xref: /aosp_15_r20/external/mesa3d/src/broadcom/compiler/v3d_nir_lower_logic_ops.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright © 2019 Broadcom
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 /**
25  * Implements lowering for logical operations.
26  *
27  * V3D doesn't have any hardware support for logic ops.  Instead, you read the
28  * current contents of the destination from the tile buffer, then do math using
29  * your output color and that destination value, and update the output color
30  * appropriately.
31  */
32 
33 #include "util/format/u_format.h"
34 #include "compiler/nir/nir_builder.h"
35 #include "compiler/nir/nir_format_convert.h"
36 #include "v3d_compiler.h"
37 
38 
39 typedef nir_def *(*nir_pack_func)(nir_builder *b, nir_def *c);
40 typedef nir_def *(*nir_unpack_func)(nir_builder *b, nir_def *c);
41 
42 static bool
logicop_depends_on_dst_color(int logicop_func)43 logicop_depends_on_dst_color(int logicop_func)
44 {
45         switch (logicop_func) {
46         case PIPE_LOGICOP_SET:
47         case PIPE_LOGICOP_CLEAR:
48         case PIPE_LOGICOP_COPY:
49         case PIPE_LOGICOP_COPY_INVERTED:
50                 return false;
51         default:
52                 return true;
53         }
54 }
55 
56 static nir_def *
v3d_logicop(nir_builder * b,int logicop_func,nir_def * src,nir_def * dst)57 v3d_logicop(nir_builder *b, int logicop_func,
58             nir_def *src, nir_def *dst)
59 {
60         switch (logicop_func) {
61         case PIPE_LOGICOP_CLEAR:
62                 return nir_imm_int(b, 0);
63         case PIPE_LOGICOP_NOR:
64                 return nir_inot(b, nir_ior(b, src, dst));
65         case PIPE_LOGICOP_AND_INVERTED:
66                 return nir_iand(b, nir_inot(b, src), dst);
67         case PIPE_LOGICOP_COPY_INVERTED:
68                 return nir_inot(b, src);
69         case PIPE_LOGICOP_AND_REVERSE:
70                 return nir_iand(b, src, nir_inot(b, dst));
71         case PIPE_LOGICOP_INVERT:
72                 return nir_inot(b, dst);
73         case PIPE_LOGICOP_XOR:
74                 return nir_ixor(b, src, dst);
75         case PIPE_LOGICOP_NAND:
76                 return nir_inot(b, nir_iand(b, src, dst));
77         case PIPE_LOGICOP_AND:
78                 return nir_iand(b, src, dst);
79         case PIPE_LOGICOP_EQUIV:
80                 return nir_inot(b, nir_ixor(b, src, dst));
81         case PIPE_LOGICOP_NOOP:
82                 return dst;
83         case PIPE_LOGICOP_OR_INVERTED:
84                 return nir_ior(b, nir_inot(b, src), dst);
85         case PIPE_LOGICOP_OR_REVERSE:
86                 return nir_ior(b, src, nir_inot(b, dst));
87         case PIPE_LOGICOP_OR:
88                 return nir_ior(b, src, dst);
89         case PIPE_LOGICOP_SET:
90                 return nir_imm_int(b, ~0);
91         default:
92                 fprintf(stderr, "Unknown logic op %d\n", logicop_func);
93                 FALLTHROUGH;
94         case PIPE_LOGICOP_COPY:
95                 return src;
96         }
97 }
98 
99 static nir_def *
v3d_nir_get_swizzled_channel(nir_builder * b,nir_def ** srcs,int swiz)100 v3d_nir_get_swizzled_channel(nir_builder *b, nir_def **srcs, int swiz)
101 {
102         switch (swiz) {
103         default:
104         case PIPE_SWIZZLE_NONE:
105                 fprintf(stderr, "warning: unknown swizzle\n");
106                 FALLTHROUGH;
107         case PIPE_SWIZZLE_0:
108                 return nir_imm_float(b, 0.0);
109         case PIPE_SWIZZLE_1:
110                 return nir_imm_float(b, 1.0);
111         case PIPE_SWIZZLE_X:
112         case PIPE_SWIZZLE_Y:
113         case PIPE_SWIZZLE_Z:
114         case PIPE_SWIZZLE_W:
115                 return srcs[swiz];
116         }
117 }
118 
119 static nir_def *
v3d_nir_swizzle_and_pack(nir_builder * b,nir_def ** chans,const uint8_t * swiz,nir_pack_func pack_func)120 v3d_nir_swizzle_and_pack(nir_builder *b, nir_def **chans,
121                          const uint8_t *swiz, nir_pack_func pack_func)
122 {
123         nir_def *c[4];
124         for (int i = 0; i < 4; i++)
125                 c[i] = v3d_nir_get_swizzled_channel(b, chans, swiz[i]);
126 
127         return pack_func(b, nir_vec4(b, c[0], c[1], c[2], c[3]));
128 }
129 
130 static nir_def *
v3d_nir_unpack_and_swizzle(nir_builder * b,nir_def * packed,const uint8_t * swiz,nir_unpack_func unpack_func)131 v3d_nir_unpack_and_swizzle(nir_builder *b, nir_def *packed,
132                            const uint8_t *swiz, nir_unpack_func unpack_func)
133 {
134         nir_def *unpacked = unpack_func(b, packed);
135 
136         nir_def *unpacked_chans[4];
137         for (int i = 0; i < 4; i++)
138                 unpacked_chans[i] = nir_channel(b, unpacked, i);
139 
140         nir_def *c[4];
141         for (int i = 0; i < 4; i++)
142                 c[i] = v3d_nir_get_swizzled_channel(b, unpacked_chans, swiz[i]);
143 
144         return nir_vec4(b, c[0], c[1], c[2], c[3]);
145 }
146 
147 static nir_def *
pack_unorm_rgb10a2(nir_builder * b,nir_def * c)148 pack_unorm_rgb10a2(nir_builder *b, nir_def *c)
149 {
150         static const unsigned bits[4] = { 10, 10, 10, 2 };
151         nir_def *unorm = nir_format_float_to_unorm(b, c, bits);
152 
153         nir_def *chans[4];
154         for (int i = 0; i < 4; i++)
155                 chans[i] = nir_channel(b, unorm, i);
156 
157         nir_def *result = nir_mov(b, chans[0]);
158         int offset = bits[0];
159         for (int i = 1; i < 4; i++) {
160                 nir_def *shifted_chan =
161                         nir_ishl_imm(b, chans[i], offset);
162                 result = nir_ior(b, result, shifted_chan);
163                 offset += bits[i];
164         }
165         return result;
166 }
167 
168 static nir_def *
unpack_unorm_rgb10a2(nir_builder * b,nir_def * c)169 unpack_unorm_rgb10a2(nir_builder *b, nir_def *c)
170 {
171         static const unsigned bits[4] = { 10, 10, 10, 2 };
172         const unsigned masks[4] = { BITFIELD_MASK(bits[0]),
173                                     BITFIELD_MASK(bits[1]),
174                                     BITFIELD_MASK(bits[2]),
175                                     BITFIELD_MASK(bits[3]) };
176 
177         nir_def *chans[4];
178         for (int i = 0; i < 4; i++) {
179                 nir_def *unorm = nir_iand_imm(b, c, masks[i]);
180                 chans[i] = nir_format_unorm_to_float(b, unorm, &bits[i]);
181                 c = nir_ushr_imm(b, c, bits[i]);
182         }
183 
184         return nir_vec4(b, chans[0], chans[1], chans[2], chans[3]);
185 }
186 
187 static const uint8_t *
v3d_get_format_swizzle_for_rt(struct v3d_compile * c,int rt)188 v3d_get_format_swizzle_for_rt(struct v3d_compile *c, int rt)
189 {
190         static const uint8_t ident[4] = { 0, 1, 2, 3 };
191 
192         /* We will automatically swap R and B channels for BGRA formats
193          * on tile loads and stores (see 'swap_rb' field in v3d_resource) so
194          * we want to treat these surfaces as if they were regular RGBA formats.
195          */
196         if (c->fs_key->color_fmt[rt].swizzle[0] == 2 &&
197             c->fs_key->color_fmt[rt].format != PIPE_FORMAT_B5G6R5_UNORM) {
198                 return ident;
199         } else {
200                 return  c->fs_key->color_fmt[rt].swizzle;
201         }
202 }
203 
204 static nir_def *
v3d_nir_get_tlb_color(nir_builder * b,struct v3d_compile * c,int rt,int sample)205 v3d_nir_get_tlb_color(nir_builder *b, struct v3d_compile *c, int rt, int sample)
206 {
207         uint32_t num_components =
208                 util_format_get_nr_components(c->fs_key->color_fmt[rt].format);
209 
210         nir_def *color[4];
211         for (int i = 0; i < 4; i++) {
212                 if (i < num_components) {
213                         color[i] =
214                                 nir_load_tlb_color_brcm(b, 1, 32,
215                                                         nir_imm_int(b, rt),
216                                                         .base = sample,
217                                                         .component = i);
218                 } else {
219                         /* These will be DCEd */
220                         color[i] = nir_imm_int(b, 0);
221                 }
222         }
223         return nir_vec4(b, color[0], color[1], color[2], color[3]);
224 }
225 
226 static nir_def *
v3d_emit_logic_op_raw(struct v3d_compile * c,nir_builder * b,nir_def ** src_chans,nir_def ** dst_chans,int rt,int sample)227 v3d_emit_logic_op_raw(struct v3d_compile *c, nir_builder *b,
228                       nir_def **src_chans, nir_def **dst_chans,
229                       int rt, int sample)
230 {
231         const uint8_t *fmt_swz = v3d_get_format_swizzle_for_rt(c, rt);
232 
233         nir_def *op_res[4];
234         for (int i = 0; i < 4; i++) {
235                 nir_def *src = src_chans[i];
236                 nir_def *dst =
237                         v3d_nir_get_swizzled_channel(b, dst_chans, fmt_swz[i]);
238                 op_res[i] = v3d_logicop(b, c->fs_key->logicop_func, src, dst);
239 
240                 /* We configure our integer RTs to clamp, so we need to ignore
241                  * result bits that don't fit in the destination RT component
242                  * size.
243                  */
244                 uint32_t bits =
245                         util_format_get_component_bits(
246                                 c->fs_key->color_fmt[rt].format,
247                                 UTIL_FORMAT_COLORSPACE_RGB, i);
248                 if (bits > 0 && bits < 32) {
249                         op_res[i] =
250                                 nir_iand_imm(b, op_res[i], (1u << bits) - 1);
251                 }
252         }
253 
254         nir_def *r[4];
255         for (int i = 0; i < 4; i++)
256                 r[i] = v3d_nir_get_swizzled_channel(b, op_res, fmt_swz[i]);
257 
258         return nir_vec4(b, r[0], r[1], r[2], r[3]);
259 }
260 
261 static nir_def *
v3d_emit_logic_op_unorm(struct v3d_compile * c,nir_builder * b,nir_def ** src_chans,nir_def ** dst_chans,int rt,int sample,nir_pack_func pack_func,nir_unpack_func unpack_func)262 v3d_emit_logic_op_unorm(struct v3d_compile *c, nir_builder *b,
263                         nir_def **src_chans, nir_def **dst_chans,
264                         int rt, int sample,
265                         nir_pack_func pack_func, nir_unpack_func unpack_func)
266 {
267         static const uint8_t src_swz[4] = { 0, 1, 2, 3 };
268         nir_def *packed_src =
269                 v3d_nir_swizzle_and_pack(b, src_chans, src_swz, pack_func);
270 
271         const uint8_t *fmt_swz = v3d_get_format_swizzle_for_rt(c, rt);
272         nir_def *packed_dst =
273                 v3d_nir_swizzle_and_pack(b, dst_chans, fmt_swz, pack_func);
274 
275         nir_def *packed_result =
276                 v3d_logicop(b, c->fs_key->logicop_func, packed_src, packed_dst);
277 
278         return v3d_nir_unpack_and_swizzle(b, packed_result, fmt_swz, unpack_func);
279 }
280 
281 static nir_def *
v3d_nir_emit_logic_op(struct v3d_compile * c,nir_builder * b,nir_def * src,int rt,int sample)282 v3d_nir_emit_logic_op(struct v3d_compile *c, nir_builder *b,
283                       nir_def *src, int rt, int sample)
284 {
285         nir_def *dst = v3d_nir_get_tlb_color(b, c, rt, sample);
286 
287         nir_def *src_chans[4], *dst_chans[4];
288         for (unsigned i = 0; i < 4; i++) {
289                 src_chans[i] = nir_channel(b, src, i);
290                 dst_chans[i] = nir_channel(b, dst, i);
291         }
292 
293         if (c->fs_key->color_fmt[rt].format == PIPE_FORMAT_R10G10B10A2_UNORM) {
294                 return v3d_emit_logic_op_unorm(
295                         c, b, src_chans, dst_chans, rt, 0,
296                         pack_unorm_rgb10a2, unpack_unorm_rgb10a2);
297         }
298 
299         if (util_format_is_unorm(c->fs_key->color_fmt[rt].format)) {
300                 return v3d_emit_logic_op_unorm(
301                         c, b, src_chans, dst_chans, rt, 0,
302                         nir_pack_unorm_4x8, nir_unpack_unorm_4x8);
303         }
304 
305         return v3d_emit_logic_op_raw(c, b, src_chans, dst_chans, rt, 0);
306 }
307 
308 static void
v3d_emit_ms_output(nir_builder * b,nir_def * color,nir_src * offset,nir_alu_type type,int rt,int sample)309 v3d_emit_ms_output(nir_builder *b,
310                    nir_def *color, nir_src *offset,
311                    nir_alu_type type, int rt, int sample)
312 {
313         nir_store_tlb_sample_color_v3d(b, color, nir_imm_int(b, rt), .base = sample, .component = 0, .src_type = type);
314 }
315 
316 static void
v3d_nir_lower_logic_op_instr(struct v3d_compile * c,nir_builder * b,nir_intrinsic_instr * intr,int rt)317 v3d_nir_lower_logic_op_instr(struct v3d_compile *c,
318                              nir_builder *b,
319                              nir_intrinsic_instr *intr,
320                              int rt)
321 {
322         nir_def *frag_color = intr->src[0].ssa;
323 
324 
325         const int logic_op = c->fs_key->logicop_func;
326         if (c->fs_key->msaa && logicop_depends_on_dst_color(logic_op)) {
327                 c->msaa_per_sample_output = true;
328 
329                 nir_src *offset = &intr->src[1];
330                 nir_alu_type type = nir_intrinsic_src_type(intr);
331                 for (int i = 0; i < V3D_MAX_SAMPLES; i++) {
332                         nir_def *sample =
333                                 v3d_nir_emit_logic_op(c, b, frag_color, rt, i);
334 
335                         v3d_emit_ms_output(b, sample, offset, type, rt, i);
336                 }
337 
338                 nir_instr_remove(&intr->instr);
339         } else {
340                 nir_def *result =
341                         v3d_nir_emit_logic_op(c, b, frag_color, rt, 0);
342 
343                 nir_src_rewrite(&intr->src[0], result);
344                 intr->num_components = result->num_components;
345         }
346 }
347 
348 static bool
v3d_nir_lower_logic_ops_block(nir_block * block,struct v3d_compile * c)349 v3d_nir_lower_logic_ops_block(nir_block *block, struct v3d_compile *c)
350 {
351         bool progress = false;
352 
353         nir_foreach_instr_safe(instr, block) {
354                 if (instr->type != nir_instr_type_intrinsic)
355                         continue;
356 
357                 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
358                 if (intr->intrinsic != nir_intrinsic_store_output)
359                         continue;
360 
361                 nir_foreach_shader_out_variable(var, c->s) {
362                         const int driver_loc = var->data.driver_location;
363                         if (driver_loc != nir_intrinsic_base(intr))
364                                 continue;
365 
366                         const int loc = var->data.location;
367                         if (loc != FRAG_RESULT_COLOR &&
368                             (loc < FRAG_RESULT_DATA0 ||
369                              loc >= FRAG_RESULT_DATA0 + V3D_MAX_DRAW_BUFFERS)) {
370                                 continue;
371                         }
372 
373                         /* Logic operations do not apply on floating point or
374                          * sRGB enabled render targets.
375                          */
376                         const int rt = driver_loc;
377                         assert(rt < V3D_MAX_DRAW_BUFFERS);
378 
379                         const enum pipe_format format =
380                                 c->fs_key->color_fmt[rt].format;
381                         if (util_format_is_float(format) ||
382                             util_format_is_srgb(format)) {
383                                 continue;
384                         }
385 
386                         nir_builder b = nir_builder_at(nir_before_instr(&intr->instr));
387                         v3d_nir_lower_logic_op_instr(c, &b, intr, rt);
388 
389                         progress = true;
390                 }
391         }
392 
393         return progress;
394 }
395 
396 bool
v3d_nir_lower_logic_ops(nir_shader * s,struct v3d_compile * c)397 v3d_nir_lower_logic_ops(nir_shader *s, struct v3d_compile *c)
398 {
399         bool progress = false;
400 
401         /* Nothing to do if logic op is 'copy src to dst' or if logic ops are
402          * disabled (we set the logic op to copy in that case).
403          */
404         if (c->fs_key->logicop_func == PIPE_LOGICOP_COPY)
405                 return false;
406 
407         nir_foreach_function_impl(impl, s) {
408                 nir_foreach_block(block, impl)
409                         progress |= v3d_nir_lower_logic_ops_block(block, c);
410 
411                 if (progress) {
412                         nir_metadata_preserve(impl,
413                                               nir_metadata_control_flow);
414                 } else {
415                         nir_metadata_preserve(impl,
416                                               nir_metadata_all);
417                 }
418         }
419 
420         return progress;
421 }
422