiled) {
190       lp_build_tiled_sample_offset(&bld->int_coord_bld,
191                                    bld->format_desc->format,
192                                    bld->static_texture_state,
193                                    x, y, z, width, height, z_stride,
194                                    &offset, &i, &j);
195    } else {
196       lp_build_sample_offset(&bld->int_coord_bld,
197                              bld->format_desc,
198                              x, y, z, y_stride, z_stride,
199                              &offset, &i, &j);
200    }
201 
202    if (mipoffsets) {
203       offset = lp_build_add(&bld->int_coord_bld, offset, mipoffsets);
204    }
205 
206    if (use_border) {
207       /* If we can sample the border color, it means that texcoords may
208        * lie outside the bounds of the texture image.  We need to do
209        * something to prevent reading out of bounds and causing a segfault.
210        *
211        * Simply AND the texture coords with !use_border.  This will cause
212        * coords which are out of bounds to become zero.  Zero's guaranteed
213        * to be inside the texture image.
214        */
215       offset = lp_build_andnot(&bld->int_coord_bld, offset, use_border);
216    }
217 
218    if (bld->residency) {
219       LLVMValueRef real_offset = offset;
220 
221       if (!mipoffsets) {
222          mipoffsets = lp_build_get_mip_offsets(bld, ilevel);
223          real_offset = lp_build_add(&bld->int_coord_bld, real_offset, mipoffsets);
224 
225          if (use_border)
226             real_offset = lp_build_andnot(&bld->int_coord_bld, real_offset, use_border);
227       }
228 
229       lp_build_gather_resident(&bld->float_vec_bld, bld->dynamic_state,
230                                bld->resources_type, bld->resources_ptr,
231                                real_offset, &bld->resident);
232    }
233 
234    lp_build_fetch_rgba_soa(bld->gallivm,
235                            bld->format_desc,
236                            bld->texel_type, true,
237                            data_ptr, offset,
238                            i, j,
239                            bld->cache,
240                            texel_out);
241 
242    /*
243     * Note: if we find an app which frequently samples the texture border
244     * we might want to implement a true conditional here to avoid sampling
245     * the texture whenever possible (since that's quite a bit of code).
246     * Ex:
247     *   if (use_border) {
248     *      texel = border_color;
249     *   } else {
250     *      texel = sample_texture(coord);
251     *   }
252     * As it is now, we always sample the texture, then selectively replace
253     * the texel color results with the border color.
254     */
255 
256    if (use_border) {
257       /* select texel color or border color depending on use_border. */
258       const struct util_format_description *format_desc = bld->format_desc;
259       struct lp_type border_type = bld->texel_type;
260       border_type.length = 4;
261       /*
262        * Only replace channels which are actually present. The others should
263        * get optimized away eventually by sampler_view swizzle anyway but it's
264        * easier too.
265        */
266       for (unsigned chan = 0; chan < 4; chan++) {
267          unsigned chan_s;
268          /* reverse-map channel... */
269          if (util_format_has_stencil(format_desc)) {
270             if (chan == 0)
271                chan_s = 0;
272             else
273                break;
274          } else {
275             for (chan_s = 0; chan_s < 4; chan_s++) {
276                if (chan_s == format_desc->swizzle[chan]) {
277                   break;
278                }
279             }
280          }
281          if (chan_s <= 3) {
282             /* use the already clamped color */
283             LLVMValueRef idx = lp_build_const_int32(bld->gallivm, chan);
284             LLVMValueRef border_chan;
285 
286             border_chan = lp_build_extract_broadcast(bld->gallivm,
287                                                      border_type,
288                                                      bld->texel_type,
289                                                      bld->border_color_clamped,
290                                                      idx);
291             texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
292                                               border_chan, texel_out[chan]);
293          }
294       }
295    }
296 }
297 
298 static LLVMValueRef
get_first_level(struct gallivm_state * gallivm,LLVMTypeRef resources_type,LLVMValueRef resources_ptr,unsigned texture_unit,LLVMValueRef texture_unit_offset,const struct lp_static_texture_state * static_state,struct lp_sampler_dynamic_state * dynamic_state)299 get_first_level(struct gallivm_state *gallivm,
300                 LLVMTypeRef resources_type,
301                 LLVMValueRef resources_ptr,
302                 unsigned texture_unit,
303                 LLVMValueRef texture_unit_offset,
304                 const struct lp_static_texture_state *static_state,
305                 struct lp_sampler_dynamic_state *dynamic_state)
306 {
307    if (static_state->level_zero_only)
308       return lp_build_const_int32(gallivm, 0);
309    else {
310       LLVMValueRef first_level;
311 
312       first_level = dynamic_state->first_level(gallivm, resources_type,
313                                                resources_ptr, texture_unit,
314                                                texture_unit_offset);
315       first_level = LLVMBuildZExt(gallivm->builder, first_level,
316                                   LLVMInt32TypeInContext(gallivm->context), "");
317       return first_level;
318    }
319 }
320 
321 
322 static LLVMValueRef
get_last_level(struct gallivm_state * gallivm,LLVMTypeRef resources_type,LLVMValueRef resources_ptr,unsigned texture_unit,LLVMValueRef texture_unit_offset,const struct lp_static_texture_state * static_state,struct lp_sampler_dynamic_state * dynamic_state)323 get_last_level(struct gallivm_state *gallivm,
324                LLVMTypeRef resources_type,
325                LLVMValueRef resources_ptr,
326                unsigned texture_unit,
327                LLVMValueRef texture_unit_offset,
328                const struct lp_static_texture_state *static_state,
329                struct lp_sampler_dynamic_state *dynamic_state)
330 {
331    if (static_state->level_zero_only)
332       return lp_build_const_int32(gallivm, 0);
333    else {
334       LLVMValueRef last_level;
335 
336       last_level = dynamic_state->last_level(gallivm, resources_type,
337                                              resources_ptr, texture_unit,
338                                              texture_unit_offset);
339       last_level = LLVMBuildZExt(gallivm->builder, last_level,
340                                  LLVMInt32TypeInContext(gallivm->context), "");
341       return last_level;
342    }
343 }
344 
345 /**
346  * Helper to compute the mirror function for the PIPE_WRAP_MIRROR_REPEAT mode.
347  * (Note that with pot sizes could do this much more easily post-scale
348  * with some bit arithmetic.)
349  */
350 static LLVMValueRef
lp_build_coord_mirror(struct lp_build_sample_context * bld,LLVMValueRef coord,bool posOnly)351 lp_build_coord_mirror(struct lp_build_sample_context *bld,
352                       LLVMValueRef coord, bool posOnly)
353 {
354    struct lp_build_context *coord_bld = &bld->coord_bld;
355    LLVMValueRef fract;
356    LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
357 
358    /*
359     * We can just use 2*(x - round(0.5*x)) to do all the mirroring,
360     * it all works out. (The result is in range [-1, 1.0], negative if
361     * the coord is in the "odd" section, otherwise positive.)
362     */
363 
364    coord = lp_build_mul(coord_bld, coord, half);
365    fract = lp_build_round(coord_bld, coord);
366    fract = lp_build_sub(coord_bld, coord, fract);
367    coord = lp_build_add(coord_bld, fract, fract);
368 
369    if (posOnly) {
370       /*
371        * Theoretically it's not quite 100% accurate because the spec says
372        * that ultimately a scaled coord of -x.0 should map to int coord
373        * -x + 1 with mirroring, not -x (this does not matter for bilinear
374        * filtering).
375        */
376       coord = lp_build_abs(coord_bld, coord);
377       /* kill off NaNs */
378       /* XXX: not safe without arch rounding, fract can be anything. */
379       coord = lp_build_max_ext(coord_bld, coord, coord_bld->zero,
380                                GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
381    }
382 
383    return coord;
384 }
385 
386 
387 /**
388  * Helper to compute the first coord and the weight for
389  * linear wrap repeat npot textures
390  */
391 void
lp_build_coord_repeat_npot_linear(struct lp_build_sample_context * bld,LLVMValueRef coord_f,LLVMValueRef length_i,LLVMValueRef length_f,LLVMValueRef * coord0_i,LLVMValueRef * weight_f)392 lp_build_coord_repeat_npot_linear(struct lp_build_sample_context *bld,
393                                   LLVMValueRef coord_f,
394                                   LLVMValueRef length_i,
395                                   LLVMValueRef length_f,
396                                   LLVMValueRef *coord0_i,
397                                   LLVMValueRef *weight_f)
398 {
399    struct lp_build_context *coord_bld = &bld->coord_bld;
400    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
401    LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
402    LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length_i,
403                                                 int_coord_bld->one);
404    LLVMValueRef mask;
405    /* wrap with normalized floats is just fract */
406    coord_f = lp_build_fract(coord_bld, coord_f);
407    /* mul by size and subtract 0.5 */
408    coord_f = lp_build_mul(coord_bld, coord_f, length_f);
409    coord_f = lp_build_sub(coord_bld, coord_f, half);
410    /*
411     * we avoided the 0.5/length division before the repeat wrap,
412     * now need to fix up edge cases with selects
413     */
414    /*
415     * Note we do a float (unordered) compare so we can eliminate NaNs.
416     * (Otherwise would need fract_safe above).
417     */
418    mask = lp_build_compare(coord_bld->gallivm, coord_bld->type,
419                            PIPE_FUNC_LESS, coord_f, coord_bld->zero);
420 
421    /* convert to int, compute lerp weight */
422    lp_build_ifloor_fract(coord_bld, coord_f, coord0_i, weight_f);
423    *coord0_i = lp_build_select(int_coord_bld, mask, length_minus_one, *coord0_i);
424 }
425 
426 
427 /**
428  * Build LLVM code for texture wrap mode for linear filtering.
429  * \param x0_out  returns first integer texcoord
430  * \param x1_out  returns second integer texcoord
431  * \param weight_out  returns linear interpolation weight
432  */
433 static void
lp_build_sample_wrap_linear(struct lp_build_sample_context * bld,bool is_gather,LLVMValueRef coord,LLVMValueRef length,LLVMValueRef length_f,LLVMValueRef offset,bool is_pot,unsigned wrap_mode,LLVMValueRef * x0_out,LLVMValueRef * x1_out,LLVMValueRef * weight_out)434 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
435                             bool is_gather,
436                             LLVMValueRef coord,
437                             LLVMValueRef length,
438                             LLVMValueRef length_f,
439                             LLVMValueRef offset,
440                             bool is_pot,
441                             unsigned wrap_mode,
442                             LLVMValueRef *x0_out,
443                             LLVMValueRef *x1_out,
444                             LLVMValueRef *weight_out)
445 {
446    struct lp_build_context *coord_bld = &bld->coord_bld;
447    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
448    LLVMBuilderRef builder = bld->gallivm->builder;
449    LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
450    LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
451    LLVMValueRef coord0, coord1, weight;
452 
453    switch (wrap_mode) {
454    case PIPE_TEX_WRAP_REPEAT:
455       if (is_pot) {
456          /* mul by size and subtract 0.5 */
457          coord = lp_build_mul(coord_bld, coord, length_f);
458          coord = lp_build_sub(coord_bld, coord, half);
459          if (offset) {
460             offset = lp_build_int_to_float(coord_bld, offset);
461             coord = lp_build_add(coord_bld, coord, offset);
462          }
463          /* convert to int, compute lerp weight */
464          lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
465          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
466          /* repeat wrap */
467          coord0 = LLVMBuildAnd(builder, coord0, length_minus_one, "");
468          coord1 = LLVMBuildAnd(builder, coord1, length_minus_one, "");
469       } else {
470          LLVMValueRef mask;
471          if (offset) {
472             offset = lp_build_int_to_float(coord_bld, offset);
473             offset = lp_build_div(coord_bld, offset, length_f);
474             coord = lp_build_add(coord_bld, coord, offset);
475          }
476          lp_build_coord_repeat_npot_linear(bld, coord,
477                                            length, length_f,
478                                            &coord0, &weight);
479          mask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
480                                  PIPE_FUNC_NOTEQUAL, coord0, length_minus_one);
481          coord1 = LLVMBuildAnd(builder,
482                                lp_build_add(int_coord_bld, coord0, int_coord_bld->one),
483                                mask, "");
484       }
485       break;
486 
487    case PIPE_TEX_WRAP_CLAMP:
488       if (bld->static_sampler_state->normalized_coords) {
489          /* scale coord to length */
490          coord = lp_build_mul(coord_bld, coord, length_f);
491       }
492       if (offset) {
493          offset = lp_build_int_to_float(coord_bld, offset);
494          coord = lp_build_add(coord_bld, coord, offset);
495       }
496 
497       /*
498        * clamp to [0, length]
499        *
500        * Unlike some other wrap modes, this should be correct for gather
501        * too. GL_CLAMP explicitly does this clamp on the coord prior to
502        * actual wrapping (which is per sample).
503        */
504       coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f);
505 
506       coord = lp_build_sub(coord_bld, coord, half);
507 
508       /* convert to int, compute lerp weight */
509       lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
510       coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
511       break;
512 
513    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
514       {
515          struct lp_build_context abs_coord_bld = bld->coord_bld;
516          abs_coord_bld.type.sign = false;
517 
518          if (bld->static_sampler_state->normalized_coords) {
519             /* mul by tex size */
520             coord = lp_build_mul(coord_bld, coord, length_f);
521          }
522          if (offset) {
523             offset = lp_build_int_to_float(coord_bld, offset);
524             coord = lp_build_add(coord_bld, coord, offset);
525          }
526 
527          /* clamp to length max */
528          coord = lp_build_min_ext(coord_bld, coord, length_f,
529                                   GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
530          if (!is_gather) {
531             /* subtract 0.5 */
532             coord = lp_build_sub(coord_bld, coord, half);
533             /* clamp to [0, length - 0.5] */
534             coord = lp_build_max(coord_bld, coord, coord_bld->zero);
535             /* convert to int, compute lerp weight */
536             lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
537             coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
538          } else {
539             /*
540              * The non-gather path will end up with coords 0, 1 if coord was
541              * smaller than 0.5 (with corresponding weight 0.0 so it doesn't
542              * really matter what the second coord is). But for gather, we
543              * really need to end up with coords 0, 0.
544              */
545             coord = lp_build_max(coord_bld, coord, coord_bld->zero);
546             coord0 = lp_build_sub(coord_bld, coord, half);
547             coord1 = lp_build_add(coord_bld, coord, half);
548             /* Values range ([-0.5, length_f - 0.5], [0.5, length_f + 0.5] */
549             coord0 = lp_build_itrunc(coord_bld, coord0);
550             coord1 = lp_build_itrunc(coord_bld, coord1);
551             weight = coord_bld->undef;
552          }
553          /* coord1 = min(coord1, length-1) */
554          coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
555          break;
556       }
557 
558    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
559       if (bld->static_sampler_state->normalized_coords) {
560          /* scale coord to length */
561          coord = lp_build_mul(coord_bld, coord, length_f);
562       }
563       if (offset) {
564          offset = lp_build_int_to_float(coord_bld, offset);
565          coord = lp_build_add(coord_bld, coord, offset);
566       }
567       /*
568        * We don't need any clamp. Technically, for very large (pos or neg)
569        * (or infinite) values, clamp against [-length, length] would be
570        * correct, but we don't need to guarantee any specific
571        * result for such coords (the ifloor will be undefined, but for modes
572        * requiring border all resulting coords are safe).
573        */
574       coord = lp_build_sub(coord_bld, coord, half);
575       /* convert to int, compute lerp weight */
576       lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
577       coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
578       break;
579 
580    case PIPE_TEX_WRAP_MIRROR_REPEAT:
581       if (offset) {
582          offset = lp_build_int_to_float(coord_bld, offset);
583          offset = lp_build_div(coord_bld, offset, length_f);
584          coord = lp_build_add(coord_bld, coord, offset);
585       }
586       if (!is_gather) {
587          /* compute mirror function */
588          coord = lp_build_coord_mirror(bld, coord, true);
589 
590          /* scale coord to length */
591          coord = lp_build_mul(coord_bld, coord, length_f);
592          coord = lp_build_sub(coord_bld, coord, half);
593 
594          /* convert to int, compute lerp weight */
595          lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
596          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
597 
598          /* coord0 = max(coord0, 0) */
599          coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
600          /* coord1 = min(coord1, length-1) */
601          coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
602       } else {
603          /*
604           * This is pretty reasonable in the end,  all what the tests care
605           * about is nasty edge cases (scaled coords x.5, so the individual
606           * coords are actually integers, which is REALLY tricky to get right
607           * due to this working differently both for negative numbers as well
608           * as for even/odd cases). But with enough magic it's not too complex
609           * after all.
610           * Maybe should try a bit arithmetic one though for POT textures...
611           */
612          LLVMValueRef isNeg;
613          /*
614           * Wrapping just once still works, even though it means we can
615           * get "wrong" sign due to performing mirror in the middle of the
616           * two coords (because this can only happen very near the odd/even
617           * edges, so both coords will actually end up as 0 or length - 1
618           * in the end).
619           * For GL4 gather with per-sample offsets we'd need to the mirroring
620           * per coord too.
621           */
622          coord = lp_build_coord_mirror(bld, coord, false);
623          coord = lp_build_mul(coord_bld, coord, length_f);
624 
625          /*
626           * NaNs should be safe here, we'll do away with them with
627           * the ones' complement plus min.
628           */
629          coord0 = lp_build_sub(coord_bld, coord, half);
630          coord0 = lp_build_ifloor(coord_bld, coord0);
631          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
632          /* ones complement for neg numbers (mirror(negX) = X - 1)  */
633          isNeg = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS,
634                               coord0, int_coord_bld->zero);
635          coord0 = lp_build_xor(int_coord_bld, coord0, isNeg);
636          isNeg = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS,
637                               coord1, int_coord_bld->zero);
638          coord1 = lp_build_xor(int_coord_bld, coord1, isNeg);
639          coord0 = lp_build_min(int_coord_bld, coord0, length_minus_one);
640          coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
641 
642          weight = coord_bld->undef;
643       }
644       break;
645 
646    case PIPE_TEX_WRAP_MIRROR_CLAMP:
647       if (bld->static_sampler_state->normalized_coords) {
648          /* scale coord to length */
649          coord = lp_build_mul(coord_bld, coord, length_f);
650       }
651       if (offset) {
652          offset = lp_build_int_to_float(coord_bld, offset);
653          coord = lp_build_add(coord_bld, coord, offset);
654       }
655       /*
656        * XXX: probably not correct for gather, albeit I'm not
657        * entirely sure as it's poorly specified. The wrapping looks
658        * correct according to the spec which is against gl 1.2.1,
659        * however negative values will be swapped - gl re-specified
660        * wrapping with newer versions (no more pre-clamp except with
661        * GL_CLAMP).
662        */
663       coord = lp_build_abs(coord_bld, coord);
664 
665       /* clamp to [0, length] */
666       coord = lp_build_min_ext(coord_bld, coord, length_f,
667                                GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
668 
669       coord = lp_build_sub(coord_bld, coord, half);
670 
671       /* convert to int, compute lerp weight */
672       lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
673       coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
674       break;
675 
676    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
677       {
678          struct lp_build_context abs_coord_bld = bld->coord_bld;
679          abs_coord_bld.type.sign = false;
680 
681          if (bld->static_sampler_state->normalized_coords) {
682             /* scale coord to length */
683             coord = lp_build_mul(coord_bld, coord, length_f);
684          }
685          if (offset) {
686             offset = lp_build_int_to_float(coord_bld, offset);
687             coord = lp_build_add(coord_bld, coord, offset);
688          }
689          if (!is_gather) {
690             coord = lp_build_abs(coord_bld, coord);
691 
692             /* clamp to length max */
693             coord = lp_build_min_ext(coord_bld, coord, length_f,
694                                      GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
695             /* subtract 0.5 */
696             coord = lp_build_sub(coord_bld, coord, half);
697             /* clamp to [0, length - 0.5] */
698             coord = lp_build_max(coord_bld, coord, coord_bld->zero);
699 
700             /* convert to int, compute lerp weight */
701             lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
702             coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
703             /* coord1 = min(coord1, length-1) */
704             coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
705          } else {
706             /*
707              * The non-gather path will swap coord0/1 if coord was negative,
708              * which is ok for filtering since the filter weight matches
709              * accordingly. Also, if coord is close to zero, coord0/1 will
710              * be 0 and 1, instead of 0 and 0 (again ok due to filter
711              * weight being 0.0). Both issues need to be fixed for gather.
712              */
713             LLVMValueRef isNeg;
714 
715             /*
716              * Actually wanted to cheat here and use:
717              * coord1 = lp_build_iround(coord_bld, coord);
718              * but it's not good enough for some tests (even piglit
719              * textureGather is set up in a way so the coords area always
720              * .5, that is right at the crossover points).
721              * So do ordinary sub/floor, then do ones' complement
722              * for negative numbers.
723              * (Note can't just do sub|add/abs/itrunc per coord neither -
724              * because the spec demands that mirror(3.0) = 3 but
725              * mirror(-3.0) = 2.)
726              */
727             coord = lp_build_sub(coord_bld, coord, half);
728             coord0 = lp_build_ifloor(coord_bld, coord);
729             coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
730             isNeg = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, coord0,
731                                  int_coord_bld->zero);
732             coord0 = lp_build_xor(int_coord_bld, isNeg, coord0);
733             coord0 = lp_build_min(int_coord_bld, coord0, length_minus_one);
734 
735             isNeg = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, coord1,
736                                  int_coord_bld->zero);
737             coord1 = lp_build_xor(int_coord_bld, isNeg, coord1);
738             coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
739 
740             weight = coord_bld->undef;
741          }
742       }
743       break;
744 
745    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
746       {
747          if (bld->static_sampler_state->normalized_coords) {
748             /* scale coord to length */
749             coord = lp_build_mul(coord_bld, coord, length_f);
750          }
751          if (offset) {
752             offset = lp_build_int_to_float(coord_bld, offset);
753             coord = lp_build_add(coord_bld, coord, offset);
754          }
755          /*
756           * XXX: probably not correct for gather due to swapped
757           * order if coord is negative (same rationale as for
758           * MIRROR_CLAMP).
759           */
760          coord = lp_build_abs(coord_bld, coord);
761 
762          /*
763           * We don't need any clamp. Technically, for very large
764           * (or infinite) values, clamp against length would be
765           * correct, but we don't need to guarantee any specific
766           * result for such coords (the ifloor will be undefined, but
767           * for modes requiring border all resulting coords are safe).
768           */
769          coord = lp_build_sub(coord_bld, coord, half);
770 
771          /* convert to int, compute lerp weight */
772          lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
773          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
774       }
775       break;
776 
777    default:
778       assert(0);
779       coord0 = NULL;
780       coord1 = NULL;
781       weight = NULL;
782    }
783 
784    *x0_out = coord0;
785    *x1_out = coord1;
786    *weight_out = weight;
787 }
788 
789 
790 /**
791  * Build LLVM code for texture wrap mode for nearest filtering.
792  * \param coord  the incoming texcoord (nominally in [0,1])
793  * \param length  the texture size along one dimension, as int vector
794  * \param length_f  the texture size along one dimension, as float vector
795  * \param offset  texel offset along one dimension (as int vector)
796  * \param is_pot  if TRUE, length is a power of two
797  * \param wrap_mode  one of PIPE_TEX_WRAP_x
798  */
799 static LLVMValueRef
lp_build_sample_wrap_nearest(struct lp_build_sample_context * bld,LLVMValueRef coord,LLVMValueRef length,LLVMValueRef length_f,LLVMValueRef offset,bool is_pot,unsigned wrap_mode)800 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
801                              LLVMValueRef coord,
802                              LLVMValueRef length,
803                              LLVMValueRef length_f,
804                              LLVMValueRef offset,
805                              bool is_pot,
806                              unsigned wrap_mode)
807 {
808    struct lp_build_context *coord_bld = &bld->coord_bld;
809    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
810    LLVMBuilderRef builder = bld->gallivm->builder;
811    LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
812    LLVMValueRef icoord;
813 
814    switch (wrap_mode) {
815    case PIPE_TEX_WRAP_REPEAT:
816       if (is_pot) {
817          coord = lp_build_mul(coord_bld, coord, length_f);
818          icoord = lp_build_ifloor(coord_bld, coord);
819          if (offset) {
820             icoord = lp_build_add(int_coord_bld, icoord, offset);
821          }
822          icoord = LLVMBuildAnd(builder, icoord, length_minus_one, "");
823       } else {
824           if (offset) {
825              offset = lp_build_int_to_float(coord_bld, offset);
826              offset = lp_build_div(coord_bld, offset, length_f);
827              coord = lp_build_add(coord_bld, coord, offset);
828           }
829           /* take fraction, unnormalize */
830           coord = lp_build_fract_safe(coord_bld, coord);
831           coord = lp_build_mul(coord_bld, coord, length_f);
832           icoord = lp_build_itrunc(coord_bld, coord);
833       }
834       break;
835 
836    case PIPE_TEX_WRAP_CLAMP:
837    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
838       if (bld->static_sampler_state->normalized_coords) {
839          /* scale coord to length */
840          coord = lp_build_mul(coord_bld, coord, length_f);
841       }
842 
843       if (offset) {
844          offset = lp_build_int_to_float(coord_bld, offset);
845          coord = lp_build_add(coord_bld, coord, offset);
846       }
847       /* floor */
848       /* use itrunc instead since we clamp to 0 anyway */
849       icoord = lp_build_itrunc(coord_bld, coord);
850 
851       /* clamp to [0, length - 1]. */
852       icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
853                               length_minus_one);
854       break;
855 
856    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
857       if (bld->static_sampler_state->normalized_coords) {
858          /* scale coord to length */
859          coord = lp_build_mul(coord_bld, coord, length_f);
860       }
861       /* no clamp necessary, border masking will handle this */
862       icoord = lp_build_ifloor(coord_bld, coord);
863       if (offset) {
864          icoord = lp_build_add(int_coord_bld, icoord, offset);
865       }
866       break;
867 
868    case PIPE_TEX_WRAP_MIRROR_REPEAT:
869       if (offset) {
870          offset = lp_build_int_to_float(coord_bld, offset);
871          offset = lp_build_div(coord_bld, offset, length_f);
872          coord = lp_build_add(coord_bld, coord, offset);
873       }
874       /* compute mirror function */
875       coord = lp_build_coord_mirror(bld, coord, true);
876 
877       /* scale coord to length */
878       assert(bld->static_sampler_state->normalized_coords);
879       coord = lp_build_mul(coord_bld, coord, length_f);
880 
881       /* itrunc == ifloor here */
882       icoord = lp_build_itrunc(coord_bld, coord);
883 
884       /* clamp to [0, length - 1] */
885       icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
886       break;
887 
888    case PIPE_TEX_WRAP_MIRROR_CLAMP:
889    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
890       if (bld->static_sampler_state->normalized_coords) {
891          /* scale coord to length */
892          coord = lp_build_mul(coord_bld, coord, length_f);
893       }
894       if (offset) {
895          offset = lp_build_int_to_float(coord_bld, offset);
896          coord = lp_build_add(coord_bld, coord, offset);
897       }
898       coord = lp_build_abs(coord_bld, coord);
899 
900       /* itrunc == ifloor here */
901       icoord = lp_build_itrunc(coord_bld, coord);
902       /*
903        * Use unsigned min due to possible undef values (NaNs, overflow)
904        */
905       {
906          struct lp_build_context abs_coord_bld = *int_coord_bld;
907          abs_coord_bld.type.sign = false;
908          /* clamp to [0, length - 1] */
909          icoord = lp_build_min(&abs_coord_bld, icoord, length_minus_one);
910       }
911       break;
912 
913    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
914       if (bld->static_sampler_state->normalized_coords) {
915          /* scale coord to length */
916          coord = lp_build_mul(coord_bld, coord, length_f);
917       }
918       if (offset) {
919          offset = lp_build_int_to_float(coord_bld, offset);
920          coord = lp_build_add(coord_bld, coord, offset);
921       }
922       coord = lp_build_abs(coord_bld, coord);
923 
924       /* itrunc == ifloor here */
925       icoord = lp_build_itrunc(coord_bld, coord);
926       break;
927 
928    default:
929       assert(0);
930       icoord = NULL;
931    }
932 
933    return icoord;
934 }
935 
936 
937 /**
938  * Do shadow test/comparison.
939  * \param p shadow ref value
940  * \param texel  the texel to compare against
941  */
942 static LLVMValueRef
lp_build_sample_comparefunc(struct lp_build_sample_context * bld,LLVMValueRef p,LLVMValueRef texel)943 lp_build_sample_comparefunc(struct lp_build_sample_context *bld,
944                             LLVMValueRef p,
945                             LLVMValueRef texel)
946 {
947    struct lp_build_context *texel_bld = &bld->texel_bld;
948    LLVMValueRef res;
949 
950    if (0) {
951       //lp_build_print_value(bld->gallivm, "shadow cmp coord", p);
952       lp_build_print_value(bld->gallivm, "shadow cmp texel", texel);
953    }
954 
955    /* result = (p FUNC texel) ? 1 : 0 */
956    /*
957     * honor d3d10 floating point rules here, which state that comparisons
958     * are ordered except NOT_EQUAL which is unordered.
959     */
960    if (bld->static_sampler_state->compare_func != PIPE_FUNC_NOTEQUAL) {
961       res = lp_build_cmp_ordered(texel_bld,
962                                  bld->static_sampler_state->compare_func,
963                                  p, texel);
964    } else {
965       res = lp_build_cmp(texel_bld, bld->static_sampler_state->compare_func,
966                          p, texel);
967    }
968    return res;
969 }
970 
971 
972 /**
973  * Generate code to sample a mipmap level with nearest filtering.
974  * If sampling a cube texture, r = cube face in [0,5].
975  */
976 static void
lp_build_sample_image_nearest(struct lp_build_sample_context * bld,LLVMValueRef size,LLVMValueRef row_stride_vec,LLVMValueRef img_stride_vec,LLVMValueRef data_ptr,LLVMValueRef mipoffsets,LLVMValueRef ilevel,const LLVMValueRef * coords,const LLVMValueRef * offsets,LLVMValueRef colors_out[4])977 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
978                               LLVMValueRef size,
979                               LLVMValueRef row_stride_vec,
980                               LLVMValueRef img_stride_vec,
981                               LLVMValueRef data_ptr,
982                               LLVMValueRef mipoffsets,
983                               LLVMValueRef ilevel,
984                               const LLVMValueRef *coords,
985                               const LLVMValueRef *offsets,
986                               LLVMValueRef colors_out[4])
987 {
988    const unsigned dims = bld->dims;
989    LLVMValueRef width_vec;
990    LLVMValueRef height_vec;
991    LLVMValueRef depth_vec;
992    LLVMValueRef flt_size;
993    LLVMValueRef flt_width_vec;
994    LLVMValueRef flt_height_vec;
995    LLVMValueRef flt_depth_vec;
996    LLVMValueRef x, y = NULL, z = NULL;
997 
998    lp_build_extract_image_sizes(bld,
999                                 &bld->int_size_bld,
1000                                 bld->int_coord_type,
1001                                 size,
1002                                 &width_vec, &height_vec, &depth_vec);
1003 
1004    flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
1005 
1006    lp_build_extract_image_sizes(bld,
1007                                 &bld->float_size_bld,
1008                                 bld->coord_type,
1009                                 flt_size,
1010                                 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
1011 
1012    /*
1013     * Compute integer texcoords.
1014     */
1015    x = lp_build_sample_wrap_nearest(bld, coords[0], width_vec,
1016                                     flt_width_vec, offsets[0],
1017                                     bld->static_texture_state->pot_width,
1018                                     bld->static_sampler_state->wrap_s);
1019    lp_build_name(x, "tex.x.wrapped");
1020 
1021    if (dims >= 2) {
1022       y = lp_build_sample_wrap_nearest(bld, coords[1], height_vec,
1023                                        flt_height_vec, offsets[1],
1024                                        bld->static_texture_state->pot_height,
1025                                        bld->static_sampler_state->wrap_t);
1026       lp_build_name(y, "tex.y.wrapped");
1027 
1028       if (dims == 3) {
1029          z = lp_build_sample_wrap_nearest(bld, coords[2], depth_vec,
1030                                           flt_depth_vec, offsets[2],
1031                                           bld->static_texture_state->pot_depth,
1032                                           bld->static_sampler_state->wrap_r);
1033          lp_build_name(z, "tex.z.wrapped");
1034       }
1035    }
1036    if (has_layer_coord(bld->static_texture_state->target)) {
1037       if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) {
1038          /* add cube layer to face */
1039          z = lp_build_add(&bld->int_coord_bld, coords[2], coords[3]);
1040       } else {
1041          z = coords[2];
1042       }
1043       lp_build_name(z, "tex.z.layer");
1044    }
1045 
1046    /*
1047     * Get texture colors.
1048     */
1049    lp_build_sample_texel_soa(bld,
1050                              width_vec, height_vec, depth_vec,
1051                              x, y, z,
1052                              row_stride_vec, img_stride_vec,
1053                              data_ptr, mipoffsets, ilevel, colors_out);
1054 
1055    if (bld->static_sampler_state->compare_mode != PIPE_TEX_COMPARE_NONE) {
1056       LLVMValueRef cmpval;
1057       cmpval = lp_build_sample_comparefunc(bld, coords[4], colors_out[0]);
1058       /* this is really just a AND 1.0, cmpval but llvm is clever enough */
1059       colors_out[0] = lp_build_select(&bld->texel_bld, cmpval,
1060                                       bld->texel_bld.one, bld->texel_bld.zero);
1061       colors_out[1] = colors_out[2] = colors_out[3] = colors_out[0];
1062    }
1063 
1064 }
1065 
1066 
1067 /**
1068  * Like a lerp, but inputs are 0/~0 masks, so can simplify slightly.
1069  */
1070 static LLVMValueRef
lp_build_masklerp(struct lp_build_context * bld,LLVMValueRef weight,LLVMValueRef mask0,LLVMValueRef mask1)1071 lp_build_masklerp(struct lp_build_context *bld,
1072                  LLVMValueRef weight,
1073                  LLVMValueRef mask0,
1074                  LLVMValueRef mask1)
1075 {
1076    struct gallivm_state *gallivm = bld->gallivm;
1077    LLVMBuilderRef builder = gallivm->builder;
1078    LLVMValueRef weight2;
1079 
1080    weight2 = lp_build_sub(bld, bld->one, weight);
1081    weight = LLVMBuildBitCast(builder, weight,
1082                               lp_build_int_vec_type(gallivm, bld->type), "");
1083    weight2 = LLVMBuildBitCast(builder, weight2,
1084                               lp_build_int_vec_type(gallivm, bld->type), "");
1085    weight = LLVMBuildAnd(builder, weight, mask1, "");
1086    weight2 = LLVMBuildAnd(builder, weight2, mask0, "");
1087    weight = LLVMBuildBitCast(builder, weight, bld->vec_type, "");
1088    weight2 = LLVMBuildBitCast(builder, weight2, bld->vec_type, "");
1089    return lp_build_add(bld, weight, weight2);
1090 }
1091 
1092 /**
1093  * Like a 2d lerp, but inputs are 0/~0 masks, so can simplify slightly.
1094  */
1095 static LLVMValueRef
lp_build_masklerp2d(struct lp_build_context * bld,LLVMValueRef weight0,LLVMValueRef weight1,LLVMValueRef mask00,LLVMValueRef mask01,LLVMValueRef mask10,LLVMValueRef mask11)1096 lp_build_masklerp2d(struct lp_build_context *bld,
1097                     LLVMValueRef weight0,
1098                     LLVMValueRef weight1,
1099                     LLVMValueRef mask00,
1100                     LLVMValueRef mask01,
1101                     LLVMValueRef mask10,
1102                     LLVMValueRef mask11)
1103 {
1104    LLVMValueRef val0 = lp_build_masklerp(bld, weight0, mask00, mask01);
1105    LLVMValueRef val1 = lp_build_masklerp(bld, weight0, mask10, mask11);
1106    return lp_build_lerp(bld, weight1, val0, val1, 0);
1107 }
1108 
1109 /*
1110  * this is a bit excessive code for something OpenGL just recommends
1111  * but does not require.
1112  */
1113 #define ACCURATE_CUBE_CORNERS 1
1114 
1115 /**
1116  * Generate code to sample a mipmap level with linear filtering.
1117  * If sampling a cube texture, r = cube face in [0,5].
1118  * If linear_mask is present, only pixels having their mask set
1119  * will receive linear filtering, the rest will use nearest.
1120  */
1121 static void
lp_build_sample_image_linear(struct lp_build_sample_context * bld,bool is_gather,LLVMValueRef size,LLVMValueRef linear_mask,LLVMValueRef row_stride_vec,LLVMValueRef img_stride_vec,LLVMValueRef data_ptr,LLVMValueRef mipoffsets,LLVMValueRef ilevel,const LLVMValueRef * coords,const LLVMValueRef * offsets,LLVMValueRef colors_out[4])1122 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
1123                              bool is_gather,
1124                              LLVMValueRef size,
1125                              LLVMValueRef linear_mask,
1126                              LLVMValueRef row_stride_vec,
1127                              LLVMValueRef img_stride_vec,
1128                              LLVMValueRef data_ptr,
1129                              LLVMValueRef mipoffsets,
1130                              LLVMValueRef ilevel,
1131                              const LLVMValueRef *coords,
1132                              const LLVMValueRef *offsets,
1133                              LLVMValueRef colors_out[4])
1134 {
1135    LLVMBuilderRef builder = bld->gallivm->builder;
1136    struct lp_build_context *ivec_bld = &bld->int_coord_bld;
1137    struct lp_build_context *coord_bld = &bld->coord_bld;
1138    struct lp_build_context *texel_bld = &bld->texel_bld;
1139    const unsigned dims = bld->dims;
1140    LLVMValueRef width_vec;
1141    LLVMValueRef height_vec;
1142    LLVMValueRef depth_vec;
1143    LLVMValueRef flt_size;
1144    LLVMValueRef flt_width_vec;
1145    LLVMValueRef flt_height_vec;
1146    LLVMValueRef flt_depth_vec;
1147    LLVMValueRef fall_off[4] = { 0 }, have_corners = NULL;
1148    LLVMValueRef z1 = NULL;
1149    LLVMValueRef z00 = NULL, z01 = NULL, z10 = NULL, z11 = NULL;
1150    LLVMValueRef x00 = NULL, x01 = NULL, x10 = NULL, x11 = NULL;
1151    LLVMValueRef y00 = NULL, y01 = NULL, y10 = NULL, y11 = NULL;
1152    LLVMValueRef s_fpart, t_fpart = NULL, r_fpart = NULL;
1153    LLVMValueRef xs[4], ys[4], zs[4];
1154    LLVMValueRef neighbors[2][2][4];
1155    bool seamless_cube_filter, accurate_cube_corners;
1156    unsigned chan_swiz = bld->static_texture_state->swizzle_r;
1157 
1158    if (is_gather) {
1159       switch (bld->gather_comp) {
1160       case 0: chan_swiz = bld->static_texture_state->swizzle_r; break;
1161       case 1: chan_swiz = bld->static_texture_state->swizzle_g; break;
1162       case 2: chan_swiz = bld->static_texture_state->swizzle_b; break;
1163       case 3: chan_swiz = bld->static_texture_state->swizzle_a; break;
1164       default:
1165          break;
1166       }
1167    }
1168 
1169    seamless_cube_filter = (bld->static_texture_state->target == PIPE_TEXTURE_CUBE ||
1170                            bld->static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) &&
1171                           bld->static_sampler_state->seamless_cube_map;
1172 
1173    /*
1174     * Disable accurate cube corners for integer textures, which should only
1175     * get here in the gather path.
1176     */
1177    accurate_cube_corners = ACCURATE_CUBE_CORNERS && seamless_cube_filter &&
1178      !util_format_is_pure_integer(bld->static_texture_state->format);
1179 
1180    lp_build_extract_image_sizes(bld,
1181                                 &bld->int_size_bld,
1182                                 bld->int_coord_type,
1183                                 size,
1184                                 &width_vec, &height_vec, &depth_vec);
1185 
1186    flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
1187 
1188    lp_build_extract_image_sizes(bld,
1189                                 &bld->float_size_bld,
1190                                 bld->coord_type,
1191                                 flt_size,
1192                                 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
1193 
1194    LLVMTypeRef int1t = LLVMInt1TypeInContext(bld->gallivm->context);
1195 
1196    /*
1197     * Compute integer texcoords.
1198     */
1199 
1200    if (!seamless_cube_filter) {
1201       lp_build_sample_wrap_linear(bld, is_gather, coords[0], width_vec,
1202                                   flt_width_vec, offsets[0],
1203                                   bld->static_texture_state->pot_width,
1204                                   bld->static_sampler_state->wrap_s,
1205                                   &x00, &x01, &s_fpart);
1206       lp_build_name(x00, "tex.x0.wrapped");
1207       lp_build_name(x01, "tex.x1.wrapped");
1208       x10 = x00;
1209       x11 = x01;
1210 
1211       if (dims >= 2) {
1212          lp_build_sample_wrap_linear(bld, is_gather, coords[1], height_vec,
1213                                      flt_height_vec, offsets[1],
1214                                      bld->static_texture_state->pot_height,
1215                                      bld->static_sampler_state->wrap_t,
1216                                      &y00, &y10, &t_fpart);
1217          lp_build_name(y00, "tex.y0.wrapped");
1218          lp_build_name(y10, "tex.y1.wrapped");
1219          y01 = y00;
1220          y11 = y10;
1221 
1222          if (dims == 3) {
1223             lp_build_sample_wrap_linear(bld, is_gather, coords[2], depth_vec,
1224                                         flt_depth_vec, offsets[2],
1225                                         bld->static_texture_state->pot_depth,
1226                                         bld->static_sampler_state->wrap_r,
1227                                         &z00, &z1, &r_fpart);
1228             z01 = z10 = z11 = z00;
1229             lp_build_name(z00, "tex.z0.wrapped");
1230             lp_build_name(z1, "tex.z1.wrapped");
1231          }
1232       }
1233       if (has_layer_coord(bld->static_texture_state->target)) {
1234          if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) {
1235             /* add cube layer to face */
1236             z00 = z01 = z10 = z11 = z1 =
1237                lp_build_add(&bld->int_coord_bld, coords[2], coords[3]);
1238          } else {
1239             z00 = z01 = z10 = z11 = z1 = coords[2];  /* cube face or layer */
1240          }
1241          lp_build_name(z00, "tex.z0.layer");
1242          lp_build_name(z1, "tex.z1.layer");
1243       }
1244    } else {
1245       struct lp_build_if_state edge_if;
1246       LLVMValueRef new_faces[4], new_xcoords[4][2], new_ycoords[4][2];
1247       LLVMValueRef coord0, coord1, have_edge, have_corner;
1248       LLVMValueRef fall_off_ym_notxm, fall_off_ym_notxp, fall_off_x, fall_off_y;
1249       LLVMValueRef fall_off_yp_notxm, fall_off_yp_notxp;
1250       LLVMValueRef x0, x1, y0, y1, y0_clamped, y1_clamped;
1251       LLVMValueRef face = coords[2];
1252       LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5f);
1253       LLVMValueRef length_minus_one = lp_build_sub(ivec_bld, width_vec, ivec_bld->one);
1254       /* XXX drop height calcs. Could (should) do this without seamless filtering too */
1255       height_vec = width_vec;
1256       flt_height_vec = flt_width_vec;
1257 
1258       /* XXX the overflow logic is actually sort of duplicated with trilinear,
1259        * since an overflow in one mip should also have a corresponding overflow
1260        * in another.
1261        */
1262       /* should always have normalized coords, and offsets are undefined */
1263       assert(bld->static_sampler_state->normalized_coords);
1264       /*
1265        * The coords should all be between [0,1] however we can have NaNs,
1266        * which will wreak havoc. In particular the y1_clamped value below
1267        * can be -INT_MAX (on x86) and be propagated right through (probably
1268        * other values might be bogus in the end too).
1269        * So kill off the NaNs here.
1270        */
1271       coord0 = lp_build_max_ext(coord_bld, coords[0], coord_bld->zero,
1272                                 GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
1273       coord0 = lp_build_mul(coord_bld, coord0, flt_width_vec);
1274       /* instead of clamp, build mask if overflowed */
1275       coord0 = lp_build_sub(coord_bld, coord0, half);
1276       /* convert to int, compute lerp weight */
1277       /* not ideal with AVX (and no AVX2) */
1278       lp_build_ifloor_fract(coord_bld, coord0, &x0, &s_fpart);
1279       x1 = lp_build_add(ivec_bld, x0, ivec_bld->one);
1280       coord1 = lp_build_max_ext(coord_bld, coords[1], coord_bld->zero,
1281                                 GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN);
1282       coord1 = lp_build_mul(coord_bld, coord1, flt_height_vec);
1283       coord1 = lp_build_sub(coord_bld, coord1, half);
1284       lp_build_ifloor_fract(coord_bld, coord1, &y0, &t_fpart);
1285       y1 = lp_build_add(ivec_bld, y0, ivec_bld->one);
1286 
1287       fall_off[0] = lp_build_cmp(ivec_bld, PIPE_FUNC_LESS, x0, ivec_bld->zero);
1288       fall_off[1] = lp_build_cmp(ivec_bld, PIPE_FUNC_GREATER, x1, length_minus_one);
1289       fall_off[2] = lp_build_cmp(ivec_bld, PIPE_FUNC_LESS, y0, ivec_bld->zero);
1290       fall_off[3] = lp_build_cmp(ivec_bld, PIPE_FUNC_GREATER, y1, length_minus_one);
1291 
1292       fall_off_x = lp_build_or(ivec_bld, fall_off[0], fall_off[1]);
1293       fall_off_y = lp_build_or(ivec_bld, fall_off[2], fall_off[3]);
1294       have_edge = lp_build_or(ivec_bld, fall_off_x, fall_off_y);
1295       have_edge = lp_build_any_true_range(ivec_bld, ivec_bld->type.length, have_edge);
1296 
1297       /* needed for accurate corner filtering branch later, rely on 0 init */
1298       have_corners = lp_build_alloca(bld->gallivm, int1t, "have_corner");
1299 
1300       for (unsigned texel_index = 0; texel_index < 4; texel_index++) {
1301          xs[texel_index] = lp_build_alloca(bld->gallivm, ivec_bld->vec_type, "xs");
1302          ys[texel_index] = lp_build_alloca(bld->gallivm, ivec_bld->vec_type, "ys");
1303          zs[texel_index] = lp_build_alloca(bld->gallivm, ivec_bld->vec_type, "zs");
1304       }
1305 
1306       lp_build_if(&edge_if, bld->gallivm, have_edge);
1307 
1308       have_corner = lp_build_and(ivec_bld, fall_off_x, fall_off_y);
1309       have_corner = lp_build_any_true_range(ivec_bld, ivec_bld->type.length, have_corner);
1310       LLVMBuildStore(builder, have_corner, have_corners);
1311 
1312       /*
1313        * Need to feed clamped values here for cheap corner handling,
1314        * but only for y coord (as when falling off both edges we only
1315        * fall off the x one) - this should be sufficient.
1316        */
1317       y0_clamped = lp_build_max(ivec_bld, y0, ivec_bld->zero);
1318       y1_clamped = lp_build_min(ivec_bld, y1, length_minus_one);
1319 
1320       /*
1321        * Get all possible new coords.
1322        */
1323       lp_build_cube_new_coords(ivec_bld, face,
1324                                x0, x1, y0_clamped, y1_clamped,
1325                                length_minus_one,
1326                                new_faces, new_xcoords, new_ycoords);
1327 
1328       /* handle fall off x-, x+ direction */
1329       /* determine new coords, face (not both fall_off vars can be true at same time) */
1330       x00 = lp_build_select(ivec_bld, fall_off[0], new_xcoords[0][0], x0);
1331       y00 = lp_build_select(ivec_bld, fall_off[0], new_ycoords[0][0], y0_clamped);
1332       x10 = lp_build_select(ivec_bld, fall_off[0], new_xcoords[0][1], x0);
1333       y10 = lp_build_select(ivec_bld, fall_off[0], new_ycoords[0][1], y1_clamped);
1334       x01 = lp_build_select(ivec_bld, fall_off[1], new_xcoords[1][0], x1);
1335       y01 = lp_build_select(ivec_bld, fall_off[1], new_ycoords[1][0], y0_clamped);
1336       x11 = lp_build_select(ivec_bld, fall_off[1], new_xcoords[1][1], x1);
1337       y11 = lp_build_select(ivec_bld, fall_off[1], new_ycoords[1][1], y1_clamped);
1338 
1339       z00 = z10 = lp_build_select(ivec_bld, fall_off[0], new_faces[0], face);
1340       z01 = z11 = lp_build_select(ivec_bld, fall_off[1], new_faces[1], face);
1341 
1342       /* handle fall off y-, y+ direction */
1343       /*
1344        * Cheap corner logic: just hack up things so a texel doesn't fall
1345        * off both sides (which means filter weights will be wrong but we'll only
1346        * use valid texels in the filter).
1347        * This means however (y) coords must additionally be clamped (see above).
1348        * This corner handling should be fully OpenGL (but not d3d10) compliant.
1349        */
1350       fall_off_ym_notxm = lp_build_andnot(ivec_bld, fall_off[2], fall_off[0]);
1351       fall_off_ym_notxp = lp_build_andnot(ivec_bld, fall_off[2], fall_off[1]);
1352       fall_off_yp_notxm = lp_build_andnot(ivec_bld, fall_off[3], fall_off[0]);
1353       fall_off_yp_notxp = lp_build_andnot(ivec_bld, fall_off[3], fall_off[1]);
1354 
1355       x00 = lp_build_select(ivec_bld, fall_off_ym_notxm, new_xcoords[2][0], x00);
1356       y00 = lp_build_select(ivec_bld, fall_off_ym_notxm, new_ycoords[2][0], y00);
1357       x01 = lp_build_select(ivec_bld, fall_off_ym_notxp, new_xcoords[2][1], x01);
1358       y01 = lp_build_select(ivec_bld, fall_off_ym_notxp, new_ycoords[2][1], y01);
1359       x10 = lp_build_select(ivec_bld, fall_off_yp_notxm, new_xcoords[3][0], x10);
1360       y10 = lp_build_select(ivec_bld, fall_off_yp_notxm, new_ycoords[3][0], y10);
1361       x11 = lp_build_select(ivec_bld, fall_off_yp_notxp, new_xcoords[3][1], x11);
1362       y11 = lp_build_select(ivec_bld, fall_off_yp_notxp, new_ycoords[3][1], y11);
1363 
1364       z00 = lp_build_select(ivec_bld, fall_off_ym_notxm, new_faces[2], z00);
1365       z01 = lp_build_select(ivec_bld, fall_off_ym_notxp, new_faces[2], z01);
1366       z10 = lp_build_select(ivec_bld, fall_off_yp_notxm, new_faces[3], z10);
1367       z11 = lp_build_select(ivec_bld, fall_off_yp_notxp, new_faces[3], z11);
1368 
1369       if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) {
1370          /* now can add cube layer to face (per sample) */
1371          z00 = lp_build_add(ivec_bld, z00, coords[3]);
1372          z01 = lp_build_add(ivec_bld, z01, coords[3]);
1373          z10 = lp_build_add(ivec_bld, z10, coords[3]);
1374          z11 = lp_build_add(ivec_bld, z11, coords[3]);
1375       }
1376 
1377       LLVMBuildStore(builder, x00, xs[0]);
1378       LLVMBuildStore(builder, x01, xs[1]);
1379       LLVMBuildStore(builder, x10, xs[2]);
1380       LLVMBuildStore(builder, x11, xs[3]);
1381       LLVMBuildStore(builder, y00, ys[0]);
1382       LLVMBuildStore(builder, y01, ys[1]);
1383       LLVMBuildStore(builder, y10, ys[2]);
1384       LLVMBuildStore(builder, y11, ys[3]);
1385       LLVMBuildStore(builder, z00, zs[0]);
1386       LLVMBuildStore(builder, z01, zs[1]);
1387       LLVMBuildStore(builder, z10, zs[2]);
1388       LLVMBuildStore(builder, z11, zs[3]);
1389 
1390       lp_build_else(&edge_if);
1391 
1392       LLVMBuildStore(builder, x0, xs[0]);
1393       LLVMBuildStore(builder, x1, xs[1]);
1394       LLVMBuildStore(builder, x0, xs[2]);
1395       LLVMBuildStore(builder, x1, xs[3]);
1396       LLVMBuildStore(builder, y0, ys[0]);
1397       LLVMBuildStore(builder, y0, ys[1]);
1398       LLVMBuildStore(builder, y1, ys[2]);
1399       LLVMBuildStore(builder, y1, ys[3]);
1400       if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) {
1401          LLVMValueRef cube_layer = lp_build_add(ivec_bld, face, coords[3]);
1402          LLVMBuildStore(builder, cube_layer, zs[0]);
1403          LLVMBuildStore(builder, cube_layer, zs[1]);
1404          LLVMBuildStore(builder, cube_layer, zs[2]);
1405          LLVMBuildStore(builder, cube_layer, zs[3]);
1406       } else {
1407          LLVMBuildStore(builder, face, zs[0]);
1408          LLVMBuildStore(builder, face, zs[1]);
1409          LLVMBuildStore(builder, face, zs[2]);
1410          LLVMBuildStore(builder, face, zs[3]);
1411       }
1412 
1413       lp_build_endif(&edge_if);
1414 
1415       LLVMTypeRef type = ivec_bld->vec_type;
1416       x00 = LLVMBuildLoad2(builder, type, xs[0], "");
1417       x01 = LLVMBuildLoad2(builder, type, xs[1], "");
1418       x10 = LLVMBuildLoad2(builder, type, xs[2], "");
1419       x11 = LLVMBuildLoad2(builder, type, xs[3], "");
1420       y00 = LLVMBuildLoad2(builder, type, ys[0], "");
1421       y01 = LLVMBuildLoad2(builder, type, ys[1], "");
1422       y10 = LLVMBuildLoad2(builder, type, ys[2], "");
1423       y11 = LLVMBuildLoad2(builder, type, ys[3], "");
1424       z00 = LLVMBuildLoad2(builder, type, zs[0], "");
1425       z01 = LLVMBuildLoad2(builder, type, zs[1], "");
1426       z10 = LLVMBuildLoad2(builder, type, zs[2], "");
1427       z11 = LLVMBuildLoad2(builder, type, zs[3], "");
1428    }
1429 
1430    if (linear_mask) {
1431       /*
1432        * Whack filter weights into place. Whatever texel had more weight is
1433        * the one which should have been selected by nearest filtering hence
1434        * just use 100% weight for it.
1435        */
1436       struct lp_build_context *c_bld = &bld->coord_bld;
1437       LLVMValueRef w1_mask, w1_weight;
1438       LLVMValueRef half = lp_build_const_vec(bld->gallivm, c_bld->type, 0.5f);
1439 
1440       w1_mask = lp_build_cmp(c_bld, PIPE_FUNC_GREATER, s_fpart, half);
1441       /* this select is really just a "and" */
1442       w1_weight = lp_build_select(c_bld, w1_mask, c_bld->one, c_bld->zero);
1443       s_fpart = lp_build_select(c_bld, linear_mask, s_fpart, w1_weight);
1444       if (dims >= 2) {
1445          w1_mask = lp_build_cmp(c_bld, PIPE_FUNC_GREATER, t_fpart, half);
1446          w1_weight = lp_build_select(c_bld, w1_mask, c_bld->one, c_bld->zero);
1447          t_fpart = lp_build_select(c_bld, linear_mask, t_fpart, w1_weight);
1448          if (dims == 3) {
1449             w1_mask = lp_build_cmp(c_bld, PIPE_FUNC_GREATER, r_fpart, half);
1450             w1_weight = lp_build_select(c_bld, w1_mask, c_bld->one, c_bld->zero);
1451             r_fpart = lp_build_select(c_bld, linear_mask, r_fpart, w1_weight);
1452          }
1453       }
1454    }
1455 
1456    /*
1457     * Get texture colors.
1458     */
1459    /* get x0/x1 texels */
1460    lp_build_sample_texel_soa(bld,
1461                              width_vec, height_vec, depth_vec,
1462                              x00, y00, z00,
1463                              row_stride_vec, img_stride_vec,
1464                              data_ptr, mipoffsets, ilevel, neighbors[0][0]);
1465    lp_build_sample_texel_soa(bld,
1466                              width_vec, height_vec, depth_vec,
1467                              x01, y01, z01,
1468                              row_stride_vec, img_stride_vec,
1469                              data_ptr, mipoffsets, ilevel, neighbors[0][1]);
1470 
1471    if (dims == 1) {
1472       assert(!is_gather);
1473       if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE) {
1474          lp_build_reduce_filter(texel_bld,
1475                                 bld->static_sampler_state->reduction_mode,
1476                                 0,
1477                                 4,
1478                                 s_fpart,
1479                                 neighbors[0][0],
1480                                 neighbors[0][1],
1481                                 colors_out);
1482       } else {
1483          LLVMValueRef cmpval0, cmpval1;
1484          cmpval0 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][0][0]);
1485          cmpval1 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][1][0]);
1486          /* simplified lerp, AND mask with weight and add */
1487          colors_out[0] = lp_build_masklerp(texel_bld, s_fpart,
1488                                            cmpval0, cmpval1);
1489          colors_out[1] = colors_out[2] = colors_out[3] = colors_out[0];
1490       }
1491    } else {
1492       /* 2D/3D texture */
1493       struct lp_build_if_state corner_if;
1494       LLVMValueRef colors0[4], colorss[4] = { 0 };
1495 
1496       /* get x0/x1 texels at y1 */
1497       lp_build_sample_texel_soa(bld,
1498                                 width_vec, height_vec, depth_vec,
1499                                 x10, y10, z10,
1500                                 row_stride_vec, img_stride_vec,
1501                                 data_ptr, mipoffsets, ilevel, neighbors[1][0]);
1502       lp_build_sample_texel_soa(bld,
1503                                 width_vec, height_vec, depth_vec,
1504                                 x11, y11, z11,
1505                                 row_stride_vec, img_stride_vec,
1506                                 data_ptr, mipoffsets, ilevel, neighbors[1][1]);
1507 
1508       /*
1509        * To avoid having to duplicate linear_mask / fetch code use
1510        * another branch (with corner condition though edge would work
1511        * as well) here.
1512        */
1513       if (have_corners && accurate_cube_corners &&
1514           bld->static_sampler_state->reduction_mode == PIPE_TEX_REDUCTION_WEIGHTED_AVERAGE) {
1515          LLVMValueRef c00, c01, c10, c11, c00f, c01f, c10f, c11f;
1516          LLVMValueRef have_corner, one_third;
1517 
1518          colorss[0] = lp_build_alloca(bld->gallivm, coord_bld->vec_type, "cs0");
1519          colorss[1] = lp_build_alloca(bld->gallivm, coord_bld->vec_type, "cs1");
1520          colorss[2] = lp_build_alloca(bld->gallivm, coord_bld->vec_type, "cs2");
1521          colorss[3] = lp_build_alloca(bld->gallivm, coord_bld->vec_type, "cs3");
1522 
1523          have_corner = LLVMBuildLoad2(builder, int1t, have_corners, "");
1524 
1525          lp_build_if(&corner_if, bld->gallivm, have_corner);
1526 
1527          one_third = lp_build_const_vec(bld->gallivm, coord_bld->type,
1528                                         1.0f/3.0f);
1529 
1530          /* find corner */
1531          c00 = lp_build_and(ivec_bld, fall_off[0], fall_off[2]);
1532          c00f = LLVMBuildBitCast(builder, c00, coord_bld->vec_type, "");
1533          c01 = lp_build_and(ivec_bld, fall_off[1], fall_off[2]);
1534          c01f = LLVMBuildBitCast(builder, c01, coord_bld->vec_type, "");
1535          c10 = lp_build_and(ivec_bld, fall_off[0], fall_off[3]);
1536          c10f = LLVMBuildBitCast(builder, c10, coord_bld->vec_type, "");
1537          c11 = lp_build_and(ivec_bld, fall_off[1], fall_off[3]);
1538          c11f = LLVMBuildBitCast(builder, c11, coord_bld->vec_type, "");
1539 
1540          if (!is_gather) {
1541             /*
1542              * we can't use standard 2d lerp as we need per-element weight
1543              * in case of corners, so just calculate bilinear result as
1544              * w00*s00 + w01*s01 + w10*s10 + w11*s11.
1545              * (This is actually less work than using 2d lerp, 7 vs. 9
1546              * instructions, however calculating the weights needs another 6,
1547              * so actually probably not slower than 2d lerp only for 4 channels
1548              * as weights only need to be calculated once - of course fixing
1549              * the weights has additional cost.)
1550              */
1551             LLVMValueRef w00, w01, w10, w11, wx0, wy0, c_weight, tmp;
1552             wx0 = lp_build_sub(coord_bld, coord_bld->one, s_fpart);
1553             wy0 = lp_build_sub(coord_bld, coord_bld->one, t_fpart);
1554             w00 = lp_build_mul(coord_bld, wx0, wy0);
1555             w01 = lp_build_mul(coord_bld, s_fpart, wy0);
1556             w10 = lp_build_mul(coord_bld, wx0, t_fpart);
1557             w11 = lp_build_mul(coord_bld, s_fpart, t_fpart);
1558 
1559             /* find corner weight */
1560             c_weight = lp_build_select(coord_bld, c00, w00, coord_bld->zero);
1561             c_weight = lp_build_select(coord_bld, c01, w01, c_weight);
1562             c_weight = lp_build_select(coord_bld, c10, w10, c_weight);
1563             c_weight = lp_build_select(coord_bld, c11, w11, c_weight);
1564 
1565             /*
1566              * add 1/3 of the corner weight to the weight of the 3 other
1567              * samples and null out corner weight.
1568              */
1569             c_weight = lp_build_mul(coord_bld, c_weight, one_third);
1570             w00 = lp_build_add(coord_bld, w00, c_weight);
1571             w00 = lp_build_andnot(coord_bld, w00, c00f);
1572             w01 = lp_build_add(coord_bld, w01, c_weight);
1573             w01 = lp_build_andnot(coord_bld, w01, c01f);
1574             w10 = lp_build_add(coord_bld, w10, c_weight);
1575             w10 = lp_build_andnot(coord_bld, w10, c10f);
1576             w11 = lp_build_add(coord_bld, w11, c_weight);
1577             w11 = lp_build_andnot(coord_bld, w11, c11f);
1578 
1579             if (bld->static_sampler_state->compare_mode ==
1580                 PIPE_TEX_COMPARE_NONE) {
1581                for (unsigned chan = 0; chan < 4; chan++) {
1582                   colors0[chan] = lp_build_mul(coord_bld, w00,
1583                                                neighbors[0][0][chan]);
1584                   tmp = lp_build_mul(coord_bld, w01, neighbors[0][1][chan]);
1585                   colors0[chan] = lp_build_add(coord_bld, tmp, colors0[chan]);
1586                   tmp = lp_build_mul(coord_bld, w10, neighbors[1][0][chan]);
1587                   colors0[chan] = lp_build_add(coord_bld, tmp, colors0[chan]);
1588                   tmp = lp_build_mul(coord_bld, w11, neighbors[1][1][chan]);
1589                   colors0[chan] = lp_build_add(coord_bld, tmp, colors0[chan]);
1590                }
1591             } else {
1592                LLVMValueRef cmpval00, cmpval01, cmpval10, cmpval11;
1593                cmpval00 = lp_build_sample_comparefunc(bld, coords[4],
1594                                                       neighbors[0][0][0]);
1595                cmpval01 = lp_build_sample_comparefunc(bld, coords[4],
1596                                                       neighbors[0][1][0]);
1597                cmpval10 = lp_build_sample_comparefunc(bld, coords[4],
1598                                                       neighbors[1][0][0]);
1599                cmpval11 = lp_build_sample_comparefunc(bld, coords[4],
1600                                                       neighbors[1][1][0]);
1601                /*
1602                 * inputs to interpolation are just masks so just add
1603                 * masked weights together
1604                 */
1605                cmpval00 = LLVMBuildBitCast(builder, cmpval00,
1606                                            coord_bld->vec_type, "");
1607                cmpval01 = LLVMBuildBitCast(builder, cmpval01,
1608                                            coord_bld->vec_type, "");
1609                cmpval10 = LLVMBuildBitCast(builder, cmpval10,
1610                                            coord_bld->vec_type, "");
1611                cmpval11 = LLVMBuildBitCast(builder, cmpval11,
1612                                            coord_bld->vec_type, "");
1613                colors0[0] = lp_build_and(coord_bld, w00, cmpval00);
1614                tmp = lp_build_and(coord_bld, w01, cmpval01);
1615                colors0[0] = lp_build_add(coord_bld, tmp, colors0[0]);
1616                tmp = lp_build_and(coord_bld, w10, cmpval10);
1617                colors0[0] = lp_build_add(coord_bld, tmp, colors0[0]);
1618                tmp = lp_build_and(coord_bld, w11, cmpval11);
1619                colors0[0] = lp_build_add(coord_bld, tmp, colors0[0]);
1620                colors0[1] = colors0[2] = colors0[3] = colors0[0];
1621             }
1622          } else {
1623             /*
1624              * We don't have any weights to adjust, so instead calculate
1625              * the fourth texel as simply the average of the other 3.
1626              * (This would work for non-gather too, however we'd have
1627              * a boatload more of the select stuff due to there being
1628              * 4 times as many colors as weights.)
1629              */
1630             LLVMValueRef col00, col01, col10, col11;
1631             LLVMValueRef colc, colc0, colc1;
1632             col10 = lp_build_swizzle_soa_channel(texel_bld,
1633                                                  neighbors[1][0], chan_swiz);
1634             col11 = lp_build_swizzle_soa_channel(texel_bld,
1635                                                  neighbors[1][1], chan_swiz);
1636             col01 = lp_build_swizzle_soa_channel(texel_bld,
1637                                                  neighbors[0][1], chan_swiz);
1638             col00 = lp_build_swizzle_soa_channel(texel_bld,
1639                                                  neighbors[0][0], chan_swiz);
1640 
1641             /*
1642              * The spec says for comparison filtering, the comparison
1643              * must happen before synthesizing the new value.
1644              * This means all gathered values are always 0 or 1,
1645              * except for the non-existing texel, which can be 0,1/3,2/3,1...
1646              * Seems like we'd be allowed to just return 0 or 1 too, so we
1647              * could simplify and pass down the compare mask values to the
1648              * end (using int arithmetic/compare on the mask values to
1649              * construct the fourth texel) and only there convert to floats
1650              * but it's probably not worth it (it might be easier for the cpu
1651              * but not for the code)...
1652              */
1653             if (bld->static_sampler_state->compare_mode !=
1654                 PIPE_TEX_COMPARE_NONE) {
1655                LLVMValueRef cmpval00, cmpval01, cmpval10, cmpval11;
1656                cmpval00 = lp_build_sample_comparefunc(bld, coords[4], col00);
1657                cmpval01 = lp_build_sample_comparefunc(bld, coords[4], col01);
1658                cmpval10 = lp_build_sample_comparefunc(bld, coords[4], col10);
1659                cmpval11 = lp_build_sample_comparefunc(bld, coords[4], col11);
1660                col00 = lp_build_select(texel_bld, cmpval00,
1661                                        texel_bld->one, texel_bld->zero);
1662                col01 = lp_build_select(texel_bld, cmpval01,
1663                                        texel_bld->one, texel_bld->zero);
1664                col10 = lp_build_select(texel_bld, cmpval10,
1665                                        texel_bld->one, texel_bld->zero);
1666                col11 = lp_build_select(texel_bld, cmpval11,
1667                                        texel_bld->one, texel_bld->zero);
1668             }
1669 
1670             /*
1671              * Null out corner color.
1672              */
1673             col00 = lp_build_andnot(coord_bld, col00, c00f);
1674             col01 = lp_build_andnot(coord_bld, col01, c01f);
1675             col10 = lp_build_andnot(coord_bld, col10, c10f);
1676             col11 = lp_build_andnot(coord_bld, col11, c11f);
1677 
1678             /*
1679              * New corner texel color is all colors added / 3.
1680              */
1681             colc0 = lp_build_add(coord_bld, col00, col01);
1682             colc1 = lp_build_add(coord_bld, col10, col11);
1683             colc = lp_build_add(coord_bld, colc0, colc1);
1684             colc = lp_build_mul(coord_bld, one_third, colc);
1685 
1686             /*
1687              * Replace the corner texel color with the new value.
1688              */
1689             col00 = lp_build_select(coord_bld, c00, colc, col00);
1690             col01 = lp_build_select(coord_bld, c01, colc, col01);
1691             col10 = lp_build_select(coord_bld, c10, colc, col10);
1692             col11 = lp_build_select(coord_bld, c11, colc, col11);
1693 
1694             colors0[0] = col10;
1695             colors0[1] = col11;
1696             colors0[2] = col01;
1697             colors0[3] = col00;
1698          }
1699 
1700          LLVMBuildStore(builder, colors0[0], colorss[0]);
1701          LLVMBuildStore(builder, colors0[1], colorss[1]);
1702          LLVMBuildStore(builder, colors0[2], colorss[2]);
1703          LLVMBuildStore(builder, colors0[3], colorss[3]);
1704 
1705          lp_build_else(&corner_if);
1706       }
1707 
1708       if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE) {
1709          if (is_gather) {
1710             /*
1711              * Just assign the red channel (no component selection yet).
1712              * This is a bit hackish, we usually do the swizzle at the
1713              * end of sampling (much less values to swizzle), but this
1714              * obviously cannot work when using gather.
1715              */
1716             colors0[0] = lp_build_swizzle_soa_channel(texel_bld,
1717                                                       neighbors[1][0],
1718                                                       chan_swiz);
1719             colors0[1] = lp_build_swizzle_soa_channel(texel_bld,
1720                                                       neighbors[1][1],
1721                                                       chan_swiz);
1722             colors0[2] = lp_build_swizzle_soa_channel(texel_bld,
1723                                                       neighbors[0][1],
1724                                                       chan_swiz);
1725             colors0[3] = lp_build_swizzle_soa_channel(texel_bld,
1726                                                       neighbors[0][0],
1727                                                       chan_swiz);
1728          } else {
1729             /* Bilinear interpolate the four samples from the 2D image / 3D slice */
1730             lp_build_reduce_filter_2d(texel_bld,
1731                                       bld->static_sampler_state->reduction_mode,
1732                                       0,
1733                                       4,
1734                                       s_fpart,
1735                                       t_fpart,
1736                                       neighbors[0][0],
1737                                       neighbors[0][1],
1738                                       neighbors[1][0],
1739                                       neighbors[1][1],
1740                                       colors0);
1741          }
1742       } else {
1743          LLVMValueRef cmpval00, cmpval01, cmpval10, cmpval11;
1744          cmpval00 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][0][0]);
1745          cmpval01 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][1][0]);
1746          cmpval10 = lp_build_sample_comparefunc(bld, coords[4], neighbors[1][0][0]);
1747          cmpval11 = lp_build_sample_comparefunc(bld, coords[4], neighbors[1][1][0]);
1748 
1749          if (is_gather) {
1750             /* more hacks for swizzling, should be X, ONE or ZERO... */
1751             colors0[0] = lp_build_select(texel_bld, cmpval10,
1752                                          texel_bld->one, texel_bld->zero);
1753             colors0[1] = lp_build_select(texel_bld, cmpval11,
1754                                          texel_bld->one, texel_bld->zero);
1755             colors0[2] = lp_build_select(texel_bld, cmpval01,
1756                                          texel_bld->one, texel_bld->zero);
1757             colors0[3] = lp_build_select(texel_bld, cmpval00,
1758                                          texel_bld->one, texel_bld->zero);
1759          } else {
1760             colors0[0] = lp_build_masklerp2d(texel_bld, s_fpart, t_fpart,
1761                                              cmpval00, cmpval01, cmpval10, cmpval11);
1762             colors0[1] = colors0[2] = colors0[3] = colors0[0];
1763          }
1764       }
1765 
1766       if (have_corners && accurate_cube_corners &&
1767           bld->static_sampler_state->reduction_mode == PIPE_TEX_REDUCTION_WEIGHTED_AVERAGE) {
1768          LLVMBuildStore(builder, colors0[0], colorss[0]);
1769          LLVMBuildStore(builder, colors0[1], colorss[1]);
1770          LLVMBuildStore(builder, colors0[2], colorss[2]);
1771          LLVMBuildStore(builder, colors0[3], colorss[3]);
1772 
1773          lp_build_endif(&corner_if);
1774 
1775          colors0[0] = LLVMBuildLoad2(builder, coord_bld->vec_type, colorss[0], "");
1776          colors0[1] = LLVMBuildLoad2(builder, coord_bld->vec_type, colorss[1], "");
1777          colors0[2] = LLVMBuildLoad2(builder, coord_bld->vec_type, colorss[2], "");
1778          colors0[3] = LLVMBuildLoad2(builder, coord_bld->vec_type, colorss[3], "");
1779       }
1780 
1781       if (dims == 3) {
1782          LLVMValueRef neighbors1[2][2][4];
1783          LLVMValueRef colors1[4];
1784 
1785          assert(!is_gather);
1786 
1787          /* get x0/x1/y0/y1 texels at z1 */
1788          lp_build_sample_texel_soa(bld,
1789                                    width_vec, height_vec, depth_vec,
1790                                    x00, y00, z1,
1791                                    row_stride_vec, img_stride_vec,
1792                                    data_ptr, mipoffsets, ilevel, neighbors1[0][0]);
1793          lp_build_sample_texel_soa(bld,
1794                                    width_vec, height_vec, depth_vec,
1795                                    x01, y01, z1,
1796                                    row_stride_vec, img_stride_vec,
1797                                    data_ptr, mipoffsets, ilevel, neighbors1[0][1]);
1798          lp_build_sample_texel_soa(bld,
1799                                    width_vec, height_vec, depth_vec,
1800                                    x10, y10, z1,
1801                                    row_stride_vec, img_stride_vec,
1802                                    data_ptr, mipoffsets, ilevel, neighbors1[1][0]);
1803          lp_build_sample_texel_soa(bld,
1804                                    width_vec, height_vec, depth_vec,
1805                                    x11, y11, z1,
1806                                    row_stride_vec, img_stride_vec,
1807                                    data_ptr, mipoffsets, ilevel, neighbors1[1][1]);
1808 
1809          if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE) {
1810             /* Bilinear interpolate the four samples from the second Z slice */
1811             lp_build_reduce_filter_2d(texel_bld,
1812                                       bld->static_sampler_state->reduction_mode,
1813                                       0,
1814                                       4,
1815                                       s_fpart,
1816                                       t_fpart,
1817                                       neighbors1[0][0],
1818                                       neighbors1[0][1],
1819                                       neighbors1[1][0],
1820                                       neighbors1[1][1],
1821                                       colors1);
1822 
1823             /* Linearly interpolate the two samples from the two 3D slices */
1824             lp_build_reduce_filter(texel_bld,
1825                                    bld->static_sampler_state->reduction_mode,
1826                                    0,
1827                                    4,
1828                                    r_fpart,
1829                                    colors0,
1830                                    colors1,
1831                                    colors_out);
1832          } else {
1833             LLVMValueRef cmpval00, cmpval01, cmpval10, cmpval11;
1834             cmpval00 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][0][0]);
1835             cmpval01 = lp_build_sample_comparefunc(bld, coords[4], neighbors[0][1][0]);
1836             cmpval10 = lp_build_sample_comparefunc(bld, coords[4], neighbors[1][0][0]);
1837             cmpval11 = lp_build_sample_comparefunc(bld, coords[4], neighbors[1][1][0]);
1838             colors1[0] = lp_build_masklerp2d(texel_bld, s_fpart, t_fpart,
1839                                              cmpval00, cmpval01, cmpval10, cmpval11);
1840             /* Linearly interpolate the two samples from the two 3D slices */
1841             colors_out[0] = lp_build_lerp(texel_bld,
1842                                           r_fpart,
1843                                           colors0[0], colors1[0],
1844                                           0);
1845             colors_out[1] = colors_out[2] = colors_out[3] = colors_out[0];
1846          }
1847       } else {
1848          /* 2D tex */
1849          for (unsigned chan = 0; chan < 4; chan++) {
1850             colors_out[chan] = colors0[chan];
1851          }
1852       }
1853    }
1854    if (is_gather) {
1855       /*
1856        * For gather, we can't do our usual channel swizzling done later,
1857        * so do it here. It only really matters for 0/1 swizzles in case
1858        * of comparison filtering, since in this case the results would be
1859        * wrong, without comparison it should all work out alright but it
1860        * can't hurt to do that here, since it will instantly drop all
1861        * calculations above, though it's a rather stupid idea to do
1862        * gather on a channel which will always return 0 or 1 in any case...
1863        */
1864       if (chan_swiz == PIPE_SWIZZLE_1) {
1865          for (unsigned chan = 0; chan < 4; chan++) {
1866             colors_out[chan] = texel_bld->one;
1867          }
1868       } else if (chan_swiz == PIPE_SWIZZLE_0) {
1869          for (unsigned chan = 0; chan < 4; chan++) {
1870             colors_out[chan] = texel_bld->zero;
1871          }
1872       }
1873    }
1874 }
1875 
1876 
1877 /**
1878  * Sample the texture/mipmap using given image filter and mip filter.
1879  * ilevel0 and ilevel1 indicate the two mipmap levels to sample
1880  * from (vectors or scalars).
1881  * If we're using nearest miplevel sampling the '1' values will be null/unused.
1882  */
1883 static void
lp_build_sample_mipmap(struct lp_build_sample_context * bld,unsigned img_filter,unsigned mip_filter,bool is_gather,const LLVMValueRef * coords,const LLVMValueRef * offsets,LLVMValueRef ilevel0,LLVMValueRef ilevel1,LLVMValueRef lod_fpart,LLVMValueRef * colors_out)1884 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
1885                        unsigned img_filter,
1886                        unsigned mip_filter,
1887                        bool is_gather,
1888                        const LLVMValueRef *coords,
1889                        const LLVMValueRef *offsets,
1890                        LLVMValueRef ilevel0,
1891                        LLVMValueRef ilevel1,
1892                        LLVMValueRef lod_fpart,
1893                        LLVMValueRef *colors_out)
1894 {
1895    LLVMBuilderRef builder = bld->gallivm->builder;
1896    LLVMValueRef size0 = NULL;
1897    LLVMValueRef size1 = NULL;
1898    LLVMValueRef row_stride0_vec = NULL;
1899    LLVMValueRef row_stride1_vec = NULL;
1900    LLVMValueRef img_stride0_vec = NULL;
1901    LLVMValueRef img_stride1_vec = NULL;
1902    LLVMValueRef data_ptr0 = NULL;
1903    LLVMValueRef data_ptr1 = NULL;
1904    LLVMValueRef mipoff0 = NULL;
1905    LLVMValueRef mipoff1 = NULL;
1906    LLVMValueRef colors0[4], colors1[4];
1907 
1908    /* sample the first mipmap level */
1909    lp_build_mipmap_level_sizes(bld, ilevel0,
1910                                &size0,
1911                                &row_stride0_vec, &img_stride0_vec);
1912    if (bld->num_mips == 1) {
1913       data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
1914    } else {
1915       /* This path should work for num_lods 1 too but slightly less efficient */
1916       data_ptr0 = bld->base_ptr;
1917       mipoff0 = lp_build_get_mip_offsets(bld, ilevel0);
1918    }
1919 
1920    if (img_filter == PIPE_TEX_FILTER_NEAREST) {
1921       lp_build_sample_image_nearest(bld, size0,
1922                                     row_stride0_vec, img_stride0_vec,
1923                                     data_ptr0, mipoff0, ilevel0, coords, offsets,
1924                                     colors0);
1925    } else {
1926       assert(img_filter == PIPE_TEX_FILTER_LINEAR);
1927       lp_build_sample_image_linear(bld, is_gather, size0, NULL,
1928                                    row_stride0_vec, img_stride0_vec,
1929                                    data_ptr0, mipoff0, ilevel0, coords, offsets,
1930                                    colors0);
1931    }
1932 
1933    /* Store the first level's colors in the output variables */
1934    for (unsigned chan = 0; chan < 4; chan++) {
1935        LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
1936    }
1937 
1938    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1939       struct lp_build_if_state if_ctx;
1940       LLVMValueRef need_lerp;
1941 
1942       /* need_lerp = lod_fpart > 0 */
1943       if (bld->num_lods == 1) {
1944          need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT,
1945                                    lod_fpart, bld->lodf_bld.zero,
1946                                    "need_lerp");
1947       } else {
1948          /*
1949           * We'll do mip filtering if any of the quads (or individual
1950           * pixel in case of per-pixel lod) need it.
1951           * It might be better to split the vectors here and only fetch/filter
1952           * quads which need it (if there's one lod per quad).
1953           */
1954          need_lerp = lp_build_compare(bld->gallivm, bld->lodf_bld.type,
1955                                       PIPE_FUNC_GREATER,
1956                                       lod_fpart, bld->lodf_bld.zero);
1957          need_lerp = lp_build_any_true_range(&bld->lodi_bld, bld->num_lods, need_lerp);
1958          lp_build_name(need_lerp, "need_lerp");
1959       }
1960 
1961       lp_build_if(&if_ctx, bld->gallivm, need_lerp);
1962       {
1963          /*
1964           * We unfortunately need to clamp lod_fpart here since we can get
1965           * negative values which would screw up filtering if not all
1966           * lod_fpart values have same sign.
1967           */
1968          lod_fpart = lp_build_max(&bld->lodf_bld, lod_fpart,
1969                                   bld->lodf_bld.zero);
1970          /* sample the second mipmap level */
1971          lp_build_mipmap_level_sizes(bld, ilevel1,
1972                                      &size1,
1973                                      &row_stride1_vec, &img_stride1_vec);
1974          if (bld->num_mips == 1) {
1975             data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
1976          } else {
1977             data_ptr1 = bld->base_ptr;
1978             mipoff1 = lp_build_get_mip_offsets(bld, ilevel1);
1979          }
1980          if (img_filter == PIPE_TEX_FILTER_NEAREST) {
1981             lp_build_sample_image_nearest(bld, size1,
1982                                           row_stride1_vec, img_stride1_vec,
1983                                           data_ptr1, mipoff1, ilevel1, coords, offsets,
1984                                           colors1);
1985          } else {
1986             lp_build_sample_image_linear(bld, false, size1, NULL,
1987                                          row_stride1_vec, img_stride1_vec,
1988                                          data_ptr1, mipoff1, ilevel1, coords, offsets,
1989                                          colors1);
1990          }
1991 
1992          /* interpolate samples from the two mipmap levels */
1993 
1994          if (bld->num_lods != bld->coord_type.length)
1995             lod_fpart = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
1996                                                               bld->lodf_bld.type,
1997                                                               bld->texel_bld.type,
1998                                                               lod_fpart);
1999 
2000          for (unsigned chan = 0; chan < 4; chan++) {
2001             colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
2002                                           colors0[chan], colors1[chan],
2003                                           0);
2004             LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
2005          }
2006       }
2007       lp_build_endif(&if_ctx);
2008    }
2009 }
2010 
2011 
2012 /**
2013  * Sample the texture/mipmap using given mip filter, and using
2014  * both nearest and linear filtering at the same time depending
2015  * on linear_mask.
2016  * lod can be per quad but linear_mask is always per pixel.
2017  * ilevel0 and ilevel1 indicate the two mipmap levels to sample
2018  * from (vectors or scalars).
2019  * If we're using nearest miplevel sampling the '1' values will be null/unused.
2020  */
2021 static void
lp_build_sample_mipmap_both(struct lp_build_sample_context * bld,LLVMValueRef linear_mask,unsigned mip_filter,const LLVMValueRef * coords,const LLVMValueRef * offsets,LLVMValueRef ilevel0,LLVMValueRef ilevel1,LLVMValueRef lod_fpart,LLVMValueRef lod_positive,LLVMValueRef * colors_out)2022 lp_build_sample_mipmap_both(struct lp_build_sample_context *bld,
2023                             LLVMValueRef linear_mask,
2024                             unsigned mip_filter,
2025                             const LLVMValueRef *coords,
2026                             const LLVMValueRef *offsets,
2027                             LLVMValueRef ilevel0,
2028                             LLVMValueRef ilevel1,
2029                             LLVMValueRef lod_fpart,
2030                             LLVMValueRef lod_positive,
2031                             LLVMValueRef *colors_out)
2032 {
2033    LLVMBuilderRef builder = bld->gallivm->builder;
2034    LLVMValueRef size0 = NULL;
2035    LLVMValueRef size1 = NULL;
2036    LLVMValueRef row_stride0_vec = NULL;
2037    LLVMValueRef row_stride1_vec = NULL;
2038    LLVMValueRef img_stride0_vec = NULL;
2039    LLVMValueRef img_stride1_vec = NULL;
2040    LLVMValueRef data_ptr0 = NULL;
2041    LLVMValueRef data_ptr1 = NULL;
2042    LLVMValueRef mipoff0 = NULL;
2043    LLVMValueRef mipoff1 = NULL;
2044    LLVMValueRef colors0[4], colors1[4];
2045 
2046    /* sample the first mipmap level */
2047    lp_build_mipmap_level_sizes(bld, ilevel0,
2048                                &size0,
2049                                &row_stride0_vec, &img_stride0_vec);
2050    if (bld->num_mips == 1) {
2051       data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
2052    } else {
2053       /* This path should work for num_lods 1 too but slightly less efficient */
2054       data_ptr0 = bld->base_ptr;
2055       mipoff0 = lp_build_get_mip_offsets(bld, ilevel0);
2056    }
2057 
2058    lp_build_sample_image_linear(bld, false, size0, linear_mask,
2059                                 row_stride0_vec, img_stride0_vec,
2060                                 data_ptr0, mipoff0, ilevel0, coords, offsets,
2061                                 colors0);
2062 
2063    /* Store the first level's colors in the output variables */
2064    for (unsigned chan = 0; chan < 4; chan++) {
2065        LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
2066    }
2067 
2068    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
2069       struct lp_build_if_state if_ctx;
2070       LLVMValueRef need_lerp;
2071 
2072       /*
2073        * We'll do mip filtering if any of the quads (or individual
2074        * pixel in case of per-pixel lod) need it.
2075        * Note using lod_positive here not lod_fpart since it may be the same
2076        * condition as that used in the outer "if" in the caller hence llvm
2077        * should be able to merge the branches in this case.
2078        */
2079       need_lerp = lp_build_any_true_range(&bld->lodi_bld, bld->num_lods, lod_positive);
2080       lp_build_name(need_lerp, "need_lerp");
2081 
2082       lp_build_if(&if_ctx, bld->gallivm, need_lerp);
2083       {
2084          /*
2085           * We unfortunately need to clamp lod_fpart here since we can get
2086           * negative values which would screw up filtering if not all
2087           * lod_fpart values have same sign.
2088           */
2089          lod_fpart = lp_build_max(&bld->lodf_bld, lod_fpart,
2090                                   bld->lodf_bld.zero);
2091          /* sample the second mipmap level */
2092          lp_build_mipmap_level_sizes(bld, ilevel1,
2093                                      &size1,
2094                                      &row_stride1_vec, &img_stride1_vec);
2095          if (bld->num_mips == 1) {
2096             data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
2097          } else {
2098             data_ptr1 = bld->base_ptr;
2099             mipoff1 = lp_build_get_mip_offsets(bld, ilevel1);
2100          }
2101 
2102          lp_build_sample_image_linear(bld, false, size1, linear_mask,
2103                                       row_stride1_vec, img_stride1_vec,
2104                                       data_ptr1, mipoff1, ilevel1, coords, offsets,
2105                                       colors1);
2106 
2107          /* interpolate samples from the two mipmap levels */
2108 
2109          if (bld->num_lods != bld->coord_type.length)
2110             lod_fpart = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
2111                                                               bld->lodf_bld.type,
2112                                                               bld->texel_bld.type,
2113                                                               lod_fpart);
2114 
2115          for (unsigned chan = 0; chan < 4; chan++) {
2116             colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
2117                                           colors0[chan], colors1[chan],
2118                                           0);
2119             LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
2120          }
2121       }
2122       lp_build_endif(&if_ctx);
2123    }
2124 }
2125 
2126 
2127 /**
2128  * Build (per-coord) layer value.
2129  * Either clamp layer to valid values or fill in optional out_of_bounds
2130  * value and just return value unclamped.
2131  */
2132 static LLVMValueRef
lp_build_layer_coord(struct lp_build_sample_context * bld,unsigned texture_unit,bool is_cube_array,LLVMValueRef layer,LLVMValueRef * out_of_bounds)2133 lp_build_layer_coord(struct lp_build_sample_context *bld,
2134                      unsigned texture_unit,
2135                      bool is_cube_array,
2136                      LLVMValueRef layer,
2137                      LLVMValueRef *out_of_bounds)
2138 {
2139    LLVMValueRef num_layers;
2140    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
2141 
2142    num_layers = bld->dynamic_state->depth(bld->gallivm, bld->resources_type,
2143                                           bld->resources_ptr, texture_unit, NULL);
2144    num_layers = LLVMBuildZExt(bld->gallivm->builder, num_layers,
2145                               bld->int_bld.elem_type, "");
2146    if (out_of_bounds) {
2147       LLVMValueRef out1, out;
2148       assert(!is_cube_array);
2149       num_layers = lp_build_broadcast_scalar(int_coord_bld, num_layers);
2150       out = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, layer, int_coord_bld->zero);
2151       out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, layer, num_layers);
2152       *out_of_bounds = lp_build_or(int_coord_bld, out, out1);
2153       return layer;
2154    } else {
2155       LLVMValueRef maxlayer;
2156       LLVMValueRef s = is_cube_array ? lp_build_const_int32(bld->gallivm, 6) :
2157                                        bld->int_bld.one;
2158       maxlayer = lp_build_sub(&bld->int_bld, num_layers, s);
2159       maxlayer = lp_build_broadcast_scalar(int_coord_bld, maxlayer);
2160       return lp_build_clamp(int_coord_bld, layer, int_coord_bld->zero, maxlayer);
2161    }
2162 }
2163 
2164 static void
lp_build_sample_ms_offset(struct lp_build_context * int_coord_bld,LLVMValueRef ms_index,LLVMValueRef num_samples,LLVMValueRef sample_stride,LLVMValueRef * offset,LLVMValueRef * out_of_bounds)2165 lp_build_sample_ms_offset(struct lp_build_context *int_coord_bld,
2166                           LLVMValueRef ms_index,
2167                           LLVMValueRef num_samples,
2168                           LLVMValueRef sample_stride,
2169                           LLVMValueRef *offset,
2170                           LLVMValueRef *out_of_bounds)
2171 {
2172    LLVMValueRef out1;
2173    num_samples = lp_build_broadcast_scalar(int_coord_bld, num_samples);
2174    sample_stride = lp_build_broadcast_scalar(int_coord_bld, sample_stride);
2175    out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, ms_index, int_coord_bld->zero);
2176    *out_of_bounds = lp_build_or(int_coord_bld, *out_of_bounds, out1);
2177    out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, ms_index, num_samples);
2178    *out_of_bounds = lp_build_or(int_coord_bld, *out_of_bounds, out1);
2179    LLVMValueRef sample_offset = lp_build_mul(int_coord_bld,
2180                                              sample_stride, ms_index);
2181    *offset = lp_build_add(int_coord_bld, *offset, sample_offset);
2182 }
2183 
2184 
2185 #define WEIGHT_LUT_SIZE 1024
2186 
2187 
2188 static void
lp_build_sample_aniso(struct lp_build_sample_context * bld,unsigned img_filter,unsigned mip_filter,bool is_gather,const LLVMValueRef * coords,const LLVMValueRef * offsets,LLVMValueRef ilevel0,LLVMValueRef ilevel1,LLVMValueRef lod_fpart,LLVMValueRef * colors_out)2189 lp_build_sample_aniso(struct lp_build_sample_context *bld,
2190                       unsigned img_filter,
2191                       unsigned mip_filter,
2192                       bool is_gather,
2193                       const LLVMValueRef *coords,
2194                       const LLVMValueRef *offsets,
2195                       LLVMValueRef ilevel0,
2196                       LLVMValueRef ilevel1,
2197                       LLVMValueRef lod_fpart,
2198                       LLVMValueRef *colors_out)
2199 {
2200    struct gallivm_state *gallivm = bld->gallivm;
2201    LLVMBuilderRef builder = gallivm->builder;
2202    struct lp_build_context *coord_bld = &bld->coord_bld;
2203    struct lp_build_context *float_size_bld = &bld->float_size_in_bld;
2204    LLVMValueRef ddx_ddy = lp_build_packed_ddx_ddy_twocoord(&bld->coord_bld, coords[0], coords[1]);
2205    LLVMValueRef float_size;
2206    LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
2207    LLVMValueRef index0 = LLVMConstInt(i32t, 0, 0);
2208    LLVMValueRef index1 = LLVMConstInt(i32t, 1, 0);
2209    const unsigned length = bld->coord_bld.type.length;
2210    const unsigned num_quads = length / 4;
2211    LLVMValueRef filter_table = bld->aniso_filter_table;
2212    LLVMValueRef size0, row_stride0_vec, img_stride0_vec;
2213    LLVMValueRef data_ptr0, mipoff0 = NULL;
2214 
2215    lp_build_mipmap_level_sizes(bld, ilevel0,
2216                                &size0,
2217                                &row_stride0_vec, &img_stride0_vec);
2218    if (bld->num_mips == 1) {
2219       data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
2220    } else {
2221       /* This path should work for num_lods 1 too but slightly less efficient */
2222       data_ptr0 = bld->base_ptr;
2223       mipoff0 = lp_build_get_mip_offsets(bld, ilevel0);
2224    }
2225 
2226    float_size = lp_build_int_to_float(&bld->float_size_in_bld, bld->int_size);
2227 
2228    LLVMValueRef float_size_lvl = lp_build_int_to_float(&bld->float_size_bld, size0);
2229    /* extract width and height into vectors for use later */
2230    static const unsigned char swizzle15[] = { /* no-op swizzle */
2231       1, 1, 1, 1, 5, 5, 5, 5
2232    };
2233    static const unsigned char swizzle04[] = { /* no-op swizzle */
2234       0, 0, 0, 0, 4, 4, 4, 4
2235    };
2236    LLVMValueRef width_dim, height_dim;
2237 
2238    width_dim = lp_build_swizzle_aos_n(gallivm, float_size_lvl, swizzle04,
2239                                       bld->float_size_bld.type.length,
2240                                       bld->coord_bld.type.length);
2241    height_dim = lp_build_swizzle_aos_n(gallivm, float_size_lvl, swizzle15,
2242                                        bld->float_size_bld.type.length,
2243                                        bld->coord_bld.type.length);
2244 
2245 
2246    /* shuffle width/height for ddx/ddy calculations. */
2247    LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 4];
2248 
2249    for (unsigned i = 0; i < num_quads; i++) {
2250       shuffles[i*4+0] = shuffles[i*4+1] = index0;
2251       shuffles[i*4+2] = shuffles[i*4+3] = index1;
2252    }
2253 
2254    LLVMValueRef floatdim =
2255       LLVMBuildShuffleVector(builder, float_size, float_size,
2256                              LLVMConstVector(shuffles, length), "");
2257 
2258    ddx_ddy = lp_build_mul(coord_bld, ddx_ddy, floatdim);
2259 
2260    LLVMValueRef scaling =
2261       lp_build_shl(&bld->leveli_bld, bld->leveli_bld.one, ilevel0);
2262    scaling = lp_build_int_to_float(&bld->levelf_bld, scaling);
2263    scaling = lp_build_rcp(&bld->levelf_bld, scaling);
2264 
2265    if (bld->levelf_bld.type.length != length) {
2266       if (bld->levelf_bld.type.length == 1) {
2267          scaling = lp_build_broadcast_scalar(coord_bld,
2268                                              scaling);
2269       } else {
2270          scaling = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
2271                                                          bld->levelf_bld.type,
2272                                                          coord_bld->type,
2273                                                          scaling);
2274       }
2275    }
2276 
2277    ddx_ddy = lp_build_mul(coord_bld, ddx_ddy, scaling);
2278 
2279    static const unsigned char swizzle01[] = { /* no-op swizzle */
2280       0, 1, 0, 1,
2281    };
2282    static const unsigned char swizzle23[] = {
2283       2, 3, 2, 3,
2284    };
2285 
2286    LLVMValueRef ddx_ddys, ddx_ddyt;
2287    ddx_ddys = lp_build_swizzle_aos(coord_bld, ddx_ddy, swizzle01);
2288    ddx_ddyt = lp_build_swizzle_aos(coord_bld, ddx_ddy, swizzle23);
2289 
2290    /* compute ellipse coefficients */
2291    /* * A*x*x + B*x*y + C*y*y = F.*/
2292    /* float A = vx*vx+vy*vy+1; */
2293    LLVMValueRef A = lp_build_mul(coord_bld, ddx_ddyt, ddx_ddyt);
2294 
2295    LLVMValueRef Ay = lp_build_swizzle_aos(coord_bld, A, swizzle15);
2296    A = lp_build_add(coord_bld, A, Ay);
2297    A = lp_build_add(coord_bld, A, coord_bld->one);
2298    A = lp_build_swizzle_aos(coord_bld, A, swizzle04);
2299 
2300    /* float B = -2*(ux*vx+uy*vy); */
2301    LLVMValueRef B = lp_build_mul(coord_bld, ddx_ddys, ddx_ddyt);
2302    LLVMValueRef By = lp_build_swizzle_aos(coord_bld, B, swizzle15);
2303    B = lp_build_add(coord_bld, B, By);
2304    B = lp_build_mul_imm(coord_bld, B, -2);
2305    B = lp_build_swizzle_aos(coord_bld, B, swizzle04);
2306 
2307    /* float C = ux*ux+uy*uy+1; */
2308    LLVMValueRef C = lp_build_mul(coord_bld, ddx_ddys, ddx_ddys);
2309    LLVMValueRef Cy = lp_build_swizzle_aos(coord_bld, C, swizzle15);
2310    C = lp_build_add(coord_bld, C, Cy);
2311    C = lp_build_add(coord_bld, C, coord_bld->one);
2312    C = lp_build_swizzle_aos(coord_bld, C, swizzle04);
2313 
2314    /* float F = A*C-B*B/4.0f; */
2315    LLVMValueRef F = lp_build_mul(coord_bld, B, B);
2316    F = lp_build_div(coord_bld, F, lp_build_const_vec(gallivm, coord_bld->type, 4.0));
2317    LLVMValueRef F_p2 = lp_build_mul(coord_bld, A, C);
2318    F = lp_build_sub(coord_bld, F_p2, F);
2319 
2320    /* compute ellipse bounding box in texture space */
2321    /* const float d = -B*B+4.0f*C*A; */
2322    LLVMValueRef d = lp_build_sub(coord_bld, coord_bld->zero, lp_build_mul(coord_bld, B, B));
2323    LLVMValueRef d_p2 = lp_build_mul(coord_bld, A, C);
2324    d_p2 = lp_build_mul_imm(coord_bld, d_p2, 4);
2325    d = lp_build_add(coord_bld, d, d_p2);
2326 
2327    /* const float box_u = 2.0f / d * sqrtf(d*C*F); */
2328    /* box_u -> half of bbox with   */
2329    LLVMValueRef temp;
2330    temp = lp_build_mul(coord_bld, d, C);
2331    temp = lp_build_mul(coord_bld, temp, F);
2332    temp = lp_build_sqrt(coord_bld, temp);
2333 
2334    LLVMValueRef box_u = lp_build_div(coord_bld, lp_build_const_vec(gallivm, coord_bld->type, 2.0), d);
2335    box_u = lp_build_mul(coord_bld, box_u, temp);
2336 
2337    /* const float box_v = 2.0f / d * sqrtf(A*d*F); */
2338    /* box_v -> half of bbox height */
2339    temp = lp_build_mul(coord_bld, A, d);
2340    temp = lp_build_mul(coord_bld, temp, F);
2341    temp = lp_build_sqrt(coord_bld, temp);
2342 
2343    LLVMValueRef box_v = lp_build_div(coord_bld, lp_build_const_vec(gallivm, coord_bld->type, 2.0), d);
2344    box_v = lp_build_mul(coord_bld, box_v, temp);
2345 
2346    /* Scale ellipse formula to directly index the Filter Lookup Table.
2347     * i.e. scale so that F = WEIGHT_LUT_SIZE-1
2348     */
2349    LLVMValueRef formScale = lp_build_div(coord_bld, lp_build_const_vec(gallivm, coord_bld->type, WEIGHT_LUT_SIZE - 1), F);
2350 
2351    A = lp_build_mul(coord_bld, A, formScale);
2352    B = lp_build_mul(coord_bld, B, formScale);
2353    C = lp_build_mul(coord_bld, C, formScale);
2354    /* F *= formScale; */ /* no need to scale F as we don't use it below here */
2355 
2356    LLVMValueRef ddq = lp_build_mul_imm(coord_bld, A, 2);
2357 
2358    /* Heckbert MS thesis, p. 59; scan over the bounding box of the ellipse
2359     * and incrementally update the value of Ax^2+Bxy*Cy^2; when this
2360     * value, q, is less than F, we're inside the ellipse
2361     */
2362 
2363    LLVMValueRef float_size0 = lp_build_int_to_float(float_size_bld, bld->int_size);
2364    LLVMValueRef width0 = lp_build_extract_broadcast(gallivm,
2365                                                     float_size_bld->type,
2366                                                     coord_bld->type,
2367                                                     float_size0, index0);
2368    LLVMValueRef height0 = lp_build_extract_broadcast(gallivm,
2369                                                      float_size_bld->type,
2370                                                      coord_bld->type,
2371                                                      float_size0, index1);
2372 
2373    /* texture->width0 * scaling */
2374    width0 = lp_build_mul(coord_bld, width0, scaling);
2375    /* texture->height0 * scaling */
2376    height0 = lp_build_mul(coord_bld, height0, scaling);
2377 
2378    /* tex_u = -0.5f * s[j] * texture->width0 * scaling */
2379    LLVMValueRef tex_u = lp_build_mul(coord_bld, coords[0], width0);
2380    tex_u = lp_build_add(coord_bld, tex_u, lp_build_const_vec(gallivm, coord_bld->type, -0.5f));
2381 
2382    /* tex_v = -0.5f * t[j] * texture->height0 * scaling */
2383    LLVMValueRef tex_v = lp_build_mul(coord_bld, coords[1], height0);
2384    tex_v = lp_build_add(coord_bld, tex_v, lp_build_const_vec(gallivm, coord_bld->type, -0.5f));
2385 
2386    /* const int u0 = (int) floorf(tex_u - box_u); */
2387    LLVMValueRef u0 = lp_build_itrunc(coord_bld, lp_build_floor(coord_bld, lp_build_sub(coord_bld, tex_u, box_u)));
2388    /* const int u1 = (int) ceilf(tex_u + box_u); */
2389    LLVMValueRef u1 = lp_build_itrunc(coord_bld, lp_build_ceil(coord_bld, lp_build_add(coord_bld, tex_u, box_u)));
2390 
2391    /* const int v0 = (int) floorf(tex_v - box_v); */
2392    LLVMValueRef v0 = lp_build_itrunc(coord_bld, lp_build_floor(coord_bld, lp_build_sub(coord_bld, tex_v, box_v)));
2393    /* const int v1 = (int) ceilf(tex_v + box_v); */
2394    LLVMValueRef v1 = lp_build_itrunc(coord_bld, lp_build_ceil(coord_bld, lp_build_add(coord_bld, tex_v, box_v)));
2395 
2396    /* const float U = u0 - tex_u; */
2397    LLVMValueRef U = lp_build_sub(coord_bld, lp_build_int_to_float(coord_bld, u0), tex_u);
2398 
2399    /* A * (2 * U + 1) */
2400    LLVMValueRef dq_base = lp_build_mul_imm(coord_bld, U, 2);
2401    dq_base = lp_build_add(coord_bld, dq_base, coord_bld->one);
2402    dq_base = lp_build_mul(coord_bld, dq_base, A);
2403 
2404    /* A * U * U */
2405    LLVMValueRef q_base = lp_build_mul(coord_bld, U, U);
2406    q_base = lp_build_mul(coord_bld, q_base, A);
2407 
2408    LLVMValueRef colors0[4];
2409    LLVMValueRef den_store = lp_build_alloca(gallivm, bld->texel_bld.vec_type, "den");
2410 
2411    for (unsigned chan = 0; chan < 4; chan++)
2412       colors0[chan] = lp_build_alloca(gallivm, bld->texel_bld.vec_type, "colors");
2413 
2414    LLVMValueRef q_store, dq_store;
2415    q_store = lp_build_alloca(gallivm, bld->coord_bld.vec_type, "q");
2416    dq_store = lp_build_alloca(gallivm, bld->coord_bld.vec_type, "dq");
2417 
2418    LLVMValueRef v_limiter = lp_build_alloca(gallivm, bld->int_coord_bld.vec_type, "v_limiter");
2419    LLVMValueRef u_limiter = lp_build_alloca(gallivm, bld->int_coord_bld.vec_type, "u_limiter");
2420 
2421    LLVMBuildStore(builder, v0, v_limiter);
2422 
2423    /* create an LLVM loop block for the V iterator */
2424    LLVMBasicBlockRef v_loop_block = lp_build_insert_new_block(gallivm, "vloop");
2425 
2426    LLVMBuildBr(builder, v_loop_block);
2427    LLVMPositionBuilderAtEnd(builder, v_loop_block);
2428 
2429    LLVMValueRef v_val = LLVMBuildLoad2(builder, bld->int_coord_bld.vec_type, v_limiter, "");
2430    LLVMValueRef v_mask = LLVMBuildICmp(builder, LLVMIntSLE, v_val, v1, "");
2431 
2432    /* loop over V values. */
2433    {
2434       /*  const float V = v - tex_v; */
2435       LLVMValueRef V =
2436          lp_build_sub(coord_bld,
2437                       lp_build_int_to_float(coord_bld, v_val), tex_v);
2438 
2439       /* float dq = dq_base + B * V; */
2440       LLVMValueRef dq = lp_build_mul(coord_bld, V, B);
2441       dq = lp_build_add(coord_bld, dq, dq_base);
2442 
2443       /* float q = (C * V + B * U) * V + q_base */
2444       LLVMValueRef q = lp_build_mul(coord_bld, C, V);
2445       q = lp_build_add(coord_bld, q, lp_build_mul(coord_bld, B, U));
2446       q = lp_build_mul(coord_bld, q, V);
2447       q = lp_build_add(coord_bld, q, q_base);
2448 
2449       LLVMBuildStore(builder, q, q_store);
2450       LLVMBuildStore(builder, dq, dq_store);
2451 
2452       LLVMBuildStore(builder, u0, u_limiter);
2453 
2454       /* create an LLVM loop block for the V iterator */
2455       LLVMBasicBlockRef u_loop_block = lp_build_insert_new_block(gallivm, "uloop");
2456 
2457       LLVMBuildBr(builder, u_loop_block);
2458       LLVMPositionBuilderAtEnd(builder, u_loop_block);
2459 
2460       LLVMValueRef u_val = LLVMBuildLoad2(builder, bld->int_coord_bld.vec_type,
2461                                           u_limiter, "");
2462       LLVMValueRef u_mask = LLVMBuildICmp(builder,
2463                                           LLVMIntSLE,
2464                                           u_val,
2465                                           u1, "");
2466 
2467       /* loop over U values */
2468       {
2469          /* q = (int)q */
2470          q = lp_build_itrunc(coord_bld,
2471                              LLVMBuildLoad2(builder, bld->coord_bld.vec_type,
2472                                             q_store, ""));
2473 
2474          /*
2475           * avoid OOB access to filter table, generate a mask for q > 1024,
2476           * then truncate it.
2477           */
2478          LLVMValueRef q_mask = LLVMBuildICmp(builder,
2479                                              LLVMIntSLE,
2480                                              q,
2481                                              lp_build_const_int_vec(gallivm, bld->int_coord_bld.type, 0x3ff), "");
2482          q_mask = LLVMBuildSExt(builder, q_mask, bld->int_coord_bld.vec_type, "");
2483 
2484          q = lp_build_max(&bld->int_coord_bld, q, bld->int_coord_bld.zero);
2485          q = lp_build_and(&bld->int_coord_bld, q, lp_build_const_int_vec(gallivm, bld->int_coord_bld.type, 0x3ff));
2486 
2487          /* update the offsets to deal with float size. */
2488          q = lp_build_mul_imm(&bld->int_coord_bld, q, 4);
2489          filter_table = LLVMBuildBitCast(gallivm->builder, filter_table, LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0), "");
2490 
2491          /* Lookup weights in filter table */
2492          LLVMValueRef weights = lp_build_gather(gallivm, coord_bld->type.length,
2493                                                 coord_bld->type.width,
2494                                                 lp_elem_type(coord_bld->type),
2495                                                 true, filter_table, q, true);
2496 
2497          /*
2498           * Mask off the weights here which should ensure no-op for loops
2499           * where some of the u/v values are not being calculated.
2500           */
2501          weights = LLVMBuildBitCast(builder, weights, bld->int_coord_bld.vec_type, "");
2502          weights = lp_build_and(&bld->int_coord_bld, weights, LLVMBuildSExt(builder, u_mask, bld->int_coord_bld.vec_type, ""));
2503          weights = lp_build_and(&bld->int_coord_bld, weights, LLVMBuildSExt(builder, v_mask, bld->int_coord_bld.vec_type, ""));
2504          weights = lp_build_and(&bld->int_coord_bld, weights, q_mask);
2505          weights = LLVMBuildBitCast(builder, weights, bld->coord_bld.vec_type, "");
2506 
2507          /* if the weights are all 0 avoid doing the sampling at all. */
2508          struct lp_build_if_state noloadw0;
2509 
2510          LLVMValueRef wnz = LLVMBuildFCmp(gallivm->builder, LLVMRealUNE,
2511                                           weights, bld->coord_bld.zero, "");
2512          wnz = LLVMBuildSExt(builder, wnz, bld->int_coord_bld.vec_type, "");
2513          wnz = lp_build_any_true_range(&bld->coord_bld, bld->coord_bld.type.length, wnz);
2514          lp_build_if(&noloadw0, gallivm, wnz);
2515          LLVMValueRef new_coords[4];
2516          new_coords[0] = lp_build_div(coord_bld, lp_build_int_to_float(coord_bld, u_val), width_dim);
2517          new_coords[1] = lp_build_div(coord_bld, lp_build_int_to_float(coord_bld, v_val), height_dim);
2518          new_coords[2] = coords[2];
2519          new_coords[3] = coords[3];
2520 
2521          /* lookup q in filter table */
2522          LLVMValueRef temp_colors[4];
2523          lp_build_sample_image_nearest(bld, size0,
2524                                        row_stride0_vec, img_stride0_vec,
2525                                        data_ptr0, mipoff0, ilevel0, new_coords, offsets,
2526                                        temp_colors);
2527 
2528          for (unsigned chan = 0; chan < 4; chan++) {
2529             LLVMValueRef tcolor = LLVMBuildLoad2(builder, bld->texel_bld.vec_type, colors0[chan], "");
2530 
2531             tcolor = lp_build_add(&bld->texel_bld, tcolor, lp_build_mul(&bld->texel_bld, temp_colors[chan], weights));
2532             LLVMBuildStore(builder, tcolor, colors0[chan]);
2533          }
2534 
2535          /* multiple colors by weight and add in. */
2536          /* den += weight; */
2537          LLVMValueRef den = LLVMBuildLoad2(builder, bld->texel_bld.vec_type, den_store, "");
2538          den = lp_build_add(&bld->texel_bld, den, weights);
2539          LLVMBuildStore(builder, den, den_store);
2540 
2541          lp_build_endif(&noloadw0);
2542          /* q += dq; */
2543          /* dq += ddq; */
2544          q = LLVMBuildLoad2(builder, bld->texel_bld.vec_type, q_store, "");
2545          dq = LLVMBuildLoad2(builder, bld->texel_bld.vec_type, dq_store, "");
2546          q = lp_build_add(coord_bld, q, dq);
2547          dq = lp_build_add(coord_bld, dq, ddq);
2548          LLVMBuildStore(builder, q, q_store);
2549          LLVMBuildStore(builder, dq, dq_store);
2550       }
2551       /* u += 1 */
2552       u_val = LLVMBuildLoad2(builder, bld->int_coord_bld.vec_type, u_limiter, "");
2553       u_val = lp_build_add(&bld->int_coord_bld, u_val, bld->int_coord_bld.one);
2554       LLVMBuildStore(builder, u_val, u_limiter);
2555 
2556       u_mask = LLVMBuildICmp(builder,
2557                              LLVMIntSLE,
2558                              u_val,
2559                              u1, "");
2560       LLVMValueRef u_end_cond = LLVMBuildSExt(builder, u_mask, bld->int_coord_bld.vec_type, "");
2561       u_end_cond = lp_build_any_true_range(&bld->coord_bld, bld->coord_bld.type.length, u_end_cond);
2562 
2563       LLVMBasicBlockRef u_end_loop = lp_build_insert_new_block(gallivm, "u_end_loop");
2564 
2565       LLVMBuildCondBr(builder, u_end_cond,
2566                       u_loop_block, u_end_loop);
2567 
2568       LLVMPositionBuilderAtEnd(builder, u_end_loop);
2569 
2570    }
2571 
2572    /* v += 1 */
2573    v_val = LLVMBuildLoad2(builder, bld->int_coord_bld.vec_type, v_limiter, "");
2574    v_val = lp_build_add(&bld->int_coord_bld, v_val, bld->int_coord_bld.one);
2575    LLVMBuildStore(builder, v_val, v_limiter);
2576 
2577    v_mask = LLVMBuildICmp(builder,
2578                           LLVMIntSLE,
2579                           v_val,
2580                           v1, "");
2581    LLVMValueRef v_end_cond = LLVMBuildSExt(builder, v_mask,
2582                                            bld->int_coord_bld.vec_type, "");
2583    v_end_cond = lp_build_any_true_range(&bld->coord_bld,
2584                                         bld->coord_bld.type.length, v_end_cond);
2585 
2586    LLVMBasicBlockRef v_end_loop = lp_build_insert_new_block(gallivm, "v_end_loop");
2587 
2588    LLVMBuildCondBr(builder, v_end_cond,
2589                    v_loop_block, v_end_loop);
2590 
2591    LLVMPositionBuilderAtEnd(builder, v_end_loop);
2592 
2593    LLVMValueRef den = LLVMBuildLoad2(builder, bld->texel_bld.vec_type, den_store, "");
2594 
2595    for (unsigned chan = 0; chan < 4; chan++) {
2596       colors0[chan] =
2597          lp_build_div(&bld->texel_bld,
2598                       LLVMBuildLoad2(builder, bld->texel_bld.vec_type,
2599                                      colors0[chan], ""), den);
2600    }
2601 
2602    LLVMValueRef den0 = lp_build_cmp(&bld->coord_bld, PIPE_FUNC_EQUAL,
2603                                     den, bld->coord_bld.zero);
2604 
2605    LLVMValueRef den0_any =
2606       lp_build_any_true_range(&bld->coord_bld,
2607                               bld->coord_bld.type.length, den0);
2608 
2609    struct lp_build_if_state den0_fallback;
2610    lp_build_if(&den0_fallback, gallivm, den0_any);
2611    {
2612       LLVMValueRef colors_den0[4];
2613       lp_build_sample_image_linear(bld, false, size0, NULL,
2614                                    row_stride0_vec, img_stride0_vec,
2615                                    data_ptr0, mipoff0, ilevel0, coords, offsets,
2616                                    colors_den0);
2617       for (unsigned chan = 0; chan < 4; chan++) {
2618          LLVMValueRef chan_val =
2619             lp_build_select(&bld->texel_bld, den0,
2620                             colors_den0[chan], colors0[chan]);
2621          LLVMBuildStore(builder, chan_val, colors_out[chan]);
2622       }
2623    }
2624    lp_build_else(&den0_fallback);
2625    {
2626       for (unsigned chan = 0; chan < 4; chan++) {
2627          LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
2628       }
2629    }
2630    lp_build_endif(&den0_fallback);
2631 }
2632 
2633 
2634 /**
2635  * Calculate cube face, lod, mip levels.
2636  */
2637 static void
lp_build_sample_common(struct lp_build_sample_context * bld,bool is_lodq,unsigned texture_index,unsigned sampler_index,LLVMValueRef * coords,const struct lp_derivatives * derivs,LLVMValueRef lod_bias,LLVMValueRef explicit_lod,LLVMValueRef * lod_pos_or_zero,LLVMValueRef * lod,LLVMValueRef * lod_fpart,LLVMValueRef * ilevel0,LLVMValueRef * ilevel1)2638 lp_build_sample_common(struct lp_build_sample_context *bld,
2639                        bool is_lodq,
2640                        unsigned texture_index,
2641                        unsigned sampler_index,
2642                        LLVMValueRef *coords,
2643                        const struct lp_derivatives *derivs, /* optional */
2644                        LLVMValueRef lod_bias, /* optional */
2645                        LLVMValueRef explicit_lod, /* optional */
2646                        LLVMValueRef *lod_pos_or_zero,
2647                        LLVMValueRef *lod,
2648                        LLVMValueRef *lod_fpart,
2649                        LLVMValueRef *ilevel0,
2650                        LLVMValueRef *ilevel1)
2651 {
2652    const unsigned mip_filter = bld->static_sampler_state->min_mip_filter;
2653    const unsigned min_filter = bld->static_sampler_state->min_img_filter;
2654    const unsigned mag_filter = bld->static_sampler_state->mag_img_filter;
2655    const unsigned target = bld->static_texture_state->target;
2656    const bool aniso = bld->static_sampler_state->aniso;
2657    LLVMValueRef first_level, last_level;
2658    LLVMValueRef lod_ipart = NULL;
2659    struct lp_derivatives cube_derivs;
2660 
2661    /*
2662    printf("%s mip %d  min %d  mag %d\n", __func__,
2663           mip_filter, min_filter, mag_filter);
2664    */
2665 
2666    first_level = get_first_level(bld->gallivm,
2667                                  bld->resources_type,
2668                                  bld->resources_ptr,
2669                                  texture_index, NULL,
2670                                  bld->static_texture_state,
2671                                  bld->dynamic_state);
2672    last_level = get_last_level(bld->gallivm,
2673                                bld->resources_type,
2674                                bld->resources_ptr,
2675                                texture_index, NULL,
2676                                bld->static_texture_state,
2677                                bld->dynamic_state);
2678 
2679    /*
2680     * Choose cube face, recompute texcoords for the chosen face and
2681     * calculate / transform derivatives.
2682     */
2683    if (target == PIPE_TEXTURE_CUBE || target == PIPE_TEXTURE_CUBE_ARRAY) {
2684       bool need_derivs = ((min_filter != mag_filter ||
2685                            mip_filter != PIPE_TEX_MIPFILTER_NONE) &&
2686                           !bld->static_sampler_state->min_max_lod_equal &&
2687                           !explicit_lod);
2688       lp_build_cube_lookup(bld, coords, derivs, &cube_derivs, need_derivs);
2689       if (need_derivs)
2690          derivs = &cube_derivs;
2691 
2692       if (target == PIPE_TEXTURE_CUBE_ARRAY && !is_lodq) {
2693          /* calculate cube layer coord now */
2694          LLVMValueRef layer = lp_build_iround(&bld->coord_bld, coords[3]);
2695          LLVMValueRef six = lp_build_const_int_vec(bld->gallivm, bld->int_coord_type, 6);
2696          layer = lp_build_mul(&bld->int_coord_bld, layer, six);
2697          coords[3] = lp_build_layer_coord(bld, texture_index, true, layer, NULL);
2698          /* because of seamless filtering can't add it to face (coords[2]) here. */
2699       }
2700    } else if ((target == PIPE_TEXTURE_1D_ARRAY ||
2701              target == PIPE_TEXTURE_2D_ARRAY) && !is_lodq) {
2702       coords[2] = lp_build_iround(&bld->coord_bld, coords[2]);
2703       coords[2] = lp_build_layer_coord(bld, texture_index, false, coords[2], NULL);
2704    }
2705 
2706    if (bld->static_sampler_state->compare_mode != PIPE_TEX_COMPARE_NONE) {
2707       /*
2708        * Clamp p coords to [0,1] for fixed function depth texture format here.
2709        * Technically this is not entirely correct for unorm depth as the ref
2710        * value should be converted to the depth format (quantization!) and
2711        * comparison then done in texture format. This would actually help
2712        * performance (since only need to do it once and could save the
2713        * per-sample conversion of texels to floats instead), but it would need
2714        * more messy code (would need to push at least some bits down to actual
2715        * fetch so conversion could be skipped, and would have ugly interaction
2716        * with border color, would need to convert border color to that format
2717        * too or do some other tricks to make it work).
2718        */
2719       const struct util_format_description *format_desc = bld->format_desc;
2720       /* not entirely sure we couldn't end up with non-valid swizzle here */
2721       const enum util_format_type chan_type =
2722          format_desc->swizzle[0] <= PIPE_SWIZZLE_W
2723            ? format_desc->channel[format_desc->swizzle[0]].type
2724            : UTIL_FORMAT_TYPE_FLOAT;
2725       if (chan_type != UTIL_FORMAT_TYPE_FLOAT) {
2726          coords[4] = lp_build_clamp(&bld->coord_bld, coords[4],
2727                                     bld->coord_bld.zero, bld->coord_bld.one);
2728       }
2729    }
2730 
2731    /*
2732     * Compute the level of detail (float).
2733     */
2734    if (min_filter != mag_filter ||
2735        mip_filter != PIPE_TEX_MIPFILTER_NONE || is_lodq) {
2736       LLVMValueRef max_aniso = NULL;
2737 
2738       if (aniso)
2739          max_aniso = bld->dynamic_state->max_aniso(bld->gallivm,
2740                                                    bld->resources_type,
2741                                                    bld->resources_ptr,
2742                                                    sampler_index);
2743 
2744       /* Need to compute lod either to choose mipmap levels or to
2745        * distinguish between minification/magnification with one mipmap level.
2746        */
2747       LLVMValueRef first_level_vec =
2748          lp_build_broadcast_scalar(&bld->int_size_in_bld, first_level);
2749       lp_build_lod_selector(bld, is_lodq, sampler_index,
2750                             first_level_vec,
2751                             coords[0], coords[1], coords[2],
2752                             derivs, lod_bias, explicit_lod,
2753                             mip_filter, max_aniso, lod,
2754                             &lod_ipart, lod_fpart, lod_pos_or_zero);
2755       if (is_lodq) {
2756          last_level = lp_build_sub(&bld->int_bld, last_level, first_level);
2757          last_level = lp_build_int_to_float(&bld->float_bld, last_level);
2758          last_level = lp_build_broadcast_scalar(&bld->lodf_bld, last_level);
2759 
2760          switch (mip_filter) {
2761          case PIPE_TEX_MIPFILTER_NONE:
2762             *lod_fpart = bld->lodf_bld.zero;
2763             break;
2764          case PIPE_TEX_MIPFILTER_NEAREST:
2765             *lod_fpart = lp_build_round(&bld->lodf_bld, *lod_fpart);
2766             FALLTHROUGH;
2767          case PIPE_TEX_MIPFILTER_LINEAR:
2768             *lod_fpart = lp_build_clamp(&bld->lodf_bld, *lod_fpart,
2769                                         bld->lodf_bld.zero, last_level);
2770             break;
2771          }
2772          return;
2773       }
2774    } else {
2775       lod_ipart = bld->lodi_bld.zero;
2776       *lod_pos_or_zero = bld->lodi_bld.zero;
2777    }
2778 
2779    if ((bld->num_lods != bld->num_mips || bld->num_lods == 1) &&
2780        bld->lodi_bld.type.length != 1) {
2781       /* only makes sense if there's just a single mip level */
2782       assert(bld->num_mips == 1);
2783       lod_ipart = lp_build_extract_range(bld->gallivm, lod_ipart, 0, 1);
2784    }
2785 
2786    first_level = lp_build_broadcast_scalar(&bld->leveli_bld, first_level);
2787    last_level = lp_build_broadcast_scalar(&bld->leveli_bld, last_level);
2788 
2789    /*
2790     * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
2791     */
2792 
2793    if (aniso) {
2794       lp_build_nearest_mip_level(bld,
2795                                  first_level, last_level,
2796                                  lod_ipart, ilevel0, NULL);
2797       return;
2798    }
2799 
2800    switch (mip_filter) {
2801    default:
2802       unreachable("Bad mip_filter value in lp_build_sample_soa()");
2803    case PIPE_TEX_MIPFILTER_NONE:
2804       /* always use mip level 0 */
2805       *ilevel0 = first_level;
2806       break;
2807    case PIPE_TEX_MIPFILTER_NEAREST:
2808       assert(lod_ipart);
2809       lp_build_nearest_mip_level(bld,
2810                                  first_level, last_level,
2811                                  lod_ipart, ilevel0, NULL);
2812       break;
2813    case PIPE_TEX_MIPFILTER_LINEAR:
2814       assert(lod_ipart);
2815       assert(*lod_fpart);
2816 
2817       lp_build_linear_mip_levels(bld, texture_index,
2818                                  first_level, last_level,
2819                                  lod_ipart, lod_fpart,
2820                                  ilevel0, ilevel1);
2821       break;
2822    }
2823 }
2824 
2825 
2826 static void
lp_build_clamp_border_color(struct lp_build_sample_context * bld,unsigned sampler_unit)2827 lp_build_clamp_border_color(struct lp_build_sample_context *bld,
2828                             unsigned sampler_unit)
2829 {
2830    struct gallivm_state *gallivm = bld->gallivm;
2831    LLVMBuilderRef builder = gallivm->builder;
2832    LLVMValueRef border_color_ptr =
2833       bld->dynamic_state->border_color(gallivm,
2834                                        bld->resources_type,
2835                                        bld->resources_ptr, sampler_unit);
2836    LLVMValueRef border_color;
2837    const struct util_format_description *format_desc = bld->format_desc;
2838    struct lp_type vec4_type = bld->texel_type;
2839    struct lp_build_context vec4_bld;
2840    LLVMValueRef min_clamp = NULL;
2841    LLVMValueRef max_clamp = NULL;
2842 
2843    /*
2844     * For normalized format need to clamp border color (technically
2845     * probably should also quantize the data). Really sucks doing this
2846     * here but can't avoid at least for now since this is part of
2847     * sampler state and texture format is part of sampler_view state.
2848     * GL expects also expects clamping for uint/sint formats too so
2849     * do that as well (d3d10 can't end up here with uint/sint since it
2850     * only supports them with ld).
2851     */
2852    vec4_type.length = 4;
2853    lp_build_context_init(&vec4_bld, gallivm, vec4_type);
2854 
2855    /*
2856     * Vectorized clamping of border color. Loading is a bit of a hack since
2857     * we just cast the pointer to float array to pointer to vec4
2858     * (int or float).
2859     */
2860    LLVMTypeRef border_color_type = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
2861    border_color_ptr = lp_build_array_get_ptr2(gallivm, border_color_type, border_color_ptr,
2862                                               lp_build_const_int32(gallivm, 0));
2863    border_color_ptr = LLVMBuildBitCast(builder, border_color_ptr,
2864                                        LLVMPointerType(vec4_bld.vec_type, 0), "");
2865    border_color = LLVMBuildLoad2(builder, vec4_bld.vec_type, border_color_ptr, "");
2866    /* we don't have aligned type in the dynamic state unfortunately */
2867    LLVMSetAlignment(border_color, 4);
2868 
2869    /*
2870     * Instead of having some incredibly complex logic which will try to figure
2871     * out clamping necessary for each channel, simply use the first channel,
2872     * and treat mixed signed/unsigned normalized formats specially.  (Mixed
2873     * non-normalized, which wouldn't work at all here, do not exist for a good
2874     * reason.)
2875     */
2876    if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN) {
2877       int chan;
2878       /* d/s needs special handling because both present means just sampling depth */
2879       if (util_format_is_depth_and_stencil(format_desc->format)) {
2880          chan = format_desc->swizzle[0];
2881       } else {
2882          chan = util_format_get_first_non_void_channel(format_desc->format);
2883       }
2884       if (chan >= 0 && chan <= PIPE_SWIZZLE_W) {
2885          unsigned chan_type = format_desc->channel[chan].type;
2886          unsigned chan_norm = format_desc->channel[chan].normalized;
2887          unsigned chan_pure = format_desc->channel[chan].pure_integer;
2888          if (chan_type == UTIL_FORMAT_TYPE_SIGNED) {
2889             if (chan_norm) {
2890                min_clamp = lp_build_const_vec(gallivm, vec4_type, -1.0F);
2891                max_clamp = vec4_bld.one;
2892             } else if (chan_pure) {
2893                /*
2894                 * Border color was stored as int, hence need min/max clamp
2895                 * only if chan has less than 32 bits..
2896                 */
2897                unsigned chan_size = format_desc->channel[chan].size;
2898                if (chan_size < 32) {
2899                   min_clamp = lp_build_const_int_vec(gallivm, vec4_type,
2900                                                      0 - (1 << (chan_size - 1)));
2901                   max_clamp = lp_build_const_int_vec(gallivm, vec4_type,
2902                                                      (1 << (chan_size - 1)) - 1);
2903                }
2904             }
2905             /* TODO: no idea about non-pure, non-normalized! */
2906          } else if (chan_type == UTIL_FORMAT_TYPE_UNSIGNED) {
2907             if (chan_norm) {
2908                min_clamp = vec4_bld.zero;
2909                max_clamp = vec4_bld.one;
2910             } else if (chan_pure) {
2911                /*
2912                 * Need a ugly hack here, because we don't have Z32_FLOAT_X8X24
2913                 * we use Z32_FLOAT_S8X24 to imply sampling depth component and
2914                 * ignoring stencil, which will blow up here if we try to do a
2915                 * uint clamp in a float texel build...  And even if we had
2916                 * that format, mesa st also thinks using z24s8 means depth
2917                 * sampling ignoring stencil.
2918                 */
2919 
2920                /*
2921                 * Border color was stored as uint, hence never need min clamp,
2922                 * and only need max clamp if chan has less than 32 bits.
2923                 */
2924                unsigned chan_size = format_desc->channel[chan].size;
2925                if (chan_size < 32) {
2926                   max_clamp = lp_build_const_int_vec(gallivm, vec4_type,
2927                                                      (1 << chan_size) - 1);
2928                }
2929                /* TODO: no idea about non-pure, non-normalized! */
2930             }
2931          } else if (chan_type == UTIL_FORMAT_TYPE_FIXED) {
2932             /* TODO: I have no idea what clamp this would need if any! */
2933          }
2934       }
2935       /* mixed plain formats (or different pure size) */
2936       switch (format_desc->format) {
2937       case PIPE_FORMAT_B10G10R10A2_UINT:
2938       case PIPE_FORMAT_R10G10B10A2_UINT:
2939          {
2940             unsigned max10 = (1 << 10) - 1;
2941             max_clamp = lp_build_const_aos(gallivm, vec4_type, max10, max10,
2942                                            max10, (1 << 2) - 1, NULL);
2943          }
2944          break;
2945       case PIPE_FORMAT_R10SG10SB10SA2U_NORM:
2946          min_clamp = lp_build_const_aos(gallivm, vec4_type, -1.0F, -1.0F,
2947                                         -1.0F, 0.0F, NULL);
2948          max_clamp = vec4_bld.one;
2949          break;
2950       case PIPE_FORMAT_R8SG8SB8UX8U_NORM:
2951       case PIPE_FORMAT_R5SG5SB6U_NORM:
2952          min_clamp = lp_build_const_aos(gallivm, vec4_type, -1.0F, -1.0F,
2953                                         0.0F, 0.0F, NULL);
2954          max_clamp = vec4_bld.one;
2955          break;
2956       default:
2957          break;
2958       }
2959    } else {
2960       /* cannot figure this out from format description */
2961       if (format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
2962          /* s3tc formats are always unorm */
2963          min_clamp = vec4_bld.zero;
2964          max_clamp = vec4_bld.one;
2965       } else if (format_desc->layout == UTIL_FORMAT_LAYOUT_RGTC ||
2966                  format_desc->layout == UTIL_FORMAT_LAYOUT_ETC ||
2967                  format_desc->layout == UTIL_FORMAT_LAYOUT_BPTC) {
2968          switch (format_desc->format) {
2969          case PIPE_FORMAT_RGTC1_UNORM:
2970          case PIPE_FORMAT_RGTC2_UNORM:
2971          case PIPE_FORMAT_LATC1_UNORM:
2972          case PIPE_FORMAT_LATC2_UNORM:
2973          case PIPE_FORMAT_ETC1_RGB8:
2974          case PIPE_FORMAT_BPTC_RGBA_UNORM:
2975          case PIPE_FORMAT_BPTC_SRGBA:
2976             min_clamp = vec4_bld.zero;
2977             max_clamp = vec4_bld.one;
2978             break;
2979          case PIPE_FORMAT_RGTC1_SNORM:
2980          case PIPE_FORMAT_RGTC2_SNORM:
2981          case PIPE_FORMAT_LATC1_SNORM:
2982          case PIPE_FORMAT_LATC2_SNORM:
2983             min_clamp = lp_build_const_vec(gallivm, vec4_type, -1.0F);
2984             max_clamp = vec4_bld.one;
2985             break;
2986          case PIPE_FORMAT_BPTC_RGB_FLOAT:
2987             /* not sure if we should clamp to max half float? */
2988             break;
2989          case PIPE_FORMAT_BPTC_RGB_UFLOAT:
2990             min_clamp = vec4_bld.zero;
2991             break;
2992          default:
2993             assert(0);
2994             break;
2995          }
2996       } else if (format_desc->colorspace != UTIL_FORMAT_COLORSPACE_YUV){
2997          /*
2998           * all others from subsampled/other group, though we don't care
2999           * about yuv (and should not have any from zs here)
3000           */
3001          switch (format_desc->format) {
3002          case PIPE_FORMAT_R8G8_B8G8_UNORM:
3003          case PIPE_FORMAT_G8R8_G8B8_UNORM:
3004          case PIPE_FORMAT_G8R8_B8R8_UNORM:
3005          case PIPE_FORMAT_R8G8_R8B8_UNORM:
3006          case PIPE_FORMAT_G8B8_G8R8_UNORM:
3007          case PIPE_FORMAT_B8G8_R8G8_UNORM:
3008          case PIPE_FORMAT_R1_UNORM: /* doesn't make sense but ah well */
3009             min_clamp = vec4_bld.zero;
3010             max_clamp = vec4_bld.one;
3011             break;
3012          case PIPE_FORMAT_R8G8Bx_SNORM:
3013             min_clamp = lp_build_const_vec(gallivm, vec4_type, -1.0F);
3014             max_clamp = vec4_bld.one;
3015             break;
3016             /*
3017              * Note smallfloat formats usually don't need clamping
3018              * (they still have infinite range) however this is not
3019              * true for r11g11b10 and r9g9b9e5, which can't represent
3020              * negative numbers (and additionally r9g9b9e5 can't represent
3021              * very large numbers). d3d10 seems happy without clamping in
3022              * this case, but gl spec is pretty clear: "for floating
3023              * point and integer formats, border values are clamped to
3024              * the representable range of the format" so do that here.
3025              */
3026          case PIPE_FORMAT_R11G11B10_FLOAT:
3027             min_clamp = vec4_bld.zero;
3028             break;
3029          case PIPE_FORMAT_R9G9B9E5_FLOAT:
3030             min_clamp = vec4_bld.zero;
3031             max_clamp = lp_build_const_vec(gallivm, vec4_type, MAX_RGB9E5);
3032             break;
3033          default:
3034             assert(0);
3035             break;
3036          }
3037       }
3038    }
3039 
3040    if (min_clamp) {
3041       border_color = lp_build_max(&vec4_bld, border_color, min_clamp);
3042    }
3043    if (max_clamp) {
3044       border_color = lp_build_min(&vec4_bld, border_color, max_clamp);
3045    }
3046 
3047    bld->border_color_clamped = border_color;
3048 }
3049 
3050 
3051 /**
3052  * General texture sampling codegen.
3053  * This function handles texture sampling for all texture targets (1D,
3054  * 2D, 3D, cube) and all filtering modes.
3055  */
3056 static void
lp_build_sample_general(struct lp_build_sample_context * bld,unsigned sampler_unit,bool is_gather,const LLVMValueRef * coords,const LLVMValueRef * offsets,LLVMValueRef lod_positive,LLVMValueRef lod_fpart,LLVMValueRef ilevel0,LLVMValueRef ilevel1,LLVMValueRef * colors_out)3057 lp_build_sample_general(struct lp_build_sample_context *bld,
3058                         unsigned sampler_unit,
3059                         bool is_gather,
3060                         const LLVMValueRef *coords,
3061                         const LLVMValueRef *offsets,
3062                         LLVMValueRef lod_positive,
3063                         LLVMValueRef lod_fpart,
3064                         LLVMValueRef ilevel0,
3065                         LLVMValueRef ilevel1,
3066                         LLVMValueRef *colors_out)
3067 {
3068    LLVMBuilderRef builder = bld->gallivm->builder;
3069    const struct lp_static_sampler_state *sampler_state = bld->static_sampler_state;
3070    const unsigned mip_filter = sampler_state->min_mip_filter;
3071    const unsigned min_filter = sampler_state->min_img_filter;
3072    const unsigned mag_filter = sampler_state->mag_img_filter;
3073    LLVMValueRef texels[4];
3074    unsigned chan;
3075 
3076    /* if we need border color, (potentially) clamp it now */
3077    if (lp_sampler_wrap_mode_uses_border_color(sampler_state->wrap_s,
3078                                               min_filter,
3079                                               mag_filter) ||
3080        (bld->dims > 1 &&
3081            lp_sampler_wrap_mode_uses_border_color(sampler_state->wrap_t,
3082                                                   min_filter,
3083                                                   mag_filter)) ||
3084        (bld->dims > 2 &&
3085            lp_sampler_wrap_mode_uses_border_color(sampler_state->wrap_r,
3086                                                   min_filter,
3087                                                   mag_filter))) {
3088       lp_build_clamp_border_color(bld, sampler_unit);
3089    }
3090 
3091 
3092    /*
3093     * Get/interpolate texture colors.
3094     */
3095 
3096    for (chan = 0; chan < 4; ++chan) {
3097      texels[chan] = lp_build_alloca(bld->gallivm, bld->texel_bld.vec_type, "");
3098      lp_build_name(texels[chan], "sampler%u_texel_%c_var", sampler_unit, "xyzw"[chan]);
3099    }
3100 
3101    if (sampler_state->aniso) {
3102       lp_build_sample_aniso(bld, PIPE_TEX_FILTER_NEAREST, mip_filter,
3103                             false, coords, offsets, ilevel0,
3104                             ilevel1, lod_fpart, texels);
3105    } else if (min_filter == mag_filter) {
3106       /* no need to distinguish between minification and magnification */
3107       lp_build_sample_mipmap(bld, min_filter, mip_filter,
3108                              is_gather,
3109                              coords, offsets,
3110                              ilevel0, ilevel1, lod_fpart,
3111                              texels);
3112    } else {
3113       /*
3114        * Could also get rid of the if-logic and always use mipmap_both, both
3115        * for the single lod and multi-lod case if nothing really uses this.
3116        */
3117       if (bld->num_lods == 1) {
3118          /* Emit conditional to choose min image filter or mag image filter
3119           * depending on the lod being > 0 or <= 0, respectively.
3120           */
3121          struct lp_build_if_state if_ctx;
3122 
3123          lod_positive = LLVMBuildTrunc(builder, lod_positive,
3124                                        LLVMInt1TypeInContext(bld->gallivm->context),
3125                                        "lod_pos");
3126 
3127          lp_build_if(&if_ctx, bld->gallivm, lod_positive);
3128          {
3129             /* Use the minification filter */
3130             lp_build_sample_mipmap(bld, min_filter, mip_filter, false,
3131                                    coords, offsets,
3132                                    ilevel0, ilevel1, lod_fpart,
3133                                    texels);
3134          }
3135          lp_build_else(&if_ctx);
3136          {
3137             /* Use the magnification filter */
3138             lp_build_sample_mipmap(bld, mag_filter, PIPE_TEX_MIPFILTER_NONE,
3139                                    false,
3140                                    coords, offsets,
3141                                    ilevel0, NULL, NULL,
3142                                    texels);
3143          }
3144          lp_build_endif(&if_ctx);
3145       } else {
3146          LLVMValueRef need_linear, linear_mask;
3147          unsigned mip_filter_for_nearest;
3148          struct lp_build_if_state if_ctx;
3149 
3150          if (min_filter == PIPE_TEX_FILTER_LINEAR) {
3151             linear_mask = lod_positive;
3152             mip_filter_for_nearest = PIPE_TEX_MIPFILTER_NONE;
3153          } else {
3154             linear_mask = lp_build_not(&bld->lodi_bld, lod_positive);
3155             mip_filter_for_nearest = mip_filter;
3156          }
3157          need_linear = lp_build_any_true_range(&bld->lodi_bld, bld->num_lods,
3158                                                linear_mask);
3159          lp_build_name(need_linear, "need_linear");
3160 
3161          if (bld->num_lods != bld->coord_type.length) {
3162             linear_mask = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
3163                                                                 bld->lodi_type,
3164                                                                 bld->int_coord_type,
3165                                                                 linear_mask);
3166          }
3167 
3168          lp_build_if(&if_ctx, bld->gallivm, need_linear);
3169          {
3170             /*
3171              * Do sampling with both filters simultaneously. This means using
3172              * a linear filter and doing some tricks (with weights) for the
3173              * pixels which need nearest filter.
3174              * Note that it's probably rare some pixels need nearest and some
3175              * linear filter but the fixups required for the nearest pixels
3176              * aren't all that complicated so just always run a combined path
3177              * if at least some pixels require linear.
3178              */
3179             lp_build_sample_mipmap_both(bld, linear_mask, mip_filter,
3180                                         coords, offsets,
3181                                         ilevel0, ilevel1,
3182                                         lod_fpart, lod_positive,
3183                                         texels);
3184          }
3185          lp_build_else(&if_ctx);
3186          {
3187             /*
3188              * All pixels require just nearest filtering, which is way
3189              * cheaper than linear, hence do a separate path for that.
3190              */
3191             lp_build_sample_mipmap(bld, PIPE_TEX_FILTER_NEAREST,
3192                                    mip_filter_for_nearest, false,
3193                                    coords, offsets,
3194                                    ilevel0, ilevel1, lod_fpart,
3195                                    texels);
3196          }
3197          lp_build_endif(&if_ctx);
3198       }
3199    }
3200 
3201    for (chan = 0; chan < 4; ++chan) {
3202      colors_out[chan] = LLVMBuildLoad2(builder, bld->texel_bld.vec_type, texels[chan], "");
3203      lp_build_name(colors_out[chan], "sampler%u_texel_%c", sampler_unit, "xyzw"[chan]);
3204    }
3205 }
3206 
3207 
3208 /**
3209  * Texel fetch function.  In contrast to general sampling there is no
3210  * filtering, no coord minification, lod (if any) is always explicit uint,
3211  * coords are uints (in terms of texel units) directly to be applied to the
3212  * selected mip level (after adding texel offsets).  This function handles
3213  * texel fetch for all targets where texel fetch is supported (no cube maps,
3214  * but 1d, 2d, 3d are supported, arrays and buffers should be too).
3215  */
3216 static void
lp_build_fetch_texel(struct lp_build_sample_context * bld,unsigned texture_unit,LLVMValueRef ms_index,const LLVMValueRef * coords,LLVMValueRef explicit_lod,const LLVMValueRef * offsets,LLVMValueRef * colors_out)3217 lp_build_fetch_texel(struct lp_build_sample_context *bld,
3218                      unsigned texture_unit,
3219                      LLVMValueRef ms_index,
3220                      const LLVMValueRef *coords,
3221                      LLVMValueRef explicit_lod,
3222                      const LLVMValueRef *offsets,
3223                      LLVMValueRef *colors_out)
3224 {
3225    struct lp_build_context *perquadi_bld = &bld->lodi_bld;
3226    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
3227    unsigned dims = bld->dims, chan;
3228    unsigned target = bld->static_texture_state->target;
3229    bool out_of_bound_ret_zero = true;
3230    LLVMValueRef size, ilevel;
3231    LLVMValueRef row_stride_vec = NULL, img_stride_vec = NULL;
3232    LLVMValueRef x = coords[0], y = coords[1], z = coords[2];
3233    LLVMValueRef width, height, depth, i, j;
3234    LLVMValueRef offset, out_of_bounds, out1;
3235 
3236    LLVMValueRef first_level;
3237 
3238    first_level = get_first_level(bld->gallivm,
3239                                  bld->resources_type,
3240                                  bld->resources_ptr,
3241                                  texture_unit, NULL,
3242                                  bld->static_texture_state,
3243                                  bld->dynamic_state);
3244    out_of_bounds = int_coord_bld->zero;
3245 
3246    if (explicit_lod && bld->static_texture_state->target != PIPE_BUFFER) {
3247       if (bld->num_mips != int_coord_bld->type.length) {
3248          ilevel = lp_build_pack_aos_scalars(bld->gallivm, int_coord_bld->type,
3249                                             perquadi_bld->type, explicit_lod, 0);
3250       } else {
3251          ilevel = explicit_lod;
3252       }
3253 
3254       LLVMValueRef last_level;
3255 
3256       last_level = get_last_level(bld->gallivm,
3257                                   bld->resources_type,
3258                                   bld->resources_ptr,
3259                                   texture_unit, NULL,
3260                                   bld->static_texture_state,
3261                                   bld->dynamic_state);
3262 
3263       first_level = lp_build_broadcast_scalar(&bld->leveli_bld, first_level);
3264       last_level = lp_build_broadcast_scalar(&bld->leveli_bld, last_level);
3265       lp_build_nearest_mip_level(bld,
3266                                  first_level, last_level,
3267                                  ilevel, &ilevel,
3268                                  out_of_bound_ret_zero ? &out_of_bounds : NULL);
3269    } else {
3270       assert(bld->num_mips == 1);
3271       if (bld->static_texture_state->target != PIPE_BUFFER) {
3272          ilevel = first_level;
3273       } else {
3274          ilevel = lp_build_const_int32(bld->gallivm, 0);
3275       }
3276    }
3277    lp_build_mipmap_level_sizes(bld, ilevel,
3278                                &size,
3279                                &row_stride_vec, &img_stride_vec);
3280    lp_build_extract_image_sizes(bld, &bld->int_size_bld, int_coord_bld->type,
3281                                 size, &width, &height, &depth);
3282 
3283    if (target == PIPE_TEXTURE_1D_ARRAY ||
3284        target == PIPE_TEXTURE_2D_ARRAY) {
3285       if (out_of_bound_ret_zero) {
3286          z = lp_build_layer_coord(bld, texture_unit, false, z, &out1);
3287          out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
3288       } else {
3289          z = lp_build_layer_coord(bld, texture_unit, false, z, NULL);
3290       }
3291    }
3292 
3293    /* This is a lot like border sampling */
3294    if (offsets[0]) {
3295       /*
3296        * coords are really unsigned, offsets are signed, but I don't think
3297        * exceeding 31 bits is possible
3298        */
3299       x = lp_build_add(int_coord_bld, x, offsets[0]);
3300    }
3301    out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
3302    out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
3303    out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
3304    out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
3305 
3306    if (dims >= 2) {
3307       if (offsets[1]) {
3308          y = lp_build_add(int_coord_bld, y, offsets[1]);
3309       }
3310       out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
3311       out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
3312       out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
3313       out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
3314 
3315       if (dims >= 3) {
3316          if (offsets[2]) {
3317             z = lp_build_add(int_coord_bld, z, offsets[2]);
3318          }
3319          out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
3320          out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
3321          out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
3322          out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
3323       }
3324    }
3325 
3326    if (bld->static_texture_state->tiled) {
3327       lp_build_tiled_sample_offset(&bld->int_coord_bld,
3328                                    bld->format_desc->format,
3329                                    bld->static_texture_state,
3330                                    x, y, z, width, height, img_stride_vec,
3331                                    &offset, &i, &j);
3332    } else {
3333       lp_build_sample_offset(int_coord_bld,
3334                              bld->format_desc,
3335                              x, y, z, row_stride_vec, img_stride_vec,
3336                              &offset, &i, &j);
3337    }
3338 
3339    if (bld->static_texture_state->target != PIPE_BUFFER) {
3340       offset = lp_build_add(int_coord_bld, offset,
3341                             lp_build_get_mip_offsets(bld, ilevel));
3342    }
3343 
3344    if (bld->fetch_ms && bld->static_texture_state->level_zero_only) {
3345       LLVMValueRef num_samples = bld->dynamic_state->last_level(bld->gallivm,
3346                                                                 bld->resources_type,
3347                                                                 bld->resources_ptr,
3348                                                                 texture_unit, NULL);
3349       num_samples = LLVMBuildZExt(bld->gallivm->builder, num_samples,
3350                                   bld->int_bld.elem_type, "");
3351       LLVMValueRef sample_stride = lp_sample_load_mip_value(bld->gallivm,
3352                                                             bld->mip_offsets_type,
3353                                                             bld->mip_offsets,
3354                                                             lp_build_const_int32(bld->gallivm, LP_JIT_TEXTURE_SAMPLE_STRIDE));
3355       lp_build_sample_ms_offset(int_coord_bld, ms_index, num_samples, sample_stride,
3356                                 &offset, &out_of_bounds);
3357    }
3358 
3359    if (bld->residency) {
3360       lp_build_gather_resident(&bld->float_vec_bld, bld->dynamic_state,
3361                                bld->resources_type, bld->resources_ptr,
3362                                offset, &bld->resident);
3363    }
3364 
3365    offset = lp_build_andnot(int_coord_bld, offset, out_of_bounds);
3366 
3367    lp_build_fetch_rgba_soa(bld->gallivm,
3368                            bld->format_desc,
3369                            bld->texel_type, true,
3370                            bld->base_ptr, offset,
3371                            i, j,
3372                            bld->cache,
3373                            colors_out);
3374 
3375    if (out_of_bound_ret_zero) {
3376       /*
3377        * Only needed for ARB_robust_buffer_access_behavior and d3d10.
3378        * Could use min/max above instead of out-of-bounds comparisons
3379        * if we don't care about the result returned for out-of-bounds.
3380        */
3381       LLVMValueRef oob[4] = {
3382          bld->texel_bld.zero,
3383          bld->texel_bld.zero,
3384          bld->texel_bld.zero,
3385          bld->texel_bld.zero,
3386       };
3387       lp_build_format_swizzle_soa(bld->format_desc, &bld->texel_bld, oob, oob);
3388       for (chan = 0; chan < 4; chan++) {
3389          colors_out[chan] = lp_build_select(&bld->texel_bld, out_of_bounds,
3390                                             oob[chan], colors_out[chan]);
3391       }
3392    }
3393 }
3394 
3395 
3396 /**
3397  * Just set texels to white instead of actually sampling the texture.
3398  * For debugging.
3399  */
3400 void
lp_build_sample_nop(struct gallivm_state * gallivm,struct lp_type type,const LLVMValueRef * coords,LLVMValueRef texel_out[4])3401 lp_build_sample_nop(struct gallivm_state *gallivm,
3402                     struct lp_type type,
3403                     const LLVMValueRef *coords,
3404                     LLVMValueRef texel_out[4])
3405 {
3406    LLVMValueRef one = lp_build_one(gallivm, type);
3407    for (unsigned chan = 0; chan < 4; chan++) {
3408       texel_out[chan] = one;
3409    }
3410 }
3411 
3412 
3413 struct lp_type
lp_build_texel_type(struct lp_type texel_type,const struct util_format_description * format_desc)3414 lp_build_texel_type(struct lp_type texel_type,
3415                     const struct util_format_description *format_desc)
3416 {
3417    /* always using the first channel hopefully should be safe,
3418     * if not things WILL break in other places anyway.
3419     */
3420    if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
3421        format_desc->channel[0].pure_integer) {
3422       if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
3423          texel_type = lp_type_int_vec(texel_type.width, texel_type.width * texel_type.length);
3424       } else if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
3425          texel_type = lp_type_uint_vec(texel_type.width, texel_type.width * texel_type.length);
3426       }
3427    } else if (util_format_has_stencil(format_desc) &&
3428        !util_format_has_depth(format_desc)) {
3429       /* for stencil only formats, sample stencil (uint) */
3430       texel_type = lp_type_uint_vec(texel_type.width, texel_type.width * texel_type.length);
3431    }
3432    return texel_type;
3433 }
3434 
3435 
3436 /**
3437  * Build the actual texture sampling code.
3438  * 'texel' will return a vector of four LLVMValueRefs corresponding to
3439  * R, G, B, A.
3440  * \param type  vector float type to use for coords, etc.
3441  * \param sample_key
3442  * \param derivs  partial derivatives of (s,t,r,q) with respect to x and y
3443  */
3444 void
lp_build_sample_soa_code(struct gallivm_state * gallivm,const struct lp_static_texture_state * static_texture_state,const struct lp_static_sampler_state * static_sampler_state,struct lp_sampler_dynamic_state * dynamic_state,struct lp_type type,unsigned sample_key,unsigned texture_index,unsigned sampler_index,LLVMTypeRef resources_type,LLVMValueRef resources_ptr,LLVMTypeRef thread_data_type,LLVMValueRef thread_data_ptr,const LLVMValueRef * coords,const LLVMValueRef * offsets,const struct lp_derivatives * derivs,LLVMValueRef lod,LLVMValueRef ms_index,LLVMValueRef aniso_filter_table,LLVMValueRef * texel_out)3445 lp_build_sample_soa_code(struct gallivm_state *gallivm,
3446                          const struct lp_static_texture_state *static_texture_state,
3447                          const struct lp_static_sampler_state *static_sampler_state,
3448                          struct lp_sampler_dynamic_state *dynamic_state,
3449                          struct lp_type type,
3450                          unsigned sample_key,
3451                          unsigned texture_index,
3452                          unsigned sampler_index,
3453                          LLVMTypeRef resources_type,
3454                          LLVMValueRef resources_ptr,
3455                          LLVMTypeRef thread_data_type,
3456                          LLVMValueRef thread_data_ptr,
3457                          const LLVMValueRef *coords,
3458                          const LLVMValueRef *offsets,
3459                          const struct lp_derivatives *derivs, /* optional */
3460                          LLVMValueRef lod, /* optional */
3461                          LLVMValueRef ms_index, /* optional */
3462                          LLVMValueRef aniso_filter_table,
3463                          LLVMValueRef *texel_out)
3464 {
3465    assert(static_texture_state);
3466    assert(static_texture_state->format < PIPE_FORMAT_COUNT);
3467    assert(static_sampler_state);
3468 
3469    const enum pipe_texture_target target = static_texture_state->target;
3470    const unsigned dims = texture_dims(target);
3471    const unsigned num_quads = type.length == 1 ? 1 : type.length / 4;
3472    struct lp_build_sample_context bld;
3473    struct lp_static_sampler_state derived_sampler_state = *static_sampler_state;
3474    LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
3475    LLVMBuilderRef builder = gallivm->builder;
3476    const struct util_format_description *res_format_desc;
3477 
3478    if (0) {
3479       enum pipe_format fmt = static_texture_state->format;
3480       debug_printf("Sample from %s\n", util_format_name(fmt));
3481    }
3482 
3483    const enum lp_sampler_lod_property lod_property =
3484       (sample_key & LP_SAMPLER_LOD_PROPERTY_MASK) >>
3485       LP_SAMPLER_LOD_PROPERTY_SHIFT;
3486    const enum lp_sampler_lod_control lod_control =
3487       (sample_key & LP_SAMPLER_LOD_CONTROL_MASK) >>
3488       LP_SAMPLER_LOD_CONTROL_SHIFT;
3489    const enum lp_sampler_op_type op_type =
3490       (sample_key & LP_SAMPLER_OP_TYPE_MASK) >>
3491       LP_SAMPLER_OP_TYPE_SHIFT;
3492 
3493    const bool fetch_ms = !!(sample_key & LP_SAMPLER_FETCH_MS);
3494    const bool op_is_tex = op_type == LP_SAMPLER_OP_TEXTURE;
3495    const bool op_is_lodq = op_type == LP_SAMPLER_OP_LODQ;
3496    const bool op_is_gather = op_type == LP_SAMPLER_OP_GATHER;
3497 
3498    LLVMValueRef lod_bias = NULL;
3499    LLVMValueRef explicit_lod = NULL;
3500    if (lod_control == LP_SAMPLER_LOD_BIAS) {
3501       lod_bias = lod;
3502       assert(lod);
3503       assert(derivs == NULL);
3504    } else if (lod_control == LP_SAMPLER_LOD_EXPLICIT) {
3505       explicit_lod = lod;
3506       assert(lod);
3507       assert(derivs == NULL);
3508    } else if (lod_control == LP_SAMPLER_LOD_DERIVATIVES) {
3509       assert(derivs);
3510       assert(lod == NULL);
3511    } else {
3512       assert(derivs == NULL);
3513       assert(lod == NULL);
3514    }
3515 
3516    if (static_texture_state->format == PIPE_FORMAT_NONE) {
3517       /*
3518        * If there's nothing bound, format is NONE, and we must return
3519        * all zero as mandated by d3d10 in this case.
3520        */
3521       LLVMValueRef zero = lp_build_zero(gallivm, type);
3522       for (unsigned chan = 0; chan < 4; chan++) {
3523          texel_out[chan] = zero;
3524       }
3525       return;
3526    }
3527 
3528    assert(type.floating);
3529 
3530    /* Setup our build context */
3531    memset(&bld, 0, sizeof bld);
3532    bld.gallivm = gallivm;
3533    bld.resources_type = resources_type;
3534    bld.resources_ptr = resources_ptr;
3535    bld.aniso_filter_table = aniso_filter_table;
3536    bld.static_sampler_state = &derived_sampler_state;
3537    bld.static_texture_state = static_texture_state;
3538    bld.dynamic_state = dynamic_state;
3539    bld.format_desc = util_format_description(static_texture_state->format);
3540    bld.dims = dims;
3541 
3542    res_format_desc = util_format_description(static_texture_state->res_format);
3543 
3544    if (gallivm_perf & GALLIVM_PERF_NO_QUAD_LOD || op_is_lodq) {
3545       bld.no_quad_lod = true;
3546    }
3547    if (!(gallivm_perf & GALLIVM_PERF_RHO_APPROX) || op_is_lodq) {
3548       bld.no_rho_approx = true;
3549    }
3550    if (!(gallivm_perf & GALLIVM_PERF_BRILINEAR) || op_is_lodq || lod_bias || explicit_lod) {
3551       bld.no_brilinear = true;
3552    }
3553 
3554    bld.vector_width = lp_type_width(type);
3555 
3556    bld.float_type = lp_type_float(32);
3557    bld.int_type = lp_type_int(32);
3558    bld.coord_type = type;
3559    bld.int_coord_type = lp_int_type(type);
3560    bld.float_size_in_type = lp_type_float(32);
3561    bld.float_size_in_type.length = dims > 1 ? 4 : 1;
3562    bld.int_size_in_type = lp_int_type(bld.float_size_in_type);
3563 
3564    bld.texel_type = lp_build_texel_type(type, bld.format_desc);
3565 
3566    if (!static_texture_state->level_zero_only ||
3567        !static_sampler_state->max_lod_pos || op_is_lodq) {
3568       derived_sampler_state.min_mip_filter = static_sampler_state->min_mip_filter;
3569    } else {
3570       derived_sampler_state.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
3571    }
3572 
3573    if (op_is_gather) {
3574       /*
3575        * gather4 is exactly like GL_LINEAR filtering but in the end skipping
3576        * the actual filtering. Using mostly the same paths, so cube face
3577        * selection, coord wrapping etc. all naturally uses the same code.
3578        */
3579       derived_sampler_state.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
3580       derived_sampler_state.min_img_filter = PIPE_TEX_FILTER_LINEAR;
3581       derived_sampler_state.mag_img_filter = PIPE_TEX_FILTER_LINEAR;
3582    }
3583 
3584    const enum pipe_tex_mipfilter mip_filter =
3585       derived_sampler_state.min_mip_filter;
3586 
3587    if (static_texture_state->target == PIPE_TEXTURE_CUBE ||
3588        static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) {
3589       /*
3590        * Seamless filtering ignores wrap modes.
3591        * Setting to CLAMP_TO_EDGE is correct for nearest filtering, for
3592        * bilinear it's not correct but way better than using for instance
3593        * repeat.  Note we even set this for non-seamless. Technically GL
3594        * allows any wrap mode, which made sense when supporting true borders
3595        * (can get seamless effect with border and CLAMP_TO_BORDER), but
3596        * gallium doesn't support borders and d3d9 requires wrap modes to be
3597        * ignored and it's a pain to fix up the sampler state (as it makes it
3598        * texture dependent).
3599        */
3600       derived_sampler_state.wrap_s = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
3601       derived_sampler_state.wrap_t = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
3602    }
3603 
3604    /*
3605     * We could force CLAMP to CLAMP_TO_EDGE here if min/mag filter is nearest,
3606     * so AoS path could be used. Not sure it's worth the trouble...
3607     */
3608    const enum pipe_tex_filter min_img_filter =
3609       derived_sampler_state.min_img_filter;
3610    const enum pipe_tex_filter mag_img_filter =
3611       derived_sampler_state.mag_img_filter;
3612 
3613    /*
3614     * This is all a bit complicated different paths are chosen for performance
3615     * reasons.
3616     * Essentially, there can be 1 lod per element, 1 lod per quad or 1 lod for
3617     * everything (the last two options are equivalent for 4-wide case).
3618     * If there's per-quad lod but we split to 4-wide so we can use AoS, per-quad
3619     * lod is calculated then the lod value extracted afterwards so making this
3620     * case basically the same as far as lod handling is concerned for the
3621     * further sample/filter code as the 1 lod for everything case.
3622     * Different lod handling mostly shows up when building mipmap sizes
3623     * (lp_build_mipmap_level_sizes() and friends) and also in filtering
3624     * (getting the fractional part of the lod to the right texels).
3625     */
3626 
3627    /*
3628     * There are other situations where at least the multiple int lods could be
3629     * avoided like min and max lod being equal.
3630     */
3631    bld.num_mips = bld.num_lods = 1;
3632 
3633    if ((mip_filter != PIPE_TEX_MIPFILTER_NONE && op_is_tex &&
3634          (static_texture_state->target == PIPE_TEXTURE_CUBE ||
3635           static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY)) ||
3636         op_is_lodq) {
3637       /*
3638        * special case for using per-pixel lod even for implicit lod,
3639        * which is generally never required (ok by APIs) except to please
3640        * some (somewhat broken imho) tests (because per-pixel face selection
3641        * can cause derivatives to be different for pixels outside the primitive
3642        * due to the major axis division even if pre-project derivatives are
3643        * looking normal).
3644        * For lodq, we do it to simply avoid scalar pack / unpack (albeit for
3645        * cube maps we do indeed get per-pixel lod values).
3646        */
3647       bld.num_mips = type.length;
3648       bld.num_lods = type.length;
3649    } else if (lod_property == LP_SAMPLER_LOD_PER_ELEMENT ||
3650        (explicit_lod || lod_bias || derivs)) {
3651       if ((!op_is_tex && target != PIPE_BUFFER) ||
3652           (op_is_tex && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
3653          bld.num_mips = type.length;
3654          bld.num_lods = type.length;
3655       } else if (op_is_tex && min_img_filter != mag_img_filter) {
3656          bld.num_mips = 1;
3657          bld.num_lods = type.length;
3658       }
3659    }
3660    /* TODO: for true scalar_lod should only use 1 lod value */
3661    else if ((!op_is_tex && explicit_lod && target != PIPE_BUFFER) ||
3662             (op_is_tex && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
3663       bld.num_mips = num_quads;
3664       bld.num_lods = num_quads;
3665    } else if (op_is_tex && min_img_filter != mag_img_filter) {
3666       bld.num_mips = 1;
3667       bld.num_lods = num_quads;
3668    }
3669 
3670    bld.fetch_ms = fetch_ms;
3671    bld.residency = !!(sample_key & LP_SAMPLER_RESIDENCY);
3672    if (op_is_gather)
3673       bld.gather_comp = (sample_key & LP_SAMPLER_GATHER_COMP_MASK) >> LP_SAMPLER_GATHER_COMP_SHIFT;
3674    bld.lodf_type = type;
3675    /* we want native vector size to be able to use our intrinsics */
3676    if (bld.num_lods != type.length) {
3677       /* TODO: this currently always has to be per-quad or per-element */
3678       bld.lodf_type.length = type.length > 4 ? ((type.length + 15) / 16) * 4 : 1;
3679    }
3680    bld.lodi_type = lp_int_type(bld.lodf_type);
3681    bld.levelf_type = bld.lodf_type;
3682    if (bld.num_mips == 1) {
3683       bld.levelf_type.length = 1;
3684    }
3685    bld.leveli_type = lp_int_type(bld.levelf_type);
3686    bld.float_size_type = bld.float_size_in_type;
3687 
3688    /* Note: size vectors may not be native. They contain minified w/h/d/_
3689     * values, with per-element lod that is w0/h0/d0/_/w1/h1/d1_/... so up to
3690     * 8x4f32
3691     */
3692    if (bld.num_mips > 1) {
3693       bld.float_size_type.length = bld.num_mips == type.length ?
3694                                       bld.num_mips * bld.float_size_in_type.length :
3695                                       type.length;
3696    }
3697    bld.int_size_type = lp_int_type(bld.float_size_type);
3698 
3699    lp_build_context_init(&bld.float_bld, gallivm, bld.float_type);
3700    lp_build_context_init(&bld.float_vec_bld, gallivm, type);
3701    lp_build_context_init(&bld.int_bld, gallivm, bld.int_type);
3702    lp_build_context_init(&bld.coord_bld, gallivm, bld.coord_type);
3703    lp_build_context_init(&bld.int_coord_bld, gallivm, bld.int_coord_type);
3704    lp_build_context_init(&bld.int_size_in_bld, gallivm, bld.int_size_in_type);
3705    lp_build_context_init(&bld.float_size_in_bld, gallivm, bld.float_size_in_type);
3706    lp_build_context_init(&bld.int_size_bld, gallivm, bld.int_size_type);
3707    lp_build_context_init(&bld.float_size_bld, gallivm, bld.float_size_type);
3708    lp_build_context_init(&bld.texel_bld, gallivm, bld.texel_type);
3709    lp_build_context_init(&bld.levelf_bld, gallivm, bld.levelf_type);
3710    lp_build_context_init(&bld.leveli_bld, gallivm, bld.leveli_type);
3711    lp_build_context_init(&bld.lodf_bld, gallivm, bld.lodf_type);
3712    lp_build_context_init(&bld.lodi_bld, gallivm, bld.lodi_type);
3713 
3714    /* Get the dynamic state */
3715    LLVMValueRef tex_width = dynamic_state->width(gallivm, resources_type,
3716                                                  resources_ptr, texture_index,
3717                                                  NULL);
3718    bld.row_stride_array = dynamic_state->row_stride(gallivm, resources_type,
3719                                                     resources_ptr, texture_index, NULL,
3720                                                     &bld.row_stride_type);
3721    bld.img_stride_array = dynamic_state->img_stride(gallivm, resources_type,
3722                                                     resources_ptr, texture_index, NULL,
3723                                                     &bld.img_stride_type);
3724    bld.base_ptr = dynamic_state->base_ptr(gallivm, resources_type,
3725                                           resources_ptr, texture_index, NULL);
3726    bld.mip_offsets = dynamic_state->mip_offsets(gallivm, resources_type,
3727                                                 resources_ptr, texture_index, NULL,
3728                                                 &bld.mip_offsets_type);
3729 
3730    /* Note that mip_offsets is an array[level] of offsets to texture images */
3731 
3732    if (dynamic_state->cache_ptr && thread_data_ptr) {
3733       bld.cache = dynamic_state->cache_ptr(gallivm, thread_data_type,
3734                                            thread_data_ptr, texture_index);
3735    }
3736 
3737    uint32_t res_bw = res_format_desc->block.width;
3738    uint32_t res_bh = res_format_desc->block.height;
3739    uint32_t bw = bld.format_desc->block.width;
3740    uint32_t bh = bld.format_desc->block.height;
3741 
3742    /* only scale if the blocksizes are different. */
3743    if (res_bw == bw)
3744       res_bw = bw = 1;
3745    if (res_bh == bh)
3746       res_bh = bh = 1;
3747 
3748    /* width, height, depth as single int vector */
3749    if (dims <= 1) {
3750       bld.int_size = tex_width;
3751       bld.int_tex_blocksize = LLVMConstInt(i32t, res_bw, 0);
3752       bld.int_tex_blocksize_log2 = LLVMConstInt(i32t, util_logbase2(res_bw), 0);
3753       bld.int_view_blocksize = LLVMConstInt(i32t, bw, 0);
3754    } else {
3755       bld.int_size = LLVMBuildInsertElement(builder, bld.int_size_in_bld.undef,
3756                                             tex_width,
3757                                             LLVMConstInt(i32t, 0, 0), "");
3758       bld.int_tex_blocksize = LLVMBuildInsertElement(builder, bld.int_size_in_bld.undef,
3759                                                      LLVMConstInt(i32t, res_bw, 0),
3760                                                      LLVMConstInt(i32t, 0, 0), "");
3761       bld.int_tex_blocksize_log2 = LLVMBuildInsertElement(builder, bld.int_size_in_bld.undef,
3762                                                           LLVMConstInt(i32t, util_logbase2(res_bw), 0),
3763                                                           LLVMConstInt(i32t, 0, 0), "");
3764       bld.int_view_blocksize = LLVMBuildInsertElement(builder, bld.int_size_in_bld.undef,
3765                                                       LLVMConstInt(i32t, bw, 0),
3766                                                       LLVMConstInt(i32t, 0, 0), "");
3767       if (dims >= 2) {
3768          LLVMValueRef tex_height =
3769             dynamic_state->height(gallivm, resources_type,
3770                                   resources_ptr, texture_index, NULL);
3771          tex_height = LLVMBuildZExt(gallivm->builder, tex_height,
3772                                     bld.int_bld.elem_type, "");
3773          bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
3774                                                tex_height,
3775                                                LLVMConstInt(i32t, 1, 0), "");
3776          bld.int_tex_blocksize = LLVMBuildInsertElement(builder, bld.int_tex_blocksize,
3777                                                         LLVMConstInt(i32t, res_bh, 0),
3778                                                         LLVMConstInt(i32t, 1, 0), "");
3779          bld.int_tex_blocksize_log2 = LLVMBuildInsertElement(builder, bld.int_tex_blocksize_log2,
3780                                                              LLVMConstInt(i32t, util_logbase2(res_bh), 0),
3781                                                         LLVMConstInt(i32t, 1, 0), "");
3782          bld.int_view_blocksize = LLVMBuildInsertElement(builder, bld.int_view_blocksize,
3783                                                          LLVMConstInt(i32t, bh, 0),
3784                                                          LLVMConstInt(i32t, 1, 0), "");
3785          if (dims >= 3) {
3786             LLVMValueRef tex_depth =
3787                dynamic_state->depth(gallivm, resources_type, resources_ptr,
3788                                     texture_index, NULL);
3789             tex_depth = LLVMBuildZExt(gallivm->builder, tex_depth,
3790                                       bld.int_bld.elem_type, "");
3791             bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
3792                                                   tex_depth,
3793                                                   LLVMConstInt(i32t, 2, 0), "");
3794             bld.int_tex_blocksize = LLVMBuildInsertElement(builder, bld.int_tex_blocksize,
3795                                                            LLVMConstInt(i32t, 1, 0),
3796                                                            LLVMConstInt(i32t, 2, 0), "");
3797             bld.int_tex_blocksize_log2 = LLVMBuildInsertElement(builder, bld.int_tex_blocksize_log2,
3798                                                            LLVMConstInt(i32t, 0, 0),
3799                                                            LLVMConstInt(i32t, 2, 0), "");
3800             bld.int_view_blocksize = LLVMBuildInsertElement(builder, bld.int_view_blocksize,
3801                                                             LLVMConstInt(i32t, 1, 0),
3802                                                             LLVMConstInt(i32t, 2, 0), "");
3803          }
3804       }
3805    }
3806 
3807    LLVMValueRef newcoords[5];
3808    for (unsigned i = 0; i < 5; i++) {
3809       newcoords[i] = coords[i];
3810    }
3811 
3812    if (util_format_is_pure_integer(static_texture_state->format) &&
3813        !util_format_has_depth(bld.format_desc) && op_is_tex &&
3814        (static_sampler_state->min_mip_filter == PIPE_TEX_MIPFILTER_LINEAR ||
3815         static_sampler_state->min_img_filter == PIPE_TEX_FILTER_LINEAR ||
3816         static_sampler_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR)) {
3817       /*
3818        * Bail if impossible filtering is specified (the awkard additional
3819        * depth check is because it is legal in gallium to have things like
3820        * S8Z24 here which would say it's pure int despite such formats should
3821        * sample the depth component).
3822        * In GL such filters make the texture incomplete, this makes it robust
3823        * against gallium frontends which set this up regardless (we'd crash in
3824        * the lerp later otherwise).
3825        * At least in some apis it may be legal to use such filters with lod
3826        * queries and/or gather (at least for gather d3d10 says only the wrap
3827        * bits are really used hence filter bits are likely simply ignored).
3828        * For fetch, we don't get valid samplers either way here.
3829        */
3830       LLVMValueRef zero = lp_build_zero(gallivm, type);
3831       for (unsigned chan = 0; chan < 4; chan++) {
3832          texel_out[chan] = zero;
3833       }
3834       return;
3835    }
3836 
3837    if (0) {
3838       /* For debug: no-op texture sampling */
3839       lp_build_sample_nop(gallivm,
3840                           bld.texel_type,
3841                           newcoords,
3842                           texel_out);
3843    } else if (op_type == LP_SAMPLER_OP_FETCH) {
3844       lp_build_fetch_texel(&bld, texture_index, ms_index, newcoords,
3845                            lod, offsets, texel_out);
3846       if (bld.residency)
3847          texel_out[4] = bld.resident;
3848    } else {
3849       LLVMValueRef lod_fpart = NULL, lod_positive = NULL;
3850       LLVMValueRef ilevel0 = NULL, ilevel1 = NULL, lod = NULL;
3851       bool use_aos = util_format_fits_8unorm(bld.format_desc) &&
3852                 op_is_tex &&
3853                 /* not sure this is strictly needed or simply impossible */
3854                 derived_sampler_state.compare_mode == PIPE_TEX_COMPARE_NONE &&
3855                 derived_sampler_state.aniso == 0 &&
3856                 lp_is_simple_wrap_mode(derived_sampler_state.wrap_s);
3857 
3858       use_aos &= bld.num_lods <= num_quads ||
3859                  derived_sampler_state.min_img_filter ==
3860                     derived_sampler_state.mag_img_filter;
3861 
3862       use_aos &= !static_texture_state->tiled;
3863 
3864       if (gallivm_perf & GALLIVM_PERF_NO_AOS_SAMPLING) {
3865          use_aos = 0;
3866       }
3867 
3868       if (dims > 1) {
3869          use_aos &= lp_is_simple_wrap_mode(derived_sampler_state.wrap_t);
3870          if (dims > 2) {
3871             use_aos &= lp_is_simple_wrap_mode(derived_sampler_state.wrap_r);
3872          }
3873       }
3874       if ((static_texture_state->target == PIPE_TEXTURE_CUBE ||
3875            static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) &&
3876           derived_sampler_state.seamless_cube_map &&
3877           (derived_sampler_state.min_img_filter == PIPE_TEX_FILTER_LINEAR ||
3878            derived_sampler_state.mag_img_filter == PIPE_TEX_FILTER_LINEAR)) {
3879          /* theoretically possible with AoS filtering but not implemented (complex!) */
3880          use_aos = 0;
3881       }
3882 
3883       if ((gallivm_debug & GALLIVM_DEBUG_PERF) &&
3884           !use_aos && util_format_fits_8unorm(bld.format_desc)) {
3885          debug_printf("%s: using floating point linear filtering for %s\n",
3886                       __func__, bld.format_desc->short_name);
3887          debug_printf("  min_img %d  mag_img %d  mip %d  target %d  seamless %d"
3888                       "  wraps %d  wrapt %d  wrapr %d\n",
3889                       derived_sampler_state.min_img_filter,
3890                       derived_sampler_state.mag_img_filter,
3891                       derived_sampler_state.min_mip_filter,
3892                       static_texture_state->target,
3893                       derived_sampler_state.seamless_cube_map,
3894                       derived_sampler_state.wrap_s,
3895                       derived_sampler_state.wrap_t,
3896                       derived_sampler_state.wrap_r);
3897       }
3898 
3899       lp_build_sample_common(&bld, op_is_lodq, texture_index, sampler_index,
3900                              newcoords, derivs, lod_bias, explicit_lod,
3901                              &lod_positive, &lod, &lod_fpart,
3902                              &ilevel0, &ilevel1);
3903 
3904       if (op_is_lodq) {
3905          texel_out[0] = lod_fpart;
3906          texel_out[1] = lod;
3907          texel_out[2] = texel_out[3] = bld.coord_bld.zero;
3908          if (bld.residency)
3909             texel_out[4] = bld.resident;
3910          return;
3911       }
3912 
3913       if (use_aos && static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) {
3914          /* The aos path doesn't do seamless filtering so simply add cube layer
3915           * to face now.
3916           */
3917          newcoords[2] = lp_build_add(&bld.int_coord_bld, newcoords[2], newcoords[3]);
3918       }
3919 
3920       /*
3921        * we only try 8-wide sampling with soa or if we have AVX2
3922        * as it appears to be a loss with just AVX)
3923        */
3924       if (num_quads == 1 || !use_aos ||
3925           (util_get_cpu_caps()->has_avx2 &&
3926            (bld.num_lods == 1 ||
3927             derived_sampler_state.min_img_filter == derived_sampler_state.mag_img_filter))) {
3928          if (use_aos) {
3929             /* do sampling/filtering with fixed pt arithmetic */
3930             lp_build_sample_aos(&bld,
3931                                 newcoords[0], newcoords[1],
3932                                 newcoords[2],
3933                                 offsets, lod_positive, lod_fpart,
3934                                 ilevel0, ilevel1,
3935                                 texel_out);
3936          } else {
3937             lp_build_sample_general(&bld, sampler_index,
3938                                     op_type == LP_SAMPLER_OP_GATHER,
3939                                     newcoords, offsets,
3940                                     lod_positive, lod_fpart,
3941                                     ilevel0, ilevel1,
3942                                     texel_out);
3943             if (bld.residency)
3944                texel_out[4] = bld.resident;
3945          }
3946       } else {
3947          struct lp_build_sample_context bld4;
3948          struct lp_type type4 = type;
3949          LLVMValueRef texelout4[4];
3950          LLVMValueRef texelouttmp[4][LP_MAX_VECTOR_LENGTH/16];
3951 
3952          type4.length = 4;
3953 
3954          /* Setup our build context */
3955          memset(&bld4, 0, sizeof bld4);
3956          bld4.no_quad_lod = bld.no_quad_lod;
3957          bld4.no_rho_approx = bld.no_rho_approx;
3958          bld4.no_brilinear = bld.no_brilinear;
3959          bld4.gallivm = bld.gallivm;
3960          bld4.resources_type = bld.resources_type;
3961          bld4.resources_ptr = bld.resources_ptr;
3962          bld4.aniso_filter_table = aniso_filter_table;
3963          bld4.static_texture_state = bld.static_texture_state;
3964          bld4.static_sampler_state = bld.static_sampler_state;
3965          bld4.dynamic_state = bld.dynamic_state;
3966          bld4.format_desc = bld.format_desc;
3967          bld4.dims = bld.dims;
3968          bld4.row_stride_type = bld.row_stride_type;
3969          bld4.row_stride_array = bld.row_stride_array;
3970          bld4.img_stride_type = bld.img_stride_type;
3971          bld4.img_stride_array = bld.img_stride_array;
3972          bld4.base_ptr = bld.base_ptr;
3973          bld4.mip_offsets_type = bld.mip_offsets_type;
3974          bld4.mip_offsets = bld.mip_offsets;
3975          bld4.int_size = bld.int_size;
3976          bld4.int_tex_blocksize = bld.int_tex_blocksize;
3977          bld4.int_tex_blocksize_log2 = bld.int_tex_blocksize_log2;
3978          bld4.int_view_blocksize = bld.int_view_blocksize;
3979          bld4.cache = bld.cache;
3980 
3981          bld4.vector_width = lp_type_width(type4);
3982 
3983          bld4.float_type = lp_type_float(32);
3984          bld4.int_type = lp_type_int(32);
3985          bld4.coord_type = type4;
3986          bld4.int_coord_type = lp_int_type(type4);
3987          bld4.float_size_in_type = lp_type_float(32);
3988          bld4.float_size_in_type.length = dims > 1 ? 4 : 1;
3989          bld4.int_size_in_type = lp_int_type(bld4.float_size_in_type);
3990          bld4.texel_type = bld.texel_type;
3991          bld4.texel_type.length = 4;
3992 
3993          bld4.num_mips = bld4.num_lods = 1;
3994          if (bld4.no_quad_lod && bld4.no_rho_approx &&
3995              (static_texture_state->target == PIPE_TEXTURE_CUBE ||
3996               static_texture_state->target == PIPE_TEXTURE_CUBE_ARRAY) &&
3997              (op_is_tex && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
3998             bld4.num_mips = type4.length;
3999             bld4.num_lods = type4.length;
4000          }
4001          if (lod_property == LP_SAMPLER_LOD_PER_ELEMENT &&
4002              (explicit_lod || lod_bias || derivs)) {
4003             if ((!op_is_tex && target != PIPE_BUFFER) ||
4004                 (op_is_tex && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
4005                bld4.num_mips = type4.length;
4006                bld4.num_lods = type4.length;
4007             } else if (op_is_tex && min_img_filter != mag_img_filter) {
4008                bld4.num_mips = 1;
4009                bld4.num_lods = type4.length;
4010             }
4011          }
4012 
4013          /* we want native vector size to be able to use our intrinsics */
4014          bld4.lodf_type = type4;
4015          if (bld4.num_lods != type4.length) {
4016             bld4.lodf_type.length = 1;
4017          }
4018          bld4.lodi_type = lp_int_type(bld4.lodf_type);
4019          bld4.levelf_type = type4;
4020          if (bld4.num_mips != type4.length) {
4021             bld4.levelf_type.length = 1;
4022          }
4023          bld4.leveli_type = lp_int_type(bld4.levelf_type);
4024          bld4.float_size_type = bld4.float_size_in_type;
4025          if (bld4.num_mips > 1) {
4026             bld4.float_size_type.length = bld4.num_mips == type4.length ?
4027                                             bld4.num_mips * bld4.float_size_in_type.length :
4028                                             type4.length;
4029          }
4030          bld4.int_size_type = lp_int_type(bld4.float_size_type);
4031 
4032          lp_build_context_init(&bld4.float_bld, gallivm, bld4.float_type);
4033          lp_build_context_init(&bld4.float_vec_bld, gallivm, type4);
4034          lp_build_context_init(&bld4.int_bld, gallivm, bld4.int_type);
4035          lp_build_context_init(&bld4.coord_bld, gallivm, bld4.coord_type);
4036          lp_build_context_init(&bld4.int_coord_bld, gallivm, bld4.int_coord_type);
4037          lp_build_context_init(&bld4.int_size_in_bld, gallivm, bld4.int_size_in_type);
4038          lp_build_context_init(&bld4.float_size_in_bld, gallivm, bld4.float_size_in_type);
4039          lp_build_context_init(&bld4.int_size_bld, gallivm, bld4.int_size_type);
4040          lp_build_context_init(&bld4.float_size_bld, gallivm, bld4.float_size_type);
4041          lp_build_context_init(&bld4.texel_bld, gallivm, bld4.texel_type);
4042          lp_build_context_init(&bld4.levelf_bld, gallivm, bld4.levelf_type);
4043          lp_build_context_init(&bld4.leveli_bld, gallivm, bld4.leveli_type);
4044          lp_build_context_init(&bld4.lodf_bld, gallivm, bld4.lodf_type);
4045          lp_build_context_init(&bld4.lodi_bld, gallivm, bld4.lodi_type);
4046 
4047          for (unsigned i = 0; i < num_quads; i++) {
4048             LLVMValueRef s4, t4, r4;
4049             LLVMValueRef lod_positive4, lod_fpart4 = NULL;
4050             LLVMValueRef ilevel04, ilevel14 = NULL;
4051             LLVMValueRef offsets4[4] = { NULL };
4052             unsigned num_lods = bld4.num_lods;
4053 
4054             s4 = lp_build_extract_range(gallivm, newcoords[0], 4*i, 4);
4055             t4 = lp_build_extract_range(gallivm, newcoords[1], 4*i, 4);
4056             r4 = lp_build_extract_range(gallivm, newcoords[2], 4*i, 4);
4057 
4058             if (offsets[0]) {
4059                offsets4[0] = lp_build_extract_range(gallivm, offsets[0], 4*i, 4);
4060                if (dims > 1) {
4061                   offsets4[1] = lp_build_extract_range(gallivm, offsets[1], 4*i, 4);
4062                   if (dims > 2) {
4063                      offsets4[2] = lp_build_extract_range(gallivm, offsets[2], 4*i, 4);
4064                   }
4065                }
4066             }
4067             lod_positive4 = lp_build_extract_range(gallivm, lod_positive, num_lods * i, num_lods);
4068             ilevel04 = bld.num_mips == 1 ? ilevel0 :
4069                           lp_build_extract_range(gallivm, ilevel0, num_lods * i, num_lods);
4070             if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
4071                ilevel14 = lp_build_extract_range(gallivm, ilevel1, num_lods * i, num_lods);
4072                lod_fpart4 = lp_build_extract_range(gallivm, lod_fpart, num_lods * i, num_lods);
4073             }
4074 
4075             if (use_aos) {
4076                /* do sampling/filtering with fixed pt arithmetic */
4077                lp_build_sample_aos(&bld4,
4078                                    s4, t4, r4, offsets4,
4079                                    lod_positive4, lod_fpart4,
4080                                    ilevel04, ilevel14,
4081                                    texelout4);
4082             } else {
4083                /* this path is currently unreachable and hence might break easily... */
4084                LLVMValueRef newcoords4[5];
4085                newcoords4[0] = s4;
4086                newcoords4[1] = t4;
4087                newcoords4[2] = r4;
4088                newcoords4[3] = lp_build_extract_range(gallivm, newcoords[3], 4*i, 4);
4089                newcoords4[4] = lp_build_extract_range(gallivm, newcoords[4], 4*i, 4);
4090 
4091                lp_build_sample_general(&bld4, sampler_index,
4092                                        op_type == LP_SAMPLER_OP_GATHER,
4093                                        newcoords4, offsets4,
4094                                        lod_positive4, lod_fpart4,
4095                                        ilevel04, ilevel14,
4096                                        texelout4);
4097             }
4098             for (unsigned j = 0; j < 4; j++) {
4099                texelouttmp[j][i] = texelout4[j];
4100             }
4101          }
4102 
4103          for (unsigned j = 0; j < 4; j++) {
4104             texel_out[j] = lp_build_concat(gallivm, texelouttmp[j], type4, num_quads);
4105          }
4106       }
4107    }
4108 
4109    if (target != PIPE_BUFFER && op_type != LP_SAMPLER_OP_GATHER) {
4110       apply_sampler_swizzle(&bld, texel_out);
4111    }
4112 
4113    /*
4114     * texel type can be a (32bit) int/uint (for pure int formats only),
4115     * however we are expected to always return floats (storage is untyped).
4116     */
4117    if (!bld.texel_type.floating) {
4118       unsigned chan;
4119       for (chan = 0; chan < 4; chan++) {
4120          texel_out[chan] = LLVMBuildBitCast(builder, texel_out[chan],
4121                                             lp_build_vec_type(gallivm, type), "");
4122       }
4123    }
4124 }
4125 
4126 
4127 #define USE_TEX_FUNC_CALL 1
4128 
4129 static inline void
get_target_info(enum pipe_texture_target target,unsigned * num_coords,unsigned * num_derivs,unsigned * num_offsets,unsigned * layer)4130 get_target_info(enum pipe_texture_target target,
4131                 unsigned *num_coords, unsigned *num_derivs,
4132                 unsigned *num_offsets, unsigned *layer)
4133 {
4134    unsigned dims = texture_dims(target);
4135    *num_coords = dims;
4136    *num_offsets = dims;
4137    *num_derivs = (target == PIPE_TEXTURE_CUBE ||
4138                   target == PIPE_TEXTURE_CUBE_ARRAY) ? 3 : dims;
4139    *layer = has_layer_coord(target) ? 2: 0;
4140    if (target == PIPE_TEXTURE_CUBE_ARRAY) {
4141       /*
4142        * dims doesn't include r coord for cubes - this is handled
4143        * by layer instead, but need to fix up for cube arrays...
4144        */
4145       *layer = 3;
4146       *num_coords = 3;
4147    }
4148 }
4149 
4150 
4151 /**
4152  * Generate the function body for a texture sampling function.
4153  */
4154 static void
lp_build_sample_gen_func(struct gallivm_state * gallivm,const struct lp_static_texture_state * static_texture_state,const struct lp_static_sampler_state * static_sampler_state,struct lp_sampler_dynamic_state * dynamic_state,struct lp_type type,LLVMTypeRef resources_type,LLVMTypeRef thread_data_type,unsigned texture_index,unsigned sampler_index,LLVMValueRef function,unsigned num_args,unsigned sample_key,bool has_aniso_filter_table)4155 lp_build_sample_gen_func(struct gallivm_state *gallivm,
4156                          const struct lp_static_texture_state *static_texture_state,
4157                          const struct lp_static_sampler_state *static_sampler_state,
4158                          struct lp_sampler_dynamic_state *dynamic_state,
4159                          struct lp_type type,
4160                          LLVMTypeRef resources_type,
4161                          LLVMTypeRef thread_data_type,
4162                          unsigned texture_index,
4163                          unsigned sampler_index,
4164                          LLVMValueRef function,
4165                          unsigned num_args,
4166                          unsigned sample_key,
4167                          bool has_aniso_filter_table)
4168 {
4169    LLVMBuilderRef old_builder;
4170    LLVMBasicBlockRef block;
4171    LLVMValueRef coords[5];
4172    LLVMValueRef offsets[3] = { NULL };
4173    LLVMValueRef lod = NULL;
4174    LLVMValueRef ms_index = NULL;
4175    LLVMValueRef resources_ptr;
4176    LLVMValueRef thread_data_ptr = NULL;
4177    LLVMValueRef aniso_filter_table = NULL;
4178    LLVMValueRef texel_out[4];
4179    struct lp_derivatives derivs;
4180    struct lp_derivatives *deriv_ptr = NULL;
4181    unsigned num_param = 0;
4182    unsigned num_coords, num_derivs, num_offsets, layer;
4183    bool need_cache = false;
4184 
4185    const enum lp_sampler_lod_control lod_control =
4186        (sample_key & LP_SAMPLER_LOD_CONTROL_MASK)
4187        >> LP_SAMPLER_LOD_CONTROL_SHIFT;
4188 
4189    const enum lp_sampler_op_type op_type =
4190       (sample_key & LP_SAMPLER_OP_TYPE_MASK) >> LP_SAMPLER_OP_TYPE_SHIFT;
4191 
4192    get_target_info(static_texture_state->target,
4193                    &num_coords, &num_derivs, &num_offsets, &layer);
4194 
4195    /* lod query doesn't take a layer */
4196    if (layer && op_type == LP_SAMPLER_OP_LODQ)
4197       layer = 0;
4198 
4199    if (dynamic_state->cache_ptr) {
4200       const struct util_format_description *format_desc;
4201       format_desc = util_format_description(static_texture_state->format);
4202       if (format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
4203          need_cache = true;
4204       }
4205    }
4206 
4207    /* "unpack" arguments */
4208    resources_ptr = LLVMGetParam(function, num_param++);
4209    if (has_aniso_filter_table)
4210       aniso_filter_table = LLVMGetParam(function, num_param++);
4211    if (need_cache) {
4212       thread_data_ptr = LLVMGetParam(function, num_param++);
4213    }
4214    for (unsigned i = 0; i < num_coords; i++) {
4215       coords[i] = LLVMGetParam(function, num_param++);
4216    }
4217    for (unsigned i = num_coords; i < 5; i++) {
4218       /* This is rather unfortunate... */
4219       coords[i] = lp_build_undef(gallivm, type);
4220    }
4221    if (layer) {
4222       coords[layer] = LLVMGetParam(function, num_param++);
4223    }
4224    if (sample_key & LP_SAMPLER_SHADOW) {
4225       coords[4] = LLVMGetParam(function, num_param++);
4226    }
4227    if (sample_key & LP_SAMPLER_FETCH_MS) {
4228       ms_index = LLVMGetParam(function, num_param++);
4229    }
4230    if (sample_key & LP_SAMPLER_OFFSETS) {
4231       for (unsigned i = 0; i < num_offsets; i++) {
4232          offsets[i] = LLVMGetParam(function, num_param++);
4233       }
4234    }
4235    if (lod_control == LP_SAMPLER_LOD_BIAS ||
4236        lod_control == LP_SAMPLER_LOD_EXPLICIT) {
4237       lod = LLVMGetParam(function, num_param++);
4238    } else if (lod_control == LP_SAMPLER_LOD_DERIVATIVES) {
4239       for (unsigned i = 0; i < num_derivs; i++) {
4240          derivs.ddx[i] = LLVMGetParam(function, num_param++);
4241          derivs.ddy[i] = LLVMGetParam(function, num_param++);
4242       }
4243       deriv_ptr = &derivs;
4244    }
4245 
4246    assert(num_args == num_param);
4247 
4248    /*
4249     * Function body
4250     */
4251 
4252    old_builder = gallivm->builder;
4253    block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
4254    gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
4255    LLVMPositionBuilderAtEnd(gallivm->builder, block);
4256 
4257    lp_build_sample_soa_code(gallivm,
4258                             static_texture_state,
4259                             static_sampler_state,
4260                             dynamic_state,
4261                             type,
4262                             sample_key,
4263                             texture_index,
4264                             sampler_index,
4265                             resources_type,
4266                             resources_ptr,
4267                             thread_data_type,
4268                             thread_data_ptr,
4269                             coords,
4270                             offsets,
4271                             deriv_ptr,
4272                             lod,
4273                             ms_index,
4274                             aniso_filter_table,
4275                             texel_out);
4276 
4277    LLVMBuildAggregateRet(gallivm->builder, texel_out, 4);
4278 
4279    LLVMDisposeBuilder(gallivm->builder);
4280    gallivm->builder = old_builder;
4281 
4282    gallivm_verify_function(gallivm, function);
4283 }
4284 
4285 
4286 /**
4287  * Call the matching function for texture sampling.
4288  * If there's no match, generate a new one.
4289  */
4290 static void
lp_build_sample_soa_func(struct gallivm_state * gallivm,const struct lp_static_texture_state * static_texture_state,const struct lp_static_sampler_state * static_sampler_state,struct lp_sampler_dynamic_state * dynamic_state,const struct lp_sampler_params * params,unsigned texture_index,unsigned sampler_index,LLVMValueRef * tex_ret)4291 lp_build_sample_soa_func(struct gallivm_state *gallivm,
4292                          const struct lp_static_texture_state *static_texture_state,
4293                          const struct lp_static_sampler_state *static_sampler_state,
4294                          struct lp_sampler_dynamic_state *dynamic_state,
4295                          const struct lp_sampler_params *params,
4296                          unsigned texture_index, unsigned sampler_index,
4297                          LLVMValueRef *tex_ret)
4298 {
4299    LLVMBuilderRef builder = gallivm->builder;
4300    LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(
4301                              LLVMGetInsertBlock(builder)));
4302    LLVMValueRef args[LP_MAX_TEX_FUNC_ARGS];
4303    unsigned sample_key = params->sample_key;
4304    const LLVMValueRef *coords = params->coords;
4305    const LLVMValueRef *offsets = params->offsets;
4306    const struct lp_derivatives *derivs = params->derivs;
4307 
4308    const enum lp_sampler_lod_control lod_control =
4309       (sample_key & LP_SAMPLER_LOD_CONTROL_MASK) >>
4310       LP_SAMPLER_LOD_CONTROL_SHIFT;
4311 
4312    const enum lp_sampler_op_type op_type =
4313       (sample_key & LP_SAMPLER_OP_TYPE_MASK) >> LP_SAMPLER_OP_TYPE_SHIFT;
4314 
4315    unsigned num_coords, num_derivs, num_offsets, layer;
4316    get_target_info(static_texture_state->target,
4317                    &num_coords, &num_derivs, &num_offsets, &layer);
4318 
4319    /* lod query doesn't take a layer */
4320    if (layer && op_type == LP_SAMPLER_OP_LODQ)
4321       layer = 0;
4322 
4323    bool need_cache = false;
4324    if (dynamic_state->cache_ptr) {
4325       const struct util_format_description *format_desc;
4326       format_desc = util_format_description(static_texture_state->format);
4327       if (format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
4328          need_cache = true;
4329       }
4330    }
4331 
4332    /*
4333     * texture function matches are found by name.
4334     * Thus the name has to include both the texture and sampler unit
4335     * (which covers all static state) plus the actual texture function
4336     * (including things like offsets, shadow coord, lod control).
4337     * Additionally lod_property has to be included too.
4338     */
4339    char func_name[64];
4340    snprintf(func_name, sizeof(func_name), "texfunc_res_%d_sam_%d_%x",
4341             texture_index, sampler_index, sample_key);
4342 
4343    LLVMValueRef function = LLVMGetNamedFunction(module, func_name);
4344    LLVMTypeRef arg_types[LP_MAX_TEX_FUNC_ARGS];
4345    LLVMTypeRef ret_type;
4346    LLVMTypeRef val_type[4];
4347    unsigned num_param = 0;
4348 
4349    /*
4350     * Generate the function prototype.
4351     */
4352 
4353    arg_types[num_param++] = LLVMTypeOf(params->resources_ptr);
4354    if (params->aniso_filter_table)
4355       arg_types[num_param++] = LLVMTypeOf(params->aniso_filter_table);
4356    if (need_cache) {
4357       arg_types[num_param++] = LLVMTypeOf(params->thread_data_ptr);
4358    }
4359    for (unsigned i = 0; i < num_coords; i++) {
4360       arg_types[num_param++] = LLVMTypeOf(coords[0]);
4361       assert(LLVMTypeOf(coords[0]) == LLVMTypeOf(coords[i]));
4362    }
4363    if (layer) {
4364       arg_types[num_param++] = LLVMTypeOf(coords[layer]);
4365       assert(LLVMTypeOf(coords[0]) == LLVMTypeOf(coords[layer]));
4366    }
4367    if (sample_key & LP_SAMPLER_SHADOW) {
4368       arg_types[num_param++] = LLVMTypeOf(coords[0]);
4369    }
4370    if (sample_key & LP_SAMPLER_FETCH_MS) {
4371       arg_types[num_param++] = LLVMTypeOf(params->ms_index);
4372    }
4373    if (sample_key & LP_SAMPLER_OFFSETS) {
4374       for (unsigned i = 0; i < num_offsets; i++) {
4375          arg_types[num_param++] = LLVMTypeOf(offsets[0]);
4376          assert(LLVMTypeOf(offsets[0]) == LLVMTypeOf(offsets[i]));
4377       }
4378    }
4379    if (lod_control == LP_SAMPLER_LOD_BIAS ||
4380        lod_control == LP_SAMPLER_LOD_EXPLICIT) {
4381       arg_types[num_param++] = LLVMTypeOf(params->lod);
4382    } else if (lod_control == LP_SAMPLER_LOD_DERIVATIVES) {
4383       for (unsigned i = 0; i < num_derivs; i++) {
4384          arg_types[num_param++] = LLVMTypeOf(derivs->ddx[i]);
4385          arg_types[num_param++] = LLVMTypeOf(derivs->ddy[i]);
4386          assert(LLVMTypeOf(derivs->ddx[0]) == LLVMTypeOf(derivs->ddx[i]));
4387          assert(LLVMTypeOf(derivs->ddy[0]) == LLVMTypeOf(derivs->ddy[i]));
4388       }
4389    }
4390 
4391    val_type[0] = val_type[1] = val_type[2] = val_type[3] =
4392          lp_build_vec_type(gallivm, params->type);
4393    ret_type = LLVMStructTypeInContext(gallivm->context, val_type, 4, 0);
4394    LLVMTypeRef function_type = LLVMFunctionType(ret_type, arg_types, num_param, 0);
4395 
4396    if (!function) {
4397       function = LLVMAddFunction(module, func_name, function_type);
4398 
4399       for (unsigned i = 0; i < num_param; ++i) {
4400          if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) {
4401 
4402             lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
4403          }
4404       }
4405 
4406       LLVMSetFunctionCallConv(function, LLVMFastCallConv);
4407       LLVMSetLinkage(function, LLVMInternalLinkage);
4408 
4409       lp_build_sample_gen_func(gallivm,
4410                                static_texture_state,
4411                                static_sampler_state,
4412                                dynamic_state,
4413                                params->type,
4414                                params->resources_type,
4415                                params->thread_data_type,
4416                                texture_index,
4417                                sampler_index,
4418                                function,
4419                                num_param,
4420                                sample_key,
4421                                params->aniso_filter_table ? true : false);
4422    }
4423 
4424    unsigned num_args = 0;
4425    args[num_args++] = params->resources_ptr;
4426    if (params->aniso_filter_table)
4427       args[num_args++] = params->aniso_filter_table;
4428    if (need_cache) {
4429       args[num_args++] = params->thread_data_ptr;
4430    }
4431    for (unsigned i = 0; i < num_coords; i++) {
4432       args[num_args++] = coords[i];
4433    }
4434    if (layer) {
4435       args[num_args++] = coords[layer];
4436    }
4437    if (sample_key & LP_SAMPLER_SHADOW) {
4438       args[num_args++] = coords[4];
4439    }
4440    if (sample_key & LP_SAMPLER_FETCH_MS) {
4441       args[num_args++] = params->ms_index;
4442    }
4443    if (sample_key & LP_SAMPLER_OFFSETS) {
4444       for (unsigned i = 0; i < num_offsets; i++) {
4445          args[num_args++] = offsets[i];
4446       }
4447    }
4448    if (lod_control == LP_SAMPLER_LOD_BIAS ||
4449        lod_control == LP_SAMPLER_LOD_EXPLICIT) {
4450       args[num_args++] = params->lod;
4451    } else if (lod_control == LP_SAMPLER_LOD_DERIVATIVES) {
4452       for (unsigned i = 0; i < num_derivs; i++) {
4453          args[num_args++] = derivs->ddx[i];
4454          args[num_args++] = derivs->ddy[i];
4455       }
4456    }
4457 
4458    assert(num_args <= LP_MAX_TEX_FUNC_ARGS);
4459 
4460    *tex_ret = LLVMBuildCall2(builder, function_type, function, args, num_args, "");
4461    LLVMBasicBlockRef bb = LLVMGetInsertBlock(builder);
4462    LLVMValueRef inst = LLVMGetLastInstruction(bb);
4463    LLVMSetInstructionCallConv(inst, LLVMFastCallConv);
4464 }
4465 
4466 
4467 /**
4468  * Build texture sampling code.
4469  * Either via a function call or inline it directly.
4470  */
4471 void
lp_build_sample_soa(const struct lp_static_texture_state * static_texture_state,const struct lp_static_sampler_state * static_sampler_state,struct lp_sampler_dynamic_state * dynamic_state,struct gallivm_state * gallivm,const struct lp_sampler_params * params)4472 lp_build_sample_soa(const struct lp_static_texture_state *static_texture_state,
4473                     const struct lp_static_sampler_state *static_sampler_state,
4474                     struct lp_sampler_dynamic_state *dynamic_state,
4475                     struct gallivm_state *gallivm,
4476                     const struct lp_sampler_params *params)
4477 {
4478    bool use_tex_func = false;
4479 
4480    /*
4481     * Do not use a function call if the sampling is "simple enough".
4482     * We define this by
4483     * a) format
4484     * b) no mips (either one level only or no mip filter)
4485     * No mips will definitely make the code smaller, though
4486     * the format requirement is a bit iffy - there's some (SoA) formats
4487     * which definitely generate less code. This does happen to catch
4488     * some important cases though which are hurt quite a bit by using
4489     * a call (though not really because of the call overhead but because
4490     * they are reusing the same texture unit with some of the same
4491     * parameters).
4492     * Ideally we'd let llvm recognize this stuff by doing IPO passes.
4493     */
4494 
4495    if (USE_TEX_FUNC_CALL) {
4496       const struct util_format_description *format_desc =
4497          util_format_description(static_texture_state->format);
4498       const bool simple_format =
4499          (util_format_is_rgba8_variant(format_desc) &&
4500          format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB);
4501       const enum lp_sampler_op_type op_type =
4502          (params->sample_key & LP_SAMPLER_OP_TYPE_MASK) >>
4503          LP_SAMPLER_OP_TYPE_SHIFT;
4504       const bool simple_tex =
4505          op_type != LP_SAMPLER_OP_TEXTURE ||
4506            ((static_sampler_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE ||
4507              static_texture_state->level_zero_only == true) &&
4508             static_sampler_state->min_img_filter == static_sampler_state->mag_img_filter);
4509 
4510       use_tex_func = !(simple_format && simple_tex);
4511    }
4512 
4513    if (use_tex_func) {
4514       LLVMValueRef tex_ret;
4515       lp_build_sample_soa_func(gallivm,
4516                                static_texture_state,
4517                                static_sampler_state,
4518                                dynamic_state,
4519                                params, params->texture_index,
4520                                params->sampler_index, &tex_ret);
4521 
4522       for (unsigned i = 0; i < 4; i++) {
4523          params->texel[i] =
4524             LLVMBuildExtractValue(gallivm->builder, tex_ret, i, "");
4525       }
4526    } else {
4527       lp_build_sample_soa_code(gallivm,
4528                                static_texture_state,
4529                                static_sampler_state,
4530                                dynamic_state,
4531                                params->type,
4532                                params->sample_key,
4533                                params->texture_index,
4534                                params->sampler_index,
4535                                params->resources_type,
4536                                params->resources_ptr,
4537                                params->thread_data_type,
4538                                params->thread_data_ptr,
4539                                params->coords,
4540                                params->offsets,
4541                                params->derivs,
4542                                params->lod,
4543                                params->ms_index,
4544                                params->aniso_filter_table,
4545                                params->texel);
4546    }
4547 }
4548 
4549 
4550 void
lp_build_size_query_soa(struct gallivm_state * gallivm,const struct lp_static_texture_state * static_state,struct lp_sampler_dynamic_state * dynamic_state,const struct lp_sampler_size_query_params * params)4551 lp_build_size_query_soa(struct gallivm_state *gallivm,
4552                         const struct lp_static_texture_state *static_state,
4553                         struct lp_sampler_dynamic_state *dynamic_state,
4554                         const struct lp_sampler_size_query_params *params)
4555 {
4556    LLVMValueRef first_level = NULL;
4557    const unsigned num_lods = 1;
4558    LLVMTypeRef resources_type = params->resources_type;
4559    LLVMValueRef resources_ptr = params->resources_ptr;
4560    const unsigned texture_unit = params->texture_unit;
4561    const enum pipe_texture_target target = params->target;
4562    LLVMValueRef texture_unit_offset = params->texture_unit_offset;
4563    const struct util_format_description *format_desc =
4564       util_format_description(static_state->format);
4565    const struct util_format_description *res_format_desc =
4566       util_format_description(static_state->res_format);
4567 
4568    if (static_state->format == PIPE_FORMAT_NONE) {
4569       /*
4570        * If there's nothing bound, format is NONE, and we must return
4571        * all zero as mandated by d3d10 in this case.
4572        */
4573       LLVMValueRef zero = lp_build_const_vec(gallivm, params->int_type, 0.0F);
4574       for (unsigned chan = 0; chan < 4; chan++) {
4575          params->sizes_out[chan] = zero;
4576       }
4577       return;
4578    }
4579 
4580    /*
4581     * Do some sanity verification about bound texture and shader dcl target.
4582     * Not entirely sure what's possible but assume array/non-array
4583     * always compatible (probably not ok for OpenGL but d3d10 has no
4584     * distinction of arrays at the resource level).
4585     * Everything else looks bogus (though not entirely sure about rect/2d).
4586     * Currently disabled because it causes assertion failures if there's
4587     * nothing bound (or rather a dummy texture, not that this case would
4588     * return the right values).
4589     */
4590    if (0 && static_state->target != target) {
4591       if (static_state->target == PIPE_TEXTURE_1D)
4592          assert(target == PIPE_TEXTURE_1D_ARRAY);
4593       else if (static_state->target == PIPE_TEXTURE_1D_ARRAY)
4594          assert(target == PIPE_TEXTURE_1D);
4595       else if (static_state->target == PIPE_TEXTURE_2D)
4596          assert(target == PIPE_TEXTURE_2D_ARRAY);
4597       else if (static_state->target == PIPE_TEXTURE_2D_ARRAY)
4598          assert(target == PIPE_TEXTURE_2D);
4599       else if (static_state->target == PIPE_TEXTURE_CUBE)
4600          assert(target == PIPE_TEXTURE_CUBE_ARRAY);
4601       else if (static_state->target == PIPE_TEXTURE_CUBE_ARRAY)
4602          assert(target == PIPE_TEXTURE_CUBE);
4603       else
4604          assert(0);
4605    }
4606 
4607    const unsigned dims = texture_dims(target);
4608 
4609    const bool has_array = has_layer_coord(target);
4610 
4611    assert(!params->int_type.floating);
4612 
4613    struct lp_build_context bld_int_vec4;
4614    lp_build_context_init(&bld_int_vec4, gallivm, lp_type_int_vec(32, 128));
4615 
4616    if (params->samples_only) {
4617       LLVMValueRef num_samples;
4618       if (params->ms && static_state->level_zero_only) {
4619          /* multisample never has levels. */
4620          num_samples = dynamic_state->last_level(gallivm,
4621                                                  resources_type,
4622                                                  resources_ptr,
4623                                                  texture_unit,
4624                                                  texture_unit_offset);
4625          num_samples = LLVMBuildZExt(gallivm->builder, num_samples,
4626                                      bld_int_vec4.elem_type, "");
4627       } else {
4628          num_samples = lp_build_const_int32(gallivm, 0);
4629       }
4630       params->sizes_out[0] =
4631          lp_build_broadcast(gallivm,
4632                             lp_build_vec_type(gallivm, params->int_type),
4633                             num_samples);
4634       return;
4635    }
4636 
4637    LLVMValueRef lod;
4638    LLVMValueRef level = 0;
4639    if (params->explicit_lod) {
4640       /* FIXME: this needs to honor per-element lod */
4641       lod = LLVMBuildExtractElement(gallivm->builder, params->explicit_lod,
4642                                     lp_build_const_int32(gallivm, 0), "");
4643       first_level = get_first_level(gallivm, resources_type, resources_ptr,
4644                                     texture_unit, texture_unit_offset,
4645                                     static_state, dynamic_state);
4646       level = LLVMBuildAdd(gallivm->builder, lod, first_level, "level");
4647       lod = lp_build_broadcast_scalar(&bld_int_vec4, level);
4648    } else {
4649       lod = bld_int_vec4.zero;
4650    }
4651 
4652    LLVMValueRef size = bld_int_vec4.undef;
4653    LLVMValueRef tex_blocksize = bld_int_vec4.undef;
4654    LLVMValueRef tex_blocksize_log2 = bld_int_vec4.undef;
4655    LLVMValueRef view_blocksize = bld_int_vec4.undef;
4656 
4657    uint32_t res_bw = res_format_desc->block.width;
4658    uint32_t res_bh = res_format_desc->block.height;
4659    uint32_t bw = format_desc->block.width;
4660    uint32_t bh = format_desc->block.height;
4661 
4662    /* only scale if the blocksizes are different. */
4663    if (res_bw == bw)
4664       res_bw = bw = 1;
4665    if (res_bh == bh)
4666       res_bh = bh = 1;
4667 
4668    LLVMValueRef tex_width = dynamic_state->width(gallivm,
4669                                                  resources_type,
4670                                                  resources_ptr,
4671                                                  texture_unit,
4672                                                  texture_unit_offset);
4673    size = LLVMBuildInsertElement(gallivm->builder, size,
4674                                  tex_width,
4675                                  lp_build_const_int32(gallivm, 0), "");
4676    tex_blocksize = LLVMBuildInsertElement(gallivm->builder, tex_blocksize,
4677                                           lp_build_const_int32(gallivm, res_bw),
4678                                           lp_build_const_int32(gallivm, 0), "");
4679    tex_blocksize_log2 = LLVMBuildInsertElement(gallivm->builder, tex_blocksize_log2,
4680                                                lp_build_const_int32(gallivm, util_logbase2(res_bw)),
4681                                                lp_build_const_int32(gallivm, 0), "");
4682    view_blocksize = LLVMBuildInsertElement(gallivm->builder, view_blocksize,
4683                                            lp_build_const_int32(gallivm, bw),
4684                                            lp_build_const_int32(gallivm, 0), "");
4685    if (dims >= 2) {
4686       LLVMValueRef tex_height =
4687          dynamic_state->height(gallivm, resources_type,
4688                                resources_ptr, texture_unit, texture_unit_offset);
4689       tex_height = LLVMBuildZExt(gallivm->builder, tex_height,
4690                                  bld_int_vec4.elem_type, "");
4691       size = LLVMBuildInsertElement(gallivm->builder, size, tex_height,
4692                                     lp_build_const_int32(gallivm, 1), "");
4693       tex_blocksize = LLVMBuildInsertElement(gallivm->builder, tex_blocksize,
4694                                              lp_build_const_int32(gallivm, res_bh),
4695                                              lp_build_const_int32(gallivm, 1), "");
4696       tex_blocksize_log2 = LLVMBuildInsertElement(gallivm->builder, tex_blocksize_log2,
4697                                                   lp_build_const_int32(gallivm, util_logbase2(res_bh)),
4698                                                   lp_build_const_int32(gallivm, 1), "");
4699       view_blocksize = LLVMBuildInsertElement(gallivm->builder, view_blocksize,
4700                                               lp_build_const_int32(gallivm, bh),
4701                                               lp_build_const_int32(gallivm, 1), "");
4702    }
4703 
4704    if (dims >= 3) {
4705       LLVMValueRef tex_depth  =
4706          dynamic_state->depth(gallivm, resources_type,
4707                               resources_ptr, texture_unit, texture_unit_offset);
4708       tex_depth = LLVMBuildZExt(gallivm->builder, tex_depth,
4709                                 bld_int_vec4.elem_type, "");
4710       size = LLVMBuildInsertElement(gallivm->builder, size, tex_depth,
4711                                     lp_build_const_int32(gallivm, 2), "");
4712       tex_blocksize = LLVMBuildInsertElement(gallivm->builder, tex_blocksize,
4713                                              lp_build_const_int32(gallivm, 1),
4714                                              lp_build_const_int32(gallivm, 2), "");
4715       tex_blocksize_log2 = LLVMBuildInsertElement(gallivm->builder, tex_blocksize_log2,
4716                                                   lp_build_const_int32(gallivm, 0),
4717                                                   lp_build_const_int32(gallivm, 2), "");
4718       view_blocksize = LLVMBuildInsertElement(gallivm->builder, view_blocksize,
4719                                               lp_build_const_int32(gallivm, 1),
4720                                               lp_build_const_int32(gallivm, 2), "");
4721    }
4722 
4723    size = lp_build_minify(&bld_int_vec4, size, lod, true);
4724    size = lp_build_scale_view_dims(&bld_int_vec4, size, tex_blocksize,
4725                                    tex_blocksize_log2, view_blocksize);
4726 
4727    if (has_array) {
4728       LLVMValueRef layers = dynamic_state->depth(gallivm, resources_type,
4729                                                  resources_ptr, texture_unit,
4730                                                  texture_unit_offset);
4731       layers = LLVMBuildZExt(gallivm->builder, layers,
4732                              bld_int_vec4.elem_type, "");
4733       if (target == PIPE_TEXTURE_CUBE_ARRAY) {
4734          /*
4735           * It looks like GL wants number of cubes, d3d10.1 has it undefined?
4736           * Could avoid this by passing in number of cubes instead of total
4737           * number of layers (might make things easier elsewhere too).
4738           */
4739          LLVMValueRef six = lp_build_const_int32(gallivm, 6);
4740          layers = LLVMBuildSDiv(gallivm->builder, layers, six, "");
4741       }
4742       size = LLVMBuildInsertElement(gallivm->builder, size, layers,
4743                                     lp_build_const_int32(gallivm, dims), "");
4744    }
4745 
4746    /*
4747     * d3d10 requires zero for x/y/z values (but not w, i.e. mip levels)
4748     * if level is out of bounds (note this can't cover unbound texture
4749     * here, which also requires returning zero).
4750     */
4751    if (params->explicit_lod && params->is_sviewinfo) {
4752       LLVMValueRef last_level, out, out1;
4753       struct lp_build_context leveli_bld;
4754 
4755       /* everything is scalar for now */
4756       lp_build_context_init(&leveli_bld, gallivm, lp_type_int_vec(32, 32));
4757       last_level = get_last_level(gallivm, resources_type, resources_ptr,
4758                                   texture_unit, texture_unit_offset,
4759                                   static_state, dynamic_state);
4760 
4761       out = lp_build_cmp(&leveli_bld, PIPE_FUNC_LESS, level, first_level);
4762       out1 = lp_build_cmp(&leveli_bld, PIPE_FUNC_GREATER, level, last_level);
4763       out = lp_build_or(&leveli_bld, out, out1);
4764       if (num_lods == 1) {
4765          out = lp_build_broadcast_scalar(&bld_int_vec4, out);
4766       } else {
4767          /* TODO */
4768          assert(0);
4769       }
4770       size = lp_build_andnot(&bld_int_vec4, size, out);
4771    }
4772 
4773    unsigned i;
4774    for (i = 0; i < dims + (has_array ? 1 : 0); i++) {
4775       params->sizes_out[i] =
4776          lp_build_extract_broadcast(gallivm, bld_int_vec4.type,
4777                                     params->int_type,
4778                                     size,
4779                                     lp_build_const_int32(gallivm, i));
4780    }
4781    if (params->is_sviewinfo) {
4782       for (; i < 4; i++) {
4783          params->sizes_out[i] = lp_build_const_vec(gallivm,
4784                                                    params->int_type, 0.0);
4785       }
4786    }
4787 
4788    /*
4789     * if there's no explicit_lod (buffers, rects) queries requiring nr of
4790     * mips would be illegal.
4791     */
4792    if (params->is_sviewinfo && params->explicit_lod) {
4793       struct lp_build_context bld_int_scalar;
4794       lp_build_context_init(&bld_int_scalar, gallivm, lp_type_int(32));
4795 
4796       LLVMValueRef num_levels;
4797       if (static_state->level_zero_only) {
4798          num_levels = bld_int_scalar.one;
4799       } else {
4800          LLVMValueRef last_level;
4801          last_level = get_last_level(gallivm, resources_type, resources_ptr,
4802                                      texture_unit, texture_unit_offset,
4803                                      static_state, dynamic_state);
4804          num_levels = lp_build_sub(&bld_int_scalar, last_level, first_level);
4805          num_levels = lp_build_add(&bld_int_scalar, num_levels,
4806                                    bld_int_scalar.one);
4807       }
4808       params->sizes_out[3] =
4809          lp_build_broadcast(gallivm,
4810                             lp_build_vec_type(gallivm, params->int_type),
4811                             num_levels);
4812    }
4813 
4814    if (target == PIPE_BUFFER) {
4815       struct lp_build_context bld_int;
4816       lp_build_context_init(&bld_int, gallivm, params->int_type);
4817 
4818       params->sizes_out[0] = lp_build_min(&bld_int, params->sizes_out[0],
4819          lp_build_const_int_vec(gallivm, params->int_type, LP_MAX_TEXEL_BUFFER_ELEMENTS));
4820    }
4821 }
4822 
4823 
4824 static void
lp_build_do_atomic_soa(struct gallivm_state * gallivm,const struct util_format_description * format_desc,struct lp_type type,LLVMValueRef exec_mask,LLVMValueRef base_ptr,LLVMValueRef offset,LLVMValueRef out_of_bounds,unsigned img_op,LLVMAtomicRMWBinOp op,const LLVMValueRef rgba_in[4],const LLVMValueRef rgba2_in[4],LLVMValueRef atomic_result[4])4825 lp_build_do_atomic_soa(struct gallivm_state *gallivm,
4826                        const struct util_format_description *format_desc,
4827                        struct lp_type type,
4828                        LLVMValueRef exec_mask,
4829                        LLVMValueRef base_ptr,
4830                        LLVMValueRef offset,
4831                        LLVMValueRef out_of_bounds,
4832                        unsigned img_op,
4833                        LLVMAtomicRMWBinOp op,
4834                        const LLVMValueRef rgba_in[4],
4835                        const LLVMValueRef rgba2_in[4],
4836                        LLVMValueRef atomic_result[4])
4837 {
4838    const enum pipe_format format = format_desc->format;
4839 
4840    bool valid = format == PIPE_FORMAT_R32_UINT ||
4841                 format == PIPE_FORMAT_R32_SINT ||
4842                 format == PIPE_FORMAT_R32_FLOAT;
4843 
4844    bool integer = format != PIPE_FORMAT_R32_FLOAT;
4845    if (img_op == LP_IMG_ATOMIC) {
4846       switch (op) {
4847       case LLVMAtomicRMWBinOpAdd:
4848       case LLVMAtomicRMWBinOpSub:
4849       case LLVMAtomicRMWBinOpAnd:
4850       case LLVMAtomicRMWBinOpNand:
4851       case LLVMAtomicRMWBinOpOr:
4852       case LLVMAtomicRMWBinOpXor:
4853       case LLVMAtomicRMWBinOpMax:
4854       case LLVMAtomicRMWBinOpMin:
4855       case LLVMAtomicRMWBinOpUMax:
4856       case LLVMAtomicRMWBinOpUMin:
4857          valid &= integer;
4858          break;
4859       case LLVMAtomicRMWBinOpFAdd:
4860       case LLVMAtomicRMWBinOpFSub:
4861 #if LLVM_VERSION_MAJOR >= 15
4862          case LLVMAtomicRMWBinOpFMax:
4863          case LLVMAtomicRMWBinOpFMin:
4864 #endif
4865          valid &= !integer;
4866          break;
4867       default:
4868          break;
4869       }
4870    } else {
4871       valid &= integer;
4872    }
4873 
4874    if (!valid) {
4875       atomic_result[0] = lp_build_zero(gallivm, type);
4876       return;
4877    }
4878 
4879    LLVMTypeRef ref_type = (format == PIPE_FORMAT_R32_FLOAT) ?
4880       LLVMFloatTypeInContext(gallivm->context) :
4881       LLVMInt32TypeInContext(gallivm->context);
4882 
4883    LLVMTypeRef atom_res_elem_type =
4884       LLVMVectorType(ref_type, type.length);
4885    LLVMValueRef atom_res = lp_build_alloca(gallivm, atom_res_elem_type, "");
4886 
4887    offset = LLVMBuildGEP2(gallivm->builder,
4888                           LLVMInt8TypeInContext(gallivm->context),
4889                           base_ptr, &offset, 1, "");
4890 
4891    struct lp_build_loop_state loop_state;
4892    lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
4893    struct lp_build_if_state ifthen;
4894    LLVMValueRef cond;
4895    LLVMValueRef packed = rgba_in[0], packed2 = rgba2_in[0];
4896 
4897    LLVMValueRef should_store_mask =
4898       LLVMBuildAnd(gallivm->builder, exec_mask,
4899                    LLVMBuildNot(gallivm->builder, out_of_bounds, ""),
4900                    "store_mask");
4901    assert(exec_mask);
4902 
4903    cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, should_store_mask,
4904                         lp_build_const_int_vec(gallivm, type, 0), "");
4905    cond = LLVMBuildExtractElement(gallivm->builder, cond,
4906                                   loop_state.counter, "");
4907    lp_build_if(&ifthen, gallivm, cond);
4908 
4909    LLVMValueRef data =
4910       LLVMBuildExtractElement(gallivm->builder, packed, loop_state.counter, "");
4911    LLVMValueRef cast_base_ptr =
4912       LLVMBuildExtractElement(gallivm->builder, offset, loop_state.counter, "");
4913    cast_base_ptr = LLVMBuildBitCast(gallivm->builder, cast_base_ptr,
4914               LLVMPointerType(ref_type, 0), "");
4915    data = LLVMBuildBitCast(gallivm->builder, data,
4916                            ref_type, "");
4917 
4918    if (img_op == LP_IMG_ATOMIC_CAS) {
4919       LLVMValueRef cas_src_ptr =
4920          LLVMBuildExtractElement(gallivm->builder, packed2,
4921                                  loop_state.counter, "");
4922       LLVMValueRef cas_src =
4923          LLVMBuildBitCast(gallivm->builder, cas_src_ptr,
4924                           ref_type, "");
4925       data = LLVMBuildAtomicCmpXchg(gallivm->builder, cast_base_ptr, data,
4926                                     cas_src,
4927                                     LLVMAtomicOrderingSequentiallyConsistent,
4928                                     LLVMAtomicOrderingSequentiallyConsistent,
4929                                     false);
4930       data = LLVMBuildExtractValue(gallivm->builder, data, 0, "");
4931    } else {
4932       data = LLVMBuildAtomicRMW(gallivm->builder, op,
4933                                 cast_base_ptr, data,
4934                                 LLVMAtomicOrderingSequentiallyConsistent,
4935                                 false);
4936    }
4937 
4938    LLVMValueRef temp_res =
4939       LLVMBuildLoad2(gallivm->builder, atom_res_elem_type, atom_res, "");
4940    temp_res = LLVMBuildInsertElement(gallivm->builder, temp_res, data,
4941                                      loop_state.counter, "");
4942    LLVMBuildStore(gallivm->builder, temp_res, atom_res);
4943 
4944    lp_build_endif(&ifthen);
4945    lp_build_loop_end_cond(&loop_state,
4946                           lp_build_const_int32(gallivm, type.length),
4947                           NULL, LLVMIntUGE);
4948    atomic_result[0] = LLVMBuildLoad2(gallivm->builder, atom_res_elem_type,
4949                                      atom_res, "");
4950 }
4951 
4952 
4953 static void
lp_build_img_op_no_format(struct gallivm_state * gallivm,const struct lp_img_params * params,LLVMValueRef outdata[4])4954 lp_build_img_op_no_format(struct gallivm_state *gallivm,
4955                           const struct lp_img_params *params,
4956                           LLVMValueRef outdata[4])
4957 {
4958    /*
4959     * If there's nothing bound, format is NONE, and we must return
4960     * all zero as mandated by d3d10 in this case.
4961     */
4962    if (params->img_op != LP_IMG_STORE) {
4963       LLVMValueRef zero = lp_build_zero(gallivm, params->type);
4964       for (unsigned chan = 0; chan < (params->img_op == LP_IMG_LOAD ? 4 : 1);
4965            chan++) {
4966          outdata[chan] = zero;
4967       }
4968    }
4969 }
4970 
4971 
4972 void
lp_build_img_op_soa(const struct lp_static_texture_state * static_texture_state,struct lp_sampler_dynamic_state * dynamic_state,struct gallivm_state * gallivm,const struct lp_img_params * params,LLVMValueRef * outdata)4973 lp_build_img_op_soa(const struct lp_static_texture_state *static_texture_state,
4974                     struct lp_sampler_dynamic_state *dynamic_state,
4975                     struct gallivm_state *gallivm,
4976                     const struct lp_img_params *params,
4977                     LLVMValueRef *outdata)
4978 {
4979    const enum pipe_texture_target target = params->target;
4980    const unsigned dims = texture_dims(target);
4981    const struct util_format_description *format_desc =
4982       util_format_description(static_texture_state->format);
4983    const struct util_format_description *res_format_desc =
4984       util_format_description(static_texture_state->res_format);
4985    LLVMValueRef x = params->coords[0], y = params->coords[1],
4986       z = params->coords[2];
4987    LLVMValueRef row_stride_vec = NULL, img_stride_vec = NULL;
4988 
4989    /** regular scalar int type */
4990    struct lp_type int_coord_type = lp_uint_type(params->type);
4991    struct lp_build_context int_coord_bld;
4992    lp_build_context_init(&int_coord_bld, gallivm, int_coord_type);
4993 
4994    if (static_texture_state->format == PIPE_FORMAT_NONE) {
4995       lp_build_img_op_no_format(gallivm, params, outdata);
4996       return;
4997 
4998    }
4999 
5000    LLVMValueRef row_stride = dynamic_state->row_stride(gallivm,
5001                                                        params->resources_type,
5002                                                        params->resources_ptr,
5003                                                        params->image_index, NULL, NULL);
5004    LLVMValueRef img_stride = dynamic_state->img_stride(gallivm,
5005                                                        params->resources_type,
5006                                                        params->resources_ptr,
5007                                                        params->image_index, NULL, NULL);
5008    LLVMValueRef base_ptr = dynamic_state->base_ptr(gallivm,
5009                                                    params->resources_type,
5010                                                    params->resources_ptr,
5011                                                    params->image_index, NULL);
5012    LLVMValueRef width = dynamic_state->width(gallivm,
5013                                              params->resources_type,
5014                                              params->resources_ptr,
5015                                              params->image_index, NULL);
5016    LLVMValueRef height = dynamic_state->height(gallivm,
5017                                                params->resources_type,
5018                                                params->resources_ptr,
5019                                                params->image_index, NULL);
5020    height = LLVMBuildZExt(gallivm->builder, height,
5021                           int_coord_bld.elem_type, "");
5022    LLVMValueRef depth = dynamic_state->depth(gallivm,
5023                                              params->resources_type,
5024                                              params->resources_ptr,
5025                                              params->image_index, NULL);
5026    depth = LLVMBuildZExt(gallivm->builder, depth,
5027                          int_coord_bld.elem_type, "");
5028    bool layer_coord = has_layer_coord(target);
5029 
5030    width = lp_build_scale_view_dim(gallivm, width, res_format_desc->block.width,
5031                                    format_desc->block.width);
5032    width = lp_build_broadcast_scalar(&int_coord_bld, width);
5033    if (dims >= 2) {
5034       height = lp_build_scale_view_dim(gallivm, height, res_format_desc->block.height,
5035                                        format_desc->block.height);
5036       height = lp_build_broadcast_scalar(&int_coord_bld, height);
5037       row_stride_vec = lp_build_broadcast_scalar(&int_coord_bld, row_stride);
5038    }
5039    if (dims >= 3 || layer_coord) {
5040       depth = lp_build_broadcast_scalar(&int_coord_bld, depth);
5041       img_stride_vec = lp_build_broadcast_scalar(&int_coord_bld, img_stride);
5042    }
5043 
5044    LLVMValueRef out_of_bounds = int_coord_bld.zero;
5045    LLVMValueRef out1 = lp_build_cmp(&int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
5046    out_of_bounds = lp_build_or(&int_coord_bld, out_of_bounds, out1);
5047 
5048    if (dims >= 2) {
5049       out1 = lp_build_cmp(&int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
5050       out_of_bounds = lp_build_or(&int_coord_bld, out_of_bounds, out1);
5051    }
5052    if (dims >= 3 || layer_coord) {
5053       out1 = lp_build_cmp(&int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
5054       out_of_bounds = lp_build_or(&int_coord_bld, out_of_bounds, out1);
5055    }
5056 
5057    LLVMValueRef offset, i, j;
5058    if (static_texture_state->tiled) {
5059       lp_build_tiled_sample_offset(&int_coord_bld,
5060                                    format_desc->format,
5061                                    static_texture_state,
5062                                    x, y, z, width, height, img_stride_vec,
5063                                    &offset, &i, &j);
5064    } else {
5065       lp_build_sample_offset(&int_coord_bld,
5066                              format_desc,
5067                              x, y, z, row_stride_vec, img_stride_vec,
5068                              &offset, &i, &j);
5069    }
5070 
5071    if (params->ms_index && static_texture_state->level_zero_only) {
5072       LLVMValueRef num_samples = dynamic_state->last_level(gallivm,
5073                                                            params->resources_type,
5074                                                            params->resources_ptr,
5075                                                            params->image_index, NULL);
5076       num_samples = LLVMBuildZExt(gallivm->builder, num_samples,
5077                                   int_coord_bld.elem_type, "");
5078       LLVMValueRef sample_stride = dynamic_state->sample_stride(gallivm,
5079                                                                 params->resources_type,
5080                                                                 params->resources_ptr,
5081                                                                 params->image_index, NULL);
5082       lp_build_sample_ms_offset(&int_coord_bld,
5083                                 params->ms_index, num_samples,
5084                                 sample_stride, &offset,
5085                                 &out_of_bounds);
5086    }
5087    if (params->img_op == LP_IMG_LOAD || params->img_op == LP_IMG_LOAD_SPARSE) {
5088       struct lp_type texel_type = lp_build_texel_type(params->type, format_desc);
5089 
5090       if (params->img_op == LP_IMG_LOAD_SPARSE && static_texture_state->tiled) {
5091          LLVMValueRef base_offset =
5092             dynamic_state->base_offset(gallivm, params->resources_type,
5093                                        params->resources_ptr, params->image_index, NULL);
5094          base_offset = lp_build_broadcast_scalar(&int_coord_bld, base_offset);
5095 
5096          LLVMValueRef full_offset = LLVMBuildAdd(gallivm->builder, base_offset, offset, "");
5097 
5098          lp_build_gather_resident(&int_coord_bld, dynamic_state,
5099                                   params->resources_type, params->resources_ptr,
5100                                   full_offset, &outdata[4]);
5101       }
5102 
5103       offset = lp_build_andnot(&int_coord_bld, offset, out_of_bounds);
5104       struct lp_build_context texel_bld;
5105       lp_build_context_init(&texel_bld, gallivm, texel_type);
5106       lp_build_fetch_rgba_soa(gallivm,
5107                               format_desc,
5108                               texel_type, true,
5109                               base_ptr, offset,
5110                               i, j,
5111                               NULL,
5112                               outdata);
5113 
5114       for (unsigned chan = 0; chan < 3; chan++) {
5115          outdata[chan] = lp_build_select(&texel_bld, out_of_bounds,
5116                                          texel_bld.zero, outdata[chan]);
5117       }
5118       if (format_desc->swizzle[3] == PIPE_SWIZZLE_1) {
5119          outdata[3] = lp_build_select(&texel_bld, out_of_bounds,
5120                                       texel_bld.one, outdata[3]);
5121       } else {
5122          outdata[3] = lp_build_select(&texel_bld, out_of_bounds,
5123                                       texel_bld.zero, outdata[3]);
5124       }
5125    } else if (params->img_op == LP_IMG_STORE) {
5126       lp_build_store_rgba_soa(gallivm, format_desc, params->type,
5127                               params->exec_mask, base_ptr, offset,
5128                               out_of_bounds, params->indata);
5129    } else {
5130       lp_build_do_atomic_soa(gallivm, format_desc, params->type,
5131                              params->exec_mask, base_ptr, offset,
5132                              out_of_bounds, params->img_op, params->op,
5133                              params->indata, params->indata2, outdata);
5134    }
5135 }
5136 
5137 
5138 /*
5139  * These functions are for indirect texture access suppoort.
5140  *
5141  * Indirect textures are implemented using a switch statement, that
5142  * takes the texture index and jumps to the sampler functions for
5143  * that texture unit.
5144  */
5145 
5146 /*
5147  * Initialise an indexed sampler switch block.
5148  *
5149  * This sets up the switch_info state and adds the LLVM flow control pieces.
5150  */
5151 void
lp_build_sample_array_init_soa(struct lp_build_sample_array_switch * switch_info,struct gallivm_state * gallivm,const struct lp_sampler_params * params,LLVMValueRef idx,unsigned base,unsigned range)5152 lp_build_sample_array_init_soa(struct lp_build_sample_array_switch *switch_info,
5153                            struct gallivm_state *gallivm,
5154                            const struct lp_sampler_params *params,
5155                            LLVMValueRef idx,
5156                            unsigned base, unsigned range)
5157 {
5158    switch_info->gallivm = gallivm;
5159    switch_info->params = *params;
5160    switch_info->base = base;
5161    switch_info->range = range;
5162 
5163    /* for generating the switch functions we don't want the texture index
5164     * offset
5165     */
5166    switch_info->params.texture_index_offset = 0;
5167 
5168    LLVMBasicBlockRef initial_block = LLVMGetInsertBlock(gallivm->builder);
5169    switch_info->merge_ref = lp_build_insert_new_block(gallivm, "texmerge");
5170 
5171    switch_info->switch_ref = LLVMBuildSwitch(gallivm->builder, idx,
5172                                              switch_info->merge_ref,
5173                                              range - base);
5174 
5175    LLVMTypeRef val_type[4];
5176    val_type[0] = val_type[1] = val_type[2] = val_type[3] =
5177       lp_build_vec_type(gallivm, params->type);
5178 
5179    LLVMTypeRef ret_type =
5180       LLVMStructTypeInContext(gallivm->context, val_type, 4, 0);
5181 
5182    LLVMValueRef undef_val = LLVMGetUndef(ret_type);
5183 
5184    LLVMPositionBuilderAtEnd(gallivm->builder, switch_info->merge_ref);
5185 
5186    switch_info->phi = LLVMBuildPhi(gallivm->builder, ret_type, "");
5187    LLVMAddIncoming(switch_info->phi, &undef_val, &initial_block, 1);
5188 }
5189 
5190 
5191 /*
5192  * Add an individual entry to the indirect texture switch.
5193  *
5194  * This builds the sample function and links a case for it into the switch
5195  * statement.
5196  */
5197 void
lp_build_sample_array_case_soa(struct lp_build_sample_array_switch * switch_info,int idx,const struct lp_static_texture_state * static_texture_state,const struct lp_static_sampler_state * static_sampler_state,struct lp_sampler_dynamic_state * dynamic_texture_state)5198 lp_build_sample_array_case_soa(struct lp_build_sample_array_switch *switch_info,
5199                            int idx,
5200                            const struct lp_static_texture_state *static_texture_state,
5201                            const struct lp_static_sampler_state *static_sampler_state,
5202                            struct lp_sampler_dynamic_state *dynamic_texture_state)
5203 {
5204    struct gallivm_state *gallivm = switch_info->gallivm;
5205    LLVMBasicBlockRef this_block = lp_build_insert_new_block(gallivm, "texblock");
5206 
5207    LLVMAddCase(switch_info->switch_ref,
5208                LLVMConstInt(LLVMInt32TypeInContext(gallivm->context), idx, 0),
5209                this_block);
5210    LLVMPositionBuilderAtEnd(gallivm->builder, this_block);
5211 
5212    LLVMValueRef tex_ret;
5213    lp_build_sample_soa_func(gallivm, static_texture_state,
5214                             static_sampler_state, dynamic_texture_state,
5215                             &switch_info->params, idx, idx, &tex_ret);
5216 
5217    LLVMAddIncoming(switch_info->phi, &tex_ret, &this_block, 1);
5218    LLVMBuildBr(gallivm->builder, switch_info->merge_ref);
5219 }
5220 
5221 
5222 /*
5223  * Finish a switch statement.
5224  *
5225  * This handles extract the results from the switch.
5226  */
5227 void
lp_build_sample_array_fini_soa(struct lp_build_sample_array_switch * switch_info)5228 lp_build_sample_array_fini_soa(struct lp_build_sample_array_switch *switch_info)
5229 {
5230    struct gallivm_state *gallivm = switch_info->gallivm;
5231 
5232    LLVMPositionBuilderAtEnd(gallivm->builder, switch_info->merge_ref);
5233    for (unsigned i = 0; i < 4; i++) {
5234       switch_info->params.texel[i] =
5235          LLVMBuildExtractValue(gallivm->builder, switch_info->phi, i, "");
5236    }
5237 }
5238 
5239 
5240 void
lp_build_image_op_switch_soa(struct lp_build_img_op_array_switch * switch_info,struct gallivm_state * gallivm,const struct lp_img_params * params,LLVMValueRef idx,unsigned base,unsigned range)5241 lp_build_image_op_switch_soa(struct lp_build_img_op_array_switch *switch_info,
5242                              struct gallivm_state *gallivm,
5243                              const struct lp_img_params *params,
5244                              LLVMValueRef idx,
5245                              unsigned base, unsigned range)
5246 {
5247    switch_info->gallivm = gallivm;
5248    switch_info->params = *params;
5249    switch_info->base = base;
5250    switch_info->range = range;
5251 
5252    /* for generating the switch functions we don't want the texture index
5253     * offset
5254     */
5255    switch_info->params.image_index_offset = 0;
5256 
5257    LLVMBasicBlockRef initial_block = LLVMGetInsertBlock(gallivm->builder);
5258    switch_info->merge_ref = lp_build_insert_new_block(gallivm, "imgmerge");
5259 
5260    switch_info->switch_ref =
5261       LLVMBuildSwitch(gallivm->builder, idx,
5262                       switch_info->merge_ref, range - base);
5263 
5264    if (params->img_op != LP_IMG_STORE) {
5265       LLVMTypeRef ret_type = lp_build_vec_type(gallivm, params->type);
5266       LLVMValueRef undef_val = LLVMGetUndef(ret_type);
5267 
5268       LLVMPositionBuilderAtEnd(gallivm->builder, switch_info->merge_ref);
5269 
5270       for (unsigned i = 0; i < ((params->img_op == LP_IMG_LOAD) ? 4 : 1); i++) {
5271          switch_info->phi[i] = LLVMBuildPhi(gallivm->builder, ret_type, "");
5272          LLVMAddIncoming(switch_info->phi[i], &undef_val, &initial_block, 1);
5273       }
5274    }
5275 }
5276 
5277 
5278 void
lp_build_image_op_array_case(struct lp_build_img_op_array_switch * switch_info,int idx,const struct lp_static_texture_state * static_texture_state,struct lp_sampler_dynamic_state * dynamic_state)5279 lp_build_image_op_array_case(struct lp_build_img_op_array_switch *switch_info,
5280                             int idx,
5281                             const struct lp_static_texture_state *static_texture_state,
5282                             struct lp_sampler_dynamic_state *dynamic_state)
5283 {
5284    struct gallivm_state *gallivm = switch_info->gallivm;
5285    LLVMBasicBlockRef this_block = lp_build_insert_new_block(gallivm, "img");
5286    LLVMValueRef tex_ret[4];
5287 
5288    LLVMAddCase(switch_info->switch_ref,
5289                lp_build_const_int32(gallivm, idx), this_block);
5290    LLVMPositionBuilderAtEnd(gallivm->builder, this_block);
5291 
5292    switch_info->params.image_index = idx;
5293 
5294    lp_build_img_op_soa(static_texture_state, dynamic_state,
5295                        switch_info->gallivm, &switch_info->params, tex_ret);
5296 
5297    if (switch_info->params.img_op != LP_IMG_STORE) {
5298       for (unsigned i = 0;
5299            i < ((switch_info->params.img_op == LP_IMG_LOAD) ? 4 : 1); i++) {
5300          tex_ret[i] =
5301             LLVMBuildBitCast(gallivm->builder, tex_ret[i],
5302                              lp_build_vec_type(gallivm,
5303                                                switch_info->params.type), "");
5304       }
5305 
5306       this_block = LLVMGetInsertBlock(gallivm->builder);
5307       for (unsigned i = 0;
5308            i < ((switch_info->params.img_op == LP_IMG_LOAD) ? 4 : 1); i++) {
5309          LLVMAddIncoming(switch_info->phi[i], &tex_ret[i], &this_block, 1);
5310       }
5311    }
5312    LLVMBuildBr(gallivm->builder, switch_info->merge_ref);
5313 }
5314 
5315 
5316 void
lp_build_image_op_array_fini_soa(struct lp_build_img_op_array_switch * switch_info)5317 lp_build_image_op_array_fini_soa(struct lp_build_img_op_array_switch *switch_info)
5318 {
5319    struct gallivm_state *gallivm = switch_info->gallivm;
5320 
5321    LLVMPositionBuilderAtEnd(gallivm->builder, switch_info->merge_ref);
5322 
5323    if (switch_info->params.img_op != LP_IMG_STORE) {
5324       for (unsigned i = 0;
5325            i < ((switch_info->params.img_op == LP_IMG_LOAD) ? 4 : 1); i++) {
5326          switch_info->params.outdata[i] = switch_info->phi[i];
5327       }
5328    }
5329 }
5330