1 /* 2 * Copyright 2011 Adam Rak <adam.rak (at) streamnovation.com> 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 * Authors: 24 * Adam Rak <adam.rak (at) streamnovation.com> 25 */ 26 27 #include <stdio.h> 28 #include <errno.h> 29 #include "pipe/p_defines.h" 30 #include "pipe/p_state.h" 31 #include "pipe/p_context.h" 32 #include "util/u_blitter.h" 33 #include "util/list.h" 34 #include "util/u_transfer.h" 35 #include "util/u_surface.h" 36 #include "util/u_pack_color.h" 37 #include "util/u_memory.h" 38 #include "util/u_inlines.h" 39 #include "util/u_framebuffer.h" 40 #include "pipebuffer/pb_buffer.h" 41 #include "evergreend.h" 42 #include "r600_shader.h" 43 #include "r600_pipe.h" 44 #include "r600_formats.h" 45 #include "evergreen_compute.h" 46 #include "evergreen_compute_internal.h" 47 #include "compute_memory_pool.h" 48 #include "sb/sb_public.h" 49 #include "radeon/radeon_elf_util.h" 50 #include <inttypes.h> 51 52 /** 53 RAT0 is for global binding write 54 VTX1 is for global binding read 55 56 for wrting images RAT1... 57 for reading images TEX2... 58 TEX2-RAT1 is paired 59 60 TEX2... consumes the same fetch resources, that VTX2... would consume 61 62 CONST0 and VTX0 is for parameters 63 CONST0 is binding smaller input parameter buffer, and for constant indexing, 64 also constant cached 65 VTX0 is for indirect/non-constant indexing, or if the input is bigger than 66 the constant cache can handle 67 68 RAT-s are limited to 12, so we can only bind at most 11 texture for writing 69 because we reserve RAT0 for global bindings. With byteaddressing enabled, 70 we should reserve another one too.=> 10 image binding for writing max. 71 72 from Nvidia OpenCL: 73 CL_DEVICE_MAX_READ_IMAGE_ARGS: 128 74 CL_DEVICE_MAX_WRITE_IMAGE_ARGS: 8 75 76 so 10 for writing is enough. 176 is the max for reading according to the docs 77 78 writable images should be listed first < 10, so their id corresponds to RAT(id+1) 79 writable images will consume TEX slots, VTX slots too because of linear indexing 80 81 */ 82 83 struct r600_resource *r600_compute_buffer_alloc_vram(struct r600_screen *screen, 84 unsigned size) 85 { 86 struct pipe_resource *buffer = NULL; 87 assert(size); 88 89 buffer = pipe_buffer_create((struct pipe_screen*) screen, 90 0, PIPE_USAGE_IMMUTABLE, size); 91 92 return (struct r600_resource *)buffer; 93 } 94 95 96 static void evergreen_set_rat(struct r600_pipe_compute *pipe, 97 unsigned id, 98 struct r600_resource *bo, 99 int start, 100 int size) 101 { 102 struct pipe_surface rat_templ; 103 struct r600_surface *surf = NULL; 104 struct r600_context *rctx = NULL; 105 106 assert(id < 12); 107 assert((size & 3) == 0); 108 assert((start & 0xFF) == 0); 109 110 rctx = pipe->ctx; 111 112 COMPUTE_DBG(rctx->screen, "bind rat: %i \n", id); 113 114 /* Create the RAT surface */ 115 memset(&rat_templ, 0, sizeof(rat_templ)); 116 rat_templ.format = PIPE_FORMAT_R32_UINT; 117 rat_templ.u.tex.level = 0; 118 rat_templ.u.tex.first_layer = 0; 119 rat_templ.u.tex.last_layer = 0; 120 121 /* Add the RAT the list of color buffers */ 122 pipe->ctx->framebuffer.state.cbufs[id] = pipe->ctx->b.b.create_surface( 123 (struct pipe_context *)pipe->ctx, 124 (struct pipe_resource *)bo, &rat_templ); 125 126 /* Update the number of color buffers */ 127 pipe->ctx->framebuffer.state.nr_cbufs = 128 MAX2(id + 1, pipe->ctx->framebuffer.state.nr_cbufs); 129 130 /* Update the cb_target_mask 131 * XXX: I think this is a potential spot for bugs once we start doing 132 * GL interop. cb_target_mask may be modified in the 3D sections 133 * of this driver. */ 134 pipe->ctx->compute_cb_target_mask |= (0xf << (id * 4)); 135 136 surf = (struct r600_surface*)pipe->ctx->framebuffer.state.cbufs[id]; 137 evergreen_init_color_surface_rat(rctx, surf); 138 } 139 140 static void evergreen_cs_set_vertex_buffer(struct r600_context *rctx, 141 unsigned vb_index, 142 unsigned offset, 143 struct pipe_resource *buffer) 144 { 145 struct r600_vertexbuf_state *state = &rctx->cs_vertex_buffer_state; 146 struct pipe_vertex_buffer *vb = &state->vb[vb_index]; 147 vb->stride = 1; 148 vb->buffer_offset = offset; 149 vb->buffer = buffer; 150 vb->user_buffer = NULL; 151 152 /* The vertex instructions in the compute shaders use the texture cache, 153 * so we need to invalidate it. */ 154 rctx->b.flags |= R600_CONTEXT_INV_VERTEX_CACHE; 155 state->enabled_mask |= 1 << vb_index; 156 state->dirty_mask |= 1 << vb_index; 157 r600_mark_atom_dirty(rctx, &state->atom); 158 } 159 160 static void evergreen_cs_set_constant_buffer(struct r600_context *rctx, 161 unsigned cb_index, 162 unsigned offset, 163 unsigned size, 164 struct pipe_resource *buffer) 165 { 166 struct pipe_constant_buffer cb; 167 cb.buffer_size = size; 168 cb.buffer_offset = offset; 169 cb.buffer = buffer; 170 cb.user_buffer = NULL; 171 172 rctx->b.b.set_constant_buffer(&rctx->b.b, PIPE_SHADER_COMPUTE, cb_index, &cb); 173 } 174 175 /* We need to define these R600 registers here, because we can't include 176 * evergreend.h and r600d.h. 177 */ 178 #define R_028868_SQ_PGM_RESOURCES_VS 0x028868 179 #define R_028850_SQ_PGM_RESOURCES_PS 0x028850 180 181 #ifdef HAVE_OPENCL 182 183 static void r600_shader_binary_read_config(const struct radeon_shader_binary *binary, 184 struct r600_bytecode *bc, 185 uint64_t symbol_offset, 186 boolean *use_kill) 187 { 188 unsigned i; 189 const unsigned char *config = 190 radeon_shader_binary_config_start(binary, symbol_offset); 191 192 for (i = 0; i < binary->config_size_per_symbol; i+= 8) { 193 unsigned reg = 194 util_le32_to_cpu(*(uint32_t*)(config + i)); 195 unsigned value = 196 util_le32_to_cpu(*(uint32_t*)(config + i + 4)); 197 switch (reg) { 198 /* R600 / R700 */ 199 case R_028850_SQ_PGM_RESOURCES_PS: 200 case R_028868_SQ_PGM_RESOURCES_VS: 201 /* Evergreen / Northern Islands */ 202 case R_028844_SQ_PGM_RESOURCES_PS: 203 case R_028860_SQ_PGM_RESOURCES_VS: 204 case R_0288D4_SQ_PGM_RESOURCES_LS: 205 bc->ngpr = MAX2(bc->ngpr, G_028844_NUM_GPRS(value)); 206 bc->nstack = MAX2(bc->nstack, G_028844_STACK_SIZE(value)); 207 break; 208 case R_02880C_DB_SHADER_CONTROL: 209 *use_kill = G_02880C_KILL_ENABLE(value); 210 break; 211 case R_0288E8_SQ_LDS_ALLOC: 212 bc->nlds_dw = value; 213 break; 214 } 215 } 216 } 217 218 static unsigned r600_create_shader(struct r600_bytecode *bc, 219 const struct radeon_shader_binary *binary, 220 boolean *use_kill) 221 222 { 223 assert(binary->code_size % 4 == 0); 224 bc->bytecode = CALLOC(1, binary->code_size); 225 memcpy(bc->bytecode, binary->code, binary->code_size); 226 bc->ndw = binary->code_size / 4; 227 228 r600_shader_binary_read_config(binary, bc, 0, use_kill); 229 return 0; 230 } 231 232 #endif 233 234 static void r600_destroy_shader(struct r600_bytecode *bc) 235 { 236 FREE(bc->bytecode); 237 } 238 239 static void *evergreen_create_compute_state(struct pipe_context *ctx, 240 const struct pipe_compute_state *cso) 241 { 242 struct r600_context *rctx = (struct r600_context *)ctx; 243 struct r600_pipe_compute *shader = CALLOC_STRUCT(r600_pipe_compute); 244 #ifdef HAVE_OPENCL 245 const struct pipe_llvm_program_header *header; 246 const char *code; 247 void *p; 248 boolean use_kill; 249 250 COMPUTE_DBG(rctx->screen, "*** evergreen_create_compute_state\n"); 251 header = cso->prog; 252 code = cso->prog + sizeof(struct pipe_llvm_program_header); 253 radeon_shader_binary_init(&shader->binary); 254 radeon_elf_read(code, header->num_bytes, &shader->binary); 255 r600_create_shader(&shader->bc, &shader->binary, &use_kill); 256 257 /* Upload code + ROdata */ 258 shader->code_bo = r600_compute_buffer_alloc_vram(rctx->screen, 259 shader->bc.ndw * 4); 260 p = r600_buffer_map_sync_with_rings(&rctx->b, shader->code_bo, PIPE_TRANSFER_WRITE); 261 //TODO: use util_memcpy_cpu_to_le32 ? 262 memcpy(p, shader->bc.bytecode, shader->bc.ndw * 4); 263 rctx->b.ws->buffer_unmap(shader->code_bo->buf); 264 #endif 265 266 shader->ctx = rctx; 267 shader->local_size = cso->req_local_mem; 268 shader->private_size = cso->req_private_mem; 269 shader->input_size = cso->req_input_mem; 270 271 return shader; 272 } 273 274 static void evergreen_delete_compute_state(struct pipe_context *ctx, void *state) 275 { 276 struct r600_context *rctx = (struct r600_context *)ctx; 277 struct r600_pipe_compute *shader = state; 278 279 COMPUTE_DBG(rctx->screen, "*** evergreen_delete_compute_state\n"); 280 281 if (!shader) 282 return; 283 284 radeon_shader_binary_clean(&shader->binary); 285 r600_destroy_shader(&shader->bc); 286 287 /* TODO destroy shader->code_bo, shader->const_bo 288 * we'll need something like r600_buffer_free */ 289 FREE(shader); 290 } 291 292 static void evergreen_bind_compute_state(struct pipe_context *ctx, void *state) 293 { 294 struct r600_context *rctx = (struct r600_context *)ctx; 295 296 COMPUTE_DBG(rctx->screen, "*** evergreen_bind_compute_state\n"); 297 298 rctx->cs_shader_state.shader = (struct r600_pipe_compute *)state; 299 } 300 301 /* The kernel parameters are stored a vtx buffer (ID=0), besides the explicit 302 * kernel parameters there are implicit parameters that need to be stored 303 * in the vertex buffer as well. Here is how these parameters are organized in 304 * the buffer: 305 * 306 * DWORDS 0-2: Number of work groups in each dimension (x,y,z) 307 * DWORDS 3-5: Number of global work items in each dimension (x,y,z) 308 * DWORDS 6-8: Number of work items within each work group in each dimension 309 * (x,y,z) 310 * DWORDS 9+ : Kernel parameters 311 */ 312 static void evergreen_compute_upload_input(struct pipe_context *ctx, 313 const struct pipe_grid_info *info) 314 { 315 struct r600_context *rctx = (struct r600_context *)ctx; 316 struct r600_pipe_compute *shader = rctx->cs_shader_state.shader; 317 unsigned i; 318 /* We need to reserve 9 dwords (36 bytes) for implicit kernel 319 * parameters. 320 */ 321 unsigned input_size = shader->input_size + 36; 322 uint32_t *num_work_groups_start; 323 uint32_t *global_size_start; 324 uint32_t *local_size_start; 325 uint32_t *kernel_parameters_start; 326 struct pipe_box box; 327 struct pipe_transfer *transfer = NULL; 328 329 if (shader->input_size == 0) { 330 return; 331 } 332 333 if (!shader->kernel_param) { 334 /* Add space for the grid dimensions */ 335 shader->kernel_param = (struct r600_resource *) 336 pipe_buffer_create(ctx->screen, 0, 337 PIPE_USAGE_IMMUTABLE, input_size); 338 } 339 340 u_box_1d(0, input_size, &box); 341 num_work_groups_start = ctx->transfer_map(ctx, 342 (struct pipe_resource*)shader->kernel_param, 343 0, PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD_RANGE, 344 &box, &transfer); 345 global_size_start = num_work_groups_start + (3 * (sizeof(uint) /4)); 346 local_size_start = global_size_start + (3 * (sizeof(uint)) / 4); 347 kernel_parameters_start = local_size_start + (3 * (sizeof(uint)) / 4); 348 349 /* Copy the work group size */ 350 memcpy(num_work_groups_start, info->grid, 3 * sizeof(uint)); 351 352 /* Copy the global size */ 353 for (i = 0; i < 3; i++) { 354 global_size_start[i] = info->grid[i] * info->block[i]; 355 } 356 357 /* Copy the local dimensions */ 358 memcpy(local_size_start, info->block, 3 * sizeof(uint)); 359 360 /* Copy the kernel inputs */ 361 memcpy(kernel_parameters_start, info->input, shader->input_size); 362 363 for (i = 0; i < (input_size / 4); i++) { 364 COMPUTE_DBG(rctx->screen, "input %i : %u\n", i, 365 ((unsigned*)num_work_groups_start)[i]); 366 } 367 368 ctx->transfer_unmap(ctx, transfer); 369 370 /* ID=0 and ID=3 are reserved for the parameters. 371 * LLVM will preferably use ID=0, but it does not work for dynamic 372 * indices. */ 373 evergreen_cs_set_vertex_buffer(rctx, 3, 0, 374 (struct pipe_resource*)shader->kernel_param); 375 evergreen_cs_set_constant_buffer(rctx, 0, 0, input_size, 376 (struct pipe_resource*)shader->kernel_param); 377 } 378 379 static void evergreen_emit_dispatch(struct r600_context *rctx, 380 const struct pipe_grid_info *info) 381 { 382 int i; 383 struct radeon_winsys_cs *cs = rctx->b.gfx.cs; 384 struct r600_pipe_compute *shader = rctx->cs_shader_state.shader; 385 unsigned num_waves; 386 unsigned num_pipes = rctx->screen->b.info.r600_max_quad_pipes; 387 unsigned wave_divisor = (16 * num_pipes); 388 int group_size = 1; 389 int grid_size = 1; 390 unsigned lds_size = shader->local_size / 4 + 391 shader->bc.nlds_dw; 392 393 394 /* Calculate group_size/grid_size */ 395 for (i = 0; i < 3; i++) { 396 group_size *= info->block[i]; 397 } 398 399 for (i = 0; i < 3; i++) { 400 grid_size *= info->grid[i]; 401 } 402 403 /* num_waves = ceil((tg_size.x * tg_size.y, tg_size.z) / (16 * num_pipes)) */ 404 num_waves = (info->block[0] * info->block[1] * info->block[2] + 405 wave_divisor - 1) / wave_divisor; 406 407 COMPUTE_DBG(rctx->screen, "Using %u pipes, " 408 "%u wavefronts per thread block, " 409 "allocating %u dwords lds.\n", 410 num_pipes, num_waves, lds_size); 411 412 radeon_set_config_reg(cs, R_008970_VGT_NUM_INDICES, group_size); 413 414 radeon_set_config_reg_seq(cs, R_00899C_VGT_COMPUTE_START_X, 3); 415 radeon_emit(cs, 0); /* R_00899C_VGT_COMPUTE_START_X */ 416 radeon_emit(cs, 0); /* R_0089A0_VGT_COMPUTE_START_Y */ 417 radeon_emit(cs, 0); /* R_0089A4_VGT_COMPUTE_START_Z */ 418 419 radeon_set_config_reg(cs, R_0089AC_VGT_COMPUTE_THREAD_GROUP_SIZE, 420 group_size); 421 422 radeon_compute_set_context_reg_seq(cs, R_0286EC_SPI_COMPUTE_NUM_THREAD_X, 3); 423 radeon_emit(cs, info->block[0]); /* R_0286EC_SPI_COMPUTE_NUM_THREAD_X */ 424 radeon_emit(cs, info->block[1]); /* R_0286F0_SPI_COMPUTE_NUM_THREAD_Y */ 425 radeon_emit(cs, info->block[2]); /* R_0286F4_SPI_COMPUTE_NUM_THREAD_Z */ 426 427 if (rctx->b.chip_class < CAYMAN) { 428 assert(lds_size <= 8192); 429 } else { 430 /* Cayman appears to have a slightly smaller limit, see the 431 * value of CM_R_0286FC_SPI_LDS_MGMT.NUM_LS_LDS */ 432 assert(lds_size <= 8160); 433 } 434 435 radeon_compute_set_context_reg(cs, R_0288E8_SQ_LDS_ALLOC, 436 lds_size | (num_waves << 14)); 437 438 /* Dispatch packet */ 439 radeon_emit(cs, PKT3C(PKT3_DISPATCH_DIRECT, 3, 0)); 440 radeon_emit(cs, info->grid[0]); 441 radeon_emit(cs, info->grid[1]); 442 radeon_emit(cs, info->grid[2]); 443 /* VGT_DISPATCH_INITIATOR = COMPUTE_SHADER_EN */ 444 radeon_emit(cs, 1); 445 } 446 447 static void compute_emit_cs(struct r600_context *rctx, 448 const struct pipe_grid_info *info) 449 { 450 struct radeon_winsys_cs *cs = rctx->b.gfx.cs; 451 unsigned i; 452 453 /* make sure that the gfx ring is only one active */ 454 if (radeon_emitted(rctx->b.dma.cs, 0)) { 455 rctx->b.dma.flush(rctx, RADEON_FLUSH_ASYNC, NULL); 456 } 457 458 /* Initialize all the compute-related registers. 459 * 460 * See evergreen_init_atom_start_compute_cs() in this file for the list 461 * of registers initialized by the start_compute_cs_cmd atom. 462 */ 463 r600_emit_command_buffer(cs, &rctx->start_compute_cs_cmd); 464 465 /* emit config state */ 466 if (rctx->b.chip_class == EVERGREEN) 467 r600_emit_atom(rctx, &rctx->config_state.atom); 468 469 rctx->b.flags |= R600_CONTEXT_WAIT_3D_IDLE | R600_CONTEXT_FLUSH_AND_INV; 470 r600_flush_emit(rctx); 471 472 /* Emit colorbuffers. */ 473 /* XXX support more than 8 colorbuffers (the offsets are not a multiple of 0x3C for CB8-11) */ 474 for (i = 0; i < 8 && i < rctx->framebuffer.state.nr_cbufs; i++) { 475 struct r600_surface *cb = (struct r600_surface*)rctx->framebuffer.state.cbufs[i]; 476 unsigned reloc = radeon_add_to_buffer_list(&rctx->b, &rctx->b.gfx, 477 (struct r600_resource*)cb->base.texture, 478 RADEON_USAGE_READWRITE, 479 RADEON_PRIO_SHADER_RW_BUFFER); 480 481 radeon_compute_set_context_reg_seq(cs, R_028C60_CB_COLOR0_BASE + i * 0x3C, 7); 482 radeon_emit(cs, cb->cb_color_base); /* R_028C60_CB_COLOR0_BASE */ 483 radeon_emit(cs, cb->cb_color_pitch); /* R_028C64_CB_COLOR0_PITCH */ 484 radeon_emit(cs, cb->cb_color_slice); /* R_028C68_CB_COLOR0_SLICE */ 485 radeon_emit(cs, cb->cb_color_view); /* R_028C6C_CB_COLOR0_VIEW */ 486 radeon_emit(cs, cb->cb_color_info); /* R_028C70_CB_COLOR0_INFO */ 487 radeon_emit(cs, cb->cb_color_attrib); /* R_028C74_CB_COLOR0_ATTRIB */ 488 radeon_emit(cs, cb->cb_color_dim); /* R_028C78_CB_COLOR0_DIM */ 489 490 radeon_emit(cs, PKT3(PKT3_NOP, 0, 0)); /* R_028C60_CB_COLOR0_BASE */ 491 radeon_emit(cs, reloc); 492 493 radeon_emit(cs, PKT3(PKT3_NOP, 0, 0)); /* R_028C74_CB_COLOR0_ATTRIB */ 494 radeon_emit(cs, reloc); 495 } 496 for (; i < 8 ; i++) 497 radeon_compute_set_context_reg(cs, R_028C70_CB_COLOR0_INFO + i * 0x3C, 498 S_028C70_FORMAT(V_028C70_COLOR_INVALID)); 499 for (; i < 12; i++) 500 radeon_compute_set_context_reg(cs, R_028E50_CB_COLOR8_INFO + (i - 8) * 0x1C, 501 S_028C70_FORMAT(V_028C70_COLOR_INVALID)); 502 503 /* Set CB_TARGET_MASK XXX: Use cb_misc_state */ 504 radeon_compute_set_context_reg(cs, R_028238_CB_TARGET_MASK, 505 rctx->compute_cb_target_mask); 506 507 508 /* Emit vertex buffer state */ 509 rctx->cs_vertex_buffer_state.atom.num_dw = 12 * util_bitcount(rctx->cs_vertex_buffer_state.dirty_mask); 510 r600_emit_atom(rctx, &rctx->cs_vertex_buffer_state.atom); 511 512 /* Emit constant buffer state */ 513 r600_emit_atom(rctx, &rctx->constbuf_state[PIPE_SHADER_COMPUTE].atom); 514 515 /* Emit sampler state */ 516 r600_emit_atom(rctx, &rctx->samplers[PIPE_SHADER_COMPUTE].states.atom); 517 518 /* Emit sampler view (texture resource) state */ 519 r600_emit_atom(rctx, &rctx->samplers[PIPE_SHADER_COMPUTE].views.atom); 520 521 /* Emit compute shader state */ 522 r600_emit_atom(rctx, &rctx->cs_shader_state.atom); 523 524 /* Emit dispatch state and dispatch packet */ 525 evergreen_emit_dispatch(rctx, info); 526 527 /* XXX evergreen_flush_emit() hardcodes the CP_COHER_SIZE to 0xffffffff 528 */ 529 rctx->b.flags |= R600_CONTEXT_INV_CONST_CACHE | 530 R600_CONTEXT_INV_VERTEX_CACHE | 531 R600_CONTEXT_INV_TEX_CACHE; 532 r600_flush_emit(rctx); 533 rctx->b.flags = 0; 534 535 if (rctx->b.chip_class >= CAYMAN) { 536 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0)); 537 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_CS_PARTIAL_FLUSH) | EVENT_INDEX(4)); 538 /* DEALLOC_STATE prevents the GPU from hanging when a 539 * SURFACE_SYNC packet is emitted some time after a DISPATCH_DIRECT 540 * with any of the CB*_DEST_BASE_ENA or DB_DEST_BASE_ENA bits set. 541 */ 542 radeon_emit(cs, PKT3C(PKT3_DEALLOC_STATE, 0, 0)); 543 radeon_emit(cs, 0); 544 } 545 546 #if 0 547 COMPUTE_DBG(rctx->screen, "cdw: %i\n", cs->cdw); 548 for (i = 0; i < cs->cdw; i++) { 549 COMPUTE_DBG(rctx->screen, "%4i : 0x%08X\n", i, cs->buf[i]); 550 } 551 #endif 552 553 } 554 555 556 /** 557 * Emit function for r600_cs_shader_state atom 558 */ 559 void evergreen_emit_cs_shader(struct r600_context *rctx, 560 struct r600_atom *atom) 561 { 562 struct r600_cs_shader_state *state = 563 (struct r600_cs_shader_state*)atom; 564 struct r600_pipe_compute *shader = state->shader; 565 struct radeon_winsys_cs *cs = rctx->b.gfx.cs; 566 uint64_t va; 567 struct r600_resource *code_bo; 568 unsigned ngpr, nstack; 569 570 code_bo = shader->code_bo; 571 va = shader->code_bo->gpu_address + state->pc; 572 ngpr = shader->bc.ngpr; 573 nstack = shader->bc.nstack; 574 575 radeon_compute_set_context_reg_seq(cs, R_0288D0_SQ_PGM_START_LS, 3); 576 radeon_emit(cs, va >> 8); /* R_0288D0_SQ_PGM_START_LS */ 577 radeon_emit(cs, /* R_0288D4_SQ_PGM_RESOURCES_LS */ 578 S_0288D4_NUM_GPRS(ngpr) 579 | S_0288D4_STACK_SIZE(nstack)); 580 radeon_emit(cs, 0); /* R_0288D8_SQ_PGM_RESOURCES_LS_2 */ 581 582 radeon_emit(cs, PKT3C(PKT3_NOP, 0, 0)); 583 radeon_emit(cs, radeon_add_to_buffer_list(&rctx->b, &rctx->b.gfx, 584 code_bo, RADEON_USAGE_READ, 585 RADEON_PRIO_SHADER_BINARY)); 586 } 587 588 static void evergreen_launch_grid(struct pipe_context *ctx, 589 const struct pipe_grid_info *info) 590 { 591 struct r600_context *rctx = (struct r600_context *)ctx; 592 #ifdef HAVE_OPENCL 593 struct r600_pipe_compute *shader = rctx->cs_shader_state.shader; 594 boolean use_kill; 595 596 rctx->cs_shader_state.pc = info->pc; 597 /* Get the config information for this kernel. */ 598 r600_shader_binary_read_config(&shader->binary, &shader->bc, 599 info->pc, &use_kill); 600 #endif 601 602 COMPUTE_DBG(rctx->screen, "*** evergreen_launch_grid: pc = %u\n", info->pc); 603 604 605 evergreen_compute_upload_input(ctx, info); 606 compute_emit_cs(rctx, info); 607 } 608 609 static void evergreen_set_compute_resources(struct pipe_context *ctx, 610 unsigned start, unsigned count, 611 struct pipe_surface **surfaces) 612 { 613 struct r600_context *rctx = (struct r600_context *)ctx; 614 struct r600_surface **resources = (struct r600_surface **)surfaces; 615 616 COMPUTE_DBG(rctx->screen, "*** evergreen_set_compute_resources: start = %u count = %u\n", 617 start, count); 618 619 for (unsigned i = 0; i < count; i++) { 620 /* The First four vertex buffers are reserved for parameters and 621 * global buffers. */ 622 unsigned vtx_id = 4 + i; 623 if (resources[i]) { 624 struct r600_resource_global *buffer = 625 (struct r600_resource_global*) 626 resources[i]->base.texture; 627 if (resources[i]->base.writable) { 628 assert(i+1 < 12); 629 630 evergreen_set_rat(rctx->cs_shader_state.shader, i+1, 631 (struct r600_resource *)resources[i]->base.texture, 632 buffer->chunk->start_in_dw*4, 633 resources[i]->base.texture->width0); 634 } 635 636 evergreen_cs_set_vertex_buffer(rctx, vtx_id, 637 buffer->chunk->start_in_dw * 4, 638 resources[i]->base.texture); 639 } 640 } 641 } 642 643 static void evergreen_set_global_binding(struct pipe_context *ctx, 644 unsigned first, unsigned n, 645 struct pipe_resource **resources, 646 uint32_t **handles) 647 { 648 struct r600_context *rctx = (struct r600_context *)ctx; 649 struct compute_memory_pool *pool = rctx->screen->global_pool; 650 struct r600_resource_global **buffers = 651 (struct r600_resource_global **)resources; 652 unsigned i; 653 654 COMPUTE_DBG(rctx->screen, "*** evergreen_set_global_binding first = %u n = %u\n", 655 first, n); 656 657 if (!resources) { 658 /* XXX: Unset */ 659 return; 660 } 661 662 /* We mark these items for promotion to the pool if they 663 * aren't already there */ 664 for (i = first; i < first + n; i++) { 665 struct compute_memory_item *item = buffers[i]->chunk; 666 667 if (!is_item_in_pool(item)) 668 buffers[i]->chunk->status |= ITEM_FOR_PROMOTING; 669 } 670 671 if (compute_memory_finalize_pending(pool, ctx) == -1) { 672 /* XXX: Unset */ 673 return; 674 } 675 676 for (i = first; i < first + n; i++) 677 { 678 uint32_t buffer_offset; 679 uint32_t handle; 680 assert(resources[i]->target == PIPE_BUFFER); 681 assert(resources[i]->bind & PIPE_BIND_GLOBAL); 682 683 buffer_offset = util_le32_to_cpu(*(handles[i])); 684 handle = buffer_offset + buffers[i]->chunk->start_in_dw * 4; 685 686 *(handles[i]) = util_cpu_to_le32(handle); 687 } 688 689 /* globals for writing */ 690 evergreen_set_rat(rctx->cs_shader_state.shader, 0, pool->bo, 0, pool->size_in_dw * 4); 691 /* globals for reading */ 692 evergreen_cs_set_vertex_buffer(rctx, 1, 0, 693 (struct pipe_resource*)pool->bo); 694 695 /* constants for reading, LLVM puts them in text segment */ 696 evergreen_cs_set_vertex_buffer(rctx, 2, 0, 697 (struct pipe_resource*)rctx->cs_shader_state.shader->code_bo); 698 } 699 700 /** 701 * This function initializes all the compute specific registers that need to 702 * be initialized for each compute command stream. Registers that are common 703 * to both compute and 3D will be initialized at the beginning of each compute 704 * command stream by the start_cs_cmd atom. However, since the SET_CONTEXT_REG 705 * packet requires that the shader type bit be set, we must initialize all 706 * context registers needed for compute in this function. The registers 707 * initialized by the start_cs_cmd atom can be found in evergreen_state.c in the 708 * functions evergreen_init_atom_start_cs or cayman_init_atom_start_cs depending 709 * on the GPU family. 710 */ 711 void evergreen_init_atom_start_compute_cs(struct r600_context *rctx) 712 { 713 struct r600_command_buffer *cb = &rctx->start_compute_cs_cmd; 714 int num_threads; 715 int num_stack_entries; 716 717 /* since all required registers are initialized in the 718 * start_compute_cs_cmd atom, we can EMIT_EARLY here. 719 */ 720 r600_init_command_buffer(cb, 256); 721 cb->pkt_flags = RADEON_CP_PACKET3_COMPUTE_MODE; 722 723 /* This must be first. */ 724 r600_store_value(cb, PKT3(PKT3_CONTEXT_CONTROL, 1, 0)); 725 r600_store_value(cb, 0x80000000); 726 r600_store_value(cb, 0x80000000); 727 728 /* We're setting config registers here. */ 729 r600_store_value(cb, PKT3(PKT3_EVENT_WRITE, 0, 0)); 730 r600_store_value(cb, EVENT_TYPE(EVENT_TYPE_CS_PARTIAL_FLUSH) | EVENT_INDEX(4)); 731 732 switch (rctx->b.family) { 733 case CHIP_CEDAR: 734 default: 735 num_threads = 128; 736 num_stack_entries = 256; 737 break; 738 case CHIP_REDWOOD: 739 num_threads = 128; 740 num_stack_entries = 256; 741 break; 742 case CHIP_JUNIPER: 743 num_threads = 128; 744 num_stack_entries = 512; 745 break; 746 case CHIP_CYPRESS: 747 case CHIP_HEMLOCK: 748 num_threads = 128; 749 num_stack_entries = 512; 750 break; 751 case CHIP_PALM: 752 num_threads = 128; 753 num_stack_entries = 256; 754 break; 755 case CHIP_SUMO: 756 num_threads = 128; 757 num_stack_entries = 256; 758 break; 759 case CHIP_SUMO2: 760 num_threads = 128; 761 num_stack_entries = 512; 762 break; 763 case CHIP_BARTS: 764 num_threads = 128; 765 num_stack_entries = 512; 766 break; 767 case CHIP_TURKS: 768 num_threads = 128; 769 num_stack_entries = 256; 770 break; 771 case CHIP_CAICOS: 772 num_threads = 128; 773 num_stack_entries = 256; 774 break; 775 } 776 777 /* Config Registers */ 778 if (rctx->b.chip_class < CAYMAN) 779 evergreen_init_common_regs(rctx, cb, rctx->b.chip_class, rctx->b.family, 780 rctx->screen->b.info.drm_minor); 781 else 782 cayman_init_common_regs(cb, rctx->b.chip_class, rctx->b.family, 783 rctx->screen->b.info.drm_minor); 784 785 /* The primitive type always needs to be POINTLIST for compute. */ 786 r600_store_config_reg(cb, R_008958_VGT_PRIMITIVE_TYPE, 787 V_008958_DI_PT_POINTLIST); 788 789 if (rctx->b.chip_class < CAYMAN) { 790 791 /* These registers control which simds can be used by each stage. 792 * The default for these registers is 0xffffffff, which means 793 * all simds are available for each stage. It's possible we may 794 * want to play around with these in the future, but for now 795 * the default value is fine. 796 * 797 * R_008E20_SQ_STATIC_THREAD_MGMT1 798 * R_008E24_SQ_STATIC_THREAD_MGMT2 799 * R_008E28_SQ_STATIC_THREAD_MGMT3 800 */ 801 802 /* XXX: We may need to adjust the thread and stack resource 803 * values for 3D/compute interop */ 804 805 r600_store_config_reg_seq(cb, R_008C18_SQ_THREAD_RESOURCE_MGMT_1, 5); 806 807 /* R_008C18_SQ_THREAD_RESOURCE_MGMT_1 808 * Set the number of threads used by the PS/VS/GS/ES stage to 809 * 0. 810 */ 811 r600_store_value(cb, 0); 812 813 /* R_008C1C_SQ_THREAD_RESOURCE_MGMT_2 814 * Set the number of threads used by the CS (aka LS) stage to 815 * the maximum number of threads and set the number of threads 816 * for the HS stage to 0. */ 817 r600_store_value(cb, S_008C1C_NUM_LS_THREADS(num_threads)); 818 819 /* R_008C20_SQ_STACK_RESOURCE_MGMT_1 820 * Set the Control Flow stack entries to 0 for PS/VS stages */ 821 r600_store_value(cb, 0); 822 823 /* R_008C24_SQ_STACK_RESOURCE_MGMT_2 824 * Set the Control Flow stack entries to 0 for GS/ES stages */ 825 r600_store_value(cb, 0); 826 827 /* R_008C28_SQ_STACK_RESOURCE_MGMT_3 828 * Set the Contol Flow stack entries to 0 for the HS stage, and 829 * set it to the maximum value for the CS (aka LS) stage. */ 830 r600_store_value(cb, 831 S_008C28_NUM_LS_STACK_ENTRIES(num_stack_entries)); 832 } 833 /* Give the compute shader all the available LDS space. 834 * NOTE: This only sets the maximum number of dwords that a compute 835 * shader can allocate. When a shader is executed, we still need to 836 * allocate the appropriate amount of LDS dwords using the 837 * CM_R_0288E8_SQ_LDS_ALLOC register. 838 */ 839 if (rctx->b.chip_class < CAYMAN) { 840 r600_store_config_reg(cb, R_008E2C_SQ_LDS_RESOURCE_MGMT, 841 S_008E2C_NUM_PS_LDS(0x0000) | S_008E2C_NUM_LS_LDS(8192)); 842 } else { 843 r600_store_context_reg(cb, CM_R_0286FC_SPI_LDS_MGMT, 844 S_0286FC_NUM_PS_LDS(0) | 845 S_0286FC_NUM_LS_LDS(255)); /* 255 * 32 = 8160 dwords */ 846 } 847 848 /* Context Registers */ 849 850 if (rctx->b.chip_class < CAYMAN) { 851 /* workaround for hw issues with dyn gpr - must set all limits 852 * to 240 instead of 0, 0x1e == 240 / 8 853 */ 854 r600_store_context_reg(cb, R_028838_SQ_DYN_GPR_RESOURCE_LIMIT_1, 855 S_028838_PS_GPRS(0x1e) | 856 S_028838_VS_GPRS(0x1e) | 857 S_028838_GS_GPRS(0x1e) | 858 S_028838_ES_GPRS(0x1e) | 859 S_028838_HS_GPRS(0x1e) | 860 S_028838_LS_GPRS(0x1e)); 861 } 862 863 /* XXX: Investigate setting bit 15, which is FAST_COMPUTE_MODE */ 864 r600_store_context_reg(cb, R_028A40_VGT_GS_MODE, 865 S_028A40_COMPUTE_MODE(1) | S_028A40_PARTIAL_THD_AT_EOI(1)); 866 867 r600_store_context_reg(cb, R_028B54_VGT_SHADER_STAGES_EN, 2/*CS_ON*/); 868 869 r600_store_context_reg(cb, R_0286E8_SPI_COMPUTE_INPUT_CNTL, 870 S_0286E8_TID_IN_GROUP_ENA 871 | S_0286E8_TGID_ENA 872 | S_0286E8_DISABLE_INDEX_PACK) 873 ; 874 875 /* The LOOP_CONST registers are an optimizations for loops that allows 876 * you to store the initial counter, increment value, and maximum 877 * counter value in a register so that hardware can calculate the 878 * correct number of iterations for the loop, so that you don't need 879 * to have the loop counter in your shader code. We don't currently use 880 * this optimization, so we must keep track of the counter in the 881 * shader and use a break instruction to exit loops. However, the 882 * hardware will still uses this register to determine when to exit a 883 * loop, so we need to initialize the counter to 0, set the increment 884 * value to 1 and the maximum counter value to the 4095 (0xfff) which 885 * is the maximum value allowed. This gives us a maximum of 4096 886 * iterations for our loops, but hopefully our break instruction will 887 * execute before some time before the 4096th iteration. 888 */ 889 eg_store_loop_const(cb, R_03A200_SQ_LOOP_CONST_0 + (160 * 4), 0x1000FFF); 890 } 891 892 void evergreen_init_compute_state_functions(struct r600_context *rctx) 893 { 894 rctx->b.b.create_compute_state = evergreen_create_compute_state; 895 rctx->b.b.delete_compute_state = evergreen_delete_compute_state; 896 rctx->b.b.bind_compute_state = evergreen_bind_compute_state; 897 // rctx->context.create_sampler_view = evergreen_compute_create_sampler_view; 898 rctx->b.b.set_compute_resources = evergreen_set_compute_resources; 899 rctx->b.b.set_global_binding = evergreen_set_global_binding; 900 rctx->b.b.launch_grid = evergreen_launch_grid; 901 902 } 903 904 static void *r600_compute_global_transfer_map(struct pipe_context *ctx, 905 struct pipe_resource *resource, 906 unsigned level, 907 unsigned usage, 908 const struct pipe_box *box, 909 struct pipe_transfer **ptransfer) 910 { 911 struct r600_context *rctx = (struct r600_context*)ctx; 912 struct compute_memory_pool *pool = rctx->screen->global_pool; 913 struct r600_resource_global* buffer = 914 (struct r600_resource_global*)resource; 915 916 struct compute_memory_item *item = buffer->chunk; 917 struct pipe_resource *dst = NULL; 918 unsigned offset = box->x; 919 920 if (is_item_in_pool(item)) { 921 compute_memory_demote_item(pool, item, ctx); 922 } 923 else { 924 if (item->real_buffer == NULL) { 925 item->real_buffer = 926 r600_compute_buffer_alloc_vram(pool->screen, item->size_in_dw * 4); 927 } 928 } 929 930 dst = (struct pipe_resource*)item->real_buffer; 931 932 if (usage & PIPE_TRANSFER_READ) 933 buffer->chunk->status |= ITEM_MAPPED_FOR_READING; 934 935 COMPUTE_DBG(rctx->screen, "* r600_compute_global_transfer_map()\n" 936 "level = %u, usage = %u, box(x = %u, y = %u, z = %u " 937 "width = %u, height = %u, depth = %u)\n", level, usage, 938 box->x, box->y, box->z, box->width, box->height, 939 box->depth); 940 COMPUTE_DBG(rctx->screen, "Buffer id = %"PRIi64" offset = " 941 "%u (box.x)\n", item->id, box->x); 942 943 944 assert(resource->target == PIPE_BUFFER); 945 assert(resource->bind & PIPE_BIND_GLOBAL); 946 assert(box->x >= 0); 947 assert(box->y == 0); 948 assert(box->z == 0); 949 950 ///TODO: do it better, mapping is not possible if the pool is too big 951 return pipe_buffer_map_range(ctx, dst, 952 offset, box->width, usage, ptransfer); 953 } 954 955 static void r600_compute_global_transfer_unmap(struct pipe_context *ctx, 956 struct pipe_transfer *transfer) 957 { 958 /* struct r600_resource_global are not real resources, they just map 959 * to an offset within the compute memory pool. The function 960 * r600_compute_global_transfer_map() maps the memory pool 961 * resource rather than the struct r600_resource_global passed to 962 * it as an argument and then initalizes ptransfer->resource with 963 * the memory pool resource (via pipe_buffer_map_range). 964 * When transfer_unmap is called it uses the memory pool's 965 * vtable which calls r600_buffer_transfer_map() rather than 966 * this function. 967 */ 968 assert (!"This function should not be called"); 969 } 970 971 static void r600_compute_global_transfer_flush_region(struct pipe_context *ctx, 972 struct pipe_transfer *transfer, 973 const struct pipe_box *box) 974 { 975 assert(0 && "TODO"); 976 } 977 978 static void r600_compute_global_buffer_destroy(struct pipe_screen *screen, 979 struct pipe_resource *res) 980 { 981 struct r600_resource_global* buffer = NULL; 982 struct r600_screen* rscreen = NULL; 983 984 assert(res->target == PIPE_BUFFER); 985 assert(res->bind & PIPE_BIND_GLOBAL); 986 987 buffer = (struct r600_resource_global*)res; 988 rscreen = (struct r600_screen*)screen; 989 990 compute_memory_free(rscreen->global_pool, buffer->chunk->id); 991 992 buffer->chunk = NULL; 993 free(res); 994 } 995 996 static const struct u_resource_vtbl r600_global_buffer_vtbl = 997 { 998 u_default_resource_get_handle, /* get_handle */ 999 r600_compute_global_buffer_destroy, /* resource_destroy */ 1000 r600_compute_global_transfer_map, /* transfer_map */ 1001 r600_compute_global_transfer_flush_region,/* transfer_flush_region */ 1002 r600_compute_global_transfer_unmap, /* transfer_unmap */ 1003 }; 1004 1005 struct pipe_resource *r600_compute_global_buffer_create(struct pipe_screen *screen, 1006 const struct pipe_resource *templ) 1007 { 1008 struct r600_resource_global* result = NULL; 1009 struct r600_screen* rscreen = NULL; 1010 int size_in_dw = 0; 1011 1012 assert(templ->target == PIPE_BUFFER); 1013 assert(templ->bind & PIPE_BIND_GLOBAL); 1014 assert(templ->array_size == 1 || templ->array_size == 0); 1015 assert(templ->depth0 == 1 || templ->depth0 == 0); 1016 assert(templ->height0 == 1 || templ->height0 == 0); 1017 1018 result = (struct r600_resource_global*) 1019 CALLOC(sizeof(struct r600_resource_global), 1); 1020 rscreen = (struct r600_screen*)screen; 1021 1022 COMPUTE_DBG(rscreen, "*** r600_compute_global_buffer_create\n"); 1023 COMPUTE_DBG(rscreen, "width = %u array_size = %u\n", templ->width0, 1024 templ->array_size); 1025 1026 result->base.b.vtbl = &r600_global_buffer_vtbl; 1027 result->base.b.b = *templ; 1028 result->base.b.b.screen = screen; 1029 pipe_reference_init(&result->base.b.b.reference, 1); 1030 1031 size_in_dw = (templ->width0+3) / 4; 1032 1033 result->chunk = compute_memory_alloc(rscreen->global_pool, size_in_dw); 1034 1035 if (result->chunk == NULL) 1036 { 1037 free(result); 1038 return NULL; 1039 } 1040 1041 return &result->base.b.b; 1042 } 1043