xref: /aosp_15_r20/external/mesa3d/src/intel/blorp/blorp_blit.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright © 2012 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  */
23 
24 #include "blorp_nir_builder.h"
25 #include "compiler/nir/nir_format_convert.h"
26 
27 #include "blorp_priv.h"
28 #include "dev/intel_debug.h"
29 #include "dev/intel_device_info.h"
30 
31 #include "util/format_rgb9e5.h"
32 #include "util/u_math.h"
33 
34 #define FILE_DEBUG_FLAG DEBUG_BLORP
35 
36 static const bool split_blorp_blit_debug = false;
37 
38 struct blorp_blit_vars {
39    /* Input values from blorp_wm_inputs */
40    nir_variable *v_bounds_rect;
41    nir_variable *v_rect_grid;
42    nir_variable *v_coord_transform;
43    nir_variable *v_src_z;
44    nir_variable *v_src_offset;
45    nir_variable *v_dst_offset;
46    nir_variable *v_src_inv_size;
47 };
48 
49 static void
blorp_blit_vars_init(nir_builder * b,struct blorp_blit_vars * v,const struct blorp_blit_prog_key * key)50 blorp_blit_vars_init(nir_builder *b, struct blorp_blit_vars *v,
51                          const struct blorp_blit_prog_key *key)
52 {
53 #define LOAD_INPUT(name, type)\
54    v->v_##name = BLORP_CREATE_NIR_INPUT(b->shader, name, type);
55 
56    LOAD_INPUT(bounds_rect, glsl_vec4_type())
57    LOAD_INPUT(rect_grid, glsl_vec4_type())
58    LOAD_INPUT(coord_transform, glsl_vec4_type())
59    LOAD_INPUT(src_z, glsl_float_type())
60    LOAD_INPUT(src_offset, glsl_vector_type(GLSL_TYPE_UINT, 2))
61    LOAD_INPUT(dst_offset, glsl_vector_type(GLSL_TYPE_UINT, 2))
62    LOAD_INPUT(src_inv_size, glsl_vector_type(GLSL_TYPE_FLOAT, 2))
63 
64 #undef LOAD_INPUT
65 }
66 
67 static nir_def *
blorp_blit_get_frag_coords(nir_builder * b,const struct blorp_blit_prog_key * key,struct blorp_blit_vars * v)68 blorp_blit_get_frag_coords(nir_builder *b,
69                            const struct blorp_blit_prog_key *key,
70                            struct blorp_blit_vars *v)
71 {
72    nir_def *coord = nir_f2i32(b, nir_load_frag_coord(b));
73 
74    /* Account for destination surface intratile offset
75     *
76     * Transformation parameters giving translation from destination to source
77     * coordinates don't take into account possible intra-tile destination
78     * offset.  Therefore it has to be first subtracted from the incoming
79     * coordinates.  Vertices are set up based on coordinates containing the
80     * intra-tile offset.
81     */
82    if (key->need_dst_offset)
83       coord = nir_isub(b, coord, nir_load_var(b, v->v_dst_offset));
84 
85    if (key->persample_msaa_dispatch) {
86       b->shader->info.fs.uses_sample_shading = true;
87       return nir_vec3(b, nir_channel(b, coord, 0), nir_channel(b, coord, 1),
88                       nir_load_sample_id(b));
89    } else {
90       return nir_trim_vector(b, coord, 2);
91    }
92 }
93 
94 static nir_def *
blorp_blit_get_cs_dst_coords(nir_builder * b,const struct blorp_blit_prog_key * key,struct blorp_blit_vars * v)95 blorp_blit_get_cs_dst_coords(nir_builder *b,
96                              const struct blorp_blit_prog_key *key,
97                              struct blorp_blit_vars *v)
98 {
99    nir_def *coord = nir_load_global_invocation_id(b, 32);
100 
101    /* Account for destination surface intratile offset
102     *
103     * Transformation parameters giving translation from destination to source
104     * coordinates don't take into account possible intra-tile destination
105     * offset.  Therefore it has to be first subtracted from the incoming
106     * coordinates.  Vertices are set up based on coordinates containing the
107     * intra-tile offset.
108     */
109    if (key->need_dst_offset)
110       coord = nir_isub(b, coord, nir_load_var(b, v->v_dst_offset));
111 
112    assert(!key->persample_msaa_dispatch);
113    return nir_trim_vector(b, coord, 2);
114 }
115 
116 /**
117  * Emit code to translate from destination (X, Y) coordinates to source (X, Y)
118  * coordinates.
119  */
120 static nir_def *
blorp_blit_apply_transform(nir_builder * b,nir_def * src_pos,struct blorp_blit_vars * v)121 blorp_blit_apply_transform(nir_builder *b, nir_def *src_pos,
122                            struct blorp_blit_vars *v)
123 {
124    nir_def *coord_transform = nir_load_var(b, v->v_coord_transform);
125 
126    nir_def *offset = nir_vec2(b, nir_channel(b, coord_transform, 1),
127                                      nir_channel(b, coord_transform, 3));
128    nir_def *mul = nir_vec2(b, nir_channel(b, coord_transform, 0),
129                                   nir_channel(b, coord_transform, 2));
130 
131    return nir_fadd(b, nir_fmul(b, src_pos, mul), offset);
132 }
133 
134 static nir_tex_instr *
blorp_create_nir_tex_instr(nir_builder * b,struct blorp_blit_vars * v,nir_texop op,nir_def * pos,unsigned num_srcs,nir_alu_type dst_type)135 blorp_create_nir_tex_instr(nir_builder *b, struct blorp_blit_vars *v,
136                            nir_texop op, nir_def *pos, unsigned num_srcs,
137                            nir_alu_type dst_type)
138 {
139    nir_tex_instr *tex = nir_tex_instr_create(b->shader, num_srcs);
140 
141    tex->op = op;
142 
143    tex->dest_type = dst_type | 32;
144    tex->is_array = false;
145    tex->is_shadow = false;
146 
147    tex->texture_index = BLORP_TEXTURE_BT_INDEX;
148    tex->sampler_index = BLORP_SAMPLER_INDEX;
149 
150    /* To properly handle 3-D and 2-D array textures, we pull the Z component
151     * from an input.  TODO: This is a bit magic; we should probably make this
152     * more explicit in the future.
153     */
154    assert(pos->num_components >= 2);
155    if (op == nir_texop_txf || op == nir_texop_txf_ms ||
156        op == nir_texop_txf_ms_mcs_intel) {
157       pos = nir_vec3(b, nir_channel(b, pos, 0), nir_channel(b, pos, 1),
158                         nir_f2i32(b, nir_load_var(b, v->v_src_z)));
159    } else {
160       pos = nir_vec3(b, nir_channel(b, pos, 0), nir_channel(b, pos, 1),
161                         nir_load_var(b, v->v_src_z));
162    }
163 
164    tex->src[0] = nir_tex_src_for_ssa(nir_tex_src_coord, pos);
165    tex->coord_components = 3;
166 
167    nir_def_init(&tex->instr, &tex->def, 4, 32);
168 
169    return tex;
170 }
171 
172 static nir_def *
blorp_nir_tex(nir_builder * b,struct blorp_blit_vars * v,const struct blorp_blit_prog_key * key,nir_def * pos)173 blorp_nir_tex(nir_builder *b, struct blorp_blit_vars *v,
174               const struct blorp_blit_prog_key *key, nir_def *pos)
175 {
176    if (key->need_src_offset)
177       pos = nir_fadd(b, pos, nir_i2f32(b, nir_load_var(b, v->v_src_offset)));
178 
179    /* If the sampler requires normalized coordinates, we need to compensate. */
180    if (key->src_coords_normalized)
181       pos = nir_fmul(b, pos, nir_load_var(b, v->v_src_inv_size));
182 
183    nir_tex_instr *tex =
184       blorp_create_nir_tex_instr(b, v, nir_texop_txl, pos, 2,
185                                  key->texture_data_type);
186 
187    assert(pos->num_components == 2);
188    tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
189    tex->src[1] = nir_tex_src_for_ssa(nir_tex_src_lod, nir_imm_int(b, 0));
190 
191    nir_builder_instr_insert(b, &tex->instr);
192 
193    return &tex->def;
194 }
195 
196 static nir_def *
blorp_nir_txf(nir_builder * b,struct blorp_blit_vars * v,nir_def * pos,nir_alu_type dst_type)197 blorp_nir_txf(nir_builder *b, struct blorp_blit_vars *v,
198               nir_def *pos, nir_alu_type dst_type)
199 {
200    nir_tex_instr *tex =
201       blorp_create_nir_tex_instr(b, v, nir_texop_txf, pos, 2, dst_type);
202 
203    tex->sampler_dim = GLSL_SAMPLER_DIM_3D;
204    tex->src[1] = nir_tex_src_for_ssa(nir_tex_src_lod, nir_imm_int(b, 0));
205 
206    nir_builder_instr_insert(b, &tex->instr);
207 
208    return &tex->def;
209 }
210 
211 static nir_def *
blorp_nir_txf_ms(nir_builder * b,struct blorp_blit_vars * v,nir_def * pos,nir_def * mcs,nir_alu_type dst_type)212 blorp_nir_txf_ms(nir_builder *b, struct blorp_blit_vars *v,
213                  nir_def *pos, nir_def *mcs, nir_alu_type dst_type)
214 {
215    nir_tex_instr *tex =
216       blorp_create_nir_tex_instr(b, v, nir_texop_txf_ms, pos, 3, dst_type);
217 
218    tex->sampler_dim = GLSL_SAMPLER_DIM_MS;
219 
220    tex->src[1].src_type = nir_tex_src_ms_index;
221    if (pos->num_components == 2) {
222       tex->src[1].src = nir_src_for_ssa(nir_imm_int(b, 0));
223    } else {
224       assert(pos->num_components == 3);
225       tex->src[1].src = nir_src_for_ssa(nir_channel(b, pos, 2));
226    }
227 
228    if (!mcs)
229       mcs = nir_imm_zero(b, 4, 32);
230 
231    tex->src[2] = nir_tex_src_for_ssa(nir_tex_src_ms_mcs_intel, mcs);
232 
233    nir_builder_instr_insert(b, &tex->instr);
234 
235    return &tex->def;
236 }
237 
238 static nir_def *
blorp_blit_txf_ms_mcs(nir_builder * b,struct blorp_blit_vars * v,nir_def * pos)239 blorp_blit_txf_ms_mcs(nir_builder *b, struct blorp_blit_vars *v,
240                       nir_def *pos)
241 {
242    nir_tex_instr *tex =
243       blorp_create_nir_tex_instr(b, v, nir_texop_txf_ms_mcs_intel,
244                                  pos, 1, nir_type_int);
245 
246    tex->sampler_dim = GLSL_SAMPLER_DIM_MS;
247 
248    nir_builder_instr_insert(b, &tex->instr);
249 
250    return &tex->def;
251 }
252 
253 /**
254  * Emit code to compensate for the difference between Y and W tiling.
255  *
256  * This code modifies the X and Y coordinates according to the formula:
257  *
258  *   (X', Y', S') = detile(W-MAJOR, tile(Y-MAJOR, X, Y, S))
259  *
260  * (See blorp_build_nir_shader).
261  */
262 static inline nir_def *
blorp_nir_retile_y_to_w(nir_builder * b,nir_def * pos)263 blorp_nir_retile_y_to_w(nir_builder *b, nir_def *pos)
264 {
265    assert(pos->num_components == 2);
266    nir_def *x_Y = nir_channel(b, pos, 0);
267    nir_def *y_Y = nir_channel(b, pos, 1);
268 
269    /* Given X and Y coordinates that describe an address using Y tiling,
270     * translate to the X and Y coordinates that describe the same address
271     * using W tiling.
272     *
273     * If we break down the low order bits of X and Y, using a
274     * single letter to represent each low-order bit:
275     *
276     *   X = A << 7 | 0bBCDEFGH
277     *   Y = J << 5 | 0bKLMNP                                       (1)
278     *
279     * Then we can apply the Y tiling formula to see the memory offset being
280     * addressed:
281     *
282     *   offset = (J * tile_pitch + A) << 12 | 0bBCDKLMNPEFGH       (2)
283     *
284     * If we apply the W detiling formula to this memory location, that the
285     * corresponding X' and Y' coordinates are:
286     *
287     *   X' = A << 6 | 0bBCDPFH                                     (3)
288     *   Y' = J << 6 | 0bKLMNEG
289     *
290     * Combining (1) and (3), we see that to transform (X, Y) to (X', Y'),
291     * we need to make the following computation:
292     *
293     *   X' = (X & ~0b1011) >> 1 | (Y & 0b1) << 2 | X & 0b1         (4)
294     *   Y' = (Y & ~0b1) << 1 | (X & 0b1000) >> 2 | (X & 0b10) >> 1
295     */
296    nir_def *x_W = nir_imm_int(b, 0);
297    x_W = nir_mask_shift_or(b, x_W, x_Y, 0xfffffff4, -1);
298    x_W = nir_mask_shift_or(b, x_W, y_Y, 0x1, 2);
299    x_W = nir_mask_shift_or(b, x_W, x_Y, 0x1, 0);
300 
301    nir_def *y_W = nir_imm_int(b, 0);
302    y_W = nir_mask_shift_or(b, y_W, y_Y, 0xfffffffe, 1);
303    y_W = nir_mask_shift_or(b, y_W, x_Y, 0x8, -2);
304    y_W = nir_mask_shift_or(b, y_W, x_Y, 0x2, -1);
305 
306    return nir_vec2(b, x_W, y_W);
307 }
308 
309 /**
310  * Emit code to compensate for the difference between Y and W tiling.
311  *
312  * This code modifies the X and Y coordinates according to the formula:
313  *
314  *   (X', Y', S') = detile(Y-MAJOR, tile(W-MAJOR, X, Y, S))
315  *
316  * (See blorp_build_nir_shader).
317  */
318 static inline nir_def *
blorp_nir_retile_w_to_y(nir_builder * b,nir_def * pos)319 blorp_nir_retile_w_to_y(nir_builder *b, nir_def *pos)
320 {
321    assert(pos->num_components == 2);
322    nir_def *x_W = nir_channel(b, pos, 0);
323    nir_def *y_W = nir_channel(b, pos, 1);
324 
325    /* Applying the same logic as above, but in reverse, we obtain the
326     * formulas:
327     *
328     * X' = (X & ~0b101) << 1 | (Y & 0b10) << 2 | (Y & 0b1) << 1 | X & 0b1
329     * Y' = (Y & ~0b11) >> 1 | (X & 0b100) >> 2
330     */
331    nir_def *x_Y = nir_imm_int(b, 0);
332    x_Y = nir_mask_shift_or(b, x_Y, x_W, 0xfffffffa, 1);
333    x_Y = nir_mask_shift_or(b, x_Y, y_W, 0x2, 2);
334    x_Y = nir_mask_shift_or(b, x_Y, y_W, 0x1, 1);
335    x_Y = nir_mask_shift_or(b, x_Y, x_W, 0x1, 0);
336 
337    nir_def *y_Y = nir_imm_int(b, 0);
338    y_Y = nir_mask_shift_or(b, y_Y, y_W, 0xfffffffc, -1);
339    y_Y = nir_mask_shift_or(b, y_Y, x_W, 0x4, -2);
340 
341    return nir_vec2(b, x_Y, y_Y);
342 }
343 
344 /**
345  * Emit code to compensate for the difference between MSAA and non-MSAA
346  * surfaces.
347  *
348  * This code modifies the X and Y coordinates according to the formula:
349  *
350  *   (X', Y', S') = encode_msaa(num_samples, IMS, X, Y, S)
351  */
352 static inline nir_def *
blorp_nir_encode_msaa(nir_builder * b,nir_def * pos,unsigned num_samples,enum isl_msaa_layout layout)353 blorp_nir_encode_msaa(nir_builder *b, nir_def *pos,
354                       unsigned num_samples, enum isl_msaa_layout layout)
355 {
356    assert(pos->num_components == 2 || pos->num_components == 3);
357 
358    switch (layout) {
359    case ISL_MSAA_LAYOUT_NONE:
360       assert(pos->num_components == 2);
361       return pos;
362    case ISL_MSAA_LAYOUT_ARRAY:
363       /* No translation needed */
364       return pos;
365    case ISL_MSAA_LAYOUT_INTERLEAVED: {
366       nir_def *x_in = nir_channel(b, pos, 0);
367       nir_def *y_in = nir_channel(b, pos, 1);
368       nir_def *s_in = pos->num_components == 2 ? nir_imm_int(b, 0) :
369                                                      nir_channel(b, pos, 2);
370 
371       nir_def *x_out = nir_imm_int(b, 0);
372       nir_def *y_out = nir_imm_int(b, 0);
373       switch (num_samples) {
374       case 2:
375       case 4:
376          /* encode_msaa(2, IMS, X, Y, S) = (X', Y', 0)
377           *   where X' = (X & ~0b1) << 1 | (S & 0b1) << 1 | (X & 0b1)
378           *         Y' = Y
379           *
380           * encode_msaa(4, IMS, X, Y, S) = (X', Y', 0)
381           *   where X' = (X & ~0b1) << 1 | (S & 0b1) << 1 | (X & 0b1)
382           *         Y' = (Y & ~0b1) << 1 | (S & 0b10) | (Y & 0b1)
383           */
384          x_out = nir_mask_shift_or(b, x_out, x_in, 0xfffffffe, 1);
385          x_out = nir_mask_shift_or(b, x_out, s_in, 0x1, 1);
386          x_out = nir_mask_shift_or(b, x_out, x_in, 0x1, 0);
387          if (num_samples == 2) {
388             y_out = y_in;
389          } else {
390             y_out = nir_mask_shift_or(b, y_out, y_in, 0xfffffffe, 1);
391             y_out = nir_mask_shift_or(b, y_out, s_in, 0x2, 0);
392             y_out = nir_mask_shift_or(b, y_out, y_in, 0x1, 0);
393          }
394          break;
395 
396       case 8:
397          /* encode_msaa(8, IMS, X, Y, S) = (X', Y', 0)
398           *   where X' = (X & ~0b1) << 2 | (S & 0b100) | (S & 0b1) << 1
399           *              | (X & 0b1)
400           *         Y' = (Y & ~0b1) << 1 | (S & 0b10) | (Y & 0b1)
401           */
402          x_out = nir_mask_shift_or(b, x_out, x_in, 0xfffffffe, 2);
403          x_out = nir_mask_shift_or(b, x_out, s_in, 0x4, 0);
404          x_out = nir_mask_shift_or(b, x_out, s_in, 0x1, 1);
405          x_out = nir_mask_shift_or(b, x_out, x_in, 0x1, 0);
406          y_out = nir_mask_shift_or(b, y_out, y_in, 0xfffffffe, 1);
407          y_out = nir_mask_shift_or(b, y_out, s_in, 0x2, 0);
408          y_out = nir_mask_shift_or(b, y_out, y_in, 0x1, 0);
409          break;
410 
411       case 16:
412          /* encode_msaa(16, IMS, X, Y, S) = (X', Y', 0)
413           *   where X' = (X & ~0b1) << 2 | (S & 0b100) | (S & 0b1) << 1
414           *              | (X & 0b1)
415           *         Y' = (Y & ~0b1) << 2 | (S & 0b1000) >> 1 (S & 0b10)
416           *              | (Y & 0b1)
417           */
418          x_out = nir_mask_shift_or(b, x_out, x_in, 0xfffffffe, 2);
419          x_out = nir_mask_shift_or(b, x_out, s_in, 0x4, 0);
420          x_out = nir_mask_shift_or(b, x_out, s_in, 0x1, 1);
421          x_out = nir_mask_shift_or(b, x_out, x_in, 0x1, 0);
422          y_out = nir_mask_shift_or(b, y_out, y_in, 0xfffffffe, 2);
423          y_out = nir_mask_shift_or(b, y_out, s_in, 0x8, -1);
424          y_out = nir_mask_shift_or(b, y_out, s_in, 0x2, 0);
425          y_out = nir_mask_shift_or(b, y_out, y_in, 0x1, 0);
426          break;
427 
428       default:
429          unreachable("Invalid number of samples for IMS layout");
430       }
431 
432       return nir_vec2(b, x_out, y_out);
433    }
434 
435    default:
436       unreachable("Invalid MSAA layout");
437    }
438 }
439 
440 /**
441  * Emit code to compensate for the difference between MSAA and non-MSAA
442  * surfaces.
443  *
444  * This code modifies the X and Y coordinates according to the formula:
445  *
446  *   (X', Y', S) = decode_msaa(num_samples, IMS, X, Y, S)
447  */
448 static inline nir_def *
blorp_nir_decode_msaa(nir_builder * b,nir_def * pos,unsigned num_samples,enum isl_msaa_layout layout)449 blorp_nir_decode_msaa(nir_builder *b, nir_def *pos,
450                       unsigned num_samples, enum isl_msaa_layout layout)
451 {
452    assert(pos->num_components == 2 || pos->num_components == 3);
453 
454    switch (layout) {
455    case ISL_MSAA_LAYOUT_NONE:
456       /* No translation necessary, and S should already be zero. */
457       assert(pos->num_components == 2);
458       return pos;
459    case ISL_MSAA_LAYOUT_ARRAY:
460       /* No translation necessary. */
461       return pos;
462    case ISL_MSAA_LAYOUT_INTERLEAVED: {
463       assert(pos->num_components == 2);
464 
465       nir_def *x_in = nir_channel(b, pos, 0);
466       nir_def *y_in = nir_channel(b, pos, 1);
467 
468       nir_def *x_out = nir_imm_int(b, 0);
469       nir_def *y_out = nir_imm_int(b, 0);
470       nir_def *s_out = nir_imm_int(b, 0);
471       switch (num_samples) {
472       case 2:
473       case 4:
474          /* decode_msaa(2, IMS, X, Y, 0) = (X', Y', S)
475           *   where X' = (X & ~0b11) >> 1 | (X & 0b1)
476           *         S = (X & 0b10) >> 1
477           *
478           * decode_msaa(4, IMS, X, Y, 0) = (X', Y', S)
479           *   where X' = (X & ~0b11) >> 1 | (X & 0b1)
480           *         Y' = (Y & ~0b11) >> 1 | (Y & 0b1)
481           *         S = (Y & 0b10) | (X & 0b10) >> 1
482           */
483          x_out = nir_mask_shift_or(b, x_out, x_in, 0xfffffffc, -1);
484          x_out = nir_mask_shift_or(b, x_out, x_in, 0x1, 0);
485          if (num_samples == 2) {
486             y_out = y_in;
487             s_out = nir_mask_shift_or(b, s_out, x_in, 0x2, -1);
488          } else {
489             y_out = nir_mask_shift_or(b, y_out, y_in, 0xfffffffc, -1);
490             y_out = nir_mask_shift_or(b, y_out, y_in, 0x1, 0);
491             s_out = nir_mask_shift_or(b, s_out, x_in, 0x2, -1);
492             s_out = nir_mask_shift_or(b, s_out, y_in, 0x2, 0);
493          }
494          break;
495 
496       case 8:
497          /* decode_msaa(8, IMS, X, Y, 0) = (X', Y', S)
498           *   where X' = (X & ~0b111) >> 2 | (X & 0b1)
499           *         Y' = (Y & ~0b11) >> 1 | (Y & 0b1)
500           *         S = (X & 0b100) | (Y & 0b10) | (X & 0b10) >> 1
501           */
502          x_out = nir_mask_shift_or(b, x_out, x_in, 0xfffffff8, -2);
503          x_out = nir_mask_shift_or(b, x_out, x_in, 0x1, 0);
504          y_out = nir_mask_shift_or(b, y_out, y_in, 0xfffffffc, -1);
505          y_out = nir_mask_shift_or(b, y_out, y_in, 0x1, 0);
506          s_out = nir_mask_shift_or(b, s_out, x_in, 0x4, 0);
507          s_out = nir_mask_shift_or(b, s_out, y_in, 0x2, 0);
508          s_out = nir_mask_shift_or(b, s_out, x_in, 0x2, -1);
509          break;
510 
511       case 16:
512          /* decode_msaa(16, IMS, X, Y, 0) = (X', Y', S)
513           *   where X' = (X & ~0b111) >> 2 | (X & 0b1)
514           *         Y' = (Y & ~0b111) >> 2 | (Y & 0b1)
515           *         S = (Y & 0b100) << 1 | (X & 0b100) |
516           *             (Y & 0b10) | (X & 0b10) >> 1
517           */
518          x_out = nir_mask_shift_or(b, x_out, x_in, 0xfffffff8, -2);
519          x_out = nir_mask_shift_or(b, x_out, x_in, 0x1, 0);
520          y_out = nir_mask_shift_or(b, y_out, y_in, 0xfffffff8, -2);
521          y_out = nir_mask_shift_or(b, y_out, y_in, 0x1, 0);
522          s_out = nir_mask_shift_or(b, s_out, y_in, 0x4, 1);
523          s_out = nir_mask_shift_or(b, s_out, x_in, 0x4, 0);
524          s_out = nir_mask_shift_or(b, s_out, y_in, 0x2, 0);
525          s_out = nir_mask_shift_or(b, s_out, x_in, 0x2, -1);
526          break;
527 
528       default:
529          unreachable("Invalid number of samples for IMS layout");
530       }
531 
532       return nir_vec3(b, x_out, y_out, s_out);
533    }
534 
535    default:
536       unreachable("Invalid MSAA layout");
537    }
538 }
539 
540 /**
541  * Count the number of trailing 1 bits in the given value.  For example:
542  *
543  * count_trailing_one_bits(0) == 0
544  * count_trailing_one_bits(7) == 3
545  * count_trailing_one_bits(11) == 2
546  */
count_trailing_one_bits(unsigned value)547 static inline int count_trailing_one_bits(unsigned value)
548 {
549 #ifdef HAVE___BUILTIN_CTZ
550    return __builtin_ctz(~value);
551 #else
552    return util_bitcount(value & ~(value + 1));
553 #endif
554 }
555 
556 static nir_def *
blorp_nir_combine_samples(nir_builder * b,struct blorp_blit_vars * v,nir_def * pos,unsigned tex_samples,enum isl_aux_usage tex_aux_usage,nir_alu_type dst_type,enum blorp_filter filter)557 blorp_nir_combine_samples(nir_builder *b, struct blorp_blit_vars *v,
558                           nir_def *pos, unsigned tex_samples,
559                           enum isl_aux_usage tex_aux_usage,
560                           nir_alu_type dst_type,
561                           enum blorp_filter filter)
562 {
563    nir_variable *color =
564       nir_local_variable_create(b->impl, glsl_vec4_type(), "color");
565 
566    nir_def *mcs = NULL;
567    if (isl_aux_usage_has_mcs(tex_aux_usage))
568       mcs = blorp_blit_txf_ms_mcs(b, v, pos);
569 
570    nir_op combine_op;
571    switch (filter) {
572    case BLORP_FILTER_AVERAGE:
573       assert(dst_type == nir_type_float);
574       combine_op = nir_op_fadd;
575       break;
576 
577    case BLORP_FILTER_MIN_SAMPLE:
578       switch (dst_type) {
579       case nir_type_int:   combine_op = nir_op_imin;  break;
580       case nir_type_uint:  combine_op = nir_op_umin;  break;
581       case nir_type_float: combine_op = nir_op_fmin;  break;
582       default: unreachable("Invalid dst_type");
583       }
584       break;
585 
586    case BLORP_FILTER_MAX_SAMPLE:
587       switch (dst_type) {
588       case nir_type_int:   combine_op = nir_op_imax;  break;
589       case nir_type_uint:  combine_op = nir_op_umax;  break;
590       case nir_type_float: combine_op = nir_op_fmax;  break;
591       default: unreachable("Invalid dst_type");
592       }
593       break;
594 
595    default:
596       unreachable("Invalid filter");
597    }
598 
599    /* If true, we inserted an if statement that we need to pop at at the end.
600     */
601    bool inserted_if = false;
602 
603    /* We add together samples using a binary tree structure, e.g. for 4x MSAA:
604     *
605     *   result = ((sample[0] + sample[1]) + (sample[2] + sample[3])) / 4
606     *
607     * This ensures that when all samples have the same value, no numerical
608     * precision is lost, since each addition operation always adds two equal
609     * values, and summing two equal floating point values does not lose
610     * precision.
611     *
612     * We perform this computation by treating the texture_data array as a
613     * stack and performing the following operations:
614     *
615     * - push sample 0 onto stack
616     * - push sample 1 onto stack
617     * - add top two stack entries
618     * - push sample 2 onto stack
619     * - push sample 3 onto stack
620     * - add top two stack entries
621     * - add top two stack entries
622     * - divide top stack entry by 4
623     *
624     * Note that after pushing sample i onto the stack, the number of add
625     * operations we do is equal to the number of trailing 1 bits in i.  This
626     * works provided the total number of samples is a power of two, which it
627     * always is for i965.
628     *
629     * For integer formats, we replace the add operations with average
630     * operations and skip the final division.
631     */
632    nir_def *texture_data[5];
633    texture_data[0] = NULL; /* Avoid maybe-uninitialized warning with GCC 10 */
634    unsigned stack_depth = 0;
635    for (unsigned i = 0; i < tex_samples; ++i) {
636       assert(stack_depth == util_bitcount(i)); /* Loop invariant */
637 
638       /* Push sample i onto the stack */
639       assert(stack_depth < ARRAY_SIZE(texture_data));
640 
641       nir_def *ms_pos = nir_vec3(b, nir_channel(b, pos, 0),
642                                         nir_channel(b, pos, 1),
643                                         nir_imm_int(b, i));
644       texture_data[stack_depth++] = blorp_nir_txf_ms(b, v, ms_pos, mcs, dst_type);
645 
646       if (i == 0 && isl_aux_usage_has_mcs(tex_aux_usage)) {
647          /* The Ivy Bridge PRM, Vol4 Part1 p27 (Multisample Control Surface)
648           * suggests an optimization:
649           *
650           *     "A simple optimization with probable large return in
651           *     performance is to compare the MCS value to zero (indicating
652           *     all samples are on sample slice 0), and sample only from
653           *     sample slice 0 using ld2dss if MCS is zero."
654           *
655           * Note that in the case where the MCS value is zero, sampling from
656           * sample slice 0 using ld2dss and sampling from sample 0 using
657           * ld2dms are equivalent (since all samples are on sample slice 0).
658           * Since we have already sampled from sample 0, all we need to do is
659           * skip the remaining fetches and averaging if MCS is zero.
660           *
661           * It's also trivial to detect when the MCS has the magic clear color
662           * value.  In this case, the txf we did on sample 0 will return the
663           * clear color and we can skip the remaining fetches just like we do
664           * when MCS == 0.
665           */
666          nir_def *mcs_zero = nir_ieq_imm(b, nir_channel(b, mcs, 0), 0);
667          if (tex_samples == 16) {
668             mcs_zero = nir_iand(b, mcs_zero,
669                nir_ieq_imm(b, nir_channel(b, mcs, 1), 0));
670          }
671          nir_def *mcs_clear =
672             blorp_nir_mcs_is_clear_color(b, mcs, tex_samples);
673 
674          nir_push_if(b, nir_ior(b, mcs_zero, mcs_clear));
675          nir_store_var(b, color, texture_data[0], 0xf);
676 
677          nir_push_else(b, NULL);
678          inserted_if = true;
679       }
680 
681       for (int j = 0; j < count_trailing_one_bits(i); j++) {
682          assert(stack_depth >= 2);
683          --stack_depth;
684 
685          texture_data[stack_depth - 1] =
686             nir_build_alu(b, combine_op,
687                              texture_data[stack_depth - 1],
688                              texture_data[stack_depth],
689                              NULL, NULL);
690       }
691    }
692 
693    /* We should have just 1 sample on the stack now. */
694    assert(stack_depth == 1);
695 
696    if (filter == BLORP_FILTER_AVERAGE) {
697       assert(dst_type == nir_type_float);
698       texture_data[0] = nir_fmul_imm(b, texture_data[0],
699                                      1.0 / tex_samples);
700    }
701 
702    nir_store_var(b, color, texture_data[0], 0xf);
703 
704    if (inserted_if)
705       nir_pop_if(b, NULL);
706 
707    return nir_load_var(b, color);
708 }
709 
710 static nir_def *
blorp_nir_manual_blend_bilinear(nir_builder * b,nir_def * pos,unsigned tex_samples,const struct blorp_blit_prog_key * key,struct blorp_blit_vars * v)711 blorp_nir_manual_blend_bilinear(nir_builder *b, nir_def *pos,
712                                 unsigned tex_samples,
713                                 const struct blorp_blit_prog_key *key,
714                                 struct blorp_blit_vars *v)
715 {
716    nir_def *pos_xy = nir_trim_vector(b, pos, 2);
717    nir_def *rect_grid = nir_load_var(b, v->v_rect_grid);
718    nir_def *scale = nir_imm_vec2(b, key->x_scale, key->y_scale);
719 
720    /* Translate coordinates to lay out the samples in a rectangular  grid
721     * roughly corresponding to sample locations.
722     */
723    pos_xy = nir_fmul(b, pos_xy, scale);
724    /* Adjust coordinates so that integers represent pixel centers rather
725     * than pixel edges.
726     */
727    pos_xy = nir_fadd_imm(b, pos_xy, -0.5);
728    /* Clamp the X, Y texture coordinates to properly handle the sampling of
729     * texels on texture edges.
730     */
731    pos_xy = nir_fmin(b, nir_fmax(b, pos_xy, nir_imm_float(b, 0.0)),
732                         nir_trim_vector(b, rect_grid, 2));
733 
734    /* Store the fractional parts to be used as bilinear interpolation
735     * coefficients.
736     */
737    nir_def *frac_xy = nir_ffract(b, pos_xy);
738    /* Round the float coordinates down to nearest integer */
739    pos_xy = nir_fdiv(b, nir_ftrunc(b, pos_xy), scale);
740 
741    nir_def *tex_data[4];
742    for (unsigned i = 0; i < 4; ++i) {
743       float sample_off_x = (float)(i & 0x1) / key->x_scale;
744       float sample_off_y = (float)((i >> 1) & 0x1) / key->y_scale;
745       nir_def *sample_off = nir_imm_vec2(b, sample_off_x, sample_off_y);
746 
747       nir_def *sample_coords = nir_fadd(b, pos_xy, sample_off);
748       nir_def *sample_coords_int = nir_f2i32(b, sample_coords);
749 
750       /* The MCS value we fetch has to match up with the pixel that we're
751        * sampling from. Since we sample from different pixels in each
752        * iteration of this "for" loop, the call to mcs_fetch() should be
753        * here inside the loop after computing the pixel coordinates.
754        */
755       nir_def *mcs = NULL;
756       if (isl_aux_usage_has_mcs(key->tex_aux_usage))
757          mcs = blorp_blit_txf_ms_mcs(b, v, sample_coords_int);
758 
759       /* Compute sample index and map the sample index to a sample number.
760        * Sample index layout shows the numbering of slots in a rectangular
761        * grid of samples with in a pixel. Sample number layout shows the
762        * rectangular grid of samples roughly corresponding to the real sample
763        * locations with in a pixel.
764        *
765        * In the case of 2x MSAA, the layout of sample indices is reversed from
766        * the layout of sample numbers:
767        *
768        * sample index layout :  ---------    sample number layout :  ---------
769        *                        | 0 | 1 |                            | 1 | 0 |
770        *                        ---------                            ---------
771        *
772        * In case of 4x MSAA, layout of sample indices matches the layout of
773        * sample numbers:
774        *           ---------
775        *           | 0 | 1 |
776        *           ---------
777        *           | 2 | 3 |
778        *           ---------
779        *
780        * In case of 8x MSAA the two layouts don't match.
781        * sample index layout :  ---------    sample number layout :  ---------
782        *                        | 0 | 1 |                            | 3 | 7 |
783        *                        ---------                            ---------
784        *                        | 2 | 3 |                            | 5 | 0 |
785        *                        ---------                            ---------
786        *                        | 4 | 5 |                            | 1 | 2 |
787        *                        ---------                            ---------
788        *                        | 6 | 7 |                            | 4 | 6 |
789        *                        ---------                            ---------
790        *
791        * Fortunately, this can be done fairly easily as:
792        * S' = (0x17306425 >> (S * 4)) & 0xf
793        *
794        * In the case of 16x MSAA the two layouts don't match.
795        * Sample index layout:                Sample number layout:
796        * ---------------------               ---------------------
797        * |  0 |  1 |  2 |  3 |               | 15 | 10 |  9 |  7 |
798        * ---------------------               ---------------------
799        * |  4 |  5 |  6 |  7 |               |  4 |  1 |  3 | 13 |
800        * ---------------------               ---------------------
801        * |  8 |  9 | 10 | 11 |               | 12 |  2 |  0 |  6 |
802        * ---------------------               ---------------------
803        * | 12 | 13 | 14 | 15 |               | 11 |  8 |  5 | 14 |
804        * ---------------------               ---------------------
805        *
806        * This is equivalent to
807        * S' = (0xe58b602cd31479af >> (S * 4)) & 0xf
808        */
809       nir_def *frac = nir_ffract(b, sample_coords);
810       nir_def *sample =
811          nir_fdot2(b, frac, nir_imm_vec2(b, key->x_scale,
812                                             key->x_scale * key->y_scale));
813       sample = nir_f2i32(b, sample);
814 
815       if (tex_samples == 2) {
816          sample = nir_isub_imm(b, 1, sample);
817       } else if (tex_samples == 8) {
818          sample = nir_iand_imm(b, nir_ishr(b, nir_imm_int(b, 0x64210573),
819                                            nir_ishl_imm(b, sample, 2)),
820                                0xf);
821       } else if (tex_samples == 16) {
822          nir_def *sample_low =
823             nir_iand_imm(b, nir_ishr(b, nir_imm_int(b, 0xd31479af),
824                                      nir_ishl_imm(b, sample, 2)),
825                          0xf);
826          nir_def *sample_high =
827             nir_iand_imm(b, nir_ishr(b, nir_imm_int(b, 0xe58b602c),
828                                      nir_ishl_imm(b, nir_iadd_imm(b, sample, -8),
829                                                   2)),
830                          0xf);
831 
832          sample = nir_bcsel(b, nir_ilt_imm(b, sample, 8),
833                             sample_low, sample_high);
834       }
835       nir_def *pos_ms = nir_vec3(b, nir_channel(b, sample_coords_int, 0),
836                                         nir_channel(b, sample_coords_int, 1),
837                                         sample);
838       tex_data[i] = blorp_nir_txf_ms(b, v, pos_ms, mcs, key->texture_data_type);
839    }
840 
841    nir_def *frac_x = nir_channel(b, frac_xy, 0);
842    nir_def *frac_y = nir_channel(b, frac_xy, 1);
843    return nir_flrp(b, nir_flrp(b, tex_data[0], tex_data[1], frac_x),
844                       nir_flrp(b, tex_data[2], tex_data[3], frac_x),
845                       frac_y);
846 }
847 
848 /** Perform a color bit-cast operation
849  *
850  * For copy operations involving CCS, we may need to use different formats for
851  * the source and destination surfaces.  The two formats must both be UINT
852  * formats and must have the same size but may have different bit layouts.
853  * For instance, we may be copying from R8G8B8A8_UINT to R32_UINT or R32_UINT
854  * to R16G16_UINT.  This function generates code to shuffle bits around to get
855  * us from one to the other.
856  */
857 static nir_def *
bit_cast_color(struct nir_builder * b,nir_def * color,const struct blorp_blit_prog_key * key)858 bit_cast_color(struct nir_builder *b, nir_def *color,
859                const struct blorp_blit_prog_key *key)
860 {
861    if (key->src_format == key->dst_format)
862       return color;
863 
864    const struct isl_format_layout *src_fmtl =
865       isl_format_get_layout(key->src_format);
866    const struct isl_format_layout *dst_fmtl =
867       isl_format_get_layout(key->dst_format);
868 
869    /* They must be formats with the same bit size */
870    assert(src_fmtl->bpb == dst_fmtl->bpb);
871 
872    if (src_fmtl->bpb <= 32) {
873       assert(src_fmtl->channels.r.type == ISL_UINT ||
874              src_fmtl->channels.r.type == ISL_UNORM);
875       assert(dst_fmtl->channels.r.type == ISL_UINT ||
876              dst_fmtl->channels.r.type == ISL_UNORM);
877 
878       nir_def *packed = nir_imm_int(b, 0);
879       for (unsigned c = 0; c < 4; c++) {
880          if (src_fmtl->channels_array[c].bits == 0)
881             continue;
882 
883          const unsigned chan_start_bit = src_fmtl->channels_array[c].start_bit;
884          const unsigned chan_bits = src_fmtl->channels_array[c].bits;
885 
886          nir_def *chan =  nir_channel(b, color, c);
887          if (src_fmtl->channels_array[c].type == ISL_UNORM)
888             chan = nir_format_float_to_unorm(b, chan, &chan_bits);
889 
890          packed = nir_ior(b, packed, nir_shift_imm(b, chan, chan_start_bit));
891       }
892 
893       nir_def *chans[4] = { };
894       for (unsigned c = 0; c < 4; c++) {
895          if (dst_fmtl->channels_array[c].bits == 0) {
896             chans[c] = nir_imm_int(b, 0);
897             continue;
898          }
899 
900          const unsigned chan_start_bit = dst_fmtl->channels_array[c].start_bit;
901          const unsigned chan_bits = dst_fmtl->channels_array[c].bits;
902          chans[c] = nir_iand_imm(b, nir_shift_imm(b, packed, -(int)chan_start_bit),
903                                     BITFIELD_MASK(chan_bits));
904 
905          if (dst_fmtl->channels_array[c].type == ISL_UNORM)
906             chans[c] = nir_format_unorm_to_float(b, chans[c], &chan_bits);
907       }
908       color = nir_vec(b, chans, 4);
909    } else {
910       /* This path only supports UINT formats */
911       assert(src_fmtl->channels.r.type == ISL_UINT);
912       assert(dst_fmtl->channels.r.type == ISL_UINT);
913 
914       const unsigned src_bpc = src_fmtl->channels.r.bits;
915       const unsigned dst_bpc = dst_fmtl->channels.r.bits;
916 
917       assert(src_fmtl->channels.g.bits == 0 ||
918              src_fmtl->channels.g.bits == src_fmtl->channels.r.bits);
919       assert(src_fmtl->channels.b.bits == 0 ||
920              src_fmtl->channels.b.bits == src_fmtl->channels.r.bits);
921       assert(src_fmtl->channels.a.bits == 0 ||
922              src_fmtl->channels.a.bits == src_fmtl->channels.r.bits);
923       assert(dst_fmtl->channels.g.bits == 0 ||
924              dst_fmtl->channels.g.bits == dst_fmtl->channels.r.bits);
925       assert(dst_fmtl->channels.b.bits == 0 ||
926              dst_fmtl->channels.b.bits == dst_fmtl->channels.r.bits);
927       assert(dst_fmtl->channels.a.bits == 0 ||
928              dst_fmtl->channels.a.bits == dst_fmtl->channels.r.bits);
929 
930       /* Restrict to only the channels we actually have */
931       const unsigned src_channels =
932          isl_format_get_num_channels(key->src_format);
933       color = nir_trim_vector(b, color, src_channels);
934 
935       color = nir_format_bitcast_uvec_unmasked(b, color, src_bpc, dst_bpc);
936    }
937 
938    /* Blorp likes to assume that colors are vec4s */
939    nir_def *u = nir_undef(b, 1, 32);
940    nir_def *chans[4] = { u, u, u, u };
941    for (unsigned i = 0; i < color->num_components; i++)
942       chans[i] = nir_channel(b, color, i);
943    return nir_vec4(b, chans[0], chans[1], chans[2], chans[3]);
944 }
945 
946 static nir_def *
select_color_channel(struct nir_builder * b,nir_def * color,nir_alu_type data_type,enum isl_channel_select chan)947 select_color_channel(struct nir_builder *b, nir_def *color,
948                      nir_alu_type data_type,
949                      enum isl_channel_select chan)
950 {
951    if (chan == ISL_CHANNEL_SELECT_ZERO) {
952       return nir_imm_int(b, 0);
953    } else if (chan == ISL_CHANNEL_SELECT_ONE) {
954       switch (data_type) {
955       case nir_type_int:
956       case nir_type_uint:
957          return nir_imm_int(b, 1);
958       case nir_type_float:
959          return nir_imm_float(b, 1);
960       default:
961          unreachable("Invalid data type");
962       }
963    } else {
964       assert((unsigned)(chan - ISL_CHANNEL_SELECT_RED) < 4);
965       return nir_channel(b, color, chan - ISL_CHANNEL_SELECT_RED);
966    }
967 }
968 
969 static nir_def *
swizzle_color(struct nir_builder * b,nir_def * color,struct isl_swizzle swizzle,nir_alu_type data_type)970 swizzle_color(struct nir_builder *b, nir_def *color,
971               struct isl_swizzle swizzle, nir_alu_type data_type)
972 {
973    return nir_vec4(b,
974                    select_color_channel(b, color, data_type, swizzle.r),
975                    select_color_channel(b, color, data_type, swizzle.g),
976                    select_color_channel(b, color, data_type, swizzle.b),
977                    select_color_channel(b, color, data_type, swizzle.a));
978 }
979 
980 static nir_def *
convert_color(struct nir_builder * b,nir_def * color,const struct blorp_blit_prog_key * key)981 convert_color(struct nir_builder *b, nir_def *color,
982               const struct blorp_blit_prog_key *key)
983 {
984    /* All of our color conversions end up generating a single-channel color
985     * value that we need to write out.
986     */
987    nir_def *value;
988 
989    if (key->dst_format == ISL_FORMAT_R24_UNORM_X8_TYPELESS) {
990       /* The destination image is bound as R32_UINT but the data needs to be
991        * in R24_UNORM_X8_TYPELESS.  The bottom 24 are the actual data and the
992        * top 8 need to be zero.  We can accomplish this by simply multiplying
993        * by a factor to scale things down.
994        */
995       unsigned factor = (1 << 24) - 1;
996       value = nir_fsat(b, nir_channel(b, color, 0));
997       value = nir_f2i32(b, nir_fmul_imm(b, value, factor));
998    } else if (key->dst_format == ISL_FORMAT_L8_UNORM_SRGB) {
999       value = nir_format_linear_to_srgb(b, nir_channel(b, color, 0));
1000    } else if (key->dst_format == ISL_FORMAT_R8G8B8_UNORM_SRGB) {
1001       value = nir_format_linear_to_srgb(b, color);
1002    } else if (key->dst_format == ISL_FORMAT_R9G9B9E5_SHAREDEXP) {
1003       value = nir_format_pack_r9g9b9e5(b, color);
1004    } else {
1005       unreachable("Unsupported format conversion");
1006    }
1007 
1008    nir_def *out_comps[4];
1009    for (unsigned i = 0; i < 4; i++) {
1010       if (i < value->num_components)
1011          out_comps[i] = nir_channel(b, value, i);
1012       else
1013          out_comps[i] = nir_undef(b, 1, 32);
1014    }
1015    return nir_vec(b, out_comps, 4);
1016 }
1017 
1018 /**
1019  * Generator for WM programs used in BLORP blits.
1020  *
1021  * The bulk of the work done by the WM program is to wrap and unwrap the
1022  * coordinate transformations used by the hardware to store surfaces in
1023  * memory.  The hardware transforms a pixel location (X, Y, S) (where S is the
1024  * sample index for a multisampled surface) to a memory offset by the
1025  * following formulas:
1026  *
1027  *   offset = tile(tiling_format, encode_msaa(num_samples, layout, X, Y, S))
1028  *   (X, Y, S) = decode_msaa(num_samples, layout, detile(tiling_format, offset))
1029  *
1030  * For a single-sampled surface, or for a multisampled surface using
1031  * INTEL_MSAA_LAYOUT_UMS, encode_msaa() and decode_msaa are the identity
1032  * function:
1033  *
1034  *   encode_msaa(1, NONE, X, Y, 0) = (X, Y, 0)
1035  *   decode_msaa(1, NONE, X, Y, 0) = (X, Y, 0)
1036  *   encode_msaa(n, UMS, X, Y, S) = (X, Y, S)
1037  *   decode_msaa(n, UMS, X, Y, S) = (X, Y, S)
1038  *
1039  * For a 4x multisampled surface using INTEL_MSAA_LAYOUT_IMS, encode_msaa()
1040  * embeds the sample number into bit 1 of the X and Y coordinates:
1041  *
1042  *   encode_msaa(4, IMS, X, Y, S) = (X', Y', 0)
1043  *     where X' = (X & ~0b1) << 1 | (S & 0b1) << 1 | (X & 0b1)
1044  *           Y' = (Y & ~0b1 ) << 1 | (S & 0b10) | (Y & 0b1)
1045  *   decode_msaa(4, IMS, X, Y, 0) = (X', Y', S)
1046  *     where X' = (X & ~0b11) >> 1 | (X & 0b1)
1047  *           Y' = (Y & ~0b11) >> 1 | (Y & 0b1)
1048  *           S = (Y & 0b10) | (X & 0b10) >> 1
1049  *
1050  * For an 8x multisampled surface using INTEL_MSAA_LAYOUT_IMS, encode_msaa()
1051  * embeds the sample number into bits 1 and 2 of the X coordinate and bit 1 of
1052  * the Y coordinate:
1053  *
1054  *   encode_msaa(8, IMS, X, Y, S) = (X', Y', 0)
1055  *     where X' = (X & ~0b1) << 2 | (S & 0b100) | (S & 0b1) << 1 | (X & 0b1)
1056  *           Y' = (Y & ~0b1) << 1 | (S & 0b10) | (Y & 0b1)
1057  *   decode_msaa(8, IMS, X, Y, 0) = (X', Y', S)
1058  *     where X' = (X & ~0b111) >> 2 | (X & 0b1)
1059  *           Y' = (Y & ~0b11) >> 1 | (Y & 0b1)
1060  *           S = (X & 0b100) | (Y & 0b10) | (X & 0b10) >> 1
1061  *
1062  * For X tiling, tile() combines together the low-order bits of the X and Y
1063  * coordinates in the pattern 0byyyxxxxxxxxx, creating 4k tiles that are 512
1064  * bytes wide and 8 rows high:
1065  *
1066  *   tile(x_tiled, X, Y, S) = A
1067  *     where A = tile_num << 12 | offset
1068  *           tile_num = (Y' >> 3) * tile_pitch + (X' >> 9)
1069  *           offset = (Y' & 0b111) << 9
1070  *                    | (X & 0b111111111)
1071  *           X' = X * cpp
1072  *           Y' = Y + S * qpitch
1073  *   detile(x_tiled, A) = (X, Y, S)
1074  *     where X = X' / cpp
1075  *           Y = Y' % qpitch
1076  *           S = Y' / qpitch
1077  *           Y' = (tile_num / tile_pitch) << 3
1078  *                | (A & 0b111000000000) >> 9
1079  *           X' = (tile_num % tile_pitch) << 9
1080  *                | (A & 0b111111111)
1081  *
1082  * (In all tiling formulas, cpp is the number of bytes occupied by a single
1083  * sample ("chars per pixel"), tile_pitch is the number of 4k tiles required
1084  * to fill the width of the surface, and qpitch is the spacing (in rows)
1085  * between array slices).
1086  *
1087  * For Y tiling, tile() combines together the low-order bits of the X and Y
1088  * coordinates in the pattern 0bxxxyyyyyxxxx, creating 4k tiles that are 128
1089  * bytes wide and 32 rows high:
1090  *
1091  *   tile(y_tiled, X, Y, S) = A
1092  *     where A = tile_num << 12 | offset
1093  *           tile_num = (Y' >> 5) * tile_pitch + (X' >> 7)
1094  *           offset = (X' & 0b1110000) << 5
1095  *                    | (Y' & 0b11111) << 4
1096  *                    | (X' & 0b1111)
1097  *           X' = X * cpp
1098  *           Y' = Y + S * qpitch
1099  *   detile(y_tiled, A) = (X, Y, S)
1100  *     where X = X' / cpp
1101  *           Y = Y' % qpitch
1102  *           S = Y' / qpitch
1103  *           Y' = (tile_num / tile_pitch) << 5
1104  *                | (A & 0b111110000) >> 4
1105  *           X' = (tile_num % tile_pitch) << 7
1106  *                | (A & 0b111000000000) >> 5
1107  *                | (A & 0b1111)
1108  *
1109  * For W tiling, tile() combines together the low-order bits of the X and Y
1110  * coordinates in the pattern 0bxxxyyyyxyxyx, creating 4k tiles that are 64
1111  * bytes wide and 64 rows high (note that W tiling is only used for stencil
1112  * buffers, which always have cpp = 1 and S=0):
1113  *
1114  *   tile(w_tiled, X, Y, S) = A
1115  *     where A = tile_num << 12 | offset
1116  *           tile_num = (Y' >> 6) * tile_pitch + (X' >> 6)
1117  *           offset = (X' & 0b111000) << 6
1118  *                    | (Y' & 0b111100) << 3
1119  *                    | (X' & 0b100) << 2
1120  *                    | (Y' & 0b10) << 2
1121  *                    | (X' & 0b10) << 1
1122  *                    | (Y' & 0b1) << 1
1123  *                    | (X' & 0b1)
1124  *           X' = X * cpp = X
1125  *           Y' = Y + S * qpitch
1126  *   detile(w_tiled, A) = (X, Y, S)
1127  *     where X = X' / cpp = X'
1128  *           Y = Y' % qpitch = Y'
1129  *           S = Y / qpitch = 0
1130  *           Y' = (tile_num / tile_pitch) << 6
1131  *                | (A & 0b111100000) >> 3
1132  *                | (A & 0b1000) >> 2
1133  *                | (A & 0b10) >> 1
1134  *           X' = (tile_num % tile_pitch) << 6
1135  *                | (A & 0b111000000000) >> 6
1136  *                | (A & 0b10000) >> 2
1137  *                | (A & 0b100) >> 1
1138  *                | (A & 0b1)
1139  *
1140  * Finally, for a non-tiled surface, tile() simply combines together the X and
1141  * Y coordinates in the natural way:
1142  *
1143  *   tile(untiled, X, Y, S) = A
1144  *     where A = Y * pitch + X'
1145  *           X' = X * cpp
1146  *           Y' = Y + S * qpitch
1147  *   detile(untiled, A) = (X, Y, S)
1148  *     where X = X' / cpp
1149  *           Y = Y' % qpitch
1150  *           S = Y' / qpitch
1151  *           X' = A % pitch
1152  *           Y' = A / pitch
1153  *
1154  * (In these formulas, pitch is the number of bytes occupied by a single row
1155  * of samples).
1156  */
1157 static nir_shader *
blorp_build_nir_shader(struct blorp_context * blorp,struct blorp_batch * batch,void * mem_ctx,const struct blorp_blit_prog_key * key)1158 blorp_build_nir_shader(struct blorp_context *blorp,
1159                            struct blorp_batch *batch, void *mem_ctx,
1160                            const struct blorp_blit_prog_key *key)
1161 {
1162    const struct intel_device_info *devinfo = blorp->isl_dev->info;
1163    nir_def *src_pos, *dst_pos, *color;
1164 
1165    /* Sanity checks */
1166    if (key->dst_tiled_w && key->rt_samples > 1) {
1167       /* If the destination image is W tiled and multisampled, then the thread
1168        * must be dispatched once per sample, not once per pixel.  This is
1169        * necessary because after conversion between W and Y tiling, there's no
1170        * guarantee that all samples corresponding to a single pixel will still
1171        * be together.
1172        */
1173       assert(key->persample_msaa_dispatch);
1174    }
1175 
1176    if (key->persample_msaa_dispatch) {
1177       /* It only makes sense to do persample dispatch if the render target is
1178        * configured as multisampled.
1179        */
1180       assert(key->rt_samples > 0);
1181    }
1182 
1183    /* Make sure layout is consistent with sample count */
1184    assert((key->tex_layout == ISL_MSAA_LAYOUT_NONE) ==
1185           (key->tex_samples <= 1));
1186    assert((key->rt_layout == ISL_MSAA_LAYOUT_NONE) ==
1187           (key->rt_samples <= 1));
1188    assert((key->src_layout == ISL_MSAA_LAYOUT_NONE) ==
1189           (key->src_samples <= 1));
1190    assert((key->dst_layout == ISL_MSAA_LAYOUT_NONE) ==
1191           (key->dst_samples <= 1));
1192 
1193    nir_builder b;
1194    const bool compute =
1195       key->base.shader_pipeline == BLORP_SHADER_PIPELINE_COMPUTE;
1196    gl_shader_stage stage =
1197       compute ? MESA_SHADER_COMPUTE : MESA_SHADER_FRAGMENT;
1198    blorp_nir_init_shader(&b, blorp, mem_ctx, stage, NULL);
1199 
1200    struct blorp_blit_vars v;
1201    blorp_blit_vars_init(&b, &v, key);
1202 
1203    dst_pos = compute ?
1204       blorp_blit_get_cs_dst_coords(&b, key, &v) :
1205       blorp_blit_get_frag_coords(&b, key, &v);
1206 
1207    /* Render target and texture hardware don't support W tiling until Gfx8. */
1208    const bool rt_tiled_w = false;
1209    const bool tex_tiled_w = devinfo->ver >= 8 && key->src_tiled_w;
1210 
1211    /* The address that data will be written to is determined by the
1212     * coordinates supplied to the WM thread and the tiling and sample count of
1213     * the render target, according to the formula:
1214     *
1215     * (X, Y, S) = decode_msaa(rt_samples, detile(rt_tiling, offset))
1216     *
1217     * If the actual tiling and sample count of the destination surface are not
1218     * the same as the configuration of the render target, then these
1219     * coordinates are wrong and we have to adjust them to compensate for the
1220     * difference.
1221     */
1222    if (rt_tiled_w != key->dst_tiled_w ||
1223        key->rt_samples != key->dst_samples ||
1224        key->rt_layout != key->dst_layout) {
1225       dst_pos = blorp_nir_encode_msaa(&b, dst_pos, key->rt_samples,
1226                                       key->rt_layout);
1227       /* Now (X, Y, S) = detile(rt_tiling, offset) */
1228       if (rt_tiled_w != key->dst_tiled_w)
1229          dst_pos = blorp_nir_retile_y_to_w(&b, dst_pos);
1230       /* Now (X, Y, S) = detile(rt_tiling, offset) */
1231       dst_pos = blorp_nir_decode_msaa(&b, dst_pos, key->dst_samples,
1232                                       key->dst_layout);
1233    }
1234 
1235    nir_def *comp = NULL;
1236    if (key->dst_rgb) {
1237       /* The destination image is bound as a red texture three times as wide
1238        * as the actual image.  Our shader is effectively running one color
1239        * component at a time.  We need to save off the component and adjust
1240        * the destination position.
1241        */
1242       assert(dst_pos->num_components == 2);
1243       nir_def *dst_x = nir_channel(&b, dst_pos, 0);
1244       comp = nir_umod_imm(&b, dst_x, 3);
1245       dst_pos = nir_vec2(&b, nir_idiv(&b, dst_x, nir_imm_int(&b, 3)),
1246                              nir_channel(&b, dst_pos, 1));
1247    }
1248 
1249    /* Now (X, Y, S) = decode_msaa(dst_samples, detile(dst_tiling, offset)).
1250     *
1251     * That is: X, Y and S now contain the true coordinates and sample index of
1252     * the data that the WM thread should output.
1253     *
1254     * If we need to kill pixels that are outside the destination rectangle,
1255     * now is the time to do it.
1256     */
1257    nir_if *bounds_if = NULL;
1258    if (key->use_kill) {
1259       nir_def *bounds_rect = nir_load_var(&b, v.v_bounds_rect);
1260       nir_def *in_bounds = blorp_check_in_bounds(&b, bounds_rect,
1261                                                      dst_pos);
1262       if (!compute)
1263          nir_discard_if(&b, nir_inot(&b, in_bounds));
1264       else
1265          bounds_if = nir_push_if(&b, in_bounds);
1266    }
1267 
1268    src_pos = blorp_blit_apply_transform(&b, nir_i2f32(&b, dst_pos), &v);
1269    if (dst_pos->num_components == 3) {
1270       /* The sample coordinate is an integer that we want left alone but
1271        * blorp_blit_apply_transform() blindly applies the transform to all
1272        * three coordinates.  Grab the original sample index.
1273        */
1274       src_pos = nir_vec3(&b, nir_channel(&b, src_pos, 0),
1275                              nir_channel(&b, src_pos, 1),
1276                              nir_channel(&b, dst_pos, 2));
1277    }
1278 
1279    /* If the source image is not multisampled, then we want to fetch sample
1280     * number 0, because that's the only sample there is.
1281     */
1282    if (key->src_samples == 1)
1283       src_pos = nir_trim_vector(&b, src_pos, 2);
1284 
1285    /* X, Y, and S are now the coordinates of the pixel in the source image
1286     * that we want to texture from.  Exception: if we are blending, then S is
1287     * irrelevant, because we are going to fetch all samples.
1288     */
1289    switch (key->filter) {
1290    case BLORP_FILTER_NONE:
1291    case BLORP_FILTER_NEAREST:
1292    case BLORP_FILTER_SAMPLE_0:
1293       /* We're going to use texelFetch, so we need integers */
1294       if (src_pos->num_components == 2) {
1295          src_pos = nir_f2i32(&b, src_pos);
1296       } else {
1297          assert(src_pos->num_components == 3);
1298          src_pos = nir_vec3(&b, nir_channel(&b, nir_f2i32(&b, src_pos), 0),
1299                                 nir_channel(&b, nir_f2i32(&b, src_pos), 1),
1300                                 nir_channel(&b, src_pos, 2));
1301       }
1302 
1303       /* We aren't blending, which means we just want to fetch a single
1304        * sample from the source surface.  The address that we want to fetch
1305        * from is related to the X, Y and S values according to the formula:
1306        *
1307        * (X, Y, S) = decode_msaa(src_samples, detile(src_tiling, offset)).
1308        *
1309        * If the actual tiling and sample count of the source surface are
1310        * not the same as the configuration of the texture, then we need to
1311        * adjust the coordinates to compensate for the difference.
1312        */
1313       if (tex_tiled_w != key->src_tiled_w ||
1314           key->tex_samples != key->src_samples ||
1315           key->tex_layout != key->src_layout) {
1316          src_pos = blorp_nir_encode_msaa(&b, src_pos, key->src_samples,
1317                                          key->src_layout);
1318          /* Now (X, Y, S) = detile(src_tiling, offset) */
1319          if (tex_tiled_w != key->src_tiled_w)
1320             src_pos = blorp_nir_retile_w_to_y(&b, src_pos);
1321          /* Now (X, Y, S) = detile(tex_tiling, offset) */
1322          src_pos = blorp_nir_decode_msaa(&b, src_pos, key->tex_samples,
1323                                          key->tex_layout);
1324       }
1325 
1326       if (key->need_src_offset)
1327          src_pos = nir_iadd(&b, src_pos, nir_load_var(&b, v.v_src_offset));
1328 
1329       /* Now (X, Y, S) = decode_msaa(tex_samples, detile(tex_tiling, offset)).
1330        *
1331        * In other words: X, Y, and S now contain values which, when passed to
1332        * the texturing unit, will cause data to be read from the correct
1333        * memory location.  So we can fetch the texel now.
1334        */
1335       if (key->src_samples == 1) {
1336          color = blorp_nir_txf(&b, &v, src_pos, key->texture_data_type);
1337       } else {
1338          nir_def *mcs = NULL;
1339          if (isl_aux_usage_has_mcs(key->tex_aux_usage))
1340             mcs = blorp_blit_txf_ms_mcs(&b, &v, src_pos);
1341 
1342          color = blorp_nir_txf_ms(&b, &v, src_pos, mcs, key->texture_data_type);
1343       }
1344       break;
1345 
1346    case BLORP_FILTER_BILINEAR:
1347       assert(!key->src_tiled_w);
1348       assert(key->tex_samples == key->src_samples);
1349       assert(key->tex_layout == key->src_layout);
1350 
1351       if (key->src_samples == 1) {
1352          color = blorp_nir_tex(&b, &v, key, src_pos);
1353       } else {
1354          assert(!key->use_kill);
1355          color = blorp_nir_manual_blend_bilinear(&b, src_pos, key->src_samples,
1356                                                  key, &v);
1357       }
1358       break;
1359 
1360    case BLORP_FILTER_AVERAGE:
1361    case BLORP_FILTER_MIN_SAMPLE:
1362    case BLORP_FILTER_MAX_SAMPLE:
1363       assert(!key->src_tiled_w);
1364       assert(key->tex_samples == key->src_samples);
1365       assert(key->tex_layout == key->src_layout);
1366 
1367       /* Resolves (effecively) use texelFetch, so we need integers and we
1368        * don't care about the sample index if we got one.
1369        */
1370       src_pos = nir_f2i32(&b, nir_trim_vector(&b, src_pos, 2));
1371 
1372       if (devinfo->ver == 6) {
1373          /* Because gfx6 only supports 4x interleved MSAA, we can do all the
1374           * blending we need with a single linear-interpolated texture lookup
1375           * at the center of the sample. The texture coordinates to be odd
1376           * integers so that they correspond to the center of a 2x2 block
1377           * representing the four samples that maxe up a pixel.  So we need
1378           * to multiply our X and Y coordinates each by 2 and then add 1.
1379           */
1380          assert(key->src_coords_normalized);
1381          assert(key->filter == BLORP_FILTER_AVERAGE);
1382          src_pos = nir_fadd_imm(&b,
1383                                 nir_i2f32(&b, src_pos),
1384                                 0.5f);
1385          color = blorp_nir_tex(&b, &v, key, src_pos);
1386       } else {
1387          /* Gfx7+ hardware doesn't automatically blend. */
1388          color = blorp_nir_combine_samples(&b, &v, src_pos, key->src_samples,
1389                                            key->tex_aux_usage,
1390                                            key->texture_data_type,
1391                                            key->filter);
1392       }
1393       break;
1394 
1395    default:
1396       unreachable("Invalid blorp filter");
1397    }
1398 
1399    if (!isl_swizzle_is_identity(key->src_swizzle)) {
1400       color = swizzle_color(&b, color, key->src_swizzle,
1401                             key->texture_data_type);
1402    }
1403 
1404    if (!isl_swizzle_is_identity(key->dst_swizzle)) {
1405       color = swizzle_color(&b, color, isl_swizzle_invert(key->dst_swizzle),
1406                             nir_type_int);
1407    }
1408 
1409    if (key->format_bit_cast) {
1410       assert(isl_swizzle_is_identity(key->src_swizzle));
1411       assert(isl_swizzle_is_identity(key->dst_swizzle));
1412       color = bit_cast_color(&b, color, key);
1413    } else if (key->dst_format) {
1414       color = convert_color(&b, color, key);
1415    } else if (key->uint32_to_sint) {
1416       /* Normally the hardware will take care of converting values from/to
1417        * the source and destination formats.  But a few cases need help.
1418        *
1419        * The Skylake PRM, volume 07, page 658 has a programming note:
1420        *
1421        *    "When using SINT or UINT rendertarget surface formats, Blending
1422        *     must be DISABLED. The Pre-Blend Color Clamp Enable and Color
1423        *     Clamp Range fields are ignored, and an implied clamp to the
1424        *     rendertarget surface format is performed."
1425        *
1426        * For UINT to SINT blits, our sample operation gives us a uint32_t,
1427        * but our render target write expects a signed int32_t number.  If we
1428        * simply passed the value along, the hardware would interpret a value
1429        * with bit 31 set as a negative value, clamping it to the largest
1430        * negative number the destination format could represent.  But the
1431        * actual source value is a positive number, so we want to clamp it
1432        * to INT_MAX.  To fix this, we explicitly take min(color, INT_MAX).
1433        */
1434       color = nir_umin(&b, color, nir_imm_int(&b, INT32_MAX));
1435    } else if (key->sint32_to_uint) {
1436       /* Similar to above, but clamping negative numbers to zero. */
1437       color = nir_imax(&b, color, nir_imm_int(&b, 0));
1438    }
1439 
1440    if (key->dst_rgb) {
1441       /* The destination image is bound as a red texture three times as wide
1442        * as the actual image.  Our shader is effectively running one color
1443        * component at a time.  We need to pick off the appropriate component
1444        * from the source color and write that to destination red.
1445        */
1446       assert(dst_pos->num_components == 2);
1447 
1448       nir_def *color_component =
1449          nir_bcsel(&b, nir_ieq_imm(&b, comp, 0),
1450                        nir_channel(&b, color, 0),
1451                        nir_bcsel(&b, nir_ieq_imm(&b, comp, 1),
1452                                      nir_channel(&b, color, 1),
1453                                      nir_channel(&b, color, 2)));
1454 
1455       nir_def *u = nir_undef(&b, 1, 32);
1456       color = nir_vec4(&b, color_component, u, u, u);
1457    }
1458 
1459    if (compute) {
1460       nir_def *store_pos = nir_load_global_invocation_id(&b, 32);
1461       nir_image_store(&b, nir_imm_int(&b, 0),
1462                       nir_pad_vector_imm_int(&b, store_pos, 0, 4),
1463                       nir_imm_int(&b, 0),
1464                       nir_pad_vector_imm_int(&b, color, 0, 4),
1465                       nir_imm_int(&b, 0),
1466                       .image_dim = GLSL_SAMPLER_DIM_2D,
1467                       .image_array = true,
1468                       .access = ACCESS_NON_READABLE);
1469    } else if (key->dst_usage == ISL_SURF_USAGE_RENDER_TARGET_BIT) {
1470       nir_variable *color_out =
1471          nir_variable_create(b.shader, nir_var_shader_out,
1472                              glsl_vec4_type(), "gl_FragColor");
1473       color_out->data.location = FRAG_RESULT_COLOR;
1474       nir_store_var(&b, color_out, color, 0xf);
1475    } else if (key->dst_usage == ISL_SURF_USAGE_DEPTH_BIT) {
1476       nir_variable *depth_out =
1477          nir_variable_create(b.shader, nir_var_shader_out,
1478                              glsl_float_type(), "gl_FragDepth");
1479       depth_out->data.location = FRAG_RESULT_DEPTH;
1480       nir_store_var(&b, depth_out, nir_channel(&b, color, 0), 0x1);
1481    } else if (key->dst_usage == ISL_SURF_USAGE_STENCIL_BIT) {
1482       nir_variable *stencil_out =
1483          nir_variable_create(b.shader, nir_var_shader_out,
1484                              glsl_int_type(), "gl_FragStencilRef");
1485       stencil_out->data.location = FRAG_RESULT_STENCIL;
1486       nir_store_var(&b, stencil_out, nir_channel(&b, color, 0), 0x1);
1487    } else {
1488       unreachable("Invalid destination usage");
1489    }
1490 
1491    if (bounds_if)
1492       nir_pop_if(&b, bounds_if);
1493 
1494    return b.shader;
1495 }
1496 
1497 static bool
blorp_get_blit_kernel_fs(struct blorp_batch * batch,struct blorp_params * params,const struct blorp_blit_prog_key * key)1498 blorp_get_blit_kernel_fs(struct blorp_batch *batch,
1499                          struct blorp_params *params,
1500                          const struct blorp_blit_prog_key *key)
1501 {
1502    struct blorp_context *blorp = batch->blorp;
1503 
1504    if (blorp->lookup_shader(batch, key, sizeof(*key),
1505                             &params->wm_prog_kernel, &params->wm_prog_data))
1506       return true;
1507 
1508    void *mem_ctx = ralloc_context(NULL);
1509 
1510    nir_shader *nir = blorp_build_nir_shader(blorp, batch, mem_ctx, key);
1511    nir->info.name =
1512       ralloc_strdup(nir, blorp_shader_type_to_name(key->base.shader_type));
1513 
1514    const bool multisample_fbo = key->rt_samples > 1;
1515 
1516    const struct blorp_program p =
1517       blorp_compile_fs(blorp, mem_ctx, nir, multisample_fbo, false);
1518 
1519    bool result =
1520       blorp->upload_shader(batch, MESA_SHADER_FRAGMENT,
1521                            key, sizeof(*key),
1522                            p.kernel, p.kernel_size,
1523                            p.prog_data, p.prog_data_size,
1524                            &params->wm_prog_kernel, &params->wm_prog_data);
1525 
1526    ralloc_free(mem_ctx);
1527    return result;
1528 }
1529 
1530 static bool
blorp_get_blit_kernel_cs(struct blorp_batch * batch,struct blorp_params * params,const struct blorp_blit_prog_key * prog_key)1531 blorp_get_blit_kernel_cs(struct blorp_batch *batch,
1532                          struct blorp_params *params,
1533                          const struct blorp_blit_prog_key *prog_key)
1534 {
1535    struct blorp_context *blorp = batch->blorp;
1536 
1537    if (blorp->lookup_shader(batch, prog_key, sizeof(*prog_key),
1538                             &params->cs_prog_kernel, &params->cs_prog_data))
1539       return true;
1540 
1541    void *mem_ctx = ralloc_context(NULL);
1542 
1543    nir_shader *nir = blorp_build_nir_shader(blorp, batch, mem_ctx,
1544                                                 prog_key);
1545    nir->info.name = ralloc_strdup(nir, "BLORP-gpgpu-blit");
1546    blorp_set_cs_dims(nir, prog_key->local_y);
1547 
1548    assert(prog_key->rt_samples == 1);
1549 
1550    const struct blorp_program p =
1551       blorp_compile_cs(blorp, mem_ctx, nir);
1552 
1553    bool result =
1554       blorp->upload_shader(batch, MESA_SHADER_COMPUTE,
1555                            prog_key, sizeof(*prog_key),
1556                            p.kernel, p.kernel_size,
1557                            p.prog_data, p.prog_data_size,
1558                            &params->cs_prog_kernel, &params->cs_prog_data);
1559 
1560    ralloc_free(mem_ctx);
1561    return result;
1562 }
1563 
1564 static void
blorp_setup_coord_transform(struct blorp_coord_transform * xform,float src0,float src1,float dst0,float dst1,bool mirror)1565 blorp_setup_coord_transform(struct blorp_coord_transform *xform,
1566                             float src0, float src1,
1567                             float dst0, float dst1,
1568                             bool mirror)
1569 {
1570    double scale = (double)(src1 - src0) / (double)(dst1 - dst0);
1571    if (!mirror) {
1572       /* When not mirroring a coordinate (say, X), we need:
1573        *   src_x - src_x0 = (dst_x - dst_x0 + 0.5) * scale
1574        * Therefore:
1575        *   src_x = src_x0 + (dst_x - dst_x0 + 0.5) * scale
1576        *
1577        * blorp program uses "round toward zero" to convert the
1578        * transformed floating point coordinates to integer coordinates,
1579        * whereas the behaviour we actually want is "round to nearest",
1580        * so 0.5 provides the necessary correction.
1581        */
1582       xform->multiplier = scale;
1583       xform->offset = src0 + (-(double)dst0 + 0.5) * scale;
1584    } else {
1585       /* When mirroring X we need:
1586        *   src_x - src_x0 = dst_x1 - dst_x - 0.5
1587        * Therefore:
1588        *   src_x = src_x0 + (dst_x1 -dst_x - 0.5) * scale
1589        */
1590       xform->multiplier = -scale;
1591       xform->offset = src0 + ((double)dst1 - 0.5) * scale;
1592    }
1593 }
1594 
1595 static inline void
surf_get_intratile_offset_px(struct blorp_surface_info * info,uint32_t * tile_x_px,uint32_t * tile_y_px)1596 surf_get_intratile_offset_px(struct blorp_surface_info *info,
1597                              uint32_t *tile_x_px, uint32_t *tile_y_px)
1598 {
1599    if (info->surf.msaa_layout == ISL_MSAA_LAYOUT_INTERLEAVED) {
1600       struct isl_extent2d px_size_sa =
1601          isl_get_interleaved_msaa_px_size_sa(info->surf.samples);
1602       assert(info->tile_x_sa % px_size_sa.width == 0);
1603       assert(info->tile_y_sa % px_size_sa.height == 0);
1604       *tile_x_px = info->tile_x_sa / px_size_sa.width;
1605       *tile_y_px = info->tile_y_sa / px_size_sa.height;
1606    } else {
1607       *tile_x_px = info->tile_x_sa;
1608       *tile_y_px = info->tile_y_sa;
1609    }
1610 }
1611 
1612 void
blorp_surf_convert_to_single_slice(const struct isl_device * isl_dev,struct blorp_surface_info * info)1613 blorp_surf_convert_to_single_slice(const struct isl_device *isl_dev,
1614                                    struct blorp_surface_info *info)
1615 {
1616    bool ok UNUSED;
1617 
1618    /* It would be insane to try and do this on a compressed surface */
1619    assert(info->aux_usage == ISL_AUX_USAGE_NONE);
1620 
1621    /* Just bail if we have nothing to do. */
1622    if (info->surf.dim == ISL_SURF_DIM_2D &&
1623        info->view.base_level == 0 && info->view.base_array_layer == 0 &&
1624        info->surf.levels == 1 && info->surf.logical_level0_px.array_len == 1)
1625       return;
1626 
1627    /* If this gets triggered then we've gotten here twice which.  This
1628     * shouldn't happen thanks to the above early return.
1629     */
1630    assert(info->tile_x_sa == 0 && info->tile_y_sa == 0);
1631 
1632    uint32_t layer = 0, z = 0;
1633    if (info->surf.dim == ISL_SURF_DIM_3D)
1634       z = info->view.base_array_layer + info->z_offset;
1635    else
1636       layer = info->view.base_array_layer;
1637 
1638    uint64_t offset_B;
1639    isl_surf_get_image_surf(isl_dev, &info->surf,
1640                            info->view.base_level, layer, z,
1641                            &info->surf,
1642                            &offset_B, &info->tile_x_sa, &info->tile_y_sa);
1643    info->addr.offset += offset_B;
1644 
1645    uint32_t tile_x_px, tile_y_px;
1646    surf_get_intratile_offset_px(info, &tile_x_px, &tile_y_px);
1647 
1648    /* Instead of using the X/Y Offset fields in RENDER_SURFACE_STATE, we place
1649     * the image at the tile boundary and offset our sampling or rendering.
1650     * For this reason, we need to grow the image by the offset to ensure that
1651     * the hardware doesn't think we've gone past the edge.
1652     */
1653    info->surf.logical_level0_px.w += tile_x_px;
1654    info->surf.logical_level0_px.h += tile_y_px;
1655    info->surf.phys_level0_sa.w += info->tile_x_sa;
1656    info->surf.phys_level0_sa.h += info->tile_y_sa;
1657 
1658    /* The view is also different now. */
1659    info->view.base_level = 0;
1660    info->view.levels = 1;
1661    info->view.base_array_layer = 0;
1662    info->view.array_len = 1;
1663    info->z_offset = 0;
1664 }
1665 
1666 void
blorp_surf_fake_interleaved_msaa(const struct isl_device * isl_dev,struct blorp_surface_info * info)1667 blorp_surf_fake_interleaved_msaa(const struct isl_device *isl_dev,
1668                                  struct blorp_surface_info *info)
1669 {
1670    assert(info->surf.msaa_layout == ISL_MSAA_LAYOUT_INTERLEAVED);
1671 
1672    /* First, we need to convert it to a simple 1-level 1-layer 2-D surface */
1673    blorp_surf_convert_to_single_slice(isl_dev, info);
1674 
1675    info->surf.logical_level0_px = info->surf.phys_level0_sa;
1676    info->surf.samples = 1;
1677    info->surf.msaa_layout = ISL_MSAA_LAYOUT_NONE;
1678 }
1679 
1680 void
blorp_surf_retile_w_to_y(const struct isl_device * isl_dev,struct blorp_surface_info * info)1681 blorp_surf_retile_w_to_y(const struct isl_device *isl_dev,
1682                          struct blorp_surface_info *info)
1683 {
1684    assert(info->surf.tiling == ISL_TILING_W);
1685 
1686    /* First, we need to convert it to a simple 1-level 1-layer 2-D surface */
1687    blorp_surf_convert_to_single_slice(isl_dev, info);
1688 
1689    /* On gfx7+, we don't have interleaved multisampling for color render
1690     * targets so we have to fake it.
1691     *
1692     * TODO: Are we sure we don't also need to fake it on gfx6?
1693     */
1694    if (isl_dev->info->ver > 6 &&
1695        info->surf.msaa_layout == ISL_MSAA_LAYOUT_INTERLEAVED) {
1696       blorp_surf_fake_interleaved_msaa(isl_dev, info);
1697    }
1698 
1699    if (isl_dev->info->ver == 6 || isl_dev->info->ver == 7) {
1700       /* Gfx6-7 stencil buffers have a very large alignment coming in from the
1701        * miptree.  It's out-of-bounds for what the surface state can handle.
1702        * Since we have a single layer and level, it doesn't really matter as
1703        * long as we don't pass a bogus value into isl_surf_fill_state().
1704        */
1705       info->surf.image_alignment_el = isl_extent3d(4, 2, 1);
1706    }
1707 
1708    /* Now that we've converted everything to a simple 2-D surface with only
1709     * one miplevel, we can go about retiling it.
1710     */
1711    const unsigned x_align = 8, y_align = info->surf.samples != 0 ? 8 : 4;
1712    info->surf.tiling = ISL_TILING_Y0;
1713    info->surf.logical_level0_px.width =
1714       ALIGN(info->surf.logical_level0_px.width, x_align) * 2;
1715    info->surf.logical_level0_px.height =
1716       ALIGN(info->surf.logical_level0_px.height, y_align) / 2;
1717    info->tile_x_sa *= 2;
1718    info->tile_y_sa /= 2;
1719 }
1720 
1721 static bool
can_shrink_surface(const struct blorp_surface_info * surf)1722 can_shrink_surface(const struct blorp_surface_info *surf)
1723 {
1724    /* The current code doesn't support offsets into the aux buffers. This
1725     * should be possible, but we need to make sure the offset is page
1726     * aligned for both the surface and the aux buffer surface. Generally
1727     * this mean using the page aligned offset for the aux buffer.
1728     *
1729     * Currently the cases where we must split the blit are limited to cases
1730     * where we don't have a aux buffer.
1731     */
1732    if (surf->aux_addr.buffer != NULL)
1733       return false;
1734 
1735    /* We can't support splitting the blit for gen <= 7, because the qpitch
1736     * size is calculated by the hardware based on the surface height for
1737     * gen <= 7. In gen >= 8, the qpitch is controlled by the driver.
1738     */
1739    if (surf->surf.msaa_layout == ISL_MSAA_LAYOUT_ARRAY)
1740       return false;
1741 
1742    return true;
1743 }
1744 
1745 static unsigned
get_max_surface_size(const struct intel_device_info * devinfo,const struct blorp_surface_info * surf)1746 get_max_surface_size(const struct intel_device_info *devinfo,
1747                      const struct blorp_surface_info *surf)
1748 {
1749    const unsigned max = devinfo->ver >= 7 ? 16384 : 8192;
1750    if (split_blorp_blit_debug && can_shrink_surface(surf))
1751       return max >> 4; /* A smaller restriction when debug is enabled */
1752    else
1753       return max;
1754 }
1755 
1756 struct blt_axis {
1757    double src0, src1, dst0, dst1;
1758    bool mirror;
1759 };
1760 
1761 struct blt_coords {
1762    struct blt_axis x, y;
1763 };
1764 
1765 static enum isl_format
get_red_format_for_rgb_format(enum isl_format format)1766 get_red_format_for_rgb_format(enum isl_format format)
1767 {
1768    const struct isl_format_layout *fmtl = isl_format_get_layout(format);
1769 
1770    switch (fmtl->channels.r.bits) {
1771    case 8:
1772       switch (fmtl->channels.r.type) {
1773       case ISL_UNORM:
1774          return ISL_FORMAT_R8_UNORM;
1775       case ISL_SNORM:
1776          return ISL_FORMAT_R8_SNORM;
1777       case ISL_UINT:
1778          return ISL_FORMAT_R8_UINT;
1779       case ISL_SINT:
1780          return ISL_FORMAT_R8_SINT;
1781       default:
1782          unreachable("Invalid 8-bit RGB channel type");
1783       }
1784    case 16:
1785       switch (fmtl->channels.r.type) {
1786       case ISL_UNORM:
1787          return ISL_FORMAT_R16_UNORM;
1788       case ISL_SNORM:
1789          return ISL_FORMAT_R16_SNORM;
1790       case ISL_SFLOAT:
1791          return ISL_FORMAT_R16_FLOAT;
1792       case ISL_UINT:
1793          return ISL_FORMAT_R16_UINT;
1794       case ISL_SINT:
1795          return ISL_FORMAT_R16_SINT;
1796       default:
1797          unreachable("Invalid 8-bit RGB channel type");
1798       }
1799    case 32:
1800       switch (fmtl->channels.r.type) {
1801       case ISL_SFLOAT:
1802          return ISL_FORMAT_R32_FLOAT;
1803       case ISL_UINT:
1804          return ISL_FORMAT_R32_UINT;
1805       case ISL_SINT:
1806          return ISL_FORMAT_R32_SINT;
1807       default:
1808          unreachable("Invalid 8-bit RGB channel type");
1809       }
1810    default:
1811       unreachable("Invalid number of red channel bits");
1812    }
1813 }
1814 
1815 void
surf_fake_rgb_with_red(const struct isl_device * isl_dev,struct blorp_surface_info * info)1816 surf_fake_rgb_with_red(const struct isl_device *isl_dev,
1817                        struct blorp_surface_info *info)
1818 {
1819    blorp_surf_convert_to_single_slice(isl_dev, info);
1820 
1821    info->surf.logical_level0_px.width *= 3;
1822    info->surf.phys_level0_sa.width *= 3;
1823    info->tile_x_sa *= 3;
1824 
1825    enum isl_format red_format =
1826       get_red_format_for_rgb_format(info->view.format);
1827 
1828    assert(isl_format_get_layout(red_format)->channels.r.type ==
1829           isl_format_get_layout(info->view.format)->channels.r.type);
1830    assert(isl_format_get_layout(red_format)->channels.r.bits ==
1831           isl_format_get_layout(info->view.format)->channels.r.bits);
1832 
1833    info->surf.format = info->view.format = red_format;
1834 
1835    if (isl_dev->info->verx10 >= 125) {
1836       /* The horizontal alignment is in units of texels for NPOT formats, and
1837        * bytes for other formats. Since the only allowed alignment units are
1838        * powers of two, there's no way to convert the alignment.
1839        *
1840        * Thankfully, the value doesn't matter since we're only a single slice.
1841        * Pick one allowed by isl_gfx125_choose_image_alignment_el.
1842        */
1843       info->surf.image_alignment_el.w =
1844          128 / (isl_format_get_layout(red_format)->bpb / 8);
1845    }
1846 }
1847 
1848 enum blit_shrink_status {
1849    BLIT_NO_SHRINK = 0,
1850    BLIT_SRC_WIDTH_SHRINK   = (1 << 0),
1851    BLIT_DST_WIDTH_SHRINK   = (1 << 1),
1852    BLIT_SRC_HEIGHT_SHRINK  = (1 << 2),
1853    BLIT_DST_HEIGHT_SHRINK  = (1 << 3),
1854 };
1855 
1856 /* Try to blit. If the surface parameters exceed the size allowed by hardware,
1857  * then enum blit_shrink_status will be returned. If BLIT_NO_SHRINK is
1858  * returned, then the blit was successful.
1859  */
1860 static enum blit_shrink_status
try_blorp_blit(struct blorp_batch * batch,struct blorp_params * params,struct blorp_blit_prog_key * key,struct blt_coords * coords)1861 try_blorp_blit(struct blorp_batch *batch,
1862                struct blorp_params *params,
1863                struct blorp_blit_prog_key *key,
1864                struct blt_coords *coords)
1865 {
1866    const struct intel_device_info *devinfo = batch->blorp->isl_dev->info;
1867 
1868    if (params->dst.surf.usage & ISL_SURF_USAGE_DEPTH_BIT) {
1869       if (devinfo->ver >= 7) {
1870          /* We can render as depth on Gfx5 but there's no real advantage since
1871           * it doesn't support MSAA or HiZ.  On Gfx4, we can't always render
1872           * to depth due to issues with depth buffers and mip-mapping.  On
1873           * Gfx6, we can do everything but we have weird offsetting for HiZ
1874           * and stencil.  It's easier to just render using the color pipe
1875           * on those platforms.
1876           */
1877          key->dst_usage = ISL_SURF_USAGE_DEPTH_BIT;
1878       } else {
1879          key->dst_usage = ISL_SURF_USAGE_RENDER_TARGET_BIT;
1880       }
1881    } else if (params->dst.surf.usage & ISL_SURF_USAGE_STENCIL_BIT) {
1882       assert(params->dst.surf.format == ISL_FORMAT_R8_UINT);
1883       if (devinfo->ver >= 9 && !(batch->flags & BLORP_BATCH_USE_COMPUTE)) {
1884          key->dst_usage = ISL_SURF_USAGE_STENCIL_BIT;
1885       } else {
1886          key->dst_usage = ISL_SURF_USAGE_RENDER_TARGET_BIT;
1887       }
1888    } else {
1889       key->dst_usage = ISL_SURF_USAGE_RENDER_TARGET_BIT;
1890    }
1891 
1892    if (isl_format_has_sint_channel(params->src.view.format)) {
1893       key->texture_data_type = nir_type_int;
1894    } else if (isl_format_has_uint_channel(params->src.view.format)) {
1895       key->texture_data_type = nir_type_uint;
1896    } else {
1897       key->texture_data_type = nir_type_float;
1898    }
1899 
1900    /* src_samples and dst_samples are the true sample counts */
1901    key->src_samples = params->src.surf.samples;
1902    key->dst_samples = params->dst.surf.samples;
1903 
1904    key->tex_aux_usage = params->src.aux_usage;
1905 
1906    /* src_layout and dst_layout indicate the true MSAA layout used by src and
1907     * dst.
1908     */
1909    key->src_layout = params->src.surf.msaa_layout;
1910    key->dst_layout = params->dst.surf.msaa_layout;
1911 
1912    /* Round floating point values to nearest integer to avoid "off by one texel"
1913     * kind of errors when blitting.
1914     */
1915    params->x0 = params->wm_inputs.bounds_rect.x0 = round(coords->x.dst0);
1916    params->y0 = params->wm_inputs.bounds_rect.y0 = round(coords->y.dst0);
1917    params->x1 = params->wm_inputs.bounds_rect.x1 = round(coords->x.dst1);
1918    params->y1 = params->wm_inputs.bounds_rect.y1 = round(coords->y.dst1);
1919 
1920    blorp_setup_coord_transform(&params->wm_inputs.coord_transform[0],
1921                                    coords->x.src0, coords->x.src1,
1922                                    coords->x.dst0, coords->x.dst1,
1923                                    coords->x.mirror);
1924    blorp_setup_coord_transform(&params->wm_inputs.coord_transform[1],
1925                                    coords->y.src0, coords->y.src1,
1926                                    coords->y.dst0, coords->y.dst1,
1927                                    coords->y.mirror);
1928 
1929 
1930    if (devinfo->ver == 4) {
1931       /* The MinLOD and MinimumArrayElement don't work properly for cube maps.
1932        * Convert them to a single slice on gfx4.
1933        */
1934       if (params->dst.surf.usage & ISL_SURF_USAGE_CUBE_BIT) {
1935          blorp_surf_convert_to_single_slice(batch->blorp->isl_dev, &params->dst);
1936          key->need_dst_offset = true;
1937       }
1938 
1939       if (params->src.surf.usage & ISL_SURF_USAGE_CUBE_BIT) {
1940          blorp_surf_convert_to_single_slice(batch->blorp->isl_dev, &params->src);
1941          key->need_src_offset = true;
1942       }
1943    }
1944 
1945    if (devinfo->ver > 6 &&
1946        !isl_surf_usage_is_depth_or_stencil(key->dst_usage) &&
1947        params->dst.surf.msaa_layout == ISL_MSAA_LAYOUT_INTERLEAVED) {
1948       assert(params->dst.surf.samples > 1);
1949 
1950       /* We must expand the rectangle we send through the rendering pipeline,
1951        * to account for the fact that we are mapping the destination region as
1952        * single-sampled when it is in fact multisampled.  We must also align
1953        * it to a multiple of the multisampling pattern, because the
1954        * differences between multisampled and single-sampled surface formats
1955        * will mean that pixels are scrambled within the multisampling pattern.
1956        * TODO: what if this makes the coordinates too large?
1957        *
1958        * Note: this only works if the destination surface uses the IMS layout.
1959        * If it's UMS, then we have no choice but to set up the rendering
1960        * pipeline as multisampled.
1961        */
1962       struct isl_extent2d px_size_sa =
1963          isl_get_interleaved_msaa_px_size_sa(params->dst.surf.samples);
1964       params->x0 = ROUND_DOWN_TO(params->x0, 2) * px_size_sa.width;
1965       params->y0 = ROUND_DOWN_TO(params->y0, 2) * px_size_sa.height;
1966       params->x1 = ALIGN(params->x1, 2) * px_size_sa.width;
1967       params->y1 = ALIGN(params->y1, 2) * px_size_sa.height;
1968 
1969       blorp_surf_fake_interleaved_msaa(batch->blorp->isl_dev, &params->dst);
1970 
1971       key->use_kill = true;
1972       key->need_dst_offset = true;
1973    }
1974 
1975    if (params->dst.surf.tiling == ISL_TILING_W &&
1976        key->dst_usage != ISL_SURF_USAGE_STENCIL_BIT) {
1977       /* We must modify the rectangle we send through the rendering pipeline
1978        * (and the size and x/y offset of the destination surface), to account
1979        * for the fact that we are mapping it as Y-tiled when it is in fact
1980        * W-tiled.
1981        *
1982        * Both Y tiling and W tiling can be understood as organizations of
1983        * 32-byte sub-tiles; within each 32-byte sub-tile, the layout of pixels
1984        * is different, but the layout of the 32-byte sub-tiles within the 4k
1985        * tile is the same (8 sub-tiles across by 16 sub-tiles down, in
1986        * column-major order).  In Y tiling, the sub-tiles are 16 bytes wide
1987        * and 2 rows high; in W tiling, they are 8 bytes wide and 4 rows high.
1988        *
1989        * Therefore, to account for the layout differences within the 32-byte
1990        * sub-tiles, we must expand the rectangle so the X coordinates of its
1991        * edges are multiples of 8 (the W sub-tile width), and its Y
1992        * coordinates of its edges are multiples of 4 (the W sub-tile height).
1993        * Then we need to scale the X and Y coordinates of the rectangle to
1994        * account for the differences in aspect ratio between the Y and W
1995        * sub-tiles.  We need to modify the layer width and height similarly.
1996        *
1997        * A correction needs to be applied when MSAA is in use: since
1998        * INTEL_MSAA_LAYOUT_IMS uses an interleaving pattern whose height is 4,
1999        * we need to align the Y coordinates to multiples of 8, so that when
2000        * they are divided by two they are still multiples of 4.
2001        *
2002        * Note: Since the x/y offset of the surface will be applied using the
2003        * SURFACE_STATE command packet, it will be invisible to the swizzling
2004        * code in the shader; therefore it needs to be in a multiple of the
2005        * 32-byte sub-tile size.  Fortunately it is, since the sub-tile is 8
2006        * pixels wide and 4 pixels high (when viewed as a W-tiled stencil
2007        * buffer), and the miplevel alignment used for stencil buffers is 8
2008        * pixels horizontally and either 4 or 8 pixels vertically (see
2009        * intel_horizontal_texture_alignment_unit() and
2010        * intel_vertical_texture_alignment_unit()).
2011        *
2012        * Note: Also, since the SURFACE_STATE command packet can only apply
2013        * offsets that are multiples of 4 pixels horizontally and 2 pixels
2014        * vertically, it is important that the offsets will be multiples of
2015        * these sizes after they are converted into Y-tiled coordinates.
2016        * Fortunately they will be, since we know from above that the offsets
2017        * are a multiple of the 32-byte sub-tile size, and in Y-tiled
2018        * coordinates the sub-tile is 16 pixels wide and 2 pixels high.
2019        *
2020        * TODO: what if this makes the coordinates (or the texture size) too
2021        * large?
2022        */
2023       const unsigned x_align = 8;
2024       const unsigned y_align = params->dst.surf.samples != 0 ? 8 : 4;
2025       params->x0 = ROUND_DOWN_TO(params->x0, x_align) * 2;
2026       params->y0 = ROUND_DOWN_TO(params->y0, y_align) / 2;
2027       params->x1 = ALIGN(params->x1, x_align) * 2;
2028       params->y1 = ALIGN(params->y1, y_align) / 2;
2029 
2030       /* Retile the surface to Y-tiled */
2031       blorp_surf_retile_w_to_y(batch->blorp->isl_dev, &params->dst);
2032 
2033       key->dst_tiled_w = true;
2034       key->use_kill = true;
2035       key->need_dst_offset = true;
2036 
2037       if (params->dst.surf.samples > 1) {
2038          /* If the destination surface is a W-tiled multisampled stencil
2039           * buffer that we're mapping as Y tiled, then we need to arrange for
2040           * the WM program to run once per sample rather than once per pixel,
2041           * because the memory layout of related samples doesn't match between
2042           * W and Y tiling.
2043           */
2044          key->persample_msaa_dispatch = true;
2045       }
2046    }
2047 
2048    if (devinfo->ver < 8 && params->src.surf.tiling == ISL_TILING_W) {
2049       /* On Haswell and earlier, we have to fake W-tiled sources as Y-tiled.
2050        * Broadwell adds support for sampling from stencil.
2051        *
2052        * See the comments above concerning x/y offset alignment for the
2053        * destination surface.
2054        *
2055        * TODO: what if this makes the texture size too large?
2056        */
2057       blorp_surf_retile_w_to_y(batch->blorp->isl_dev, &params->src);
2058 
2059       key->src_tiled_w = true;
2060       key->need_src_offset = true;
2061    }
2062 
2063    /* tex_samples and rt_samples are the sample counts that are set up in
2064     * SURFACE_STATE.
2065     */
2066    key->tex_samples = params->src.surf.samples;
2067    key->rt_samples  = params->dst.surf.samples;
2068 
2069    /* tex_layout and rt_layout indicate the MSAA layout the GPU pipeline will
2070     * use to access the source and destination surfaces.
2071     */
2072    key->tex_layout = params->src.surf.msaa_layout;
2073    key->rt_layout = params->dst.surf.msaa_layout;
2074 
2075    if (params->src.surf.samples > 0 && params->dst.surf.samples > 1) {
2076       /* We are blitting from a multisample buffer to a multisample buffer, so
2077        * we must preserve samples within a pixel.  This means we have to
2078        * arrange for the WM program to run once per sample rather than once
2079        * per pixel.
2080        */
2081       key->persample_msaa_dispatch = true;
2082    }
2083 
2084    params->num_samples = params->dst.surf.samples;
2085 
2086    if ((key->filter == BLORP_FILTER_AVERAGE ||
2087         key->filter == BLORP_FILTER_BILINEAR) &&
2088        batch->blorp->isl_dev->info->ver <= 6) {
2089       /* Gfx4-5 don't support non-normalized texture coordinates */
2090       key->src_coords_normalized = true;
2091       params->wm_inputs.src_inv_size[0] =
2092          1.0f / u_minify(params->src.surf.logical_level0_px.width,
2093                          params->src.view.base_level);
2094       params->wm_inputs.src_inv_size[1] =
2095          1.0f / u_minify(params->src.surf.logical_level0_px.height,
2096                          params->src.view.base_level);
2097    }
2098 
2099    if (isl_format_get_layout(params->dst.view.format)->bpb % 3 == 0) {
2100       /* We can't render to  RGB formats natively because they aren't a
2101        * power-of-two size.  Instead, we fake them by using a red format
2102        * with the same channel type and size and emitting shader code to
2103        * only write one channel at a time.
2104        */
2105       params->x0 *= 3;
2106       params->x1 *= 3;
2107 
2108       /* If it happens to be sRGB, we need to force a conversion */
2109       if (params->dst.view.format == ISL_FORMAT_R8G8B8_UNORM_SRGB)
2110          key->dst_format = ISL_FORMAT_R8G8B8_UNORM_SRGB;
2111 
2112       surf_fake_rgb_with_red(batch->blorp->isl_dev, &params->dst);
2113 
2114       key->dst_rgb = true;
2115       key->need_dst_offset = true;
2116    } else if (isl_format_is_rgbx(params->dst.view.format)) {
2117       /* We can handle RGBX formats easily enough by treating them as RGBA */
2118       params->dst.view.format =
2119          isl_format_rgbx_to_rgba(params->dst.view.format);
2120    } else if (params->dst.view.format == ISL_FORMAT_R24_UNORM_X8_TYPELESS &&
2121               key->dst_usage != ISL_SURF_USAGE_DEPTH_BIT) {
2122       key->dst_format = params->dst.view.format;
2123       params->dst.view.format = ISL_FORMAT_R32_UINT;
2124    } else if (params->dst.view.format == ISL_FORMAT_A4B4G4R4_UNORM) {
2125       params->dst.view.swizzle =
2126          isl_swizzle_compose(params->dst.view.swizzle,
2127                              ISL_SWIZZLE(ALPHA, RED, GREEN, BLUE));
2128       params->dst.view.format = ISL_FORMAT_B4G4R4A4_UNORM;
2129    } else if (params->dst.view.format == ISL_FORMAT_L8_UNORM_SRGB) {
2130       key->dst_format = params->dst.view.format;
2131       params->dst.view.format = ISL_FORMAT_R8_UNORM;
2132    } else if (params->dst.view.format == ISL_FORMAT_R9G9B9E5_SHAREDEXP) {
2133       key->dst_format = params->dst.view.format;
2134       params->dst.view.format = ISL_FORMAT_R32_UINT;
2135    }
2136 
2137    if (devinfo->verx10 <= 70 &&
2138        !isl_swizzle_is_identity(params->src.view.swizzle)) {
2139       key->src_swizzle = params->src.view.swizzle;
2140       params->src.view.swizzle = ISL_SWIZZLE_IDENTITY;
2141    } else {
2142       key->src_swizzle = ISL_SWIZZLE_IDENTITY;
2143    }
2144 
2145    if (!isl_swizzle_supports_rendering(devinfo, params->dst.view.swizzle)) {
2146       key->dst_swizzle = params->dst.view.swizzle;
2147       params->dst.view.swizzle = ISL_SWIZZLE_IDENTITY;
2148    } else {
2149       key->dst_swizzle = ISL_SWIZZLE_IDENTITY;
2150    }
2151 
2152    if (params->src.tile_x_sa || params->src.tile_y_sa) {
2153       assert(key->need_src_offset);
2154       surf_get_intratile_offset_px(&params->src,
2155                                    &params->wm_inputs.src_offset.x,
2156                                    &params->wm_inputs.src_offset.y);
2157    }
2158 
2159    if (params->dst.tile_x_sa || params->dst.tile_y_sa) {
2160       assert(key->need_dst_offset);
2161       surf_get_intratile_offset_px(&params->dst,
2162                                    &params->wm_inputs.dst_offset.x,
2163                                    &params->wm_inputs.dst_offset.y);
2164       params->x0 += params->wm_inputs.dst_offset.x;
2165       params->y0 += params->wm_inputs.dst_offset.y;
2166       params->x1 += params->wm_inputs.dst_offset.x;
2167       params->y1 += params->wm_inputs.dst_offset.y;
2168    }
2169 
2170    /* For some texture types, we need to pass the layer through the sampler. */
2171    params->wm_inputs.src_z = params->src.z_offset;
2172 
2173    const bool compute =
2174       key->base.shader_pipeline == BLORP_SHADER_PIPELINE_COMPUTE;
2175    if (compute) {
2176       key->local_y = blorp_get_cs_local_y(params);
2177 
2178       unsigned workgroup_width = 16 / key->local_y;
2179       unsigned workgroup_height = key->local_y;
2180 
2181       /* If the rectangle being drawn isn't an exact multiple of the
2182        * workgroup size, we'll get extra invocations that should not
2183        * perform blits.  We need to set use_kill to bounds check and
2184        * prevent those invocations from blitting.
2185        */
2186       if ((params->x0 % workgroup_width) != 0 ||
2187           (params->x1 % workgroup_width) != 0 ||
2188           (params->y0 % workgroup_height) != 0 ||
2189           (params->y1 % workgroup_height) != 0)
2190          key->use_kill = true;
2191    }
2192 
2193    if (compute) {
2194       if (!blorp_get_blit_kernel_cs(batch, params, key))
2195          return 0;
2196    } else {
2197       if (!blorp_get_blit_kernel_fs(batch, params, key))
2198          return 0;
2199 
2200       if (!blorp_ensure_sf_program(batch, params))
2201          return 0;
2202    }
2203 
2204    unsigned result = 0;
2205    unsigned max_src_surface_size = get_max_surface_size(devinfo, &params->src);
2206    if (params->src.surf.logical_level0_px.width > max_src_surface_size)
2207       result |= BLIT_SRC_WIDTH_SHRINK;
2208    if (params->src.surf.logical_level0_px.height > max_src_surface_size)
2209       result |= BLIT_SRC_HEIGHT_SHRINK;
2210 
2211    unsigned max_dst_surface_size = get_max_surface_size(devinfo, &params->dst);
2212    if (params->dst.surf.logical_level0_px.width > max_dst_surface_size)
2213       result |= BLIT_DST_WIDTH_SHRINK;
2214    if (params->dst.surf.logical_level0_px.height > max_dst_surface_size)
2215       result |= BLIT_DST_HEIGHT_SHRINK;
2216 
2217    if (result == 0) {
2218       if (key->dst_usage == ISL_SURF_USAGE_DEPTH_BIT) {
2219          params->depth = params->dst;
2220          memset(&params->dst, 0, sizeof(params->dst));
2221       } else if (key->dst_usage == ISL_SURF_USAGE_STENCIL_BIT) {
2222          params->stencil = params->dst;
2223          params->stencil_mask = 0xff;
2224          memset(&params->dst, 0, sizeof(params->dst));
2225       }
2226 
2227       batch->blorp->exec(batch, params);
2228    }
2229 
2230    return result;
2231 }
2232 
2233 /* Adjust split blit source coordinates for the current destination
2234  * coordinates.
2235  */
2236 static void
adjust_split_source_coords(const struct blt_axis * orig,struct blt_axis * split_coords,double scale)2237 adjust_split_source_coords(const struct blt_axis *orig,
2238                            struct blt_axis *split_coords,
2239                            double scale)
2240 {
2241    /* When scale is greater than 0, then we are growing from the start, so
2242     * src0 uses delta0, and src1 uses delta1. When scale is less than 0, the
2243     * source range shrinks from the end. In that case src0 is adjusted by
2244     * delta1, and src1 is adjusted by delta0.
2245     */
2246    double delta0 = scale * (split_coords->dst0 - orig->dst0);
2247    double delta1 = scale * (split_coords->dst1 - orig->dst1);
2248    split_coords->src0 = orig->src0 + (scale >= 0.0 ? delta0 : delta1);
2249    split_coords->src1 = orig->src1 + (scale >= 0.0 ? delta1 : delta0);
2250 }
2251 
2252 static struct isl_extent2d
get_px_size_sa(const struct isl_surf * surf)2253 get_px_size_sa(const struct isl_surf *surf)
2254 {
2255    static const struct isl_extent2d one_to_one = { .w = 1, .h = 1 };
2256 
2257    if (surf->msaa_layout != ISL_MSAA_LAYOUT_INTERLEAVED)
2258       return one_to_one;
2259    else
2260       return isl_get_interleaved_msaa_px_size_sa(surf->samples);
2261 }
2262 
2263 static void
shrink_surface_params(const struct isl_device * dev,struct blorp_surface_info * info,double * x0,double * x1,double * y0,double * y1)2264 shrink_surface_params(const struct isl_device *dev,
2265                       struct blorp_surface_info *info,
2266                       double *x0, double *x1, double *y0, double *y1)
2267 {
2268    uint64_t offset_B;
2269    uint32_t x_offset_sa, y_offset_sa, size;
2270    struct isl_extent2d px_size_sa;
2271    int adjust;
2272 
2273    blorp_surf_convert_to_single_slice(dev, info);
2274 
2275    px_size_sa = get_px_size_sa(&info->surf);
2276 
2277    /* Because this gets called after we lower compressed images, the tile
2278     * offsets may be non-zero and we need to incorporate them in our
2279     * calculations.
2280     */
2281    x_offset_sa = (uint32_t)*x0 * px_size_sa.w + info->tile_x_sa;
2282    y_offset_sa = (uint32_t)*y0 * px_size_sa.h + info->tile_y_sa;
2283    uint32_t tile_z_sa, tile_a;
2284    isl_tiling_get_intratile_offset_sa(info->surf.tiling, info->surf.dim,
2285                                       info->surf.msaa_layout,
2286                                       info->surf.format, info->surf.samples,
2287                                       info->surf.row_pitch_B,
2288                                       info->surf.array_pitch_el_rows,
2289                                       x_offset_sa, y_offset_sa, 0, 0,
2290                                       &offset_B,
2291                                       &info->tile_x_sa, &info->tile_y_sa,
2292                                       &tile_z_sa, &tile_a);
2293    assert(tile_z_sa == 0 && tile_a == 0);
2294 
2295    info->addr.offset += offset_B;
2296 
2297    adjust = (int)info->tile_x_sa / px_size_sa.w - (int)*x0;
2298    *x0 += adjust;
2299    *x1 += adjust;
2300    info->tile_x_sa = 0;
2301 
2302    adjust = (int)info->tile_y_sa / px_size_sa.h - (int)*y0;
2303    *y0 += adjust;
2304    *y1 += adjust;
2305    info->tile_y_sa = 0;
2306 
2307    size = MIN2((uint32_t)ceil(*x1), info->surf.logical_level0_px.width);
2308    info->surf.logical_level0_px.width = size;
2309    info->surf.phys_level0_sa.width = size * px_size_sa.w;
2310 
2311    size = MIN2((uint32_t)ceil(*y1), info->surf.logical_level0_px.height);
2312    info->surf.logical_level0_px.height = size;
2313    info->surf.phys_level0_sa.height = size * px_size_sa.h;
2314 }
2315 
2316 static void
do_blorp_blit(struct blorp_batch * batch,const struct blorp_params * orig_params,struct blorp_blit_prog_key * key,const struct blt_coords * orig)2317 do_blorp_blit(struct blorp_batch *batch,
2318               const struct blorp_params *orig_params,
2319               struct blorp_blit_prog_key *key,
2320               const struct blt_coords *orig)
2321 {
2322    struct blorp_params params;
2323    struct blt_coords blit_coords;
2324    struct blt_coords split_coords = *orig;
2325    double w = orig->x.dst1 - orig->x.dst0;
2326    double h = orig->y.dst1 - orig->y.dst0;
2327    double x_scale = (orig->x.src1 - orig->x.src0) / w;
2328    double y_scale = (orig->y.src1 - orig->y.src0) / h;
2329    if (orig->x.mirror)
2330       x_scale = -x_scale;
2331    if (orig->y.mirror)
2332       y_scale = -y_scale;
2333 
2334    enum blit_shrink_status shrink = BLIT_NO_SHRINK;
2335    if (split_blorp_blit_debug) {
2336       if (can_shrink_surface(&orig_params->src))
2337          shrink |= BLIT_SRC_WIDTH_SHRINK | BLIT_SRC_HEIGHT_SHRINK;
2338       if (can_shrink_surface(&orig_params->dst))
2339          shrink |= BLIT_DST_WIDTH_SHRINK | BLIT_DST_HEIGHT_SHRINK;
2340    }
2341 
2342    bool x_done, y_done;
2343    do {
2344       params = *orig_params;
2345       blit_coords = split_coords;
2346 
2347       if (shrink & (BLIT_SRC_WIDTH_SHRINK | BLIT_SRC_HEIGHT_SHRINK)) {
2348          shrink_surface_params(batch->blorp->isl_dev, &params.src,
2349                                &blit_coords.x.src0, &blit_coords.x.src1,
2350                                &blit_coords.y.src0, &blit_coords.y.src1);
2351          key->need_src_offset = false;
2352       }
2353 
2354       if (shrink & (BLIT_DST_WIDTH_SHRINK | BLIT_DST_HEIGHT_SHRINK)) {
2355          shrink_surface_params(batch->blorp->isl_dev, &params.dst,
2356                                &blit_coords.x.dst0, &blit_coords.x.dst1,
2357                                &blit_coords.y.dst0, &blit_coords.y.dst1);
2358          key->need_dst_offset = false;
2359       }
2360 
2361       enum blit_shrink_status result =
2362          try_blorp_blit(batch, &params, key, &blit_coords);
2363 
2364       if (result & (BLIT_SRC_WIDTH_SHRINK | BLIT_SRC_HEIGHT_SHRINK))
2365          assert(can_shrink_surface(&orig_params->src));
2366 
2367       if (result & (BLIT_DST_WIDTH_SHRINK | BLIT_DST_HEIGHT_SHRINK))
2368          assert(can_shrink_surface(&orig_params->dst));
2369 
2370       if (result & (BLIT_SRC_WIDTH_SHRINK | BLIT_DST_WIDTH_SHRINK)) {
2371          w /= 2.0;
2372          assert(w >= 1.0);
2373          split_coords.x.dst1 = MIN2(split_coords.x.dst0 + w, orig->x.dst1);
2374          adjust_split_source_coords(&orig->x, &split_coords.x, x_scale);
2375       }
2376       if (result & (BLIT_SRC_HEIGHT_SHRINK | BLIT_DST_HEIGHT_SHRINK)) {
2377          h /= 2.0;
2378          assert(h >= 1.0);
2379          split_coords.y.dst1 = MIN2(split_coords.y.dst0 + h, orig->y.dst1);
2380          adjust_split_source_coords(&orig->y, &split_coords.y, y_scale);
2381       }
2382 
2383       if (result) {
2384          /* We may get less bits set on result than we had already, so make
2385           * sure we remember all the ways in which a resize is required.
2386           */
2387          shrink |= result;
2388          continue;
2389       }
2390 
2391       y_done = (orig->y.dst1 - split_coords.y.dst1 < 0.5);
2392       x_done = y_done && (orig->x.dst1 - split_coords.x.dst1 < 0.5);
2393       if (x_done) {
2394          break;
2395       } else if (y_done) {
2396          split_coords.x.dst0 += w;
2397          split_coords.x.dst1 = MIN2(split_coords.x.dst0 + w, orig->x.dst1);
2398          split_coords.y.dst0 = orig->y.dst0;
2399          split_coords.y.dst1 = MIN2(split_coords.y.dst0 + h, orig->y.dst1);
2400          adjust_split_source_coords(&orig->x, &split_coords.x, x_scale);
2401       } else {
2402          split_coords.y.dst0 += h;
2403          split_coords.y.dst1 = MIN2(split_coords.y.dst0 + h, orig->y.dst1);
2404          adjust_split_source_coords(&orig->y, &split_coords.y, y_scale);
2405       }
2406    } while (true);
2407 }
2408 
2409 bool
blorp_blit_supports_compute(struct blorp_context * blorp,const struct isl_surf * src_surf,const struct isl_surf * dst_surf,enum isl_aux_usage dst_aux_usage)2410 blorp_blit_supports_compute(struct blorp_context *blorp,
2411                             const struct isl_surf *src_surf,
2412                             const struct isl_surf *dst_surf,
2413                             enum isl_aux_usage dst_aux_usage)
2414 {
2415    /* Our compiler doesn't currently support typed image writes with MSAA.
2416     * Also, our BLORP compute shaders don't handle multisampling cases.
2417     */
2418    if (dst_surf->samples > 1 || src_surf->samples > 1)
2419       return false;
2420 
2421    if (blorp->isl_dev->info->ver >= 12) {
2422       return dst_aux_usage == ISL_AUX_USAGE_FCV_CCS_E ||
2423              dst_aux_usage == ISL_AUX_USAGE_CCS_E ||
2424              dst_aux_usage == ISL_AUX_USAGE_NONE;
2425    } else if (blorp->isl_dev->info->ver >= 7) {
2426       return dst_aux_usage == ISL_AUX_USAGE_NONE;
2427    } else {
2428       /* No compute shader support */
2429       return false;
2430    }
2431 }
2432 
2433 bool
blorp_blitter_supports_aux(const struct intel_device_info * devinfo,enum isl_aux_usage aux_usage)2434 blorp_blitter_supports_aux(const struct intel_device_info *devinfo,
2435                            enum isl_aux_usage aux_usage)
2436 {
2437    switch (aux_usage) {
2438    case ISL_AUX_USAGE_NONE:
2439       return true;
2440    case ISL_AUX_USAGE_CCS_E:
2441    case ISL_AUX_USAGE_FCV_CCS_E:
2442    case ISL_AUX_USAGE_STC_CCS:
2443       return devinfo->verx10 >= 125;
2444    default:
2445       return false;
2446    }
2447 }
2448 
2449 bool
blorp_copy_supports_blitter(struct blorp_context * blorp,const struct isl_surf * src_surf,const struct isl_surf * dst_surf,enum isl_aux_usage src_aux_usage,enum isl_aux_usage dst_aux_usage)2450 blorp_copy_supports_blitter(struct blorp_context *blorp,
2451                             const struct isl_surf *src_surf,
2452                             const struct isl_surf *dst_surf,
2453                             enum isl_aux_usage src_aux_usage,
2454                             enum isl_aux_usage dst_aux_usage)
2455 {
2456    const struct intel_device_info *devinfo = blorp->isl_dev->info;
2457 
2458    if (devinfo->ver < 12)
2459       return false;
2460 
2461    if (dst_surf->samples > 1 || src_surf->samples > 1)
2462       return false;
2463 
2464    if (!blorp_blitter_supports_aux(devinfo, dst_aux_usage))
2465       return false;
2466 
2467    if (!blorp_blitter_supports_aux(devinfo, src_aux_usage))
2468       return false;
2469 
2470    const struct isl_format_layout *fmtl =
2471       isl_format_get_layout(dst_surf->format);
2472 
2473    if (fmtl->bpb == 96) {
2474       /* XY_BLOCK_COPY_BLT mentions it doesn't support clear colors for 96bpp
2475        * formats, but none of them support CCS anyway, so it's a moot point.
2476        */
2477       assert(src_aux_usage == ISL_AUX_USAGE_NONE);
2478       assert(dst_aux_usage == ISL_AUX_USAGE_NONE);
2479 
2480       /* We can only support linear mode for 96bpp. */
2481       if (src_surf->tiling != ISL_TILING_LINEAR ||
2482           dst_surf->tiling != ISL_TILING_LINEAR)
2483          return false;
2484    }
2485 
2486    return true;
2487 }
2488 
2489 void
blorp_blit(struct blorp_batch * batch,const struct blorp_surf * src_surf,unsigned src_level,float src_layer,enum isl_format src_format,struct isl_swizzle src_swizzle,const struct blorp_surf * dst_surf,unsigned dst_level,unsigned dst_layer,enum isl_format dst_format,struct isl_swizzle dst_swizzle,float src_x0,float src_y0,float src_x1,float src_y1,float dst_x0,float dst_y0,float dst_x1,float dst_y1,enum blorp_filter filter,bool mirror_x,bool mirror_y)2490 blorp_blit(struct blorp_batch *batch,
2491            const struct blorp_surf *src_surf,
2492            unsigned src_level, float src_layer,
2493            enum isl_format src_format, struct isl_swizzle src_swizzle,
2494            const struct blorp_surf *dst_surf,
2495            unsigned dst_level, unsigned dst_layer,
2496            enum isl_format dst_format, struct isl_swizzle dst_swizzle,
2497            float src_x0, float src_y0,
2498            float src_x1, float src_y1,
2499            float dst_x0, float dst_y0,
2500            float dst_x1, float dst_y1,
2501            enum blorp_filter filter,
2502            bool mirror_x, bool mirror_y)
2503 {
2504    struct blorp_params params;
2505    blorp_params_init(&params);
2506    params.op = BLORP_OP_BLIT;
2507    const bool compute = batch->flags & BLORP_BATCH_USE_COMPUTE;
2508    if (compute) {
2509       assert(blorp_blit_supports_compute(batch->blorp,
2510                                          src_surf->surf, dst_surf->surf,
2511                                          dst_surf->aux_usage));
2512    }
2513 
2514    /* We cannot handle combined depth and stencil. */
2515    if (src_surf->surf->usage & ISL_SURF_USAGE_STENCIL_BIT)
2516       assert(src_surf->surf->format == ISL_FORMAT_R8_UINT);
2517    if (dst_surf->surf->usage & ISL_SURF_USAGE_STENCIL_BIT)
2518       assert(dst_surf->surf->format == ISL_FORMAT_R8_UINT);
2519 
2520    if (dst_surf->surf->usage & ISL_SURF_USAGE_STENCIL_BIT) {
2521       assert(src_surf->surf->usage & ISL_SURF_USAGE_STENCIL_BIT);
2522       /* Prior to Broadwell, we can't render to R8_UINT */
2523       if (batch->blorp->isl_dev->info->ver < 8) {
2524          src_format = ISL_FORMAT_R8_UNORM;
2525          dst_format = ISL_FORMAT_R8_UNORM;
2526       }
2527    }
2528 
2529    blorp_surface_info_init(batch, &params.src, src_surf, src_level,
2530                                src_layer, src_format, false);
2531    blorp_surface_info_init(batch, &params.dst, dst_surf, dst_level,
2532                                dst_layer, dst_format, true);
2533 
2534    params.src.view.swizzle = src_swizzle;
2535    params.dst.view.swizzle = dst_swizzle;
2536 
2537    const struct isl_format_layout *src_fmtl =
2538       isl_format_get_layout(params.src.view.format);
2539 
2540    struct blorp_blit_prog_key key = {
2541       .base = BLORP_BASE_KEY_INIT(BLORP_SHADER_TYPE_BLIT),
2542       .base.shader_pipeline = compute ? BLORP_SHADER_PIPELINE_COMPUTE :
2543                                         BLORP_SHADER_PIPELINE_RENDER,
2544       .filter = filter,
2545       .sint32_to_uint = src_fmtl->channels.r.bits == 32 &&
2546                         isl_format_has_sint_channel(params.src.view.format) &&
2547                         isl_format_has_uint_channel(params.dst.view.format),
2548       .uint32_to_sint = src_fmtl->channels.r.bits == 32 &&
2549                         isl_format_has_uint_channel(params.src.view.format) &&
2550                         isl_format_has_sint_channel(params.dst.view.format),
2551    };
2552 
2553    params.shader_type = key.base.shader_type;
2554    params.shader_pipeline = key.base.shader_pipeline;
2555 
2556    /* Scaling factors used for bilinear filtering in multisample scaled
2557     * blits.
2558     */
2559    if (params.src.surf.samples == 16)
2560       key.x_scale = 4.0f;
2561    else
2562       key.x_scale = 2.0f;
2563    key.y_scale = params.src.surf.samples / key.x_scale;
2564 
2565    params.wm_inputs.rect_grid.x1 =
2566       u_minify(params.src.surf.logical_level0_px.width, src_level) *
2567       key.x_scale - 1.0f;
2568    params.wm_inputs.rect_grid.y1 =
2569       u_minify(params.src.surf.logical_level0_px.height, src_level) *
2570       key.y_scale - 1.0f;
2571 
2572    struct blt_coords coords = {
2573       .x = {
2574          .src0 = src_x0,
2575          .src1 = src_x1,
2576          .dst0 = dst_x0,
2577          .dst1 = dst_x1,
2578          .mirror = mirror_x
2579       },
2580       .y = {
2581          .src0 = src_y0,
2582          .src1 = src_y1,
2583          .dst0 = dst_y0,
2584          .dst1 = dst_y1,
2585          .mirror = mirror_y
2586       }
2587    };
2588 
2589    do_blorp_blit(batch, &params, &key, &coords);
2590 }
2591 
2592 static enum isl_format
get_copy_format_for_bpb(const struct isl_device * isl_dev,unsigned bpb)2593 get_copy_format_for_bpb(const struct isl_device *isl_dev, unsigned bpb)
2594 {
2595    /* The choice of UNORM and UINT formats is very intentional here.  Most
2596     * of the time, we want to use a UINT format to avoid any rounding error
2597     * in the blit.  For stencil blits, R8_UINT is required by the hardware.
2598     * (It's the only format allowed in conjunction with W-tiling.)  Also we
2599     * intentionally use the 4-channel formats whenever we can.  This is so
2600     * that, when we do a RGB <-> RGBX copy, the two formats will line up
2601     * even though one of them is 3/4 the size of the other.  The choice of
2602     * UNORM vs. UINT is also very intentional because we don't have 8 or
2603     * 16-bit RGB UINT formats until Sky Lake so we have to use UNORM there.
2604     * Fortunately, the only time we should ever use two different formats in
2605     * the table below is for RGB -> RGBA blits and so we will never have any
2606     * UNORM/UINT mismatch.
2607     */
2608    if (ISL_GFX_VER(isl_dev) >= 9) {
2609       switch (bpb) {
2610       case 8:  return ISL_FORMAT_R8_UINT;
2611       case 16: return ISL_FORMAT_R8G8_UINT;
2612       case 24: return ISL_FORMAT_R8G8B8_UINT;
2613       case 32: return ISL_FORMAT_R8G8B8A8_UINT;
2614       case 48: return ISL_FORMAT_R16G16B16_UINT;
2615       case 64: return ISL_FORMAT_R16G16B16A16_UINT;
2616       case 96: return ISL_FORMAT_R32G32B32_UINT;
2617       case 128:return ISL_FORMAT_R32G32B32A32_UINT;
2618       default:
2619          unreachable("Unknown format bpb");
2620       }
2621    } else {
2622       switch (bpb) {
2623       case 8:  return ISL_FORMAT_R8_UINT;
2624       case 16: return ISL_FORMAT_R8G8_UINT;
2625       case 24: return ISL_FORMAT_R8G8B8_UNORM;
2626       case 32: return ISL_FORMAT_R8G8B8A8_UNORM;
2627       case 48: return ISL_FORMAT_R16G16B16_UNORM;
2628       case 64: return ISL_FORMAT_R16G16B16A16_UNORM;
2629       case 96: return ISL_FORMAT_R32G32B32_UINT;
2630       case 128:return ISL_FORMAT_R32G32B32A32_UINT;
2631       default:
2632          unreachable("Unknown format bpb");
2633       }
2634    }
2635 }
2636 
2637 /** Returns a UINT format that is CCS-compatible with the given format
2638  *
2639  * The PRM's say absolutely nothing about how render compression works.  The
2640  * only thing they provide is a list of formats on which it is and is not
2641  * supported.  Empirical testing indicates that the compression is only based
2642  * on the bit-layout of the format and the channel encoding doesn't matter.
2643  * So, while texture views don't work in general, you can create a view as
2644  * long as the bit-layout of the formats are the same.
2645  *
2646  * Fortunately, for every render compression capable format, the UINT format
2647  * with the same bit layout also supports render compression.  This means that
2648  * we only need to handle UINT formats for copy operations.  In order to do
2649  * copies between formats with different bit layouts, we attach both with a
2650  * UINT format and use bit_cast_color() to generate code to do the bit-cast
2651  * operation between the two bit layouts.
2652  */
2653 static enum isl_format
get_ccs_compatible_copy_format(const struct isl_format_layout * fmtl)2654 get_ccs_compatible_copy_format(const struct isl_format_layout *fmtl)
2655 {
2656    switch (fmtl->format) {
2657    case ISL_FORMAT_R32G32B32A32_FLOAT:
2658    case ISL_FORMAT_R32G32B32A32_SINT:
2659    case ISL_FORMAT_R32G32B32A32_UINT:
2660    case ISL_FORMAT_R32G32B32A32_UNORM:
2661    case ISL_FORMAT_R32G32B32A32_SNORM:
2662    case ISL_FORMAT_R32G32B32X32_FLOAT:
2663       return ISL_FORMAT_R32G32B32A32_UINT;
2664 
2665    case ISL_FORMAT_R16G16B16A16_UNORM:
2666    case ISL_FORMAT_R16G16B16A16_SNORM:
2667    case ISL_FORMAT_R16G16B16A16_SINT:
2668    case ISL_FORMAT_R16G16B16A16_UINT:
2669    case ISL_FORMAT_R16G16B16A16_FLOAT:
2670    case ISL_FORMAT_R16G16B16X16_UNORM:
2671    case ISL_FORMAT_R16G16B16X16_FLOAT:
2672       return ISL_FORMAT_R16G16B16A16_UINT;
2673 
2674    case ISL_FORMAT_R32G32_FLOAT:
2675    case ISL_FORMAT_R32G32_SINT:
2676    case ISL_FORMAT_R32G32_UINT:
2677    case ISL_FORMAT_R32G32_UNORM:
2678    case ISL_FORMAT_R32G32_SNORM:
2679       return ISL_FORMAT_R32G32_UINT;
2680 
2681    case ISL_FORMAT_B8G8R8A8_UNORM:
2682    case ISL_FORMAT_B8G8R8A8_UNORM_SRGB:
2683    case ISL_FORMAT_R8G8B8A8_UNORM:
2684    case ISL_FORMAT_R8G8B8A8_UNORM_SRGB:
2685    case ISL_FORMAT_R8G8B8A8_SNORM:
2686    case ISL_FORMAT_R8G8B8A8_SINT:
2687    case ISL_FORMAT_R8G8B8A8_UINT:
2688    case ISL_FORMAT_B8G8R8X8_UNORM:
2689    case ISL_FORMAT_B8G8R8X8_UNORM_SRGB:
2690    case ISL_FORMAT_R8G8B8X8_UNORM:
2691    case ISL_FORMAT_R8G8B8X8_UNORM_SRGB:
2692       return ISL_FORMAT_R8G8B8A8_UINT;
2693 
2694    case ISL_FORMAT_R16G16_UNORM:
2695    case ISL_FORMAT_R16G16_SNORM:
2696    case ISL_FORMAT_R16G16_SINT:
2697    case ISL_FORMAT_R16G16_UINT:
2698    case ISL_FORMAT_R16G16_FLOAT:
2699       return ISL_FORMAT_R16G16_UINT;
2700 
2701    case ISL_FORMAT_R32_SINT:
2702    case ISL_FORMAT_R32_UINT:
2703    case ISL_FORMAT_R32_FLOAT:
2704    case ISL_FORMAT_R32_UNORM:
2705    case ISL_FORMAT_R32_SNORM:
2706       return ISL_FORMAT_R32_UINT;
2707 
2708    case ISL_FORMAT_R11G11B10_FLOAT:
2709       return ISL_FORMAT_R8G8B8A8_UINT;
2710 
2711    case ISL_FORMAT_B10G10R10A2_UNORM:
2712    case ISL_FORMAT_B10G10R10A2_UNORM_SRGB:
2713    case ISL_FORMAT_R10G10B10A2_UNORM:
2714    case ISL_FORMAT_R10G10B10A2_UNORM_SRGB:
2715    case ISL_FORMAT_R10G10B10_FLOAT_A2_UNORM:
2716    case ISL_FORMAT_R10G10B10A2_UINT:
2717       return ISL_FORMAT_R10G10B10A2_UINT;
2718 
2719    case ISL_FORMAT_R16_UNORM:
2720    case ISL_FORMAT_R16_SNORM:
2721    case ISL_FORMAT_R16_SINT:
2722    case ISL_FORMAT_R16_UINT:
2723    case ISL_FORMAT_R16_FLOAT:
2724       return ISL_FORMAT_R16_UINT;
2725 
2726    case ISL_FORMAT_R8G8_UNORM:
2727    case ISL_FORMAT_R8G8_SNORM:
2728    case ISL_FORMAT_R8G8_SINT:
2729    case ISL_FORMAT_R8G8_UINT:
2730       return ISL_FORMAT_R8G8_UINT;
2731 
2732    case ISL_FORMAT_YCRCB_NORMAL:
2733    case ISL_FORMAT_YCRCB_SWAPY:
2734    case ISL_FORMAT_YCRCB_SWAPUV:
2735    case ISL_FORMAT_YCRCB_SWAPUVY:
2736       /* Tiger Lake starts claiming CCS_E support for certain YCRCB formats.
2737        * BLORP chooses to take the CCS-compatible format path whenever ISL
2738        * claims CCS_E support on a format, not when CCS_E is actually used.
2739        * Therefore, if these formats are going to be used with BLORP, we need
2740        * a CCS-compatible format. R8G8_UINT seems as good as any.
2741        */
2742       return ISL_FORMAT_R8G8_UINT;
2743 
2744    case ISL_FORMAT_B5G5R5X1_UNORM:
2745    case ISL_FORMAT_B5G5R5X1_UNORM_SRGB:
2746    case ISL_FORMAT_B5G5R5A1_UNORM:
2747    case ISL_FORMAT_B5G5R5A1_UNORM_SRGB:
2748       return ISL_FORMAT_B5G5R5A1_UNORM;
2749 
2750    case ISL_FORMAT_A4B4G4R4_UNORM:
2751    case ISL_FORMAT_B4G4R4A4_UNORM:
2752    case ISL_FORMAT_B4G4R4A4_UNORM_SRGB:
2753       return ISL_FORMAT_B4G4R4A4_UNORM;
2754 
2755    case ISL_FORMAT_B5G6R5_UNORM:
2756    case ISL_FORMAT_B5G6R5_UNORM_SRGB:
2757       return ISL_FORMAT_B5G6R5_UNORM;
2758 
2759    case ISL_FORMAT_A1B5G5R5_UNORM:
2760       return ISL_FORMAT_A1B5G5R5_UNORM;
2761 
2762    case ISL_FORMAT_A8_UNORM:
2763    case ISL_FORMAT_R8_UNORM:
2764    case ISL_FORMAT_R8_SNORM:
2765    case ISL_FORMAT_R8_SINT:
2766    case ISL_FORMAT_R8_UINT:
2767       return ISL_FORMAT_R8_UINT;
2768 
2769    default:
2770       unreachable("Not a compressible format");
2771    }
2772 }
2773 
2774 void
blorp_surf_convert_to_uncompressed(const struct isl_device * isl_dev,struct blorp_surface_info * info,uint32_t * x,uint32_t * y,uint32_t * width,uint32_t * height)2775 blorp_surf_convert_to_uncompressed(const struct isl_device *isl_dev,
2776                                    struct blorp_surface_info *info,
2777                                    uint32_t *x, uint32_t *y,
2778                                    uint32_t *width, uint32_t *height)
2779 {
2780    const struct isl_format_layout *fmtl =
2781       isl_format_get_layout(info->surf.format);
2782 
2783    assert(fmtl->bw > 1 || fmtl->bh > 1);
2784 
2785    /* This should be the first modification made to the surface */
2786    assert(info->tile_x_sa == 0 && info->tile_y_sa == 0);
2787 
2788    if (width && height) {
2789       ASSERTED const uint32_t level_width =
2790          u_minify(info->surf.logical_level0_px.width, info->view.base_level);
2791       ASSERTED const uint32_t level_height =
2792          u_minify(info->surf.logical_level0_px.height, info->view.base_level);
2793       assert(*width % fmtl->bw == 0 || *x + *width == level_width);
2794       assert(*height % fmtl->bh == 0 || *y + *height == level_height);
2795       *width = DIV_ROUND_UP(*width, fmtl->bw);
2796       *height = DIV_ROUND_UP(*height, fmtl->bh);
2797    }
2798 
2799    if (x && y) {
2800       assert(*x % fmtl->bw == 0);
2801       assert(*y % fmtl->bh == 0);
2802       *x /= fmtl->bw;
2803       *y /= fmtl->bh;
2804    }
2805 
2806    /* We only want one level and slice */
2807    info->view.levels = 1;
2808    info->view.array_len = 1;
2809 
2810    if (info->surf.dim == ISL_SURF_DIM_3D) {
2811       /* Roll the Z offset into the image view */
2812       info->view.base_array_layer += info->z_offset;
2813       info->z_offset = 0;
2814    }
2815 
2816    uint64_t offset_B;
2817    ASSERTED bool ok =
2818       isl_surf_get_uncompressed_surf(isl_dev, &info->surf, &info->view,
2819                                      &info->surf, &info->view, &offset_B,
2820                                      &info->tile_x_sa, &info->tile_y_sa);
2821    assert(ok);
2822    info->addr.offset += offset_B;
2823 
2824    /* BLORP doesn't use the actual intratile offsets.  Instead, it needs the
2825     * surface to be a bit bigger and we offset the vertices instead. Standard
2826     * tilings don't need intratile offsets because each subresource is aligned
2827     * to a bpb-based tile boundary or miptail slot offset.
2828     */
2829   if (isl_tiling_is_64(info->surf.tiling) ||
2830       isl_tiling_is_std_y(info->surf.tiling)) {
2831       assert(info->tile_x_sa == 0 && info->tile_y_sa == 0);
2832    } else {
2833       assert(info->surf.dim == ISL_SURF_DIM_2D);
2834       assert(info->surf.logical_level0_px.array_len == 1);
2835       info->surf.logical_level0_px.w += info->tile_x_sa;
2836       info->surf.logical_level0_px.h += info->tile_y_sa;
2837       info->surf.phys_level0_sa.w += info->tile_x_sa;
2838       info->surf.phys_level0_sa.h += info->tile_y_sa;
2839    }
2840 }
2841 
2842 bool
blorp_copy_supports_compute(struct blorp_context * blorp,const struct isl_surf * src_surf,const struct isl_surf * dst_surf,enum isl_aux_usage dst_aux_usage)2843 blorp_copy_supports_compute(struct blorp_context *blorp,
2844                             const struct isl_surf *src_surf,
2845                             const struct isl_surf *dst_surf,
2846                             enum isl_aux_usage dst_aux_usage)
2847 {
2848    return blorp_blit_supports_compute(blorp, src_surf, dst_surf, dst_aux_usage);
2849 }
2850 
2851 void
blorp_copy_get_formats(const struct isl_device * isl_dev,const struct isl_surf * src_surf,const struct isl_surf * dst_surf,enum isl_format * src_view_format,enum isl_format * dst_view_format)2852 blorp_copy_get_formats(const struct isl_device *isl_dev,
2853                        const struct isl_surf *src_surf,
2854                        const struct isl_surf *dst_surf,
2855                        enum isl_format *src_view_format,
2856                        enum isl_format *dst_view_format)
2857 {
2858    const struct isl_format_layout *src_fmtl =
2859       isl_format_get_layout(src_surf->format);
2860    const struct isl_format_layout *dst_fmtl =
2861       isl_format_get_layout(dst_surf->format);
2862 
2863    if (ISL_GFX_VER(isl_dev) >= 8 &&
2864        isl_surf_usage_is_depth(src_surf->usage)) {
2865       /* In order to use HiZ, we have to use the real format for the source.
2866        * Depth <-> Color copies are not allowed.
2867        */
2868       *src_view_format = src_surf->format;
2869       *dst_view_format = src_surf->format;
2870    } else if (ISL_GFX_VER(isl_dev) >= 7 &&
2871               isl_surf_usage_is_depth(dst_surf->usage)) {
2872       /* On Gfx7 and higher, we use actual depth writes for blits into depth
2873        * buffers so we need the real format.
2874        */
2875       *src_view_format = dst_surf->format;
2876       *dst_view_format = dst_surf->format;
2877    } else if (isl_surf_usage_is_depth(src_surf->usage) ||
2878               isl_surf_usage_is_depth(dst_surf->usage)) {
2879       assert(src_fmtl->bpb == dst_fmtl->bpb);
2880       *src_view_format =
2881       *dst_view_format =
2882          get_copy_format_for_bpb(isl_dev, dst_fmtl->bpb);
2883    } else if (ISL_GFX_VER(isl_dev) < 20 &&
2884               isl_format_supports_ccs_e(isl_dev->info, dst_surf->format)) {
2885       *dst_view_format = get_ccs_compatible_copy_format(dst_fmtl);
2886       if (isl_format_supports_ccs_e(isl_dev->info, src_surf->format)) {
2887          *src_view_format = get_ccs_compatible_copy_format(src_fmtl);
2888       } else if (src_fmtl->bpb == dst_fmtl->bpb) {
2889          *src_view_format = *dst_view_format;
2890       } else {
2891          *src_view_format = get_copy_format_for_bpb(isl_dev, src_fmtl->bpb);
2892       }
2893    } else if (ISL_GFX_VER(isl_dev) < 20 &&
2894               isl_format_supports_ccs_e(isl_dev->info, src_surf->format)) {
2895       *src_view_format = get_ccs_compatible_copy_format(src_fmtl);
2896       if (src_fmtl->bpb == dst_fmtl->bpb) {
2897          *dst_view_format = *src_view_format;
2898       } else {
2899          *dst_view_format = get_copy_format_for_bpb(isl_dev, dst_fmtl->bpb);
2900       }
2901    } else {
2902       *dst_view_format = get_copy_format_for_bpb(isl_dev, dst_fmtl->bpb);
2903       *src_view_format = get_copy_format_for_bpb(isl_dev, src_fmtl->bpb);
2904    }
2905 }
2906 
2907 
2908 void
blorp_copy(struct blorp_batch * batch,const struct blorp_surf * src_surf,unsigned src_level,unsigned src_layer,const struct blorp_surf * dst_surf,unsigned dst_level,unsigned dst_layer,uint32_t src_x,uint32_t src_y,uint32_t dst_x,uint32_t dst_y,uint32_t src_width,uint32_t src_height)2909 blorp_copy(struct blorp_batch *batch,
2910            const struct blorp_surf *src_surf,
2911            unsigned src_level, unsigned src_layer,
2912            const struct blorp_surf *dst_surf,
2913            unsigned dst_level, unsigned dst_layer,
2914            uint32_t src_x, uint32_t src_y,
2915            uint32_t dst_x, uint32_t dst_y,
2916            uint32_t src_width, uint32_t src_height)
2917 {
2918    const struct isl_device *isl_dev = batch->blorp->isl_dev;
2919    const struct intel_device_info *devinfo = isl_dev->info;
2920    struct blorp_params params;
2921 
2922    if (src_width == 0 || src_height == 0)
2923       return;
2924 
2925    blorp_params_init(&params);
2926    params.op = BLORP_OP_COPY;
2927 
2928    const bool compute = batch->flags & BLORP_BATCH_USE_COMPUTE;
2929    if (compute) {
2930       assert(blorp_copy_supports_compute(batch->blorp,
2931                                          src_surf->surf, dst_surf->surf,
2932                                          dst_surf->aux_usage));
2933    } else if (batch->flags & BLORP_BATCH_USE_BLITTER) {
2934       assert(blorp_copy_supports_blitter(batch->blorp,
2935                                          src_surf->surf, dst_surf->surf,
2936                                          src_surf->aux_usage,
2937                                          dst_surf->aux_usage));
2938    }
2939 
2940    blorp_surface_info_init(batch, &params.src, src_surf, src_level,
2941                                src_layer, ISL_FORMAT_UNSUPPORTED, false);
2942    blorp_surface_info_init(batch, &params.dst, dst_surf, dst_level,
2943                                dst_layer, ISL_FORMAT_UNSUPPORTED, true);
2944 
2945    struct blorp_blit_prog_key key = {
2946       .base = BLORP_BASE_KEY_INIT(BLORP_SHADER_TYPE_COPY),
2947       .base.shader_pipeline = compute ? BLORP_SHADER_PIPELINE_COMPUTE :
2948                                         BLORP_SHADER_PIPELINE_RENDER,
2949       .filter = BLORP_FILTER_NONE,
2950       .need_src_offset = src_surf->tile_x_sa || src_surf->tile_y_sa,
2951       .need_dst_offset = dst_surf->tile_x_sa || dst_surf->tile_y_sa,
2952    };
2953 
2954    params.shader_type = key.base.shader_type;
2955    params.shader_pipeline = key.base.shader_pipeline;
2956 
2957    const struct isl_format_layout *src_fmtl =
2958       isl_format_get_layout(params.src.surf.format);
2959    const struct isl_format_layout *dst_fmtl =
2960       isl_format_get_layout(params.dst.surf.format);
2961 
2962    assert(params.src.aux_usage == ISL_AUX_USAGE_NONE ||
2963           params.src.aux_usage == ISL_AUX_USAGE_HIZ ||
2964           params.src.aux_usage == ISL_AUX_USAGE_HIZ_CCS_WT ||
2965           params.src.aux_usage == ISL_AUX_USAGE_MCS ||
2966           params.src.aux_usage == ISL_AUX_USAGE_MCS_CCS ||
2967           params.src.aux_usage == ISL_AUX_USAGE_CCS_E ||
2968           params.src.aux_usage == ISL_AUX_USAGE_FCV_CCS_E ||
2969           params.src.aux_usage == ISL_AUX_USAGE_STC_CCS);
2970 
2971    blorp_copy_get_formats(isl_dev, &params.src.surf, &params.dst.surf,
2972                           &params.src.view.format, &params.dst.view.format);
2973 
2974    if (isl_aux_usage_has_fast_clears(params.src.aux_usage) &&
2975        isl_dev->ss.clear_color_state_size > 0) {
2976       /* Depending on the format, the sampler may change the location from
2977        * which it fetches the clear color. This can be a problem in some
2978        * cases, so make sure that the view format won't change the location.
2979        */
2980       ASSERTED enum isl_format src_view_fmt = params.src.view.format;
2981       ASSERTED enum isl_format src_surf_fmt = params.src.surf.format;
2982       assert(isl_get_sampler_clear_field_offset(devinfo, src_view_fmt) ==
2983              isl_get_sampler_clear_field_offset(devinfo, src_surf_fmt));
2984    }
2985 
2986    if (params.src.view.format != params.dst.view.format) {
2987       enum isl_format src_cast_format = params.src.view.format;
2988       enum isl_format dst_cast_format = params.dst.view.format;
2989 
2990       /* The BLORP bitcast code gets confused by RGB formats.  Just treat them
2991        * as RGBA and then everything will be happy.  This is perfectly safe
2992        * because BLORP likes to treat things as if they have vec4 colors all
2993        * the time anyway.
2994        */
2995       if (isl_format_get_layout(src_cast_format)->bpb % 3 == 0)
2996          src_cast_format = isl_format_rgb_to_rgba(src_cast_format);
2997       if (isl_format_get_layout(dst_cast_format)->bpb % 3 == 0)
2998          dst_cast_format = isl_format_rgb_to_rgba(dst_cast_format);
2999 
3000       if (src_cast_format != dst_cast_format) {
3001          key.format_bit_cast = true;
3002          key.src_format = src_cast_format;
3003          key.dst_format = dst_cast_format;
3004       }
3005    }
3006 
3007    if (src_fmtl->bw > 1 || src_fmtl->bh > 1) {
3008       blorp_surf_convert_to_uncompressed(batch->blorp->isl_dev, &params.src,
3009                                          &src_x, &src_y,
3010                                          &src_width, &src_height);
3011       key.need_src_offset = true;
3012    }
3013 
3014    if (dst_fmtl->bw > 1 || dst_fmtl->bh > 1) {
3015       blorp_surf_convert_to_uncompressed(batch->blorp->isl_dev, &params.dst,
3016                                          &dst_x, &dst_y, NULL, NULL);
3017       key.need_dst_offset = true;
3018    }
3019 
3020    /* Once both surfaces are stompped to uncompressed as needed, the
3021     * destination size is the same as the source size.
3022     */
3023    uint32_t dst_width = src_width;
3024    uint32_t dst_height = src_height;
3025 
3026    if (batch->flags & BLORP_BATCH_USE_BLITTER) {
3027       if (devinfo->verx10 < 125) {
3028          blorp_surf_convert_to_single_slice(isl_dev, &params.dst);
3029          blorp_surf_convert_to_single_slice(isl_dev, &params.src);
3030       }
3031 
3032       params.x0 = dst_x;
3033       params.x1 = dst_x + dst_width;
3034       params.y0 = dst_y;
3035       params.y1 = dst_y + dst_height;
3036       params.wm_inputs.coord_transform[0].offset = dst_x - (float)src_x;
3037       params.wm_inputs.coord_transform[1].offset = dst_y - (float)src_y;
3038       params.wm_inputs.coord_transform[0].multiplier = 1.0f;
3039       params.wm_inputs.coord_transform[1].multiplier = 1.0f;
3040 
3041       batch->blorp->exec(batch, &params);
3042       return;
3043    }
3044 
3045    struct blt_coords coords = {
3046       .x = {
3047          .src0 = src_x,
3048          .src1 = src_x + src_width,
3049          .dst0 = dst_x,
3050          .dst1 = dst_x + dst_width,
3051          .mirror = false
3052       },
3053       .y = {
3054          .src0 = src_y,
3055          .src1 = src_y + src_height,
3056          .dst0 = dst_y,
3057          .dst1 = dst_y + dst_height,
3058          .mirror = false
3059       }
3060    };
3061 
3062    do_blorp_blit(batch, &params, &key, &coords);
3063 }
3064 
3065 static enum isl_format
isl_format_for_size(unsigned size_B)3066 isl_format_for_size(unsigned size_B)
3067 {
3068    switch (size_B) {
3069    case 1:  return ISL_FORMAT_R8_UINT;
3070    case 2:  return ISL_FORMAT_R8G8_UINT;
3071    case 4:  return ISL_FORMAT_R8G8B8A8_UINT;
3072    case 8:  return ISL_FORMAT_R16G16B16A16_UINT;
3073    case 16: return ISL_FORMAT_R32G32B32A32_UINT;
3074    default:
3075       unreachable("Not a power-of-two format size");
3076    }
3077 }
3078 
3079 /**
3080  * Returns the greatest common divisor of a and b that is a power of two.
3081  */
3082 static uint64_t
gcd_pow2_u64(uint64_t a,uint64_t b)3083 gcd_pow2_u64(uint64_t a, uint64_t b)
3084 {
3085    assert(a > 0 || b > 0);
3086 
3087    unsigned a_log2 = ffsll(a) - 1;
3088    unsigned b_log2 = ffsll(b) - 1;
3089 
3090    /* If either a or b is 0, then a_log2 or b_log2 till be UINT_MAX in which
3091     * case, the MIN2() will take the other one.  If both are 0 then we will
3092     * hit the assert above.
3093     */
3094    return 1 << MIN2(a_log2, b_log2);
3095 }
3096 
3097 static void
do_buffer_copy(struct blorp_batch * batch,struct blorp_address * src,struct blorp_address * dst,int width,int height,int block_size)3098 do_buffer_copy(struct blorp_batch *batch,
3099                struct blorp_address *src,
3100                struct blorp_address *dst,
3101                int width, int height, int block_size)
3102 {
3103    /* The actual format we pick doesn't matter as blorp will throw it away.
3104     * The only thing that actually matters is the size.
3105     */
3106    enum isl_format format = isl_format_for_size(block_size);
3107 
3108    UNUSED bool ok;
3109    struct isl_surf surf;
3110    ok = isl_surf_init(batch->blorp->isl_dev, &surf,
3111                       .dim = ISL_SURF_DIM_2D,
3112                       .format = format,
3113                       .width = width,
3114                       .height = height,
3115                       .depth = 1,
3116                       .levels = 1,
3117                       .array_len = 1,
3118                       .samples = 1,
3119                       .row_pitch_B = width * block_size,
3120                       .usage = ISL_SURF_USAGE_TEXTURE_BIT |
3121                                ISL_SURF_USAGE_RENDER_TARGET_BIT,
3122                       .tiling_flags = ISL_TILING_LINEAR_BIT);
3123    assert(ok);
3124 
3125    struct blorp_surf src_blorp_surf = {
3126       .surf = &surf,
3127       .addr = *src,
3128    };
3129 
3130    struct blorp_surf dst_blorp_surf = {
3131       .surf = &surf,
3132       .addr = *dst,
3133    };
3134 
3135    blorp_copy(batch, &src_blorp_surf, 0, 0, &dst_blorp_surf, 0, 0,
3136               0, 0, 0, 0, width, height);
3137 }
3138 
3139 void
blorp_buffer_copy(struct blorp_batch * batch,struct blorp_address src,struct blorp_address dst,uint64_t size)3140 blorp_buffer_copy(struct blorp_batch *batch,
3141                   struct blorp_address src,
3142                   struct blorp_address dst,
3143                   uint64_t size)
3144 {
3145    const struct intel_device_info *devinfo = batch->blorp->isl_dev->info;
3146    uint64_t copy_size = size;
3147 
3148    /* This is maximum possible width/height our HW can handle */
3149    uint64_t max_surface_dim = 1 << (devinfo->ver >= 7 ? 14 : 13);
3150 
3151    /* First, we compute the biggest format that can be used with the
3152     * given offsets and size.
3153     */
3154    int bs = 16;
3155    bs = gcd_pow2_u64(bs, src.offset);
3156    bs = gcd_pow2_u64(bs, dst.offset);
3157    bs = gcd_pow2_u64(bs, size);
3158 
3159    /* First, we make a bunch of max-sized copies */
3160    uint64_t max_copy_size = max_surface_dim * max_surface_dim * bs;
3161    while (copy_size >= max_copy_size) {
3162       do_buffer_copy(batch, &src, &dst, max_surface_dim, max_surface_dim, bs);
3163       copy_size -= max_copy_size;
3164       src.offset += max_copy_size;
3165       dst.offset += max_copy_size;
3166    }
3167 
3168    /* Now make a max-width copy */
3169    uint64_t height = copy_size / (max_surface_dim * bs);
3170    assert(height < max_surface_dim);
3171    if (height != 0) {
3172       uint64_t rect_copy_size = height * max_surface_dim * bs;
3173       do_buffer_copy(batch, &src, &dst, max_surface_dim, height, bs);
3174       copy_size -= rect_copy_size;
3175       src.offset += rect_copy_size;
3176       dst.offset += rect_copy_size;
3177    }
3178 
3179    /* Finally, make a small copy to finish it off */
3180    if (copy_size != 0) {
3181       do_buffer_copy(batch, &src, &dst, copy_size / bs, 1, bs);
3182    }
3183 }
3184