1/*
2 * Copyright 2018 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 *
24 */
25
26#include "si_pipe.h"
27#include "util/format/u_format.h"
28#include "util/format_srgb.h"
29#include "util/u_helpers.h"
30
31static bool si_can_use_compute_blit(struct si_context *sctx, enum pipe_format format,
32                                    unsigned num_samples, bool is_store, bool has_dcc)
33{
34   /* TODO: This format fails AMD_TEST=imagecopy. */
35   if (format == PIPE_FORMAT_A8R8_UNORM && is_store)
36      return false;
37
38   if (num_samples > 1)
39      return false;
40
41   if (util_format_is_depth_or_stencil(format))
42      return false;
43
44   /* Image stores support DCC since GFX10. */
45   if (has_dcc && is_store && sctx->gfx_level < GFX10)
46      return false;
47
48   return true;
49}
50
51static void si_use_compute_copy_for_float_formats(struct si_context *sctx,
52                                                  struct pipe_resource *texture,
53                                                  unsigned level)
54{
55   struct si_texture *tex = (struct si_texture *)texture;
56
57   /* If we are uploading into FP16 or R11G11B10_FLOAT via a blit, CB clobbers NaNs,
58    * so in order to preserve them exactly, we have to use the compute blit.
59    * The compute blit is used only when the destination doesn't have DCC, so
60    * disable it here, which is kinda a hack.
61    * If we are uploading into 32-bit floats with DCC via a blit, NaNs will also get
62    * lost so we need to disable DCC as well.
63    *
64    * This makes KHR-GL45.texture_view.view_classes pass on gfx9.
65    */
66   if (vi_dcc_enabled(tex, level) &&
67       util_format_is_float(texture->format) &&
68       /* Check if disabling DCC enables the compute copy. */
69       !si_can_use_compute_blit(sctx, texture->format, texture->nr_samples, true, true) &&
70       si_can_use_compute_blit(sctx, texture->format, texture->nr_samples, true, false)) {
71      si_texture_disable_dcc(sctx, tex);
72   }
73}
74
75/* Determine the cache policy. */
76static enum si_cache_policy get_cache_policy(struct si_context *sctx, enum si_coherency coher,
77                                             uint64_t size)
78{
79   if ((sctx->gfx_level >= GFX9 && (coher == SI_COHERENCY_CB_META ||
80                                     coher == SI_COHERENCY_DB_META ||
81                                     coher == SI_COHERENCY_CP)) ||
82       (sctx->gfx_level >= GFX7 && coher == SI_COHERENCY_SHADER))
83      return L2_LRU; /* it's faster if L2 doesn't evict anything  */
84
85   return L2_BYPASS;
86}
87
88unsigned si_get_flush_flags(struct si_context *sctx, enum si_coherency coher,
89                            enum si_cache_policy cache_policy)
90{
91   switch (coher) {
92   default:
93   case SI_COHERENCY_NONE:
94   case SI_COHERENCY_CP:
95      return 0;
96   case SI_COHERENCY_SHADER:
97      return SI_CONTEXT_INV_SCACHE | SI_CONTEXT_INV_VCACHE |
98             (cache_policy == L2_BYPASS ? SI_CONTEXT_INV_L2 : 0);
99   case SI_COHERENCY_CB_META:
100      return SI_CONTEXT_FLUSH_AND_INV_CB;
101   case SI_COHERENCY_DB_META:
102      return SI_CONTEXT_FLUSH_AND_INV_DB;
103   }
104}
105
106static bool si_is_buffer_idle(struct si_context *sctx, struct si_resource *buf,
107                              unsigned usage)
108{
109   return !si_cs_is_buffer_referenced(sctx, buf->buf, usage) &&
110          sctx->ws->buffer_wait(sctx->ws, buf->buf, 0, usage);
111}
112
113static void si_improve_sync_flags(struct si_context *sctx, struct pipe_resource *dst,
114                                  struct pipe_resource *src, unsigned *flags)
115{
116   if (dst->target != PIPE_BUFFER || (src && src->target != PIPE_BUFFER))
117      return;
118
119   if (si_is_buffer_idle(sctx, si_resource(dst), RADEON_USAGE_READWRITE) &&
120       (!src || si_is_buffer_idle(sctx, si_resource(src), RADEON_USAGE_WRITE))) {
121      /* Idle buffers don't have to sync. */
122      *flags &= ~(SI_OP_SYNC_GE_BEFORE | SI_OP_SYNC_PS_BEFORE | SI_OP_SYNC_CS_BEFORE |
123                  SI_OP_SYNC_CPDMA_BEFORE);
124      return;
125   }
126
127   const unsigned cs_mask = SI_BIND_CONSTANT_BUFFER(PIPE_SHADER_COMPUTE) |
128                            SI_BIND_SHADER_BUFFER(PIPE_SHADER_COMPUTE) |
129                            SI_BIND_IMAGE_BUFFER(PIPE_SHADER_COMPUTE) |
130                            SI_BIND_SAMPLER_BUFFER(PIPE_SHADER_COMPUTE);
131
132   const unsigned ps_mask = SI_BIND_CONSTANT_BUFFER(PIPE_SHADER_FRAGMENT) |
133                            SI_BIND_SHADER_BUFFER(PIPE_SHADER_FRAGMENT) |
134                            SI_BIND_IMAGE_BUFFER(PIPE_SHADER_FRAGMENT) |
135                            SI_BIND_SAMPLER_BUFFER(PIPE_SHADER_FRAGMENT);
136
137   unsigned bind_history = si_resource(dst)->bind_history |
138                           (src ? si_resource(src)->bind_history : 0);
139
140   /* Clear SI_OP_SYNC_CS_BEFORE if the buffer has never been used with a CS. */
141   if (*flags & SI_OP_SYNC_CS_BEFORE && !(bind_history & cs_mask))
142      *flags &= ~SI_OP_SYNC_CS_BEFORE;
143
144   /* Clear SI_OP_SYNC_PS_BEFORE if the buffer has never been used with a PS. */
145   if (*flags & SI_OP_SYNC_PS_BEFORE && !(bind_history & ps_mask)) {
146      *flags &= ~SI_OP_SYNC_PS_BEFORE;
147      *flags |= SI_OP_SYNC_GE_BEFORE;
148   }
149}
150
151static void si_launch_grid_internal(struct si_context *sctx, const struct pipe_grid_info *info,
152                                    void *shader, unsigned flags)
153{
154   /* Wait for previous shaders to finish. */
155   if (flags & SI_OP_SYNC_GE_BEFORE)
156      sctx->flags |= SI_CONTEXT_VS_PARTIAL_FLUSH;
157
158   if (flags & SI_OP_SYNC_PS_BEFORE)
159      sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH;
160
161   if (flags & SI_OP_SYNC_CS_BEFORE)
162      sctx->flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
163
164   if (!(flags & SI_OP_CS_IMAGE))
165      sctx->flags |= SI_CONTEXT_PFP_SYNC_ME;
166
167   /* Invalidate L0-L1 caches. */
168   /* sL0 is never invalidated, because src resources don't use it. */
169   if (!(flags & SI_OP_SKIP_CACHE_INV_BEFORE))
170      sctx->flags |= SI_CONTEXT_INV_VCACHE;
171
172   /* Set settings for driver-internal compute dispatches. */
173   sctx->flags &= ~SI_CONTEXT_START_PIPELINE_STATS;
174   sctx->flags |= SI_CONTEXT_STOP_PIPELINE_STATS;
175
176   if (!(flags & SI_OP_CS_RENDER_COND_ENABLE))
177      sctx->render_cond_enabled = false;
178
179   /* Skip decompression to prevent infinite recursion. */
180   sctx->blitter_running = true;
181
182   /* Dispatch compute. */
183   void *saved_cs = sctx->cs_shader_state.program;
184   sctx->b.bind_compute_state(&sctx->b, shader);
185   sctx->b.launch_grid(&sctx->b, info);
186   sctx->b.bind_compute_state(&sctx->b, saved_cs);
187
188   /* Restore default settings. */
189   sctx->flags &= ~SI_CONTEXT_STOP_PIPELINE_STATS;
190   sctx->flags |= SI_CONTEXT_START_PIPELINE_STATS;
191   sctx->render_cond_enabled = sctx->render_cond;
192   sctx->blitter_running = false;
193
194   if (flags & SI_OP_SYNC_AFTER) {
195      sctx->flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
196
197      if (flags & SI_OP_CS_IMAGE) {
198         /* Make sure image stores are visible to CB, which doesn't use L2 on GFX6-8. */
199         sctx->flags |= sctx->gfx_level <= GFX8 ? SI_CONTEXT_WB_L2 : 0;
200         /* Make sure image stores are visible to all CUs. */
201         sctx->flags |= SI_CONTEXT_INV_VCACHE;
202         /* Make sure RBs see our DCC changes. */
203         if (sctx->gfx_level >= GFX10 && sctx->screen->info.tcc_rb_non_coherent) {
204            unsigned enabled_mask = sctx->images[PIPE_SHADER_COMPUTE].enabled_mask;
205            while (enabled_mask) {
206               int i = u_bit_scan(&enabled_mask);
207               if (sctx->images[PIPE_SHADER_COMPUTE].views[i].access & SI_IMAGE_ACCESS_ALLOW_DCC_STORE) {
208                  sctx->flags |= SI_CONTEXT_INV_L2;
209                  break;
210               }
211            }
212         }
213      } else {
214         /* Make sure buffer stores are visible to all CUs. */
215         sctx->flags |= SI_CONTEXT_INV_SCACHE | SI_CONTEXT_INV_VCACHE | SI_CONTEXT_PFP_SYNC_ME;
216      }
217   }
218}
219
220void si_launch_grid_internal_ssbos(struct si_context *sctx, struct pipe_grid_info *info,
221                                   void *shader, unsigned flags, enum si_coherency coher,
222                                   unsigned num_buffers, const struct pipe_shader_buffer *buffers,
223                                   unsigned writeable_bitmask)
224{
225   if (!(flags & SI_OP_SKIP_CACHE_INV_BEFORE))
226      sctx->flags |= si_get_flush_flags(sctx, coher, SI_COMPUTE_DST_CACHE_POLICY);
227
228   /* Save states. */
229   struct pipe_shader_buffer saved_sb[3] = {};
230   assert(num_buffers <= ARRAY_SIZE(saved_sb));
231   si_get_shader_buffers(sctx, PIPE_SHADER_COMPUTE, 0, num_buffers, saved_sb);
232
233   unsigned saved_writable_mask = 0;
234   for (unsigned i = 0; i < num_buffers; i++) {
235      if (sctx->const_and_shader_buffers[PIPE_SHADER_COMPUTE].writable_mask &
236          (1u << si_get_shaderbuf_slot(i)))
237         saved_writable_mask |= 1 << i;
238   }
239
240   /* Bind buffers and launch compute. */
241   si_set_shader_buffers(&sctx->b, PIPE_SHADER_COMPUTE, 0, num_buffers, buffers,
242                         writeable_bitmask,
243                         true /* don't update bind_history to prevent unnecessary syncs later */);
244   si_launch_grid_internal(sctx, info, shader, flags);
245
246   /* Do cache flushing at the end. */
247   if (get_cache_policy(sctx, coher, 0) == L2_BYPASS) {
248      if (flags & SI_OP_SYNC_AFTER)
249         sctx->flags |= SI_CONTEXT_WB_L2;
250   } else {
251      while (writeable_bitmask)
252         si_resource(buffers[u_bit_scan(&writeable_bitmask)].buffer)->TC_L2_dirty = true;
253   }
254
255   /* Restore states. */
256   sctx->b.set_shader_buffers(&sctx->b, PIPE_SHADER_COMPUTE, 0, num_buffers, saved_sb,
257                              saved_writable_mask);
258   for (int i = 0; i < num_buffers; i++)
259      pipe_resource_reference(&saved_sb[i].buffer, NULL);
260}
261
262/**
263 * Clear a buffer using read-modify-write with a 32-bit write bitmask.
264 * The clear value has 32 bits.
265 */
266void si_compute_clear_buffer_rmw(struct si_context *sctx, struct pipe_resource *dst,
267                                 unsigned dst_offset, unsigned size,
268                                 uint32_t clear_value, uint32_t writebitmask,
269                                 unsigned flags, enum si_coherency coher)
270{
271   assert(dst_offset % 4 == 0);
272   assert(size % 4 == 0);
273
274   assert(dst->target != PIPE_BUFFER || dst_offset + size <= dst->width0);
275
276   /* Use buffer_load_dwordx4 and buffer_store_dwordx4 per thread. */
277   unsigned dwords_per_instruction = 4;
278   unsigned block_size = 64; /* it's always 64x1x1 */
279   unsigned dwords_per_wave = dwords_per_instruction * block_size;
280
281   unsigned num_dwords = size / 4;
282   unsigned num_instructions = DIV_ROUND_UP(num_dwords, dwords_per_instruction);
283
284   struct pipe_grid_info info = {};
285   info.block[0] = MIN2(block_size, num_instructions);
286   info.block[1] = 1;
287   info.block[2] = 1;
288   info.grid[0] = DIV_ROUND_UP(num_dwords, dwords_per_wave);
289   info.grid[1] = 1;
290   info.grid[2] = 1;
291
292   struct pipe_shader_buffer sb = {};
293   sb.buffer = dst;
294   sb.buffer_offset = dst_offset;
295   sb.buffer_size = size;
296
297   sctx->cs_user_data[0] = clear_value & writebitmask;
298   sctx->cs_user_data[1] = ~writebitmask;
299
300   if (!sctx->cs_clear_buffer_rmw)
301      sctx->cs_clear_buffer_rmw = si_create_clear_buffer_rmw_cs(sctx);
302
303   si_launch_grid_internal_ssbos(sctx, &info, sctx->cs_clear_buffer_rmw, flags, coher,
304                                 1, &sb, 0x1);
305}
306
307static void si_compute_clear_12bytes_buffer(struct si_context *sctx, struct pipe_resource *dst,
308                                            unsigned dst_offset, unsigned size,
309                                            const uint32_t *clear_value, unsigned flags,
310                                            enum si_coherency coher)
311{
312   struct pipe_context *ctx = &sctx->b;
313
314   assert(dst_offset % 4 == 0);
315   assert(size % 4 == 0);
316   unsigned size_12 = DIV_ROUND_UP(size, 12);
317
318   struct pipe_shader_buffer sb = {0};
319   sb.buffer = dst;
320   sb.buffer_offset = dst_offset;
321   sb.buffer_size = size;
322
323   memcpy(sctx->cs_user_data, clear_value, 12);
324
325   struct pipe_grid_info info = {0};
326
327   if (!sctx->cs_clear_12bytes_buffer)
328      sctx->cs_clear_12bytes_buffer = si_clear_12bytes_buffer_shader(ctx);
329
330   info.block[0] = 64;
331   info.last_block[0] = size_12 % 64;
332   info.block[1] = 1;
333   info.block[2] = 1;
334   info.grid[0] = DIV_ROUND_UP(size_12, 64);
335   info.grid[1] = 1;
336   info.grid[2] = 1;
337
338   si_launch_grid_internal_ssbos(sctx, &info, sctx->cs_clear_12bytes_buffer, flags, coher,
339                                 1, &sb, 0x1);
340}
341
342static void si_compute_do_clear_or_copy(struct si_context *sctx, struct pipe_resource *dst,
343                                        unsigned dst_offset, struct pipe_resource *src,
344                                        unsigned src_offset, unsigned size,
345                                        const uint32_t *clear_value, unsigned clear_value_size,
346                                        unsigned flags, enum si_coherency coher)
347{
348   assert(src_offset % 4 == 0);
349   assert(dst_offset % 4 == 0);
350   assert(size % 4 == 0);
351
352   assert(dst->target != PIPE_BUFFER || dst_offset + size <= dst->width0);
353   assert(!src || src_offset + size <= src->width0);
354
355   /* The memory accesses are coalesced, meaning that the 1st instruction writes
356    * the 1st contiguous block of data for the whole wave, the 2nd instruction
357    * writes the 2nd contiguous block of data, etc.
358    */
359   unsigned dwords_per_thread =
360      src ? SI_COMPUTE_COPY_DW_PER_THREAD : SI_COMPUTE_CLEAR_DW_PER_THREAD;
361   unsigned instructions_per_thread = MAX2(1, dwords_per_thread / 4);
362   unsigned dwords_per_instruction = dwords_per_thread / instructions_per_thread;
363   /* The shader declares the block size like this: */
364   unsigned block_size = si_determine_wave_size(sctx->screen, NULL);
365   unsigned dwords_per_wave = dwords_per_thread * block_size;
366
367   unsigned num_dwords = size / 4;
368   unsigned num_instructions = DIV_ROUND_UP(num_dwords, dwords_per_instruction);
369
370   struct pipe_grid_info info = {};
371   info.block[0] = MIN2(block_size, num_instructions);
372   info.block[1] = 1;
373   info.block[2] = 1;
374   info.grid[0] = DIV_ROUND_UP(num_dwords, dwords_per_wave);
375   info.grid[1] = 1;
376   info.grid[2] = 1;
377
378   struct pipe_shader_buffer sb[2] = {};
379   sb[0].buffer = dst;
380   sb[0].buffer_offset = dst_offset;
381   sb[0].buffer_size = size;
382
383   bool shader_dst_stream_policy = SI_COMPUTE_DST_CACHE_POLICY != L2_LRU;
384
385   if (src) {
386      sb[1].buffer = src;
387      sb[1].buffer_offset = src_offset;
388      sb[1].buffer_size = size;
389
390      if (!sctx->cs_copy_buffer) {
391         sctx->cs_copy_buffer = si_create_dma_compute_shader(
392            &sctx->b, SI_COMPUTE_COPY_DW_PER_THREAD, shader_dst_stream_policy, true);
393      }
394
395      si_launch_grid_internal_ssbos(sctx, &info, sctx->cs_copy_buffer, flags, coher,
396                                    2, sb, 0x1);
397   } else {
398      assert(clear_value_size >= 4 && clear_value_size <= 16 &&
399             util_is_power_of_two_or_zero(clear_value_size));
400
401      for (unsigned i = 0; i < 4; i++)
402         sctx->cs_user_data[i] = clear_value[i % (clear_value_size / 4)];
403
404      if (!sctx->cs_clear_buffer) {
405         sctx->cs_clear_buffer = si_create_dma_compute_shader(
406            &sctx->b, SI_COMPUTE_CLEAR_DW_PER_THREAD, shader_dst_stream_policy, false);
407      }
408
409      si_launch_grid_internal_ssbos(sctx, &info, sctx->cs_clear_buffer, flags, coher,
410                                    1, sb, 0x1);
411   }
412}
413
414void si_clear_buffer(struct si_context *sctx, struct pipe_resource *dst,
415                     uint64_t offset, uint64_t size, uint32_t *clear_value,
416                     uint32_t clear_value_size, unsigned flags,
417                     enum si_coherency coher, enum si_clear_method method)
418{
419   if (!size)
420      return;
421
422   si_improve_sync_flags(sctx, dst, NULL, &flags);
423
424   ASSERTED unsigned clear_alignment = MIN2(clear_value_size, 4);
425
426   assert(clear_value_size != 3 && clear_value_size != 6); /* 12 is allowed. */
427   assert(offset % clear_alignment == 0);
428   assert(size % clear_alignment == 0);
429   assert(size < (UINT_MAX & ~0xf)); /* TODO: test 64-bit sizes in all codepaths */
430
431   uint32_t clamped;
432   if (util_lower_clearsize_to_dword(clear_value, (int*)&clear_value_size, &clamped))
433      clear_value = &clamped;
434
435   if (clear_value_size == 12) {
436      si_compute_clear_12bytes_buffer(sctx, dst, offset, size, clear_value, flags, coher);
437      return;
438   }
439
440   uint64_t aligned_size = size & ~3ull;
441   if (aligned_size >= 4) {
442      uint64_t compute_min_size;
443
444      if (sctx->gfx_level <= GFX8) {
445         /* CP DMA clears are terribly slow with GTT on GFX6-8, which can always
446          * happen due to BO evictions.
447          */
448         compute_min_size = 0;
449      } else {
450         /* Use a small enough size because CP DMA is slower than compute with bigger sizes. */
451         compute_min_size = 4 * 1024;
452      }
453
454      /* TODO: use compute for unaligned big sizes */
455      if (method == SI_AUTO_SELECT_CLEAR_METHOD && (
456           clear_value_size > 4 ||
457           (clear_value_size == 4 && offset % 4 == 0 && size > compute_min_size))) {
458         method = SI_COMPUTE_CLEAR_METHOD;
459      }
460      if (method == SI_COMPUTE_CLEAR_METHOD) {
461         si_compute_do_clear_or_copy(sctx, dst, offset, NULL, 0, aligned_size, clear_value,
462                                     clear_value_size, flags, coher);
463      } else {
464         assert(clear_value_size == 4);
465         si_cp_dma_clear_buffer(sctx, &sctx->gfx_cs, dst, offset, aligned_size, *clear_value,
466                                flags, coher, get_cache_policy(sctx, coher, size));
467      }
468
469      offset += aligned_size;
470      size -= aligned_size;
471   }
472
473   /* Handle non-dword alignment. */
474   if (size) {
475      assert(dst);
476      assert(dst->target == PIPE_BUFFER);
477      assert(size < 4);
478
479      sctx->b.buffer_subdata(&sctx->b, dst,
480                             PIPE_MAP_WRITE |
481                             /* TC forbids drivers to invalidate buffers and infer unsychronized mappings,
482                              * so suppress those optimizations. */
483                             (sctx->tc ? TC_TRANSFER_MAP_NO_INFER_UNSYNCHRONIZED |
484                                         TC_TRANSFER_MAP_NO_INVALIDATE : 0),
485                             offset, size, clear_value);
486   }
487}
488
489void si_screen_clear_buffer(struct si_screen *sscreen, struct pipe_resource *dst, uint64_t offset,
490                            uint64_t size, unsigned value, unsigned flags)
491{
492   struct si_context *ctx = si_get_aux_context(sscreen);
493   si_clear_buffer(ctx, dst, offset, size, &value, 4, flags,
494                   SI_COHERENCY_SHADER, SI_AUTO_SELECT_CLEAR_METHOD);
495   si_put_aux_context_flush(sscreen);
496}
497
498static void si_pipe_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst,
499                                 unsigned offset, unsigned size, const void *clear_value,
500                                 int clear_value_size)
501{
502   si_clear_buffer((struct si_context *)ctx, dst, offset, size, (uint32_t *)clear_value,
503                   clear_value_size, SI_OP_SYNC_BEFORE_AFTER, SI_COHERENCY_SHADER,
504                   SI_AUTO_SELECT_CLEAR_METHOD);
505}
506
507void si_copy_buffer(struct si_context *sctx, struct pipe_resource *dst, struct pipe_resource *src,
508                    uint64_t dst_offset, uint64_t src_offset, unsigned size, unsigned flags)
509{
510   if (!size)
511      return;
512
513   enum si_coherency coher = SI_COHERENCY_SHADER;
514   enum si_cache_policy cache_policy = get_cache_policy(sctx, coher, size);
515   uint64_t compute_min_size = 8 * 1024;
516
517   si_improve_sync_flags(sctx, dst, src, &flags);
518
519   /* Only use compute for VRAM copies on dGPUs. */
520   /* TODO: use compute for unaligned big sizes */
521   if (sctx->screen->info.has_dedicated_vram && si_resource(dst)->domains & RADEON_DOMAIN_VRAM &&
522       si_resource(src)->domains & RADEON_DOMAIN_VRAM && size > compute_min_size &&
523       dst_offset % 4 == 0 && src_offset % 4 == 0 && size % 4 == 0) {
524      si_compute_do_clear_or_copy(sctx, dst, dst_offset, src, src_offset, size, NULL, 0,
525                                  flags, coher);
526   } else {
527      si_cp_dma_copy_buffer(sctx, dst, src, dst_offset, src_offset, size,
528                            flags, coher, cache_policy);
529   }
530}
531
532static void
533set_work_size(struct pipe_grid_info *info, unsigned block_x, unsigned block_y, unsigned block_z,
534              unsigned work_x, unsigned work_y, unsigned work_z)
535{
536   info->block[0] = block_x;
537   info->block[1] = block_y;
538   info->block[2] = block_z;
539
540   unsigned work[3] = {work_x, work_y, work_z};
541   for (int i = 0; i < 3; ++i) {
542      info->last_block[i] = work[i] % info->block[i];
543      info->grid[i] = DIV_ROUND_UP(work[i], info->block[i]);
544   }
545}
546
547static void si_launch_grid_internal_images(struct si_context *sctx,
548                                           struct pipe_image_view *images,
549                                           unsigned num_images,
550                                           const struct pipe_grid_info *info,
551                                           void *shader, unsigned flags)
552{
553   struct pipe_image_view saved_image[2] = {};
554   assert(num_images <= ARRAY_SIZE(saved_image));
555
556   for (unsigned i = 0; i < num_images; i++) {
557      assert(sctx->b.screen->is_format_supported(sctx->b.screen, images[i].format,
558                                                 images[i].resource->target,
559                                                 images[i].resource->nr_samples,
560                                                 images[i].resource->nr_storage_samples,
561                                                 PIPE_BIND_SHADER_IMAGE));
562
563      /* Always allow DCC stores on gfx10+. */
564      if (sctx->gfx_level >= GFX10 &&
565          images[i].access & PIPE_IMAGE_ACCESS_WRITE &&
566          !(images[i].access & SI_IMAGE_ACCESS_DCC_OFF))
567         images[i].access |= SI_IMAGE_ACCESS_ALLOW_DCC_STORE;
568
569      /* Simplify the format according to what image stores support. */
570      if (images[i].access & PIPE_IMAGE_ACCESS_WRITE) {
571         images[i].format = util_format_linear(images[i].format); /* SRGB not supported */
572         images[i].format = util_format_luminance_to_red(images[i].format);
573         images[i].format = util_format_intensity_to_red(images[i].format);
574         images[i].format = util_format_rgbx_to_rgba(images[i].format); /* prevent partial writes */
575      }
576
577      /* Save the image. */
578      util_copy_image_view(&saved_image[i], &sctx->images[PIPE_SHADER_COMPUTE].views[i]);
579   }
580
581   /* This might invoke DCC decompression, so do it first. */
582   sctx->b.set_shader_images(&sctx->b, PIPE_SHADER_COMPUTE, 0, num_images, 0, images);
583
584   /* This should be done after set_shader_images. */
585   for (unsigned i = 0; i < num_images; i++) {
586      /* The driver doesn't decompress resources automatically here, so do it manually. */
587      si_decompress_subresource(&sctx->b, images[i].resource, PIPE_MASK_RGBAZS,
588                                images[i].u.tex.level, images[i].u.tex.first_layer,
589                                images[i].u.tex.last_layer,
590                                images[i].access & PIPE_IMAGE_ACCESS_WRITE);
591   }
592
593   /* This must be done before the compute shader. */
594   for (unsigned i = 0; i < num_images; i++) {
595      si_make_CB_shader_coherent(sctx, images[i].resource->nr_samples, true,
596            ((struct si_texture*)images[i].resource)->surface.u.gfx9.color.dcc.pipe_aligned);
597   }
598
599   si_launch_grid_internal(sctx, info, shader, flags | SI_OP_CS_IMAGE);
600
601   /* Restore images. */
602   sctx->b.set_shader_images(&sctx->b, PIPE_SHADER_COMPUTE, 0, num_images, 0, saved_image);
603   for (unsigned i = 0; i < num_images; i++)
604      pipe_resource_reference(&saved_image[i].resource, NULL);
605}
606
607bool si_compute_copy_image(struct si_context *sctx, struct pipe_resource *dst, unsigned dst_level,
608                           struct pipe_resource *src, unsigned src_level, unsigned dstx,
609                           unsigned dsty, unsigned dstz, const struct pipe_box *src_box,
610                           unsigned flags)
611{
612   struct si_texture *ssrc = (struct si_texture*)src;
613   struct si_texture *sdst = (struct si_texture*)dst;
614
615   si_use_compute_copy_for_float_formats(sctx, dst, dst_level);
616
617   /* The compute copy is mandatory for compressed and subsampled formats because the gfx copy
618    * doesn't support them. In all other cases, call si_can_use_compute_blit.
619    *
620    * The format is identical (we only need to check the src format) except compressed formats,
621    * which can be paired with an equivalent integer format.
622    */
623   if (!util_format_is_compressed(src->format) &&
624       !util_format_is_compressed(dst->format) &&
625       !util_format_is_subsampled_422(src->format) &&
626       (!si_can_use_compute_blit(sctx, dst->format, dst->nr_samples, true,
627                                 vi_dcc_enabled(sdst, dst_level)) ||
628        !si_can_use_compute_blit(sctx, src->format, src->nr_samples, false,
629                                 vi_dcc_enabled(ssrc, src_level))))
630      return false;
631
632   enum pipe_format src_format = util_format_linear(src->format);
633   enum pipe_format dst_format = util_format_linear(dst->format);
634   bool is_linear = ssrc->surface.is_linear || sdst->surface.is_linear;
635
636   assert(util_format_is_subsampled_422(src_format) == util_format_is_subsampled_422(dst_format));
637
638   /* Interpret as integer values to avoid NaN issues */
639   if (!vi_dcc_enabled(ssrc, src_level) &&
640       !vi_dcc_enabled(sdst, dst_level) &&
641       src_format == dst_format &&
642       util_format_is_float(src_format) &&
643       !util_format_is_compressed(src_format)) {
644      switch(util_format_get_blocksizebits(src_format)) {
645        case 16:
646          src_format = dst_format = PIPE_FORMAT_R16_UINT;
647          break;
648        case 32:
649          src_format = dst_format = PIPE_FORMAT_R32_UINT;
650          break;
651        case 64:
652          src_format = dst_format = PIPE_FORMAT_R32G32_UINT;
653          break;
654        case 128:
655          src_format = dst_format = PIPE_FORMAT_R32G32B32A32_UINT;
656          break;
657        default:
658          assert(false);
659      }
660   }
661
662   /* Interpret compressed formats as UINT. */
663   struct pipe_box new_box;
664   unsigned src_access = 0, dst_access = 0;
665
666   /* Note that staging copies do compressed<->UINT, so one of the formats is already UINT. */
667   if (util_format_is_compressed(src_format) || util_format_is_compressed(dst_format)) {
668      if (util_format_is_compressed(src_format))
669         src_access |= SI_IMAGE_ACCESS_BLOCK_FORMAT_AS_UINT;
670      if (util_format_is_compressed(dst_format))
671         dst_access |= SI_IMAGE_ACCESS_BLOCK_FORMAT_AS_UINT;
672
673      dstx = util_format_get_nblocksx(dst_format, dstx);
674      dsty = util_format_get_nblocksy(dst_format, dsty);
675
676      new_box.x = util_format_get_nblocksx(src_format, src_box->x);
677      new_box.y = util_format_get_nblocksy(src_format, src_box->y);
678      new_box.z = src_box->z;
679      new_box.width = util_format_get_nblocksx(src_format, src_box->width);
680      new_box.height = util_format_get_nblocksy(src_format, src_box->height);
681      new_box.depth = src_box->depth;
682      src_box = &new_box;
683
684      if (ssrc->surface.bpe == 8)
685         src_format = dst_format = PIPE_FORMAT_R16G16B16A16_UINT; /* 64-bit block */
686      else
687         src_format = dst_format = PIPE_FORMAT_R32G32B32A32_UINT; /* 128-bit block */
688   }
689
690   if (util_format_is_subsampled_422(src_format)) {
691      assert(src_format == dst_format);
692
693      src_access |= SI_IMAGE_ACCESS_BLOCK_FORMAT_AS_UINT;
694      dst_access |= SI_IMAGE_ACCESS_BLOCK_FORMAT_AS_UINT;
695
696      dstx = util_format_get_nblocksx(src_format, dstx);
697
698      new_box = *src_box;
699      new_box.x = util_format_get_nblocksx(src_format, src_box->x);
700      new_box.width = util_format_get_nblocksx(src_format, src_box->width);
701      src_box = &new_box;
702
703      src_format = dst_format = PIPE_FORMAT_R32_UINT;
704
705      /* Interpreting 422 subsampled format (16 bpp) as 32 bpp
706       * should force us to divide src_box->x, dstx and width by 2.
707       * But given that ac_surface allocates this format as 32 bpp
708       * and that surf_size is then modified to pack the values
709       * we must keep the original values to get the correct results.
710       */
711   }
712
713   /* SNORM blitting has precision issues. Use the SINT equivalent instead, which doesn't
714    * force DCC decompression.
715    */
716   if (util_format_is_snorm(dst_format))
717      src_format = dst_format = util_format_snorm_to_sint(dst_format);
718
719   if (src_box->width == 0 || src_box->height == 0 || src_box->depth == 0)
720      return true; /* success - nothing to do */
721
722   struct pipe_image_view image[2] = {0};
723   image[0].resource = src;
724   image[0].shader_access = image[0].access = PIPE_IMAGE_ACCESS_READ | src_access;
725   image[0].format = src_format;
726   image[0].u.tex.level = src_level;
727   image[0].u.tex.first_layer = 0;
728   image[0].u.tex.last_layer = util_max_layer(src, src_level);
729   image[1].resource = dst;
730   image[1].shader_access = image[1].access = PIPE_IMAGE_ACCESS_WRITE | dst_access;
731   image[1].format = dst_format;
732   image[1].u.tex.level = dst_level;
733   image[1].u.tex.first_layer = 0;
734   image[1].u.tex.last_layer = util_max_layer(dst, dst_level);
735
736   struct pipe_grid_info info = {0};
737
738   bool dst_is_1d = dst->target == PIPE_TEXTURE_1D ||
739                    dst->target == PIPE_TEXTURE_1D_ARRAY;
740   bool src_is_1d = src->target == PIPE_TEXTURE_1D ||
741                    src->target == PIPE_TEXTURE_1D_ARRAY;
742   int block_x, block_y;
743   int block_z = 1;
744
745   /* Choose the block dimensions based on the copy area size. */
746   if (src_box->height <= 4) {
747      block_y = util_next_power_of_two(src_box->height);
748      block_x = 64 / block_y;
749   } else if (src_box->width <= 4) {
750      block_x = util_next_power_of_two(src_box->width);
751      block_y = 64 / block_x;
752   } else if (is_linear) {
753      block_x = 64;
754      block_y = 1;
755   } else {
756      block_x = 8;
757      block_y = 8;
758   }
759
760   sctx->cs_user_data[0] = src_box->x | (dstx << 16);
761   sctx->cs_user_data[1] = src_box->y | (dsty << 16);
762   sctx->cs_user_data[2] = src_box->z | (dstz << 16);
763
764   set_work_size(&info, block_x, block_y, block_z,
765                 src_box->width, src_box->height, src_box->depth);
766
767   void **copy_image_cs_ptr = &sctx->cs_copy_image[src_is_1d][dst_is_1d];
768   if (!*copy_image_cs_ptr)
769      *copy_image_cs_ptr = si_create_copy_image_cs(sctx, src_is_1d, dst_is_1d);
770
771   assert(*copy_image_cs_ptr);
772
773   si_launch_grid_internal_images(sctx, image, 2, &info, *copy_image_cs_ptr, flags);
774   return true;
775}
776
777void si_retile_dcc(struct si_context *sctx, struct si_texture *tex)
778{
779   /* Set the DCC buffer. */
780   assert(tex->surface.meta_offset && tex->surface.meta_offset <= UINT_MAX);
781   assert(tex->surface.display_dcc_offset && tex->surface.display_dcc_offset <= UINT_MAX);
782   assert(tex->surface.display_dcc_offset < tex->surface.meta_offset);
783   assert(tex->buffer.bo_size <= UINT_MAX);
784
785   struct pipe_shader_buffer sb = {};
786   sb.buffer = &tex->buffer.b.b;
787   sb.buffer_offset = tex->surface.display_dcc_offset;
788   sb.buffer_size = tex->buffer.bo_size - sb.buffer_offset;
789
790   sctx->cs_user_data[0] = tex->surface.meta_offset - tex->surface.display_dcc_offset;
791   sctx->cs_user_data[1] = (tex->surface.u.gfx9.color.dcc_pitch_max + 1) |
792                           (tex->surface.u.gfx9.color.dcc_height << 16);
793   sctx->cs_user_data[2] = (tex->surface.u.gfx9.color.display_dcc_pitch_max + 1) |
794                           (tex->surface.u.gfx9.color.display_dcc_height << 16);
795
796   /* We have only 1 variant per bpp for now, so expect 32 bpp. */
797   assert(tex->surface.bpe == 4);
798
799   void **shader = &sctx->cs_dcc_retile[tex->surface.u.gfx9.swizzle_mode];
800   if (!*shader)
801      *shader = si_create_dcc_retile_cs(sctx, &tex->surface);
802
803   /* Dispatch compute. */
804   unsigned width = DIV_ROUND_UP(tex->buffer.b.b.width0, tex->surface.u.gfx9.color.dcc_block_width);
805   unsigned height = DIV_ROUND_UP(tex->buffer.b.b.height0, tex->surface.u.gfx9.color.dcc_block_height);
806
807   struct pipe_grid_info info = {};
808   info.block[0] = 8;
809   info.block[1] = 8;
810   info.block[2] = 1;
811   info.last_block[0] = width % info.block[0];
812   info.last_block[1] = height % info.block[1];
813   info.grid[0] = DIV_ROUND_UP(width, info.block[0]);
814   info.grid[1] = DIV_ROUND_UP(height, info.block[1]);
815   info.grid[2] = 1;
816
817   si_launch_grid_internal_ssbos(sctx, &info, *shader, SI_OP_SYNC_BEFORE,
818                                 SI_COHERENCY_CB_META, 1, &sb, 0x1);
819
820   /* Don't flush caches. L2 will be flushed by the kernel fence. */
821}
822
823void gfx9_clear_dcc_msaa(struct si_context *sctx, struct pipe_resource *res, uint32_t clear_value,
824                         unsigned flags, enum si_coherency coher)
825{
826   struct si_texture *tex = (struct si_texture*)res;
827
828   assert(sctx->gfx_level < GFX11);
829
830   /* Set the DCC buffer. */
831   assert(tex->surface.meta_offset && tex->surface.meta_offset <= UINT_MAX);
832   assert(tex->buffer.bo_size <= UINT_MAX);
833
834   struct pipe_shader_buffer sb = {};
835   sb.buffer = &tex->buffer.b.b;
836   sb.buffer_offset = tex->surface.meta_offset;
837   sb.buffer_size = tex->buffer.bo_size - sb.buffer_offset;
838
839   sctx->cs_user_data[0] = (tex->surface.u.gfx9.color.dcc_pitch_max + 1) |
840                           (tex->surface.u.gfx9.color.dcc_height << 16);
841   sctx->cs_user_data[1] = (clear_value & 0xffff) |
842                           ((uint32_t)tex->surface.tile_swizzle << 16);
843
844   /* These variables identify the shader variant. */
845   unsigned swizzle_mode = tex->surface.u.gfx9.swizzle_mode;
846   unsigned bpe_log2 = util_logbase2(tex->surface.bpe);
847   unsigned log2_samples = util_logbase2(tex->buffer.b.b.nr_samples);
848   bool fragments8 = tex->buffer.b.b.nr_storage_samples == 8;
849   bool is_array = tex->buffer.b.b.array_size > 1;
850   void **shader = &sctx->cs_clear_dcc_msaa[swizzle_mode][bpe_log2][fragments8][log2_samples - 2][is_array];
851
852   if (!*shader)
853      *shader = gfx9_create_clear_dcc_msaa_cs(sctx, tex);
854
855   /* Dispatch compute. */
856   unsigned width = DIV_ROUND_UP(tex->buffer.b.b.width0, tex->surface.u.gfx9.color.dcc_block_width);
857   unsigned height = DIV_ROUND_UP(tex->buffer.b.b.height0, tex->surface.u.gfx9.color.dcc_block_height);
858   unsigned depth = DIV_ROUND_UP(tex->buffer.b.b.array_size, tex->surface.u.gfx9.color.dcc_block_depth);
859
860   struct pipe_grid_info info = {};
861   info.block[0] = 8;
862   info.block[1] = 8;
863   info.block[2] = 1;
864   info.last_block[0] = width % info.block[0];
865   info.last_block[1] = height % info.block[1];
866   info.grid[0] = DIV_ROUND_UP(width, info.block[0]);
867   info.grid[1] = DIV_ROUND_UP(height, info.block[1]);
868   info.grid[2] = depth;
869
870   si_launch_grid_internal_ssbos(sctx, &info, *shader, flags, coher, 1, &sb, 0x1);
871}
872
873/* Expand FMASK to make it identity, so that image stores can ignore it. */
874void si_compute_expand_fmask(struct pipe_context *ctx, struct pipe_resource *tex)
875{
876   struct si_context *sctx = (struct si_context *)ctx;
877   bool is_array = tex->target == PIPE_TEXTURE_2D_ARRAY;
878   unsigned log_fragments = util_logbase2(tex->nr_storage_samples);
879   unsigned log_samples = util_logbase2(tex->nr_samples);
880   assert(tex->nr_samples >= 2);
881
882   assert(sctx->gfx_level < GFX11);
883
884   /* EQAA FMASK expansion is unimplemented. */
885   if (tex->nr_samples != tex->nr_storage_samples)
886      return;
887
888   si_make_CB_shader_coherent(sctx, tex->nr_samples, true,
889                              ((struct si_texture*)tex)->surface.u.gfx9.color.dcc.pipe_aligned);
890
891   /* Save states. */
892   struct pipe_image_view saved_image = {0};
893   util_copy_image_view(&saved_image, &sctx->images[PIPE_SHADER_COMPUTE].views[0]);
894
895   /* Bind the image. */
896   struct pipe_image_view image = {0};
897   image.resource = tex;
898   /* Don't set WRITE so as not to trigger FMASK expansion, causing
899    * an infinite loop. */
900   image.shader_access = image.access = PIPE_IMAGE_ACCESS_READ;
901   image.format = util_format_linear(tex->format);
902   if (is_array)
903      image.u.tex.last_layer = tex->array_size - 1;
904
905   ctx->set_shader_images(ctx, PIPE_SHADER_COMPUTE, 0, 1, 0, &image);
906
907   /* Bind the shader. */
908   void **shader = &sctx->cs_fmask_expand[log_samples - 1][is_array];
909   if (!*shader)
910      *shader = si_create_fmask_expand_cs(ctx, tex->nr_samples, is_array);
911
912   /* Dispatch compute. */
913   struct pipe_grid_info info = {0};
914   info.block[0] = 8;
915   info.last_block[0] = tex->width0 % 8;
916   info.block[1] = 8;
917   info.last_block[1] = tex->height0 % 8;
918   info.block[2] = 1;
919   info.grid[0] = DIV_ROUND_UP(tex->width0, 8);
920   info.grid[1] = DIV_ROUND_UP(tex->height0, 8);
921   info.grid[2] = is_array ? tex->array_size : 1;
922
923   si_launch_grid_internal(sctx, &info, *shader, SI_OP_SYNC_BEFORE_AFTER);
924
925   /* Restore previous states. */
926   ctx->set_shader_images(ctx, PIPE_SHADER_COMPUTE, 0, 1, 0, &saved_image);
927   pipe_resource_reference(&saved_image.resource, NULL);
928
929   /* Array of fully expanded FMASK values, arranged by [log2(fragments)][log2(samples)-1]. */
930#define INVALID 0 /* never used */
931   static const uint64_t fmask_expand_values[][4] = {
932      /* samples */
933      /* 2 (8 bpp) 4 (8 bpp)   8 (8-32bpp) 16 (16-64bpp)      fragments */
934      {0x02020202, 0x0E0E0E0E, 0xFEFEFEFE, 0xFFFEFFFE},      /* 1 */
935      {0x02020202, 0xA4A4A4A4, 0xAAA4AAA4, 0xAAAAAAA4},      /* 2 */
936      {INVALID, 0xE4E4E4E4, 0x44443210, 0x4444444444443210}, /* 4 */
937      {INVALID, INVALID, 0x76543210, 0x8888888876543210},    /* 8 */
938   };
939
940   /* Clear FMASK to identity. */
941   struct si_texture *stex = (struct si_texture *)tex;
942   si_clear_buffer(sctx, tex, stex->surface.fmask_offset, stex->surface.fmask_size,
943                   (uint32_t *)&fmask_expand_values[log_fragments][log_samples - 1],
944                   log_fragments >= 2 && log_samples == 4 ? 8 : 4, SI_OP_SYNC_AFTER,
945                   SI_COHERENCY_SHADER, SI_AUTO_SELECT_CLEAR_METHOD);
946}
947
948void si_init_compute_blit_functions(struct si_context *sctx)
949{
950   sctx->b.clear_buffer = si_pipe_clear_buffer;
951}
952
953/* Clear a region of a color surface to a constant value. */
954void si_compute_clear_render_target(struct pipe_context *ctx, struct pipe_surface *dstsurf,
955                                    const union pipe_color_union *color, unsigned dstx,
956                                    unsigned dsty, unsigned width, unsigned height,
957                                    bool render_condition_enabled)
958{
959   struct si_context *sctx = (struct si_context *)ctx;
960   unsigned num_layers = dstsurf->u.tex.last_layer - dstsurf->u.tex.first_layer + 1;
961   unsigned data[4 + sizeof(color->ui)] = {dstx, dsty, dstsurf->u.tex.first_layer, 0};
962
963   if (width == 0 || height == 0)
964      return;
965
966   if (util_format_is_srgb(dstsurf->format)) {
967      union pipe_color_union color_srgb;
968      for (int i = 0; i < 3; i++)
969         color_srgb.f[i] = util_format_linear_to_srgb_float(color->f[i]);
970      color_srgb.f[3] = color->f[3];
971      memcpy(data + 4, color_srgb.ui, sizeof(color->ui));
972   } else {
973      memcpy(data + 4, color->ui, sizeof(color->ui));
974   }
975
976   struct pipe_constant_buffer saved_cb = {};
977   si_get_pipe_constant_buffer(sctx, PIPE_SHADER_COMPUTE, 0, &saved_cb);
978
979   struct pipe_constant_buffer cb = {};
980   cb.buffer_size = sizeof(data);
981   cb.user_buffer = data;
982   ctx->set_constant_buffer(ctx, PIPE_SHADER_COMPUTE, 0, false, &cb);
983
984   struct pipe_image_view image = {0};
985   image.resource = dstsurf->texture;
986   image.shader_access = image.access = PIPE_IMAGE_ACCESS_WRITE;
987   image.format = util_format_linear(dstsurf->format);
988   image.u.tex.level = dstsurf->u.tex.level;
989   image.u.tex.first_layer = 0; /* 3D images ignore first_layer (BASE_ARRAY) */
990   image.u.tex.last_layer = dstsurf->u.tex.last_layer;
991
992   struct pipe_grid_info info = {0};
993   void *shader;
994
995   if (dstsurf->texture->target != PIPE_TEXTURE_1D_ARRAY) {
996      if (!sctx->cs_clear_render_target)
997         sctx->cs_clear_render_target = si_clear_render_target_shader(ctx);
998      shader = sctx->cs_clear_render_target;
999
1000      info.block[0] = 8;
1001      info.last_block[0] = width % 8;
1002      info.block[1] = 8;
1003      info.last_block[1] = height % 8;
1004      info.block[2] = 1;
1005      info.grid[0] = DIV_ROUND_UP(width, 8);
1006      info.grid[1] = DIV_ROUND_UP(height, 8);
1007      info.grid[2] = num_layers;
1008   } else {
1009      if (!sctx->cs_clear_render_target_1d_array)
1010         sctx->cs_clear_render_target_1d_array = si_clear_render_target_shader_1d_array(ctx);
1011      shader = sctx->cs_clear_render_target_1d_array;
1012
1013      info.block[0] = 64;
1014      info.last_block[0] = width % 64;
1015      info.block[1] = 1;
1016      info.block[2] = 1;
1017      info.grid[0] = DIV_ROUND_UP(width, 64);
1018      info.grid[1] = num_layers;
1019      info.grid[2] = 1;
1020   }
1021
1022   si_launch_grid_internal_images(sctx, &image, 1, &info, shader,
1023                                  SI_OP_SYNC_BEFORE_AFTER |
1024                                  (render_condition_enabled ? SI_OP_CS_RENDER_COND_ENABLE : 0));
1025
1026   ctx->set_constant_buffer(ctx, PIPE_SHADER_COMPUTE, 0, true, &saved_cb);
1027}
1028