src/gallium/drivers/radeonsi/si_compute.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
  25 #include "util/u_memory.h"
  26 #include "radeon/r600_pipe_common.h"
  27 #include "radeon/radeon_elf_util.h"
  28 #include "radeon/radeon_llvm_util.h"
  29
  30 #include "radeon/r600_cs.h"
  31 #include "si_pipe.h"
  32 #include "si_shader.h"
  33 #include "sid.h"
  34
  35 #define MAX_GLOBAL_BUFFERS 20
  36
  37 struct si_compute {
  38         struct si_context *ctx;
  39
  40         unsigned local_size;
  41         unsigned private_size;
  42         unsigned input_size;
  43         struct si_shader shader;
  44         unsigned num_user_sgprs;
  45
  46         struct r600_resource *input_buffer;
  47         struct pipe_resource *global_buffers[MAX_GLOBAL_BUFFERS];
  48
  49 #if HAVE_LLVM < 0x0306
  50         unsigned num_kernels;
  51         struct si_shader *kernels;
  52         LLVMContextRef llvm_ctx;
  53 #endif
  54 };
  55
  56 static void init_scratch_buffer(struct si_context *sctx, struct si_compute *program)
  57 {
  58         unsigned scratch_bytes = 0;
  59         uint64_t scratch_buffer_va;
  60         unsigned i;
  61
  62         /* Compute the scratch buffer size using the maximum number of waves.
  63          * This way we don't need to recompute it for each kernel launch. */
  64         unsigned scratch_waves = 32 * sctx->screen->b.info.max_compute_units;
  65         for (i = 0; i < program->shader.binary.global_symbol_count; i++) {
  66                 unsigned offset =
  67                                 program->shader.binary.global_symbol_offsets[i];
  68                 unsigned scratch_bytes_needed;
  69
  70                 si_shader_binary_read_config(&program->shader.binary,
  71                                              &program->shader.config, offset);
  72                 scratch_bytes_needed = program->shader.config.scratch_bytes_per_wave;
  73                 scratch_bytes = MAX2(scratch_bytes, scratch_bytes_needed);
  74         }
  75
  76         if (scratch_bytes == 0)
  77                 return;
  78
  79         program->shader.scratch_bo =
  80                                 si_resource_create_custom(sctx->b.b.screen,
  81                                 PIPE_USAGE_DEFAULT,
  82                                 scratch_bytes * scratch_waves);
  83
  84         scratch_buffer_va = program->shader.scratch_bo->gpu_address;
  85
  86         /* apply_scratch_relocs needs scratch_bytes_per_wave to be set
  87          * to the maximum bytes needed, so it can compute the stride
  88          * correctly.
  89          */
  90         program->shader.config.scratch_bytes_per_wave = scratch_bytes;
  91
  92         /* Patch the shader with the scratch buffer address. */
  93         si_shader_apply_scratch_relocs(sctx,
  94                                 &program->shader, scratch_buffer_va);
  95 }
  96
  97 static void *si_create_compute_state(
  98         struct pipe_context *ctx,
  99         const struct pipe_compute_state *cso)
 100 {
 101         struct si_context *sctx = (struct si_context *)ctx;
 102         struct si_compute *program = CALLOC_STRUCT(si_compute);
 103         const struct pipe_llvm_program_header *header;
 104         const char *code;
 105
 106         header = cso->prog;
 107         code = cso->prog + sizeof(struct pipe_llvm_program_header);
 108
 109         program->ctx = sctx;
 110         program->local_size = cso->req_local_mem;
 111         program->private_size = cso->req_private_mem;
 112         program->input_size = cso->req_input_mem;
 113
 114 #if HAVE_LLVM < 0x0306
 115         {
 116                 unsigned i;
 117                 program->llvm_ctx = LLVMContextCreate();
 118                 program->num_kernels = radeon_llvm_get_num_kernels(program->llvm_ctx,
 119                                         code, header->num_bytes);
 120                 program->kernels = CALLOC(sizeof(struct si_shader),
 121                                                         program->num_kernels);
 122                 for (i = 0; i < program->num_kernels; i++) {
 123                         LLVMModuleRef mod = radeon_llvm_get_kernel_module(program->llvm_ctx, i,
 124                                                         code, header->num_bytes);
 125                         si_compile_llvm(sctx->screen, &program->kernels[i], sctx->tm,
 126                                         mod, &sctx->b.debug, TGSI_PROCESSOR_COMPUTE);
 127                         si_shader_binary_upload(sctx->screen, &program->kernels[i]);
 128                         LLVMDisposeModule(mod);
 129                 }
 130         }
 131 #else
 132
 133         radeon_elf_read(code, header->num_bytes, &program->shader.binary);
 134
 135         /* init_scratch_buffer patches the shader code with the scratch address,
 136          * so we need to call it before si_shader_binary_read() which uploads
 137          * the shader code to the GPU.
 138          */
 139         init_scratch_buffer(sctx, program);
 140         si_shader_binary_read(sctx->screen, &program->shader.binary,
 141                               &program->shader.config, &sctx->b.debug,
 142                               TGSI_PROCESSOR_COMPUTE);
 143         si_shader_binary_upload(sctx->screen, &program->shader);
 144
 145 #endif
 146         program->input_buffer = si_resource_create_custom(sctx->b.b.screen,
 147                 PIPE_USAGE_IMMUTABLE, program->input_size);
 148
 149         return program;
 150 }
 151
 152 static void si_bind_compute_state(struct pipe_context *ctx, void *state)
 153 {
 154         struct si_context *sctx = (struct si_context*)ctx;
 155         sctx->cs_shader_state.program = (struct si_compute*)state;
 156 }
 157
 158 static void si_set_global_binding(
 159         struct pipe_context *ctx, unsigned first, unsigned n,
 160         struct pipe_resource **resources,
 161         uint32_t **handles)
 162 {
 163         unsigned i;
 164         struct si_context *sctx = (struct si_context*)ctx;
 165         struct si_compute *program = sctx->cs_shader_state.program;
 166
 167         if (!resources) {
 168                 for (i = first; i < first + n; i++) {
 169                         pipe_resource_reference(&program->global_buffers[i], NULL);
 170                 }
 171                 return;
 172         }
 173
 174         for (i = first; i < first + n; i++) {
 175                 uint64_t va;
 176                 uint32_t offset;
 177                 pipe_resource_reference(&program->global_buffers[i], resources[i]);
 178                 va = r600_resource(resources[i])->gpu_address;
 179                 offset = util_le32_to_cpu(*handles[i]);
 180                 va += offset;
 181                 va = util_cpu_to_le64(va);
 182                 memcpy(handles[i], &va, sizeof(va));
 183         }
 184 }
 185
 186 /**
 187  * This function computes the value for R_00B860_COMPUTE_TMPRING_SIZE.WAVES
 188  * /p block_layout is the number of threads in each work group.
 189  * /p grid layout is the number of work groups.
 190  */
 191 static unsigned compute_num_waves_for_scratch(
 192                 const struct radeon_info *info,
 193                 const uint *block_layout,
 194                 const uint *grid_layout)
 195 {
 196         unsigned num_sh = MAX2(info->max_sh_per_se, 1);
 197         unsigned num_se = MAX2(info->max_se, 1);
 198         unsigned num_blocks = 1;
 199         unsigned threads_per_block = 1;
 200         unsigned waves_per_block;
 201         unsigned waves_per_sh;
 202         unsigned waves;
 203         unsigned scratch_waves;
 204         unsigned i;
 205
 206         for (i = 0; i < 3; i++) {
 207                 threads_per_block *= block_layout[i];
 208                 num_blocks *= grid_layout[i];
 209         }
 210
 211         waves_per_block = align(threads_per_block, 64) / 64;
 212         waves = waves_per_block * num_blocks;
 213         waves_per_sh = align(waves, num_sh * num_se) / (num_sh * num_se);
 214         scratch_waves = waves_per_sh * num_sh * num_se;
 215
 216         if (waves_per_block > waves_per_sh) {
 217                 scratch_waves = waves_per_block * num_sh * num_se;
 218         }
 219
 220         return scratch_waves;
 221 }
 222
 223 static void si_launch_grid(
 224                 struct pipe_context *ctx,
 225                 const uint *block_layout, const uint *grid_layout,
 226                 uint32_t pc, const void *input)
 227 {
 228         struct si_context *sctx = (struct si_context*)ctx;
 229         struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
 230         struct si_compute *program = sctx->cs_shader_state.program;
 231         struct si_pm4_state *pm4 = CALLOC_STRUCT(si_pm4_state);
 232         struct r600_resource *input_buffer = program->input_buffer;
 233         unsigned kernel_args_size;
 234         unsigned num_work_size_bytes = 36;
 235         uint32_t kernel_args_offset = 0;
 236         uint32_t *kernel_args;
 237         uint64_t kernel_args_va;
 238         uint64_t scratch_buffer_va = 0;
 239         uint64_t shader_va;
 240         unsigned i;
 241         struct si_shader *shader = &program->shader;
 242         unsigned lds_blocks;
 243         unsigned num_waves_for_scratch;
 244
 245 #if HAVE_LLVM < 0x0306
 246         shader = &program->kernels[pc];
 247 #endif
 248
 249
 250         radeon_emit(cs, PKT3(PKT3_CONTEXT_CONTROL, 1, 0) | PKT3_SHADER_TYPE_S(1));
 251         radeon_emit(cs, 0x80000000);
 252         radeon_emit(cs, 0x80000000);
 253
 254         sctx->b.flags |= SI_CONTEXT_INV_VMEM_L1 |
 255                          SI_CONTEXT_INV_GLOBAL_L2 |
 256                          SI_CONTEXT_INV_ICACHE |
 257                          SI_CONTEXT_INV_SMEM_L1 |
 258                          SI_CONTEXT_FLUSH_WITH_INV_L2 |
 259                          SI_CONTEXT_FLAG_COMPUTE;
 260         si_emit_cache_flush(sctx, NULL);
 261
 262         pm4->compute_pkt = true;
 263
 264 #if HAVE_LLVM >= 0x0306
 265         /* Read the config information */
 266         si_shader_binary_read_config(&shader->binary, &shader->config, pc);
 267 #endif
 268
 269         /* Upload the kernel arguments */
 270
 271         /* The extra num_work_size_bytes are for work group / work item size information */
 272         kernel_args_size = program->input_size + num_work_size_bytes + 8 /* For scratch va */;
 273
 274         kernel_args = sctx->b.ws->buffer_map(input_buffer->buf,
 275                         sctx->b.gfx.cs, PIPE_TRANSFER_WRITE);
 276         for (i = 0; i < 3; i++) {
 277                 kernel_args[i] = grid_layout[i];
 278                 kernel_args[i + 3] = grid_layout[i] * block_layout[i];
 279                 kernel_args[i + 6] = block_layout[i];
 280         }
 281
 282         num_waves_for_scratch = compute_num_waves_for_scratch(
 283                 &sctx->screen->b.info, block_layout, grid_layout);
 284
 285         memcpy(kernel_args + (num_work_size_bytes / 4), input, program->input_size);
 286
 287         if (shader->config.scratch_bytes_per_wave > 0) {
 288
 289                 COMPUTE_DBG(sctx->screen, "Waves: %u; Scratch per wave: %u bytes; "
 290                             "Total Scratch: %u bytes\n", num_waves_for_scratch,
 291                             shader->config.scratch_bytes_per_wave,
 292                             shader->config.scratch_bytes_per_wave *
 293                             num_waves_for_scratch);
 294
 295                 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
 296                                           shader->scratch_bo,
 297                                           RADEON_USAGE_READWRITE,
 298                                           RADEON_PRIO_SCRATCH_BUFFER);
 299
 300                 scratch_buffer_va = shader->scratch_bo->gpu_address;
 301         }
 302
 303         for (i = 0; i < (kernel_args_size / 4); i++) {
 304                 COMPUTE_DBG(sctx->screen, "input %u : %u\n", i,
 305                         kernel_args[i]);
 306         }
 307
 308         kernel_args_va = input_buffer->gpu_address;
 309         kernel_args_va += kernel_args_offset;
 310
 311         radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, input_buffer,
 312                                   RADEON_USAGE_READ, RADEON_PRIO_CONST_BUFFER);
 313
 314         si_pm4_set_reg(pm4, R_00B900_COMPUTE_USER_DATA_0, kernel_args_va);
 315         si_pm4_set_reg(pm4, R_00B900_COMPUTE_USER_DATA_0 + 4, S_008F04_BASE_ADDRESS_HI (kernel_args_va >> 32) | S_008F04_STRIDE(0));
 316         si_pm4_set_reg(pm4, R_00B900_COMPUTE_USER_DATA_0 + 8, scratch_buffer_va);
 317         si_pm4_set_reg(pm4, R_00B900_COMPUTE_USER_DATA_0 + 12,
 318                 S_008F04_BASE_ADDRESS_HI(scratch_buffer_va >> 32)
 319                 |  S_008F04_STRIDE(shader->config.scratch_bytes_per_wave / 64));
 320
 321         si_pm4_set_reg(pm4, R_00B810_COMPUTE_START_X, 0);
 322         si_pm4_set_reg(pm4, R_00B814_COMPUTE_START_Y, 0);
 323         si_pm4_set_reg(pm4, R_00B818_COMPUTE_START_Z, 0);
 324
 325         si_pm4_set_reg(pm4, R_00B81C_COMPUTE_NUM_THREAD_X,
 326                                 S_00B81C_NUM_THREAD_FULL(block_layout[0]));
 327         si_pm4_set_reg(pm4, R_00B820_COMPUTE_NUM_THREAD_Y,
 328                                 S_00B820_NUM_THREAD_FULL(block_layout[1]));
 329         si_pm4_set_reg(pm4, R_00B824_COMPUTE_NUM_THREAD_Z,
 330                                 S_00B824_NUM_THREAD_FULL(block_layout[2]));
 331
 332         /* Global buffers */
 333         for (i = 0; i < MAX_GLOBAL_BUFFERS; i++) {
 334                 struct r600_resource *buffer =
 335                                 (struct r600_resource*)program->global_buffers[i];
 336                 if (!buffer) {
 337                         continue;
 338                 }
 339                 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, buffer,
 340                                           RADEON_USAGE_READWRITE,
 341                                           RADEON_PRIO_COMPUTE_GLOBAL);
 342         }
 343
 344         /* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
 345          * and is now per pipe, so it should be handled in the
 346          * kernel if we want to use something other than the default value,
 347          * which is now 0x22f.
 348          */
 349         if (sctx->b.chip_class <= SI) {
 350                 /* XXX: This should be:
 351                  * (number of compute units) * 4 * (waves per simd) - 1 */
 352
 353                 si_pm4_set_reg(pm4, R_00B82C_COMPUTE_MAX_WAVE_ID,
 354                                                 0x190 /* Default value */);
 355         }
 356
 357         shader_va = shader->bo->gpu_address;
 358
 359 #if HAVE_LLVM >= 0x0306
 360         shader_va += pc;
 361 #endif
 362         radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, shader->bo,
 363                                   RADEON_USAGE_READ, RADEON_PRIO_USER_SHADER);
 364         si_pm4_set_reg(pm4, R_00B830_COMPUTE_PGM_LO, shader_va >> 8);
 365         si_pm4_set_reg(pm4, R_00B834_COMPUTE_PGM_HI, shader_va >> 40);
 366
 367         si_pm4_set_reg(pm4, R_00B848_COMPUTE_PGM_RSRC1, shader->config.rsrc1);
 368
 369         lds_blocks = shader->config.lds_size;
 370         /* XXX: We are over allocating LDS.  For SI, the shader reports LDS in
 371          * blocks of 256 bytes, so if there are 4 bytes lds allocated in
 372          * the shader and 4 bytes allocated by the state tracker, then
 373          * we will set LDS_SIZE to 512 bytes rather than 256.
 374          */
 375         if (sctx->b.chip_class <= SI) {
 376                 lds_blocks += align(program->local_size, 256) >> 8;
 377         } else {
 378                 lds_blocks += align(program->local_size, 512) >> 9;
 379         }
 380
 381         assert(lds_blocks <= 0xFF);
 382
 383         shader->config.rsrc2 &= C_00B84C_LDS_SIZE;
 384         shader->config.rsrc2 |=  S_00B84C_LDS_SIZE(lds_blocks);
 385
 386         si_pm4_set_reg(pm4, R_00B84C_COMPUTE_PGM_RSRC2, shader->config.rsrc2);
 387         si_pm4_set_reg(pm4, R_00B854_COMPUTE_RESOURCE_LIMITS, 0);
 388
 389         si_pm4_set_reg(pm4, R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0,
 390                 S_00B858_SH0_CU_EN(0xffff /* Default value */)
 391                 | S_00B858_SH1_CU_EN(0xffff /* Default value */))
 392                 ;
 393
 394         si_pm4_set_reg(pm4, R_00B85C_COMPUTE_STATIC_THREAD_MGMT_SE1,
 395                 S_00B85C_SH0_CU_EN(0xffff /* Default value */)
 396                 | S_00B85C_SH1_CU_EN(0xffff /* Default value */))
 397                 ;
 398
 399         num_waves_for_scratch =
 400                 MIN2(num_waves_for_scratch,
 401                      32 * sctx->screen->b.info.max_compute_units);
 402         si_pm4_set_reg(pm4, R_00B860_COMPUTE_TMPRING_SIZE,
 403                 /* The maximum value for WAVES is 32 * num CU.
 404                  * If you program this value incorrectly, the GPU will hang if
 405                  * COMPUTE_PGM_RSRC2.SCRATCH_EN is enabled.
 406                  */
 407                 S_00B860_WAVES(num_waves_for_scratch)
 408                 | S_00B860_WAVESIZE(shader->config.scratch_bytes_per_wave >> 10))
 409                 ;
 410
 411         si_pm4_cmd_begin(pm4, PKT3_DISPATCH_DIRECT);
 412         si_pm4_cmd_add(pm4, grid_layout[0]); /* Thread groups DIM_X */
 413         si_pm4_cmd_add(pm4, grid_layout[1]); /* Thread groups DIM_Y */
 414         si_pm4_cmd_add(pm4, grid_layout[2]); /* Thread gropus DIM_Z */
 415         si_pm4_cmd_add(pm4, 1); /* DISPATCH_INITIATOR */
 416         si_pm4_cmd_end(pm4, false);
 417
 418         si_pm4_emit(sctx, pm4);
 419
 420 #if 0
 421         fprintf(stderr, "cdw: %i\n", sctx->cs->cdw);
 422         for (i = 0; i < sctx->cs->cdw; i++) {
 423                 fprintf(stderr, "%4i : 0x%08X\n", i, sctx->cs->buf[i]);
 424         }
 425 #endif
 426
 427         si_pm4_free_state(sctx, pm4, ~0);
 428
 429         sctx->b.flags |= SI_CONTEXT_CS_PARTIAL_FLUSH |
 430                          SI_CONTEXT_INV_VMEM_L1 |
 431                          SI_CONTEXT_INV_GLOBAL_L2 |
 432                          SI_CONTEXT_INV_ICACHE |
 433                          SI_CONTEXT_INV_SMEM_L1 |
 434                          SI_CONTEXT_FLAG_COMPUTE;
 435         si_emit_cache_flush(sctx, NULL);
 436 }
 437
 438
 439 static void si_delete_compute_state(struct pipe_context *ctx, void* state){
 440         struct si_compute *program = (struct si_compute *)state;
 441
 442         if (!state) {
 443                 return;
 444         }
 445
 446 #if HAVE_LLVM < 0x0306
 447         if (program->kernels) {
 448                 for (int i = 0; i < program->num_kernels; i++){
 449                         if (program->kernels[i].bo){
 450                                 si_shader_destroy(&program->kernels[i]);
 451                         }
 452                 }
 453                 FREE(program->kernels);
 454         }
 455
 456         if (program->llvm_ctx){
 457                 LLVMContextDispose(program->llvm_ctx);
 458         }
 459 #else
 460         FREE(program->shader.binary.config);
 461         FREE(program->shader.binary.rodata);
 462         FREE(program->shader.binary.global_symbol_offsets);
 463         si_shader_destroy(&program->shader);
 464 #endif
 465
 466         pipe_resource_reference(
 467                 (struct pipe_resource **)&program->input_buffer, NULL);
 468
 469         FREE(program);
 470 }
 471
 472 static void si_set_compute_resources(struct pipe_context * ctx_,
 473                 unsigned start, unsigned count,
 474                 struct pipe_surface ** surfaces) { }
 475
 476 void si_init_compute_functions(struct si_context *sctx)
 477 {
 478         sctx->b.b.create_compute_state = si_create_compute_state;
 479         sctx->b.b.delete_compute_state = si_delete_compute_state;
 480         sctx->b.b.bind_compute_state = si_bind_compute_state;
 481 /*       ctx->context.create_sampler_view = evergreen_compute_create_sampler_view; */
 482         sctx->b.b.set_compute_resources = si_set_compute_resources;
 483         sctx->b.b.set_global_binding = si_set_global_binding;
 484         sctx->b.b.launch_grid = si_launch_grid;
 485 }