src/gallium/drivers/radeonsi/si_cp_dma.c

   1 /*
   2  * Copyright 2013 Advanced Micro Devices, Inc.
   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  * on the rights to use, copy, modify, merge, publish, distribute, sub
   8  * license, and/or sell copies of the Software, and to permit persons to whom
   9  * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
  18  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 #include "si_pipe.h"
  25 #include "sid.h"
  26 #include "radeon/r600_cs.h"
  27
  28 /* Recommended maximum sizes for optimal performance.
  29  * Fall back to compute or SDMA if the size is greater.
  30  */
  31 #define CP_DMA_COPY_PERF_THRESHOLD      (64 * 1024) /* copied from Vulkan */
  32 #define CP_DMA_CLEAR_PERF_THRESHOLD     (32 * 1024) /* guess (clear is much slower) */
  33
  34 /* Set this if you want the ME to wait until CP DMA is done.
  35  * It should be set on the last CP DMA packet. */
  36 #define CP_DMA_SYNC             (1 << 0)
  37
  38 /* Set this if the source data was used as a destination in a previous CP DMA
  39  * packet. It's for preventing a read-after-write (RAW) hazard between two
  40  * CP DMA packets. */
  41 #define CP_DMA_RAW_WAIT         (1 << 1)
  42 #define CP_DMA_USE_L2           (1 << 2) /* CIK+ */
  43 #define CP_DMA_CLEAR            (1 << 3)
  44
  45 /* The max number of bytes that can be copied per packet. */
  46 static inline unsigned cp_dma_max_byte_count(struct si_context *sctx)
  47 {
  48         unsigned max = sctx->b.chip_class >= GFX9 ?
  49                                S_414_BYTE_COUNT_GFX9(~0u) :
  50                                S_414_BYTE_COUNT_GFX6(~0u);
  51
  52         /* make it aligned for optimal performance */
  53         return max & ~(SI_CPDMA_ALIGNMENT - 1);
  54 }
  55
  56
  57 /* Emit a CP DMA packet to do a copy from one buffer to another, or to clear
  58  * a buffer. The size must fit in bits [20:0]. If CP_DMA_CLEAR is set, src_va is a 32-bit
  59  * clear value.
  60  */
  61 static void si_emit_cp_dma(struct si_context *sctx, uint64_t dst_va,
  62                            uint64_t src_va, unsigned size, unsigned flags,
  63                            enum r600_coherency coher)
  64 {
  65         struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
  66         uint32_t header = 0, command = 0;
  67
  68         assert(size);
  69         assert(size <= cp_dma_max_byte_count(sctx));
  70
  71         if (sctx->b.chip_class >= GFX9)
  72                 command |= S_414_BYTE_COUNT_GFX9(size);
  73         else
  74                 command |= S_414_BYTE_COUNT_GFX6(size);
  75
  76         /* Sync flags. */
  77         if (flags & CP_DMA_SYNC)
  78                 header |= S_411_CP_SYNC(1);
  79         else {
  80                 if (sctx->b.chip_class >= GFX9)
  81                         command |= S_414_DISABLE_WR_CONFIRM_GFX9(1);
  82                 else
  83                         command |= S_414_DISABLE_WR_CONFIRM_GFX6(1);
  84         }
  85
  86         if (flags & CP_DMA_RAW_WAIT)
  87                 command |= S_414_RAW_WAIT(1);
  88
  89         /* Src and dst flags. */
  90         if (sctx->b.chip_class >= GFX9 && !(flags & CP_DMA_CLEAR) &&
  91             src_va == dst_va)
  92                 header |= S_411_DSL_SEL(V_411_NOWHERE); /* prefetch only */
  93         else if (flags & CP_DMA_USE_L2)
  94                 header |= S_411_DSL_SEL(V_411_DST_ADDR_TC_L2);
  95
  96         if (flags & CP_DMA_CLEAR)
  97                 header |= S_411_SRC_SEL(V_411_DATA);
  98         else if (flags & CP_DMA_USE_L2)
  99                 header |= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2);
 100
 101         if (sctx->b.chip_class >= CIK) {
 102                 radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, 0));
 103                 radeon_emit(cs, header);
 104                 radeon_emit(cs, src_va);        /* SRC_ADDR_LO [31:0] */
 105                 radeon_emit(cs, src_va >> 32);  /* SRC_ADDR_HI [31:0] */
 106                 radeon_emit(cs, dst_va);        /* DST_ADDR_LO [31:0] */
 107                 radeon_emit(cs, dst_va >> 32);  /* DST_ADDR_HI [31:0] */
 108                 radeon_emit(cs, command);
 109         } else {
 110                 header |= S_411_SRC_ADDR_HI(src_va >> 32);
 111
 112                 radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, 0));
 113                 radeon_emit(cs, src_va);        /* SRC_ADDR_LO [31:0] */
 114                 radeon_emit(cs, header);        /* SRC_ADDR_HI [15:0] + flags. */
 115                 radeon_emit(cs, dst_va);        /* DST_ADDR_LO [31:0] */
 116                 radeon_emit(cs, (dst_va >> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
 117                 radeon_emit(cs, command);
 118         }
 119
 120         /* CP DMA is executed in ME, but index buffers are read by PFP.
 121          * This ensures that ME (CP DMA) is idle before PFP starts fetching
 122          * indices. If we wanted to execute CP DMA in PFP, this packet
 123          * should precede it.
 124          */
 125         if (coher == R600_COHERENCY_SHADER && flags & CP_DMA_SYNC) {
 126                 radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
 127                 radeon_emit(cs, 0);
 128         }
 129 }
 130
 131 static unsigned get_flush_flags(struct si_context *sctx, enum r600_coherency coher)
 132 {
 133         switch (coher) {
 134         default:
 135         case R600_COHERENCY_NONE:
 136                 return 0;
 137         case R600_COHERENCY_SHADER:
 138                 return SI_CONTEXT_INV_SMEM_L1 |
 139                        SI_CONTEXT_INV_VMEM_L1 |
 140                        (sctx->b.chip_class == SI ? SI_CONTEXT_INV_GLOBAL_L2 : 0);
 141         case R600_COHERENCY_CB_META:
 142                 return SI_CONTEXT_FLUSH_AND_INV_CB;
 143         }
 144 }
 145
 146 static unsigned get_tc_l2_flag(struct si_context *sctx, enum r600_coherency coher)
 147 {
 148         if ((sctx->b.chip_class >= GFX9 && coher == R600_COHERENCY_CB_META) ||
 149             (sctx->b.chip_class >= CIK && coher == R600_COHERENCY_SHADER))
 150                 return CP_DMA_USE_L2;
 151
 152         return 0;
 153 }
 154
 155 static void si_cp_dma_prepare(struct si_context *sctx, struct pipe_resource *dst,
 156                               struct pipe_resource *src, unsigned byte_count,
 157                               uint64_t remaining_size, unsigned user_flags,
 158                               bool *is_first, unsigned *packet_flags)
 159 {
 160         /* Fast exit for a CPDMA prefetch. */
 161         if ((user_flags & SI_CPDMA_SKIP_ALL) == SI_CPDMA_SKIP_ALL) {
 162                 *is_first = false;
 163                 return;
 164         }
 165
 166         if (!(user_flags & SI_CPDMA_SKIP_BO_LIST_UPDATE)) {
 167                 /* Count memory usage in so that need_cs_space can take it into account. */
 168                 r600_context_add_resource_size(&sctx->b.b, dst);
 169                 if (src)
 170                         r600_context_add_resource_size(&sctx->b.b, src);
 171         }
 172
 173         if (!(user_flags & SI_CPDMA_SKIP_CHECK_CS_SPACE))
 174                 si_need_cs_space(sctx);
 175
 176         /* This must be done after need_cs_space. */
 177         if (!(user_flags & SI_CPDMA_SKIP_BO_LIST_UPDATE)) {
 178                 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
 179                                           (struct r600_resource*)dst,
 180                                           RADEON_USAGE_WRITE, RADEON_PRIO_CP_DMA);
 181                 if (src)
 182                         radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
 183                                                   (struct r600_resource*)src,
 184                                                   RADEON_USAGE_READ, RADEON_PRIO_CP_DMA);
 185         }
 186
 187         /* Flush the caches for the first copy only.
 188          * Also wait for the previous CP DMA operations.
 189          */
 190         if (!(user_flags & SI_CPDMA_SKIP_GFX_SYNC) && sctx->b.flags)
 191                 si_emit_cache_flush(sctx);
 192
 193         if (!(user_flags & SI_CPDMA_SKIP_SYNC_BEFORE) && *is_first)
 194                 *packet_flags |= CP_DMA_RAW_WAIT;
 195
 196         *is_first = false;
 197
 198         /* Do the synchronization after the last dma, so that all data
 199          * is written to memory.
 200          */
 201         if (!(user_flags & SI_CPDMA_SKIP_SYNC_AFTER) &&
 202             byte_count == remaining_size)
 203                 *packet_flags |= CP_DMA_SYNC;
 204 }
 205
 206 void si_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst,
 207                      uint64_t offset, uint64_t size, unsigned value,
 208                      enum r600_coherency coher)
 209 {
 210         struct si_context *sctx = (struct si_context*)ctx;
 211         struct radeon_winsys *ws = sctx->b.ws;
 212         struct r600_resource *rdst = r600_resource(dst);
 213         unsigned tc_l2_flag = get_tc_l2_flag(sctx, coher);
 214         unsigned flush_flags = get_flush_flags(sctx, coher);
 215         uint64_t dma_clear_size;
 216         bool is_first = true;
 217
 218         if (!size)
 219                 return;
 220
 221        dma_clear_size = size & ~3ull;
 222
 223         /* Mark the buffer range of destination as valid (initialized),
 224          * so that transfer_map knows it should wait for the GPU when mapping
 225          * that range. */
 226         util_range_add(&rdst->valid_buffer_range, offset,
 227                        offset + dma_clear_size);
 228
 229         /* dma_clear_buffer can use clear_buffer on failure. Make sure that
 230          * doesn't happen. We don't want an infinite recursion: */
 231         if (sctx->b.dma.cs &&
 232             !(dst->flags & PIPE_RESOURCE_FLAG_SPARSE) &&
 233             (offset % 4 == 0) &&
 234             /* CP DMA is very slow. Always use SDMA for big clears. This
 235              * alone improves DeusEx:MD performance by 70%. */
 236             (size > CP_DMA_CLEAR_PERF_THRESHOLD ||
 237              /* Buffers not used by the GFX IB yet will be cleared by SDMA.
 238               * This happens to move most buffer clears to SDMA, including
 239               * DCC and CMASK clears, because pipe->clear clears them before
 240               * si_emit_framebuffer_state (in a draw call) adds them.
 241               * For example, DeusEx:MD has 21 buffer clears per frame and all
 242               * of them are moved to SDMA thanks to this. */
 243              !ws->cs_is_buffer_referenced(sctx->b.gfx.cs, rdst->buf,
 244                                           RADEON_USAGE_READWRITE))) {
 245                 sctx->b.dma_clear_buffer(ctx, dst, offset, dma_clear_size, value);
 246
 247                 offset += dma_clear_size;
 248                 size -= dma_clear_size;
 249         } else if (dma_clear_size >= 4) {
 250                 uint64_t va = rdst->gpu_address + offset;
 251
 252                 offset += dma_clear_size;
 253                 size -= dma_clear_size;
 254
 255                 /* Flush the caches. */
 256                 sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
 257                                  SI_CONTEXT_CS_PARTIAL_FLUSH | flush_flags;
 258
 259                 while (dma_clear_size) {
 260                         unsigned byte_count = MIN2(dma_clear_size, cp_dma_max_byte_count(sctx));
 261                         unsigned dma_flags = tc_l2_flag  | CP_DMA_CLEAR;
 262
 263                         si_cp_dma_prepare(sctx, dst, NULL, byte_count, dma_clear_size, 0,
 264                                           &is_first, &dma_flags);
 265
 266                         /* Emit the clear packet. */
 267                         si_emit_cp_dma(sctx, va, value, byte_count, dma_flags, coher);
 268
 269                         dma_clear_size -= byte_count;
 270                         va += byte_count;
 271                 }
 272
 273                 if (tc_l2_flag)
 274                         rdst->TC_L2_dirty = true;
 275
 276                 /* If it's not a framebuffer fast clear... */
 277                 if (coher == R600_COHERENCY_SHADER)
 278                         sctx->b.num_cp_dma_calls++;
 279         }
 280
 281         if (size) {
 282                 /* Handle non-dword alignment.
 283                  *
 284                  * This function is called for embedded texture metadata clears,
 285                  * but those should always be properly aligned. */
 286                 assert(dst->target == PIPE_BUFFER);
 287                 assert(size < 4);
 288
 289                 pipe_buffer_write(ctx, dst, offset, size, &value);
 290         }
 291 }
 292
 293 static void si_pipe_clear_buffer(struct pipe_context *ctx,
 294                                  struct pipe_resource *dst,
 295                                  unsigned offset, unsigned size,
 296                                  const void *clear_value_ptr,
 297                                  int clear_value_size)
 298 {
 299         struct si_context *sctx = (struct si_context*)ctx;
 300         uint32_t dword_value;
 301         unsigned i;
 302
 303         assert(offset % clear_value_size == 0);
 304         assert(size % clear_value_size == 0);
 305
 306         if (clear_value_size > 4) {
 307                 const uint32_t *u32 = clear_value_ptr;
 308                 bool clear_dword_duplicated = true;
 309
 310                 /* See if we can lower large fills to dword fills. */
 311                 for (i = 1; i < clear_value_size / 4; i++)
 312                         if (u32[0] != u32[i]) {
 313                                 clear_dword_duplicated = false;
 314                                 break;
 315                         }
 316
 317                 if (!clear_dword_duplicated) {
 318                         /* Use transform feedback for 64-bit, 96-bit, and
 319                          * 128-bit fills.
 320                          */
 321                         union pipe_color_union clear_value;
 322
 323                         memcpy(&clear_value, clear_value_ptr, clear_value_size);
 324                         si_blitter_begin(ctx, SI_DISABLE_RENDER_COND);
 325                         util_blitter_clear_buffer(sctx->blitter, dst, offset,
 326                                                   size, clear_value_size / 4,
 327                                                   &clear_value);
 328                         si_blitter_end(ctx);
 329                         return;
 330                 }
 331         }
 332
 333         /* Expand the clear value to a dword. */
 334         switch (clear_value_size) {
 335         case 1:
 336                 dword_value = *(uint8_t*)clear_value_ptr;
 337                 dword_value |= (dword_value << 8) |
 338                                (dword_value << 16) |
 339                                (dword_value << 24);
 340                 break;
 341         case 2:
 342                 dword_value = *(uint16_t*)clear_value_ptr;
 343                 dword_value |= dword_value << 16;
 344                 break;
 345         default:
 346                 dword_value = *(uint32_t*)clear_value_ptr;
 347         }
 348
 349         si_clear_buffer(ctx, dst, offset, size, dword_value,
 350                         R600_COHERENCY_SHADER);
 351 }
 352
 353 /**
 354  * Realign the CP DMA engine. This must be done after a copy with an unaligned
 355  * size.
 356  *
 357  * \param size  Remaining size to the CP DMA alignment.
 358  */
 359 static void si_cp_dma_realign_engine(struct si_context *sctx, unsigned size,
 360                                      unsigned user_flags, bool *is_first)
 361 {
 362         uint64_t va;
 363         unsigned dma_flags = 0;
 364         unsigned scratch_size = SI_CPDMA_ALIGNMENT * 2;
 365
 366         assert(size < SI_CPDMA_ALIGNMENT);
 367
 368         /* Use the scratch buffer as the dummy buffer. The 3D engine should be
 369          * idle at this point.
 370          */
 371         if (!sctx->scratch_buffer ||
 372             sctx->scratch_buffer->b.b.width0 < scratch_size) {
 373                 r600_resource_reference(&sctx->scratch_buffer, NULL);
 374                 sctx->scratch_buffer = (struct r600_resource*)
 375                         si_aligned_buffer_create(&sctx->screen->b.b,
 376                                                    R600_RESOURCE_FLAG_UNMAPPABLE,
 377                                                    PIPE_USAGE_DEFAULT,
 378                                                    scratch_size, 256);
 379                 if (!sctx->scratch_buffer)
 380                         return;
 381
 382                 si_mark_atom_dirty(sctx, &sctx->scratch_state);
 383         }
 384
 385         si_cp_dma_prepare(sctx, &sctx->scratch_buffer->b.b,
 386                           &sctx->scratch_buffer->b.b, size, size, user_flags,
 387                           is_first, &dma_flags);
 388
 389         va = sctx->scratch_buffer->gpu_address;
 390         si_emit_cp_dma(sctx, va, va + SI_CPDMA_ALIGNMENT, size, dma_flags,
 391                        R600_COHERENCY_SHADER);
 392 }
 393
 394 /**
 395  * Do memcpy between buffers using CP DMA.
 396  *
 397  * \param user_flags    bitmask of SI_CPDMA_*
 398  */
 399 void si_copy_buffer(struct si_context *sctx,
 400                     struct pipe_resource *dst, struct pipe_resource *src,
 401                     uint64_t dst_offset, uint64_t src_offset, unsigned size,
 402                     unsigned user_flags)
 403 {
 404         uint64_t main_dst_offset, main_src_offset;
 405         unsigned skipped_size = 0;
 406         unsigned realign_size = 0;
 407         unsigned tc_l2_flag = get_tc_l2_flag(sctx, R600_COHERENCY_SHADER);
 408         unsigned flush_flags = get_flush_flags(sctx, R600_COHERENCY_SHADER);
 409         bool is_first = true;
 410
 411         if (!size)
 412                 return;
 413
 414         if (dst != src || dst_offset != src_offset) {
 415                 /* Mark the buffer range of destination as valid (initialized),
 416                  * so that transfer_map knows it should wait for the GPU when mapping
 417                  * that range. */
 418                 util_range_add(&r600_resource(dst)->valid_buffer_range, dst_offset,
 419                                dst_offset + size);
 420         }
 421
 422         dst_offset += r600_resource(dst)->gpu_address;
 423         src_offset += r600_resource(src)->gpu_address;
 424
 425         /* The workarounds aren't needed on Fiji and beyond. */
 426         if (sctx->b.family <= CHIP_CARRIZO ||
 427             sctx->b.family == CHIP_STONEY) {
 428                 /* If the size is not aligned, we must add a dummy copy at the end
 429                  * just to align the internal counter. Otherwise, the DMA engine
 430                  * would slow down by an order of magnitude for following copies.
 431                  */
 432                 if (size % SI_CPDMA_ALIGNMENT)
 433                         realign_size = SI_CPDMA_ALIGNMENT - (size % SI_CPDMA_ALIGNMENT);
 434
 435                 /* If the copy begins unaligned, we must start copying from the next
 436                  * aligned block and the skipped part should be copied after everything
 437                  * else has been copied. Only the src alignment matters, not dst.
 438                  */
 439                 if (src_offset % SI_CPDMA_ALIGNMENT) {
 440                         skipped_size = SI_CPDMA_ALIGNMENT - (src_offset % SI_CPDMA_ALIGNMENT);
 441                         /* The main part will be skipped if the size is too small. */
 442                         skipped_size = MIN2(skipped_size, size);
 443                         size -= skipped_size;
 444                 }
 445         }
 446
 447         /* Flush the caches. */
 448         if (!(user_flags & SI_CPDMA_SKIP_GFX_SYNC))
 449                 sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
 450                                  SI_CONTEXT_CS_PARTIAL_FLUSH | flush_flags;
 451
 452         /* This is the main part doing the copying. Src is always aligned. */
 453         main_dst_offset = dst_offset + skipped_size;
 454         main_src_offset = src_offset + skipped_size;
 455
 456         while (size) {
 457                 unsigned dma_flags = tc_l2_flag;
 458                 unsigned byte_count = MIN2(size, cp_dma_max_byte_count(sctx));
 459
 460                 si_cp_dma_prepare(sctx, dst, src, byte_count,
 461                                   size + skipped_size + realign_size,
 462                                   user_flags, &is_first, &dma_flags);
 463
 464                 si_emit_cp_dma(sctx, main_dst_offset, main_src_offset,
 465                                byte_count, dma_flags, R600_COHERENCY_SHADER);
 466
 467                 size -= byte_count;
 468                 main_src_offset += byte_count;
 469                 main_dst_offset += byte_count;
 470         }
 471
 472         /* Copy the part we skipped because src wasn't aligned. */
 473         if (skipped_size) {
 474                 unsigned dma_flags = tc_l2_flag;
 475
 476                 si_cp_dma_prepare(sctx, dst, src, skipped_size,
 477                                   skipped_size + realign_size, user_flags,
 478                                   &is_first, &dma_flags);
 479
 480                 si_emit_cp_dma(sctx, dst_offset, src_offset, skipped_size,
 481                                dma_flags, R600_COHERENCY_SHADER);
 482         }
 483
 484         /* Finally, realign the engine if the size wasn't aligned. */
 485         if (realign_size)
 486                 si_cp_dma_realign_engine(sctx, realign_size, user_flags,
 487                                          &is_first);
 488
 489         if (tc_l2_flag)
 490                 r600_resource(dst)->TC_L2_dirty = true;
 491
 492         /* If it's not a prefetch... */
 493         if (dst_offset != src_offset)
 494                 sctx->b.num_cp_dma_calls++;
 495 }
 496
 497 void cik_prefetch_TC_L2_async(struct si_context *sctx, struct pipe_resource *buf,
 498                               uint64_t offset, unsigned size)
 499 {
 500         assert(sctx->b.chip_class >= CIK);
 501
 502         si_copy_buffer(sctx, buf, buf, offset, offset, size, SI_CPDMA_SKIP_ALL);
 503 }
 504
 505 static void cik_prefetch_shader_async(struct si_context *sctx,
 506                                       struct si_pm4_state *state)
 507 {
 508         struct pipe_resource *bo = &state->bo[0]->b.b;
 509         assert(state->nbo == 1);
 510
 511         cik_prefetch_TC_L2_async(sctx, bo, 0, bo->width0);
 512 }
 513
 514 static void cik_prefetch_VBO_descriptors(struct si_context *sctx)
 515 {
 516         if (!sctx->vertex_elements)
 517                 return;
 518
 519         cik_prefetch_TC_L2_async(sctx, &sctx->vertex_buffers.buffer->b.b,
 520                                  sctx->vertex_buffers.gpu_address -
 521                                  sctx->vertex_buffers.buffer->gpu_address,
 522                                  sctx->vertex_elements->desc_list_byte_size);
 523 }
 524
 525 void cik_emit_prefetch_L2(struct si_context *sctx)
 526 {
 527         /* Prefetch shaders and VBO descriptors to TC L2. */
 528         if (sctx->b.chip_class >= GFX9) {
 529                 /* Choose the right spot for the VBO prefetch. */
 530                 if (sctx->tes_shader.cso) {
 531                         if (sctx->prefetch_L2_mask & SI_PREFETCH_HS)
 532                                 cik_prefetch_shader_async(sctx, sctx->queued.named.hs);
 533                         if (sctx->prefetch_L2_mask & SI_PREFETCH_VBO_DESCRIPTORS)
 534                                 cik_prefetch_VBO_descriptors(sctx);
 535                         if (sctx->prefetch_L2_mask & SI_PREFETCH_GS)
 536                                 cik_prefetch_shader_async(sctx, sctx->queued.named.gs);
 537                         if (sctx->prefetch_L2_mask & SI_PREFETCH_VS)
 538                                 cik_prefetch_shader_async(sctx, sctx->queued.named.vs);
 539                 } else if (sctx->gs_shader.cso) {
 540                         if (sctx->prefetch_L2_mask & SI_PREFETCH_GS)
 541                                 cik_prefetch_shader_async(sctx, sctx->queued.named.gs);
 542                         if (sctx->prefetch_L2_mask & SI_PREFETCH_VBO_DESCRIPTORS)
 543                                 cik_prefetch_VBO_descriptors(sctx);
 544                         if (sctx->prefetch_L2_mask & SI_PREFETCH_VS)
 545                                 cik_prefetch_shader_async(sctx, sctx->queued.named.vs);
 546                 } else {
 547                         if (sctx->prefetch_L2_mask & SI_PREFETCH_VS)
 548                                 cik_prefetch_shader_async(sctx, sctx->queued.named.vs);
 549                         if (sctx->prefetch_L2_mask & SI_PREFETCH_VBO_DESCRIPTORS)
 550                                 cik_prefetch_VBO_descriptors(sctx);
 551                 }
 552         } else {
 553                 /* SI-CI-VI */
 554                 /* Choose the right spot for the VBO prefetch. */
 555                 if (sctx->tes_shader.cso) {
 556                         if (sctx->prefetch_L2_mask & SI_PREFETCH_LS)
 557                                 cik_prefetch_shader_async(sctx, sctx->queued.named.ls);
 558                         if (sctx->prefetch_L2_mask & SI_PREFETCH_VBO_DESCRIPTORS)
 559                                 cik_prefetch_VBO_descriptors(sctx);
 560                         if (sctx->prefetch_L2_mask & SI_PREFETCH_HS)
 561                                 cik_prefetch_shader_async(sctx, sctx->queued.named.hs);
 562                         if (sctx->prefetch_L2_mask & SI_PREFETCH_ES)
 563                                 cik_prefetch_shader_async(sctx, sctx->queued.named.es);
 564                         if (sctx->prefetch_L2_mask & SI_PREFETCH_GS)
 565                                 cik_prefetch_shader_async(sctx, sctx->queued.named.gs);
 566                         if (sctx->prefetch_L2_mask & SI_PREFETCH_VS)
 567                                 cik_prefetch_shader_async(sctx, sctx->queued.named.vs);
 568                 } else if (sctx->gs_shader.cso) {
 569                         if (sctx->prefetch_L2_mask & SI_PREFETCH_ES)
 570                                 cik_prefetch_shader_async(sctx, sctx->queued.named.es);
 571                         if (sctx->prefetch_L2_mask & SI_PREFETCH_VBO_DESCRIPTORS)
 572                                 cik_prefetch_VBO_descriptors(sctx);
 573                         if (sctx->prefetch_L2_mask & SI_PREFETCH_GS)
 574                                 cik_prefetch_shader_async(sctx, sctx->queued.named.gs);
 575                         if (sctx->prefetch_L2_mask & SI_PREFETCH_VS)
 576                                 cik_prefetch_shader_async(sctx, sctx->queued.named.vs);
 577                 } else {
 578                         if (sctx->prefetch_L2_mask & SI_PREFETCH_VS)
 579                                 cik_prefetch_shader_async(sctx, sctx->queued.named.vs);
 580                         if (sctx->prefetch_L2_mask & SI_PREFETCH_VBO_DESCRIPTORS)
 581                                 cik_prefetch_VBO_descriptors(sctx);
 582                 }
 583         }
 584
 585         if (sctx->prefetch_L2_mask & SI_PREFETCH_PS)
 586                 cik_prefetch_shader_async(sctx, sctx->queued.named.ps);
 587
 588         sctx->prefetch_L2_mask = 0;
 589 }
 590
 591 void si_init_cp_dma_functions(struct si_context *sctx)
 592 {
 593         sctx->b.b.clear_buffer = si_pipe_clear_buffer;
 594         sctx->b.clear_buffer = si_clear_buffer;
 595 }