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
31/* Determine the cache policy. */
32static enum si_cache_policy get_cache_policy(struct si_context *sctx, enum si_coherency coher,
33                                             uint64_t size)
34{
35   if ((sctx->chip_class >= GFX9 && (coher == SI_COHERENCY_CB_META ||
36                                     coher == SI_COHERENCY_DB_META ||
37                                     coher == SI_COHERENCY_CP)) ||
38       (sctx->chip_class >= GFX7 && coher == SI_COHERENCY_SHADER))
39      return L2_LRU; /* it's faster if L2 doesn't evict anything  */
40
41   return L2_BYPASS;
42}
43
44unsigned si_get_flush_flags(struct si_context *sctx, enum si_coherency coher,
45                            enum si_cache_policy cache_policy)
46{
47   switch (coher) {
48   default:
49   case SI_COHERENCY_NONE:
50   case SI_COHERENCY_CP:
51      return 0;
52   case SI_COHERENCY_SHADER:
53      return SI_CONTEXT_INV_SCACHE | SI_CONTEXT_INV_VCACHE |
54             (cache_policy == L2_BYPASS ? SI_CONTEXT_INV_L2 : 0);
55   case SI_COHERENCY_CB_META:
56      return SI_CONTEXT_FLUSH_AND_INV_CB;
57   case SI_COHERENCY_DB_META:
58      return SI_CONTEXT_FLUSH_AND_INV_DB;
59   }
60}
61
62void si_launch_grid_internal(struct si_context *sctx, struct pipe_grid_info *info,
63                             void *shader, unsigned flags)
64{
65
66   /* Wait for previous shaders to finish. */
67   if (flags & SI_OP_SYNC_PS_BEFORE)
68      sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH;
69
70   if (flags & SI_OP_SYNC_CS_BEFORE)
71      sctx->flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
72
73   if (!(flags & SI_OP_CS_IMAGE))
74      sctx->flags |= SI_CONTEXT_PFP_SYNC_ME;
75
76   /* Invalidate L0-L1 caches. */
77   /* sL0 is never invalidated, because src resources don't use it. */
78   if (!(flags & SI_OP_SKIP_CACHE_INV_BEFORE))
79      sctx->flags |= SI_CONTEXT_INV_VCACHE;
80
81   /* Set settings for driver-internal compute dispatches. */
82   sctx->flags &= ~SI_CONTEXT_START_PIPELINE_STATS;
83   sctx->flags |= SI_CONTEXT_STOP_PIPELINE_STATS;
84
85   if (!(flags & SI_OP_CS_RENDER_COND_ENABLE))
86      sctx->render_cond_enabled = false;
87
88   /* Skip decompression to prevent infinite recursion. */
89   sctx->blitter_running = true;
90
91   /* Dispatch compute. */
92   void *saved_cs = sctx->cs_shader_state.program;
93   sctx->b.bind_compute_state(&sctx->b, shader);
94   sctx->b.launch_grid(&sctx->b, info);
95   sctx->b.bind_compute_state(&sctx->b, saved_cs);
96
97   /* Restore default settings. */
98   sctx->flags &= ~SI_CONTEXT_STOP_PIPELINE_STATS;
99   sctx->flags |= SI_CONTEXT_START_PIPELINE_STATS;
100   sctx->render_cond_enabled = sctx->render_cond;
101   sctx->blitter_running = false;
102
103   if (flags & SI_OP_SYNC_AFTER) {
104      sctx->flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
105
106      if (flags & SI_OP_CS_IMAGE) {
107         /* Make sure image stores are visible to CB, which doesn't use L2 on GFX6-8. */
108         sctx->flags |= sctx->chip_class <= GFX8 ? SI_CONTEXT_WB_L2 : 0;
109         /* Make sure image stores are visible to all CUs. */
110         sctx->flags |= SI_CONTEXT_INV_VCACHE;
111      } else {
112         /* Make sure buffer stores are visible to all CUs. */
113         sctx->flags |= SI_CONTEXT_INV_SCACHE | SI_CONTEXT_INV_VCACHE | SI_CONTEXT_PFP_SYNC_ME;
114      }
115   }
116}
117
118void si_launch_grid_internal_ssbos(struct si_context *sctx, struct pipe_grid_info *info,
119                                   void *shader, unsigned flags, enum si_coherency coher,
120                                   unsigned num_buffers, const struct pipe_shader_buffer *buffers,
121                                   unsigned writeable_bitmask)
122{
123   if (!(flags & SI_OP_SKIP_CACHE_INV_BEFORE))
124      sctx->flags |= si_get_flush_flags(sctx, coher, SI_COMPUTE_DST_CACHE_POLICY);
125
126   /* Save states. */
127   struct pipe_shader_buffer saved_sb[3] = {};
128   assert(num_buffers <= ARRAY_SIZE(saved_sb));
129   si_get_shader_buffers(sctx, PIPE_SHADER_COMPUTE, 0, num_buffers, saved_sb);
130
131   unsigned saved_writable_mask = 0;
132   for (unsigned i = 0; i < num_buffers; i++) {
133      if (sctx->const_and_shader_buffers[PIPE_SHADER_COMPUTE].writable_mask &
134          (1u << si_get_shaderbuf_slot(i)))
135         saved_writable_mask |= 1 << i;
136   }
137
138   /* Bind buffers and launch compute. */
139   sctx->b.set_shader_buffers(&sctx->b, PIPE_SHADER_COMPUTE, 0, num_buffers, buffers,
140                              writeable_bitmask);
141   si_launch_grid_internal(sctx, info, shader, flags);
142
143   /* Do cache flushing at the end. */
144   if (get_cache_policy(sctx, coher, 0) == L2_BYPASS) {
145      if (flags & SI_OP_SYNC_AFTER)
146         sctx->flags |= SI_CONTEXT_WB_L2;
147   } else {
148      while (writeable_bitmask)
149         si_resource(buffers[u_bit_scan(&writeable_bitmask)].buffer)->TC_L2_dirty = true;
150   }
151
152   /* Restore states. */
153   sctx->b.set_shader_buffers(&sctx->b, PIPE_SHADER_COMPUTE, 0, num_buffers, saved_sb,
154                              saved_writable_mask);
155   for (int i = 0; i < num_buffers; i++)
156      pipe_resource_reference(&saved_sb[i].buffer, NULL);
157}
158
159/**
160 * Clear a buffer using read-modify-write with a 32-bit write bitmask.
161 * The clear value has 32 bits.
162 */
163void si_compute_clear_buffer_rmw(struct si_context *sctx, struct pipe_resource *dst,
164                                 unsigned dst_offset, unsigned size,
165                                 uint32_t clear_value, uint32_t writebitmask,
166                                 unsigned flags, enum si_coherency coher)
167{
168   assert(dst_offset % 4 == 0);
169   assert(size % 4 == 0);
170
171   assert(dst->target != PIPE_BUFFER || dst_offset + size <= dst->width0);
172
173   /* Use buffer_load_dwordx4 and buffer_store_dwordx4 per thread. */
174   unsigned dwords_per_instruction = 4;
175   unsigned wave_size = sctx->screen->compute_wave_size;
176   unsigned dwords_per_wave = dwords_per_instruction * wave_size;
177
178   unsigned num_dwords = size / 4;
179   unsigned num_instructions = DIV_ROUND_UP(num_dwords, dwords_per_instruction);
180
181   struct pipe_grid_info info = {};
182   info.block[0] = MIN2(wave_size, num_instructions);
183   info.block[1] = 1;
184   info.block[2] = 1;
185   info.grid[0] = DIV_ROUND_UP(num_dwords, dwords_per_wave);
186   info.grid[1] = 1;
187   info.grid[2] = 1;
188
189   struct pipe_shader_buffer sb = {};
190   sb.buffer = dst;
191   sb.buffer_offset = dst_offset;
192   sb.buffer_size = size;
193
194   sctx->cs_user_data[0] = clear_value & writebitmask;
195   sctx->cs_user_data[1] = ~writebitmask;
196
197   if (!sctx->cs_clear_buffer_rmw)
198      sctx->cs_clear_buffer_rmw = si_create_clear_buffer_rmw_cs(&sctx->b);
199
200   si_launch_grid_internal_ssbos(sctx, &info, sctx->cs_clear_buffer_rmw, flags, coher,
201                                 1, &sb, 0x1);
202}
203
204static void si_compute_clear_12bytes_buffer(struct si_context *sctx, struct pipe_resource *dst,
205                                            unsigned dst_offset, unsigned size,
206                                            const uint32_t *clear_value, unsigned flags,
207                                            enum si_coherency coher)
208{
209   struct pipe_context *ctx = &sctx->b;
210
211   assert(dst_offset % 4 == 0);
212   assert(size % 4 == 0);
213   unsigned size_12 = DIV_ROUND_UP(size, 12);
214
215   struct pipe_shader_buffer sb = {0};
216   sb.buffer = dst;
217   sb.buffer_offset = dst_offset;
218   sb.buffer_size = size;
219
220   memcpy(sctx->cs_user_data, clear_value, 12);
221
222   struct pipe_grid_info info = {0};
223
224   if (!sctx->cs_clear_12bytes_buffer)
225      sctx->cs_clear_12bytes_buffer = si_clear_12bytes_buffer_shader(ctx);
226
227   info.block[0] = 64;
228   info.last_block[0] = size_12 % 64;
229   info.block[1] = 1;
230   info.block[2] = 1;
231   info.grid[0] = DIV_ROUND_UP(size_12, 64);
232   info.grid[1] = 1;
233   info.grid[2] = 1;
234
235   si_launch_grid_internal_ssbos(sctx, &info, sctx->cs_clear_12bytes_buffer, flags, coher,
236                                 1, &sb, 0x1);
237}
238
239static void si_compute_do_clear_or_copy(struct si_context *sctx, struct pipe_resource *dst,
240                                        unsigned dst_offset, struct pipe_resource *src,
241                                        unsigned src_offset, unsigned size,
242                                        const uint32_t *clear_value, unsigned clear_value_size,
243                                        unsigned flags, enum si_coherency coher)
244{
245   assert(src_offset % 4 == 0);
246   assert(dst_offset % 4 == 0);
247   assert(size % 4 == 0);
248
249   assert(dst->target != PIPE_BUFFER || dst_offset + size <= dst->width0);
250   assert(!src || src_offset + size <= src->width0);
251
252   /* The memory accesses are coalesced, meaning that the 1st instruction writes
253    * the 1st contiguous block of data for the whole wave, the 2nd instruction
254    * writes the 2nd contiguous block of data, etc.
255    */
256   unsigned dwords_per_thread =
257      src ? SI_COMPUTE_COPY_DW_PER_THREAD : SI_COMPUTE_CLEAR_DW_PER_THREAD;
258   unsigned instructions_per_thread = MAX2(1, dwords_per_thread / 4);
259   unsigned dwords_per_instruction = dwords_per_thread / instructions_per_thread;
260   unsigned wave_size = sctx->screen->compute_wave_size;
261   unsigned dwords_per_wave = dwords_per_thread * wave_size;
262
263   unsigned num_dwords = size / 4;
264   unsigned num_instructions = DIV_ROUND_UP(num_dwords, dwords_per_instruction);
265
266   struct pipe_grid_info info = {};
267   info.block[0] = MIN2(wave_size, num_instructions);
268   info.block[1] = 1;
269   info.block[2] = 1;
270   info.grid[0] = DIV_ROUND_UP(num_dwords, dwords_per_wave);
271   info.grid[1] = 1;
272   info.grid[2] = 1;
273
274   struct pipe_shader_buffer sb[2] = {};
275   sb[0].buffer = dst;
276   sb[0].buffer_offset = dst_offset;
277   sb[0].buffer_size = size;
278
279   bool shader_dst_stream_policy = SI_COMPUTE_DST_CACHE_POLICY != L2_LRU;
280
281   if (src) {
282      sb[1].buffer = src;
283      sb[1].buffer_offset = src_offset;
284      sb[1].buffer_size = size;
285
286      if (!sctx->cs_copy_buffer) {
287         sctx->cs_copy_buffer = si_create_dma_compute_shader(
288            &sctx->b, SI_COMPUTE_COPY_DW_PER_THREAD, shader_dst_stream_policy, true);
289      }
290
291      si_launch_grid_internal_ssbos(sctx, &info, sctx->cs_copy_buffer, flags, coher,
292                                    2, sb, 0x1);
293   } else {
294      assert(clear_value_size >= 4 && clear_value_size <= 16 &&
295             util_is_power_of_two_or_zero(clear_value_size));
296
297      for (unsigned i = 0; i < 4; i++)
298         sctx->cs_user_data[i] = clear_value[i % (clear_value_size / 4)];
299
300      if (!sctx->cs_clear_buffer) {
301         sctx->cs_clear_buffer = si_create_dma_compute_shader(
302            &sctx->b, SI_COMPUTE_CLEAR_DW_PER_THREAD, shader_dst_stream_policy, false);
303      }
304
305      si_launch_grid_internal_ssbos(sctx, &info, sctx->cs_clear_buffer, flags, coher,
306                                    1, sb, 0x1);
307   }
308}
309
310void si_clear_buffer(struct si_context *sctx, struct pipe_resource *dst,
311                     uint64_t offset, uint64_t size, uint32_t *clear_value,
312                     uint32_t clear_value_size, unsigned flags,
313                     enum si_coherency coher, enum si_clear_method method)
314{
315   if (!size)
316      return;
317
318   ASSERTED unsigned clear_alignment = MIN2(clear_value_size, 4);
319
320   assert(clear_value_size != 3 && clear_value_size != 6); /* 12 is allowed. */
321   assert(offset % clear_alignment == 0);
322   assert(size % clear_alignment == 0);
323   assert(size < (UINT_MAX & ~0xf)); /* TODO: test 64-bit sizes in all codepaths */
324
325   uint32_t clamped;
326   if (util_lower_clearsize_to_dword(clear_value, (int*)&clear_value_size, &clamped))
327      clear_value = &clamped;
328
329   if (clear_value_size == 12) {
330      si_compute_clear_12bytes_buffer(sctx, dst, offset, size, clear_value, flags, coher);
331      return;
332   }
333
334   uint64_t aligned_size = size & ~3ull;
335   if (aligned_size >= 4) {
336      uint64_t compute_min_size;
337
338      if (sctx->chip_class <= GFX8) {
339         /* CP DMA clears are terribly slow with GTT on GFX6-8, which can always
340          * happen due to BO evictions.
341          */
342         compute_min_size = 0;
343      } else {
344         /* Use a small enough size because CP DMA is slower than compute with bigger sizes. */
345         compute_min_size = 4 * 1024;
346      }
347
348      if (method == SI_AUTO_SELECT_CLEAR_METHOD && (
349           clear_value_size > 4 ||
350           (clear_value_size == 4 && offset % 4 == 0 && size > compute_min_size))) {
351         method = SI_COMPUTE_CLEAR_METHOD;
352      }
353      if (method == SI_COMPUTE_CLEAR_METHOD) {
354         si_compute_do_clear_or_copy(sctx, dst, offset, NULL, 0, aligned_size, clear_value,
355                                     clear_value_size, flags, coher);
356      } else {
357         assert(clear_value_size == 4);
358         si_cp_dma_clear_buffer(sctx, &sctx->gfx_cs, dst, offset, aligned_size, *clear_value,
359                                flags, coher, get_cache_policy(sctx, coher, size));
360      }
361
362      offset += aligned_size;
363      size -= aligned_size;
364   }
365
366   /* Handle non-dword alignment. */
367   if (size) {
368      assert(dst);
369      assert(dst->target == PIPE_BUFFER);
370      assert(size < 4);
371
372      pipe_buffer_write(&sctx->b, dst, offset, size, clear_value);
373   }
374}
375
376void si_screen_clear_buffer(struct si_screen *sscreen, struct pipe_resource *dst, uint64_t offset,
377                            uint64_t size, unsigned value, unsigned flags)
378{
379   struct si_context *ctx = (struct si_context *)sscreen->aux_context;
380
381   simple_mtx_lock(&sscreen->aux_context_lock);
382   si_clear_buffer(ctx, dst, offset, size, &value, 4, flags,
383                   SI_COHERENCY_SHADER, SI_AUTO_SELECT_CLEAR_METHOD);
384   sscreen->aux_context->flush(sscreen->aux_context, NULL, 0);
385   simple_mtx_unlock(&sscreen->aux_context_lock);
386}
387
388static void si_pipe_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst,
389                                 unsigned offset, unsigned size, const void *clear_value,
390                                 int clear_value_size)
391{
392   si_clear_buffer((struct si_context *)ctx, dst, offset, size, (uint32_t *)clear_value,
393                   clear_value_size, SI_OP_SYNC_BEFORE_AFTER, SI_COHERENCY_SHADER,
394                   SI_AUTO_SELECT_CLEAR_METHOD);
395}
396
397void si_copy_buffer(struct si_context *sctx, struct pipe_resource *dst, struct pipe_resource *src,
398                    uint64_t dst_offset, uint64_t src_offset, unsigned size, unsigned flags)
399{
400   if (!size)
401      return;
402
403   enum si_coherency coher = SI_COHERENCY_SHADER;
404   enum si_cache_policy cache_policy = get_cache_policy(sctx, coher, size);
405   uint64_t compute_min_size = 8 * 1024;
406
407   /* Only use compute for VRAM copies on dGPUs. */
408   if (sctx->screen->info.has_dedicated_vram && si_resource(dst)->domains & RADEON_DOMAIN_VRAM &&
409       si_resource(src)->domains & RADEON_DOMAIN_VRAM && size > compute_min_size &&
410       dst_offset % 4 == 0 && src_offset % 4 == 0 && size % 4 == 0) {
411      si_compute_do_clear_or_copy(sctx, dst, dst_offset, src, src_offset, size, NULL, 0,
412                                  flags, coher);
413   } else {
414      si_cp_dma_copy_buffer(sctx, dst, src, dst_offset, src_offset, size,
415                            flags, coher, cache_policy);
416   }
417}
418
419void si_compute_copy_image(struct si_context *sctx, struct pipe_resource *dst, unsigned dst_level,
420                           struct pipe_resource *src, unsigned src_level, unsigned dstx,
421                           unsigned dsty, unsigned dstz, const struct pipe_box *src_box,
422                           bool is_dcc_decompress, unsigned flags)
423{
424   struct pipe_context *ctx = &sctx->b;
425   struct si_texture *ssrc = (struct si_texture*)src;
426   struct si_texture *sdst = (struct si_texture*)dst;
427   unsigned width = src_box->width;
428   unsigned height = src_box->height;
429   unsigned depth = src_box->depth;
430   enum pipe_format src_format = util_format_linear(src->format);
431   enum pipe_format dst_format = util_format_linear(dst->format);
432   bool is_linear = ssrc->surface.is_linear || sdst->surface.is_linear;
433
434   assert(util_format_is_subsampled_422(src_format) == util_format_is_subsampled_422(dst_format));
435
436   if (!vi_dcc_enabled(ssrc, src_level) &&
437       !vi_dcc_enabled(sdst, dst_level) &&
438       src_format == dst_format &&
439       util_format_is_float(src_format) &&
440       !util_format_is_compressed(src_format)) {
441      /* Interpret as integer values to avoid NaN issues */
442      switch(util_format_get_blocksizebits(src_format)) {
443        case 16:
444          src_format = dst_format = PIPE_FORMAT_R16_UINT;
445          break;
446        case 32:
447          src_format = dst_format = PIPE_FORMAT_R32_UINT;
448          break;
449        case 64:
450          src_format = dst_format = PIPE_FORMAT_R32G32_UINT;
451          break;
452        case 128:
453          src_format = dst_format = PIPE_FORMAT_R32G32B32A32_UINT;
454          break;
455        default:
456          assert(false);
457      }
458   }
459
460   if (util_format_is_subsampled_422(src_format)) {
461      src_format = dst_format = PIPE_FORMAT_R32_UINT;
462      /* Interpreting 422 subsampled format (16 bpp) as 32 bpp
463       * should force us to divide src_box->x, dstx and width by 2.
464       * But given that ac_surface allocates this format as 32 bpp
465       * and that surf_size is then modified to pack the values
466       * we must keep the original values to get the correct results.
467       */
468   }
469
470   if (width == 0 || height == 0)
471      return;
472
473   /* The driver doesn't decompress resources automatically here. */
474   si_decompress_subresource(ctx, dst, PIPE_MASK_RGBAZS, dst_level, dstz,
475                             dstz + src_box->depth - 1);
476   si_decompress_subresource(ctx, src, PIPE_MASK_RGBAZS, src_level, src_box->z,
477                             src_box->z + src_box->depth - 1);
478
479   /* src and dst have the same number of samples. */
480   si_make_CB_shader_coherent(sctx, src->nr_samples, true,
481                              ssrc->surface.u.gfx9.color.dcc.pipe_aligned);
482   if (sctx->chip_class >= GFX10) {
483      /* GFX10+ uses DCC stores so si_make_CB_shader_coherent is required for dst too */
484      si_make_CB_shader_coherent(sctx, dst->nr_samples, true,
485                                 sdst->surface.u.gfx9.color.dcc.pipe_aligned);
486   }
487
488   struct si_images *images = &sctx->images[PIPE_SHADER_COMPUTE];
489   struct pipe_image_view saved_image[2] = {0};
490   util_copy_image_view(&saved_image[0], &images->views[0]);
491   util_copy_image_view(&saved_image[1], &images->views[1]);
492
493   struct pipe_image_view image[2] = {0};
494   image[0].resource = src;
495   image[0].shader_access = image[0].access = PIPE_IMAGE_ACCESS_READ;
496   image[0].format = src_format;
497   image[0].u.tex.level = src_level;
498   image[0].u.tex.first_layer = 0;
499   image[0].u.tex.last_layer = src->target == PIPE_TEXTURE_3D ? u_minify(src->depth0, src_level) - 1
500                                                              : (unsigned)(src->array_size - 1);
501   image[1].resource = dst;
502   image[1].shader_access = image[1].access = PIPE_IMAGE_ACCESS_WRITE;
503   image[1].format = dst_format;
504   image[1].u.tex.level = dst_level;
505   image[1].u.tex.first_layer = 0;
506   image[1].u.tex.last_layer = dst->target == PIPE_TEXTURE_3D ? u_minify(dst->depth0, dst_level) - 1
507                                                              : (unsigned)(dst->array_size - 1);
508
509   /* SNORM8 blitting has precision issues on some chips. Use the SINT
510    * equivalent instead, which doesn't force DCC decompression.
511    */
512   if (util_format_is_snorm8(dst->format)) {
513      image[0].format = image[1].format = util_format_snorm8_to_sint8(dst->format);
514   }
515
516   if (is_dcc_decompress)
517      image[1].access |= SI_IMAGE_ACCESS_DCC_OFF;
518   else if (sctx->chip_class >= GFX10)
519      image[1].access |= SI_IMAGE_ACCESS_ALLOW_DCC_STORE;
520
521   ctx->set_shader_images(ctx, PIPE_SHADER_COMPUTE, 0, 2, 0, image);
522
523   if (!is_dcc_decompress) {
524      sctx->cs_user_data[0] = src_box->x | (dstx << 16);
525      sctx->cs_user_data[1] = src_box->y | (dsty << 16);
526      sctx->cs_user_data[2] = src_box->z | (dstz << 16);
527   }
528
529   struct pipe_grid_info info = {0};
530
531   if (is_dcc_decompress) {
532      /* The DCC decompression is a normal blit where the load is compressed
533       * and the store is uncompressed. The workgroup size is either equal to
534       * the DCC block size or a multiple thereof. The shader uses a barrier
535       * between loads and stores to safely overwrite each DCC block of pixels.
536       */
537      unsigned dim[3] = {src_box->width, src_box->height, src_box->depth};
538
539      assert(src == dst);
540      assert(dst->target != PIPE_TEXTURE_1D && dst->target != PIPE_TEXTURE_1D_ARRAY);
541
542      if (!sctx->cs_dcc_decompress)
543         sctx->cs_dcc_decompress = si_create_dcc_decompress_cs(ctx);
544
545      info.block[0] = ssrc->surface.u.gfx9.color.dcc_block_width;
546      info.block[1] = ssrc->surface.u.gfx9.color.dcc_block_height;
547      info.block[2] = ssrc->surface.u.gfx9.color.dcc_block_depth;
548
549      /* Make sure the block size is at least the same as wave size. */
550      while (info.block[0] * info.block[1] * info.block[2] <
551             sctx->screen->compute_wave_size) {
552         info.block[0] *= 2;
553      }
554
555      for (unsigned i = 0; i < 3; i++) {
556         info.last_block[i] = dim[i] % info.block[i];
557         info.grid[i] = DIV_ROUND_UP(dim[i], info.block[i]);
558      }
559
560      si_launch_grid_internal(sctx, &info, sctx->cs_dcc_decompress, flags | SI_OP_CS_IMAGE);
561   } else if (dst->target == PIPE_TEXTURE_1D_ARRAY && src->target == PIPE_TEXTURE_1D_ARRAY) {
562      if (!sctx->cs_copy_image_1d_array)
563         sctx->cs_copy_image_1d_array = si_create_copy_image_compute_shader_1d_array(ctx);
564
565      info.block[0] = 64;
566      info.last_block[0] = width % 64;
567      info.block[1] = 1;
568      info.block[2] = 1;
569      info.grid[0] = DIV_ROUND_UP(width, 64);
570      info.grid[1] = depth;
571      info.grid[2] = 1;
572
573      si_launch_grid_internal(sctx, &info, sctx->cs_copy_image_1d_array, flags | SI_OP_CS_IMAGE);
574   } else {
575      if (!sctx->cs_copy_image)
576         sctx->cs_copy_image = si_create_copy_image_compute_shader(ctx);
577
578      /* This is better for access over PCIe. */
579      if (is_linear) {
580         info.block[0] = 64;
581         info.block[1] = 1;
582      } else {
583         info.block[0] = 8;
584         info.block[1] = 8;
585      }
586      info.last_block[0] = width % info.block[0];
587      info.last_block[1] = height % info.block[1];
588      info.block[2] = 1;
589      info.grid[0] = DIV_ROUND_UP(width, info.block[0]);
590      info.grid[1] = DIV_ROUND_UP(height, info.block[1]);
591      info.grid[2] = depth;
592
593      si_launch_grid_internal(sctx, &info, sctx->cs_copy_image, flags | SI_OP_CS_IMAGE);
594   }
595
596   ctx->set_shader_images(ctx, PIPE_SHADER_COMPUTE, 0, 2, 0, saved_image);
597   for (int i = 0; i < 2; i++)
598      pipe_resource_reference(&saved_image[i].resource, NULL);
599}
600
601void si_retile_dcc(struct si_context *sctx, struct si_texture *tex)
602{
603   /* Set the DCC buffer. */
604   assert(tex->surface.meta_offset && tex->surface.meta_offset <= UINT_MAX);
605   assert(tex->surface.display_dcc_offset && tex->surface.display_dcc_offset <= UINT_MAX);
606   assert(tex->surface.display_dcc_offset < tex->surface.meta_offset);
607   assert(tex->buffer.bo_size <= UINT_MAX);
608
609   struct pipe_shader_buffer sb = {};
610   sb.buffer = &tex->buffer.b.b;
611   sb.buffer_offset = tex->surface.display_dcc_offset;
612   sb.buffer_size = tex->buffer.bo_size - sb.buffer_offset;
613
614   sctx->cs_user_data[0] = tex->surface.meta_offset - tex->surface.display_dcc_offset;
615   sctx->cs_user_data[1] = (tex->surface.u.gfx9.color.dcc_pitch_max + 1) |
616                           (tex->surface.u.gfx9.color.dcc_height << 16);
617   sctx->cs_user_data[2] = (tex->surface.u.gfx9.color.display_dcc_pitch_max + 1) |
618                           (tex->surface.u.gfx9.color.display_dcc_height << 16);
619
620   /* We have only 1 variant per bpp for now, so expect 32 bpp. */
621   assert(tex->surface.bpe == 4);
622
623   void **shader = &sctx->cs_dcc_retile[tex->surface.u.gfx9.swizzle_mode];
624   if (!*shader)
625      *shader = si_create_dcc_retile_cs(sctx, &tex->surface);
626
627   /* Dispatch compute. */
628   unsigned width = DIV_ROUND_UP(tex->buffer.b.b.width0, tex->surface.u.gfx9.color.dcc_block_width);
629   unsigned height = DIV_ROUND_UP(tex->buffer.b.b.height0, tex->surface.u.gfx9.color.dcc_block_height);
630
631   struct pipe_grid_info info = {};
632   info.block[0] = 8;
633   info.block[1] = 8;
634   info.block[2] = 1;
635   info.last_block[0] = width % info.block[0];
636   info.last_block[1] = height % info.block[1];
637   info.grid[0] = DIV_ROUND_UP(width, info.block[0]);
638   info.grid[1] = DIV_ROUND_UP(height, info.block[1]);
639   info.grid[2] = 1;
640
641   si_launch_grid_internal_ssbos(sctx, &info, *shader, SI_OP_SYNC_BEFORE,
642                                 SI_COHERENCY_CB_META, 1, &sb, 0x1);
643
644   /* Don't flush caches. L2 will be flushed by the kernel fence. */
645}
646
647void gfx9_clear_dcc_msaa(struct si_context *sctx, struct pipe_resource *res, uint32_t clear_value,
648                         unsigned flags, enum si_coherency coher)
649{
650   struct si_texture *tex = (struct si_texture*)res;
651
652   /* Set the DCC buffer. */
653   assert(tex->surface.meta_offset && tex->surface.meta_offset <= UINT_MAX);
654   assert(tex->buffer.bo_size <= UINT_MAX);
655
656   struct pipe_shader_buffer sb = {};
657   sb.buffer = &tex->buffer.b.b;
658   sb.buffer_offset = tex->surface.meta_offset;
659   sb.buffer_size = tex->buffer.bo_size - sb.buffer_offset;
660
661   sctx->cs_user_data[0] = (tex->surface.u.gfx9.color.dcc_pitch_max + 1) |
662                           (tex->surface.u.gfx9.color.dcc_height << 16);
663   sctx->cs_user_data[1] = (clear_value & 0xffff) |
664                           ((uint32_t)tex->surface.tile_swizzle << 16);
665
666   /* These variables identify the shader variant. */
667   unsigned swizzle_mode = tex->surface.u.gfx9.swizzle_mode;
668   unsigned bpe_log2 = util_logbase2(tex->surface.bpe);
669   unsigned log2_samples = util_logbase2(tex->buffer.b.b.nr_samples);
670   bool fragments8 = tex->buffer.b.b.nr_storage_samples == 8;
671   bool is_array = tex->buffer.b.b.array_size > 1;
672   void **shader = &sctx->cs_clear_dcc_msaa[swizzle_mode][bpe_log2][fragments8][log2_samples - 2][is_array];
673
674   if (!*shader)
675      *shader = gfx9_create_clear_dcc_msaa_cs(sctx, tex);
676
677   /* Dispatch compute. */
678   unsigned width = DIV_ROUND_UP(tex->buffer.b.b.width0, tex->surface.u.gfx9.color.dcc_block_width);
679   unsigned height = DIV_ROUND_UP(tex->buffer.b.b.height0, tex->surface.u.gfx9.color.dcc_block_height);
680   unsigned depth = DIV_ROUND_UP(tex->buffer.b.b.array_size, tex->surface.u.gfx9.color.dcc_block_depth);
681
682   struct pipe_grid_info info = {};
683   info.block[0] = 8;
684   info.block[1] = 8;
685   info.block[2] = 1;
686   info.last_block[0] = width % info.block[0];
687   info.last_block[1] = height % info.block[1];
688   info.grid[0] = DIV_ROUND_UP(width, info.block[0]);
689   info.grid[1] = DIV_ROUND_UP(height, info.block[1]);
690   info.grid[2] = depth;
691
692   si_launch_grid_internal_ssbos(sctx, &info, *shader, flags, coher, 1, &sb, 0x1);
693}
694
695/* Expand FMASK to make it identity, so that image stores can ignore it. */
696void si_compute_expand_fmask(struct pipe_context *ctx, struct pipe_resource *tex)
697{
698   struct si_context *sctx = (struct si_context *)ctx;
699   bool is_array = tex->target == PIPE_TEXTURE_2D_ARRAY;
700   unsigned log_fragments = util_logbase2(tex->nr_storage_samples);
701   unsigned log_samples = util_logbase2(tex->nr_samples);
702   assert(tex->nr_samples >= 2);
703
704   /* EQAA FMASK expansion is unimplemented. */
705   if (tex->nr_samples != tex->nr_storage_samples)
706      return;
707
708   si_make_CB_shader_coherent(sctx, tex->nr_samples, true,
709                              ((struct si_texture*)tex)->surface.u.gfx9.color.dcc.pipe_aligned);
710
711   /* Save states. */
712   struct pipe_image_view saved_image = {0};
713   util_copy_image_view(&saved_image, &sctx->images[PIPE_SHADER_COMPUTE].views[0]);
714
715   /* Bind the image. */
716   struct pipe_image_view image = {0};
717   image.resource = tex;
718   /* Don't set WRITE so as not to trigger FMASK expansion, causing
719    * an infinite loop. */
720   image.shader_access = image.access = PIPE_IMAGE_ACCESS_READ;
721   image.format = util_format_linear(tex->format);
722   if (is_array)
723      image.u.tex.last_layer = tex->array_size - 1;
724
725   ctx->set_shader_images(ctx, PIPE_SHADER_COMPUTE, 0, 1, 0, &image);
726
727   /* Bind the shader. */
728   void **shader = &sctx->cs_fmask_expand[log_samples - 1][is_array];
729   if (!*shader)
730      *shader = si_create_fmask_expand_cs(ctx, tex->nr_samples, is_array);
731
732   /* Dispatch compute. */
733   struct pipe_grid_info info = {0};
734   info.block[0] = 8;
735   info.last_block[0] = tex->width0 % 8;
736   info.block[1] = 8;
737   info.last_block[1] = tex->height0 % 8;
738   info.block[2] = 1;
739   info.grid[0] = DIV_ROUND_UP(tex->width0, 8);
740   info.grid[1] = DIV_ROUND_UP(tex->height0, 8);
741   info.grid[2] = is_array ? tex->array_size : 1;
742
743   si_launch_grid_internal(sctx, &info, *shader, SI_OP_SYNC_BEFORE_AFTER);
744
745   /* Restore previous states. */
746   ctx->set_shader_images(ctx, PIPE_SHADER_COMPUTE, 0, 1, 0, &saved_image);
747   pipe_resource_reference(&saved_image.resource, NULL);
748
749   /* Array of fully expanded FMASK values, arranged by [log2(fragments)][log2(samples)-1]. */
750#define INVALID 0 /* never used */
751   static const uint64_t fmask_expand_values[][4] = {
752      /* samples */
753      /* 2 (8 bpp) 4 (8 bpp)   8 (8-32bpp) 16 (16-64bpp)      fragments */
754      {0x02020202, 0x0E0E0E0E, 0xFEFEFEFE, 0xFFFEFFFE},      /* 1 */
755      {0x02020202, 0xA4A4A4A4, 0xAAA4AAA4, 0xAAAAAAA4},      /* 2 */
756      {INVALID, 0xE4E4E4E4, 0x44443210, 0x4444444444443210}, /* 4 */
757      {INVALID, INVALID, 0x76543210, 0x8888888876543210},    /* 8 */
758   };
759
760   /* Clear FMASK to identity. */
761   struct si_texture *stex = (struct si_texture *)tex;
762   si_clear_buffer(sctx, tex, stex->surface.fmask_offset, stex->surface.fmask_size,
763                   (uint32_t *)&fmask_expand_values[log_fragments][log_samples - 1],
764                   log_fragments >= 2 && log_samples == 4 ? 8 : 4, SI_OP_SYNC_AFTER,
765                   SI_COHERENCY_SHADER, SI_AUTO_SELECT_CLEAR_METHOD);
766}
767
768void si_init_compute_blit_functions(struct si_context *sctx)
769{
770   sctx->b.clear_buffer = si_pipe_clear_buffer;
771}
772
773/* Clear a region of a color surface to a constant value. */
774void si_compute_clear_render_target(struct pipe_context *ctx, struct pipe_surface *dstsurf,
775                                    const union pipe_color_union *color, unsigned dstx,
776                                    unsigned dsty, unsigned width, unsigned height,
777                                    bool render_condition_enabled)
778{
779   struct si_context *sctx = (struct si_context *)ctx;
780   struct si_texture *tex = (struct si_texture*)dstsurf->texture;
781   unsigned num_layers = dstsurf->u.tex.last_layer - dstsurf->u.tex.first_layer + 1;
782   unsigned data[4 + sizeof(color->ui)] = {dstx, dsty, dstsurf->u.tex.first_layer, 0};
783
784   if (width == 0 || height == 0)
785      return;
786
787   /* The driver doesn't decompress resources automatically here. */
788   si_decompress_subresource(ctx, dstsurf->texture, PIPE_MASK_RGBA, dstsurf->u.tex.level,
789                             dstsurf->u.tex.first_layer, dstsurf->u.tex.last_layer);
790
791   if (util_format_is_srgb(dstsurf->format)) {
792      union pipe_color_union color_srgb;
793      for (int i = 0; i < 3; i++)
794         color_srgb.f[i] = util_format_linear_to_srgb_float(color->f[i]);
795      color_srgb.f[3] = color->f[3];
796      memcpy(data + 4, color_srgb.ui, sizeof(color->ui));
797   } else {
798      memcpy(data + 4, color->ui, sizeof(color->ui));
799   }
800
801   si_make_CB_shader_coherent(sctx, dstsurf->texture->nr_samples, true,
802                              tex->surface.u.gfx9.color.dcc.pipe_aligned);
803
804   struct pipe_constant_buffer saved_cb = {};
805   si_get_pipe_constant_buffer(sctx, PIPE_SHADER_COMPUTE, 0, &saved_cb);
806
807   struct si_images *images = &sctx->images[PIPE_SHADER_COMPUTE];
808   struct pipe_image_view saved_image = {0};
809   util_copy_image_view(&saved_image, &images->views[0]);
810
811   struct pipe_constant_buffer cb = {};
812   cb.buffer_size = sizeof(data);
813   cb.user_buffer = data;
814   ctx->set_constant_buffer(ctx, PIPE_SHADER_COMPUTE, 0, false, &cb);
815
816   struct pipe_image_view image = {0};
817   image.resource = dstsurf->texture;
818   image.shader_access = image.access = PIPE_IMAGE_ACCESS_WRITE | SI_IMAGE_ACCESS_ALLOW_DCC_STORE;
819   image.format = util_format_linear(dstsurf->format);
820   image.u.tex.level = dstsurf->u.tex.level;
821   image.u.tex.first_layer = 0; /* 3D images ignore first_layer (BASE_ARRAY) */
822   image.u.tex.last_layer = dstsurf->u.tex.last_layer;
823
824   ctx->set_shader_images(ctx, PIPE_SHADER_COMPUTE, 0, 1, 0, &image);
825
826   struct pipe_grid_info info = {0};
827   void *shader;
828
829   if (dstsurf->texture->target != PIPE_TEXTURE_1D_ARRAY) {
830      if (!sctx->cs_clear_render_target)
831         sctx->cs_clear_render_target = si_clear_render_target_shader(ctx);
832      shader = sctx->cs_clear_render_target;
833
834      info.block[0] = 8;
835      info.last_block[0] = width % 8;
836      info.block[1] = 8;
837      info.last_block[1] = height % 8;
838      info.block[2] = 1;
839      info.grid[0] = DIV_ROUND_UP(width, 8);
840      info.grid[1] = DIV_ROUND_UP(height, 8);
841      info.grid[2] = num_layers;
842   } else {
843      if (!sctx->cs_clear_render_target_1d_array)
844         sctx->cs_clear_render_target_1d_array = si_clear_render_target_shader_1d_array(ctx);
845      shader = sctx->cs_clear_render_target_1d_array;
846
847      info.block[0] = 64;
848      info.last_block[0] = width % 64;
849      info.block[1] = 1;
850      info.block[2] = 1;
851      info.grid[0] = DIV_ROUND_UP(width, 64);
852      info.grid[1] = num_layers;
853      info.grid[2] = 1;
854   }
855
856   si_launch_grid_internal(sctx, &info, shader, SI_OP_SYNC_BEFORE_AFTER | SI_OP_CS_IMAGE |
857                           (render_condition_enabled ? SI_OP_CS_RENDER_COND_ENABLE : 0));
858
859   ctx->set_shader_images(ctx, PIPE_SHADER_COMPUTE, 0, 1, 0, &saved_image);
860   ctx->set_constant_buffer(ctx, PIPE_SHADER_COMPUTE, 0, true, &saved_cb);
861   pipe_resource_reference(&saved_image.resource, NULL);
862}
863