evergreen_compute.c revision 257d697fb92c617161a2a90607c4cde6f7b9034a
1/* 2 * Copyright 2011 Adam Rak <adam.rak@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@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/u_double_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#ifdef HAVE_OPENCL 50#include "radeon_llvm_util.h" 51#endif 52 53/** 54RAT0 is for global binding write 55VTX1 is for global binding read 56 57for wrting images RAT1... 58for reading images TEX2... 59 TEX2-RAT1 is paired 60 61TEX2... consumes the same fetch resources, that VTX2... would consume 62 63CONST0 and VTX0 is for parameters 64 CONST0 is binding smaller input parameter buffer, and for constant indexing, 65 also constant cached 66 VTX0 is for indirect/non-constant indexing, or if the input is bigger than 67 the constant cache can handle 68 69RAT-s are limited to 12, so we can only bind at most 11 texture for writing 70because we reserve RAT0 for global bindings. With byteaddressing enabled, 71we should reserve another one too.=> 10 image binding for writing max. 72 73from Nvidia OpenCL: 74 CL_DEVICE_MAX_READ_IMAGE_ARGS: 128 75 CL_DEVICE_MAX_WRITE_IMAGE_ARGS: 8 76 77so 10 for writing is enough. 176 is the max for reading according to the docs 78 79writable images should be listed first < 10, so their id corresponds to RAT(id+1) 80writable images will consume TEX slots, VTX slots too because of linear indexing 81 82*/ 83 84struct r600_resource* r600_compute_buffer_alloc_vram( 85 struct r600_screen *screen, 86 unsigned size) 87{ 88 struct pipe_resource * buffer = NULL; 89 assert(size); 90 91 buffer = pipe_buffer_create( 92 (struct pipe_screen*) screen, 93 PIPE_BIND_CUSTOM, 94 PIPE_USAGE_IMMUTABLE, 95 size); 96 97 return (struct r600_resource *)buffer; 98} 99 100 101static void evergreen_set_rat( 102 struct r600_pipe_compute *pipe, 103 int id, 104 struct r600_resource* bo, 105 int start, 106 int size) 107{ 108 struct pipe_surface rat_templ; 109 struct r600_surface *surf = NULL; 110 struct r600_context *rctx = NULL; 111 112 assert(id < 12); 113 assert((size & 3) == 0); 114 assert((start & 0xFF) == 0); 115 116 rctx = pipe->ctx; 117 118 COMPUTE_DBG(rctx->screen, "bind rat: %i \n", id); 119 120 /* Create the RAT surface */ 121 memset(&rat_templ, 0, sizeof(rat_templ)); 122 rat_templ.format = PIPE_FORMAT_R32_UINT; 123 rat_templ.u.tex.level = 0; 124 rat_templ.u.tex.first_layer = 0; 125 rat_templ.u.tex.last_layer = 0; 126 127 /* Add the RAT the list of color buffers */ 128 pipe->ctx->framebuffer.state.cbufs[id] = pipe->ctx->b.b.create_surface( 129 (struct pipe_context *)pipe->ctx, 130 (struct pipe_resource *)bo, &rat_templ); 131 132 /* Update the number of color buffers */ 133 pipe->ctx->framebuffer.state.nr_cbufs = 134 MAX2(id + 1, pipe->ctx->framebuffer.state.nr_cbufs); 135 136 /* Update the cb_target_mask 137 * XXX: I think this is a potential spot for bugs once we start doing 138 * GL interop. cb_target_mask may be modified in the 3D sections 139 * of this driver. */ 140 pipe->ctx->compute_cb_target_mask |= (0xf << (id * 4)); 141 142 surf = (struct r600_surface*)pipe->ctx->framebuffer.state.cbufs[id]; 143 evergreen_init_color_surface_rat(rctx, surf); 144} 145 146static void evergreen_cs_set_vertex_buffer( 147 struct r600_context * rctx, 148 unsigned vb_index, 149 unsigned offset, 150 struct pipe_resource * buffer) 151{ 152 struct r600_vertexbuf_state *state = &rctx->cs_vertex_buffer_state; 153 struct pipe_vertex_buffer *vb = &state->vb[vb_index]; 154 vb->stride = 1; 155 vb->buffer_offset = offset; 156 vb->buffer = buffer; 157 vb->user_buffer = NULL; 158 159 /* The vertex instructions in the compute shaders use the texture cache, 160 * so we need to invalidate it. */ 161 rctx->b.flags |= R600_CONTEXT_INV_VERTEX_CACHE; 162 state->enabled_mask |= 1 << vb_index; 163 state->dirty_mask |= 1 << vb_index; 164 state->atom.dirty = true; 165} 166 167static void evergreen_cs_set_constant_buffer( 168 struct r600_context * rctx, 169 unsigned cb_index, 170 unsigned offset, 171 unsigned size, 172 struct pipe_resource * buffer) 173{ 174 struct pipe_constant_buffer cb; 175 cb.buffer_size = size; 176 cb.buffer_offset = offset; 177 cb.buffer = buffer; 178 cb.user_buffer = NULL; 179 180 rctx->b.b.set_constant_buffer(&rctx->b.b, PIPE_SHADER_COMPUTE, cb_index, &cb); 181} 182 183static const struct u_resource_vtbl r600_global_buffer_vtbl = 184{ 185 u_default_resource_get_handle, /* get_handle */ 186 r600_compute_global_buffer_destroy, /* resource_destroy */ 187 r600_compute_global_transfer_map, /* transfer_map */ 188 r600_compute_global_transfer_flush_region,/* transfer_flush_region */ 189 r600_compute_global_transfer_unmap, /* transfer_unmap */ 190 r600_compute_global_transfer_inline_write /* transfer_inline_write */ 191}; 192 193 194void *evergreen_create_compute_state( 195 struct pipe_context *ctx_, 196 const const struct pipe_compute_state *cso) 197{ 198 struct r600_context *ctx = (struct r600_context *)ctx_; 199 struct r600_pipe_compute *shader = CALLOC_STRUCT(r600_pipe_compute); 200 201#ifdef HAVE_OPENCL 202 const struct pipe_llvm_program_header * header; 203 const unsigned char * code; 204 unsigned i; 205 206 shader->llvm_ctx = LLVMContextCreate(); 207 208 COMPUTE_DBG(ctx->screen, "*** evergreen_create_compute_state\n"); 209 210 header = cso->prog; 211 code = cso->prog + sizeof(struct pipe_llvm_program_header); 212#endif 213 214 shader->ctx = (struct r600_context*)ctx; 215 shader->local_size = cso->req_local_mem; 216 shader->private_size = cso->req_private_mem; 217 shader->input_size = cso->req_input_mem; 218 219#ifdef HAVE_OPENCL 220 shader->num_kernels = radeon_llvm_get_num_kernels(shader->llvm_ctx, code, 221 header->num_bytes); 222 shader->kernels = CALLOC(sizeof(struct r600_kernel), shader->num_kernels); 223 224 for (i = 0; i < shader->num_kernels; i++) { 225 struct r600_kernel *kernel = &shader->kernels[i]; 226 kernel->llvm_module = radeon_llvm_get_kernel_module(shader->llvm_ctx, i, 227 code, header->num_bytes); 228 } 229#endif 230 return shader; 231} 232 233void evergreen_delete_compute_state(struct pipe_context *ctx, void* state) 234{ 235 struct r600_pipe_compute *shader = (struct r600_pipe_compute *)state; 236 237 if (!shader) 238 return; 239 240 FREE(shader->kernels); 241 242#ifdef HAVE_OPENCL 243 if (shader->llvm_ctx){ 244 LLVMContextDispose(shader->llvm_ctx); 245 } 246#endif 247 248 FREE(shader); 249} 250 251static void evergreen_bind_compute_state(struct pipe_context *ctx_, void *state) 252{ 253 struct r600_context *ctx = (struct r600_context *)ctx_; 254 255 COMPUTE_DBG(ctx->screen, "*** evergreen_bind_compute_state\n"); 256 257 ctx->cs_shader_state.shader = (struct r600_pipe_compute *)state; 258} 259 260/* The kernel parameters are stored a vtx buffer (ID=0), besides the explicit 261 * kernel parameters there are implicit parameters that need to be stored 262 * in the vertex buffer as well. Here is how these parameters are organized in 263 * the buffer: 264 * 265 * DWORDS 0-2: Number of work groups in each dimension (x,y,z) 266 * DWORDS 3-5: Number of global work items in each dimension (x,y,z) 267 * DWORDS 6-8: Number of work items within each work group in each dimension 268 * (x,y,z) 269 * DWORDS 9+ : Kernel parameters 270 */ 271void evergreen_compute_upload_input( 272 struct pipe_context *ctx_, 273 const uint *block_layout, 274 const uint *grid_layout, 275 const void *input) 276{ 277 struct r600_context *ctx = (struct r600_context *)ctx_; 278 struct r600_pipe_compute *shader = ctx->cs_shader_state.shader; 279 int i; 280 /* We need to reserve 9 dwords (36 bytes) for implicit kernel 281 * parameters. 282 */ 283 unsigned input_size = shader->input_size + 36; 284 uint32_t * num_work_groups_start; 285 uint32_t * global_size_start; 286 uint32_t * local_size_start; 287 uint32_t * kernel_parameters_start; 288 struct pipe_box box; 289 struct pipe_transfer *transfer = NULL; 290 291 if (shader->input_size == 0) { 292 return; 293 } 294 295 if (!shader->kernel_param) { 296 /* Add space for the grid dimensions */ 297 shader->kernel_param = (struct r600_resource *) 298 pipe_buffer_create(ctx_->screen, PIPE_BIND_CUSTOM, 299 PIPE_USAGE_IMMUTABLE, input_size); 300 } 301 302 u_box_1d(0, input_size, &box); 303 num_work_groups_start = ctx_->transfer_map(ctx_, 304 (struct pipe_resource*)shader->kernel_param, 305 0, PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD_RANGE, 306 &box, &transfer); 307 global_size_start = num_work_groups_start + (3 * (sizeof(uint) /4)); 308 local_size_start = global_size_start + (3 * (sizeof(uint)) / 4); 309 kernel_parameters_start = local_size_start + (3 * (sizeof(uint)) / 4); 310 311 /* Copy the work group size */ 312 memcpy(num_work_groups_start, grid_layout, 3 * sizeof(uint)); 313 314 /* Copy the global size */ 315 for (i = 0; i < 3; i++) { 316 global_size_start[i] = grid_layout[i] * block_layout[i]; 317 } 318 319 /* Copy the local dimensions */ 320 memcpy(local_size_start, block_layout, 3 * sizeof(uint)); 321 322 /* Copy the kernel inputs */ 323 memcpy(kernel_parameters_start, input, shader->input_size); 324 325 for (i = 0; i < (input_size / 4); i++) { 326 COMPUTE_DBG(ctx->screen, "input %i : %u\n", i, 327 ((unsigned*)num_work_groups_start)[i]); 328 } 329 330 ctx_->transfer_unmap(ctx_, transfer); 331 332 /* ID=0 is reserved for the parameters */ 333 evergreen_cs_set_constant_buffer(ctx, 0, 0, input_size, 334 (struct pipe_resource*)shader->kernel_param); 335} 336 337static void evergreen_emit_direct_dispatch( 338 struct r600_context *rctx, 339 const uint *block_layout, const uint *grid_layout) 340{ 341 int i; 342 struct radeon_winsys_cs *cs = rctx->b.rings.gfx.cs; 343 struct r600_pipe_compute *shader = rctx->cs_shader_state.shader; 344 unsigned num_waves; 345 unsigned num_pipes = rctx->screen->b.info.r600_max_pipes; 346 unsigned wave_divisor = (16 * num_pipes); 347 int group_size = 1; 348 int grid_size = 1; 349 unsigned lds_size = shader->local_size / 4 + shader->active_kernel->bc.nlds_dw; 350 351 /* Calculate group_size/grid_size */ 352 for (i = 0; i < 3; i++) { 353 group_size *= block_layout[i]; 354 } 355 356 for (i = 0; i < 3; i++) { 357 grid_size *= grid_layout[i]; 358 } 359 360 /* num_waves = ceil((tg_size.x * tg_size.y, tg_size.z) / (16 * num_pipes)) */ 361 num_waves = (block_layout[0] * block_layout[1] * block_layout[2] + 362 wave_divisor - 1) / wave_divisor; 363 364 COMPUTE_DBG(rctx->screen, "Using %u pipes, " 365 "%u wavefronts per thread block, " 366 "allocating %u dwords lds.\n", 367 num_pipes, num_waves, lds_size); 368 369 r600_write_config_reg(cs, R_008970_VGT_NUM_INDICES, group_size); 370 371 r600_write_config_reg_seq(cs, R_00899C_VGT_COMPUTE_START_X, 3); 372 radeon_emit(cs, 0); /* R_00899C_VGT_COMPUTE_START_X */ 373 radeon_emit(cs, 0); /* R_0089A0_VGT_COMPUTE_START_Y */ 374 radeon_emit(cs, 0); /* R_0089A4_VGT_COMPUTE_START_Z */ 375 376 r600_write_config_reg(cs, R_0089AC_VGT_COMPUTE_THREAD_GROUP_SIZE, 377 group_size); 378 379 r600_write_compute_context_reg_seq(cs, R_0286EC_SPI_COMPUTE_NUM_THREAD_X, 3); 380 radeon_emit(cs, block_layout[0]); /* R_0286EC_SPI_COMPUTE_NUM_THREAD_X */ 381 radeon_emit(cs, block_layout[1]); /* R_0286F0_SPI_COMPUTE_NUM_THREAD_Y */ 382 radeon_emit(cs, block_layout[2]); /* R_0286F4_SPI_COMPUTE_NUM_THREAD_Z */ 383 384 if (rctx->b.chip_class < CAYMAN) { 385 assert(lds_size <= 8192); 386 } else { 387 /* Cayman appears to have a slightly smaller limit, see the 388 * value of CM_R_0286FC_SPI_LDS_MGMT.NUM_LS_LDS */ 389 assert(lds_size <= 8160); 390 } 391 392 r600_write_compute_context_reg(cs, CM_R_0288E8_SQ_LDS_ALLOC, 393 lds_size | (num_waves << 14)); 394 395 /* Dispatch packet */ 396 radeon_emit(cs, PKT3C(PKT3_DISPATCH_DIRECT, 3, 0)); 397 radeon_emit(cs, grid_layout[0]); 398 radeon_emit(cs, grid_layout[1]); 399 radeon_emit(cs, grid_layout[2]); 400 /* VGT_DISPATCH_INITIATOR = COMPUTE_SHADER_EN */ 401 radeon_emit(cs, 1); 402} 403 404static void compute_emit_cs(struct r600_context *ctx, const uint *block_layout, 405 const uint *grid_layout) 406{ 407 struct radeon_winsys_cs *cs = ctx->b.rings.gfx.cs; 408 int i; 409 410 /* make sure that the gfx ring is only one active */ 411 if (ctx->b.rings.dma.cs && ctx->b.rings.dma.cs->cdw) { 412 ctx->b.rings.dma.flush(ctx, RADEON_FLUSH_ASYNC, NULL); 413 } 414 415 /* Initialize all the compute-related registers. 416 * 417 * See evergreen_init_atom_start_compute_cs() in this file for the list 418 * of registers initialized by the start_compute_cs_cmd atom. 419 */ 420 r600_emit_command_buffer(cs, &ctx->start_compute_cs_cmd); 421 422 ctx->b.flags |= R600_CONTEXT_WAIT_3D_IDLE | R600_CONTEXT_FLUSH_AND_INV; 423 r600_flush_emit(ctx); 424 425 /* Emit colorbuffers. */ 426 /* XXX support more than 8 colorbuffers (the offsets are not a multiple of 0x3C for CB8-11) */ 427 for (i = 0; i < 8 && i < ctx->framebuffer.state.nr_cbufs; i++) { 428 struct r600_surface *cb = (struct r600_surface*)ctx->framebuffer.state.cbufs[i]; 429 unsigned reloc = r600_context_bo_reloc(&ctx->b, &ctx->b.rings.gfx, 430 (struct r600_resource*)cb->base.texture, 431 RADEON_USAGE_READWRITE, 432 RADEON_PRIO_SHADER_RESOURCE_RW); 433 434 r600_write_compute_context_reg_seq(cs, R_028C60_CB_COLOR0_BASE + i * 0x3C, 7); 435 radeon_emit(cs, cb->cb_color_base); /* R_028C60_CB_COLOR0_BASE */ 436 radeon_emit(cs, cb->cb_color_pitch); /* R_028C64_CB_COLOR0_PITCH */ 437 radeon_emit(cs, cb->cb_color_slice); /* R_028C68_CB_COLOR0_SLICE */ 438 radeon_emit(cs, cb->cb_color_view); /* R_028C6C_CB_COLOR0_VIEW */ 439 radeon_emit(cs, cb->cb_color_info); /* R_028C70_CB_COLOR0_INFO */ 440 radeon_emit(cs, cb->cb_color_attrib); /* R_028C74_CB_COLOR0_ATTRIB */ 441 radeon_emit(cs, cb->cb_color_dim); /* R_028C78_CB_COLOR0_DIM */ 442 443 radeon_emit(cs, PKT3(PKT3_NOP, 0, 0)); /* R_028C60_CB_COLOR0_BASE */ 444 radeon_emit(cs, reloc); 445 446 if (!ctx->keep_tiling_flags) { 447 radeon_emit(cs, PKT3(PKT3_NOP, 0, 0)); /* R_028C70_CB_COLOR0_INFO */ 448 radeon_emit(cs, reloc); 449 } 450 451 radeon_emit(cs, PKT3(PKT3_NOP, 0, 0)); /* R_028C74_CB_COLOR0_ATTRIB */ 452 radeon_emit(cs, reloc); 453 } 454 if (ctx->keep_tiling_flags) { 455 for (; i < 8 ; i++) { 456 r600_write_compute_context_reg(cs, R_028C70_CB_COLOR0_INFO + i * 0x3C, 457 S_028C70_FORMAT(V_028C70_COLOR_INVALID)); 458 } 459 for (; i < 12; i++) { 460 r600_write_compute_context_reg(cs, R_028E50_CB_COLOR8_INFO + (i - 8) * 0x1C, 461 S_028C70_FORMAT(V_028C70_COLOR_INVALID)); 462 } 463 } 464 465 /* Set CB_TARGET_MASK XXX: Use cb_misc_state */ 466 r600_write_compute_context_reg(cs, R_028238_CB_TARGET_MASK, 467 ctx->compute_cb_target_mask); 468 469 470 /* Emit vertex buffer state */ 471 ctx->cs_vertex_buffer_state.atom.num_dw = 12 * util_bitcount(ctx->cs_vertex_buffer_state.dirty_mask); 472 r600_emit_atom(ctx, &ctx->cs_vertex_buffer_state.atom); 473 474 /* Emit constant buffer state */ 475 r600_emit_atom(ctx, &ctx->constbuf_state[PIPE_SHADER_COMPUTE].atom); 476 477 /* Emit compute shader state */ 478 r600_emit_atom(ctx, &ctx->cs_shader_state.atom); 479 480 /* Emit dispatch state and dispatch packet */ 481 evergreen_emit_direct_dispatch(ctx, block_layout, grid_layout); 482 483 /* XXX evergreen_flush_emit() hardcodes the CP_COHER_SIZE to 0xffffffff 484 */ 485 ctx->b.flags |= R600_CONTEXT_INV_CONST_CACHE | 486 R600_CONTEXT_INV_VERTEX_CACHE | 487 R600_CONTEXT_INV_TEX_CACHE; 488 r600_flush_emit(ctx); 489 ctx->b.flags = 0; 490 491 if (ctx->b.chip_class >= CAYMAN) { 492 cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 0, 0); 493 cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_CS_PARTIAL_FLUSH) | EVENT_INDEX(4); 494 /* DEALLOC_STATE prevents the GPU from hanging when a 495 * SURFACE_SYNC packet is emitted some time after a DISPATCH_DIRECT 496 * with any of the CB*_DEST_BASE_ENA or DB_DEST_BASE_ENA bits set. 497 */ 498 cs->buf[cs->cdw++] = PKT3C(PKT3_DEALLOC_STATE, 0, 0); 499 cs->buf[cs->cdw++] = 0; 500 } 501 502#if 0 503 COMPUTE_DBG(ctx->screen, "cdw: %i\n", cs->cdw); 504 for (i = 0; i < cs->cdw; i++) { 505 COMPUTE_DBG(ctx->screen, "%4i : 0x%08X\n", i, cs->buf[i]); 506 } 507#endif 508 509} 510 511 512/** 513 * Emit function for r600_cs_shader_state atom 514 */ 515void evergreen_emit_cs_shader( 516 struct r600_context *rctx, 517 struct r600_atom *atom) 518{ 519 struct r600_cs_shader_state *state = 520 (struct r600_cs_shader_state*)atom; 521 struct r600_pipe_compute *shader = state->shader; 522 struct r600_kernel *kernel = &shader->kernels[state->kernel_index]; 523 struct radeon_winsys_cs *cs = rctx->b.rings.gfx.cs; 524 uint64_t va; 525 526 va = r600_resource_va(&rctx->screen->b.b, &kernel->code_bo->b.b); 527 528 r600_write_compute_context_reg_seq(cs, R_0288D0_SQ_PGM_START_LS, 3); 529 radeon_emit(cs, va >> 8); /* R_0288D0_SQ_PGM_START_LS */ 530 radeon_emit(cs, /* R_0288D4_SQ_PGM_RESOURCES_LS */ 531 S_0288D4_NUM_GPRS(kernel->bc.ngpr) 532 | S_0288D4_STACK_SIZE(kernel->bc.nstack)); 533 radeon_emit(cs, 0); /* R_0288D8_SQ_PGM_RESOURCES_LS_2 */ 534 535 radeon_emit(cs, PKT3C(PKT3_NOP, 0, 0)); 536 radeon_emit(cs, r600_context_bo_reloc(&rctx->b, &rctx->b.rings.gfx, 537 kernel->code_bo, RADEON_USAGE_READ, 538 RADEON_PRIO_SHADER_DATA)); 539} 540 541static void evergreen_launch_grid( 542 struct pipe_context *ctx_, 543 const uint *block_layout, const uint *grid_layout, 544 uint32_t pc, const void *input) 545{ 546 struct r600_context *ctx = (struct r600_context *)ctx_; 547 548 struct r600_pipe_compute *shader = ctx->cs_shader_state.shader; 549 struct r600_kernel *kernel = &shader->kernels[pc]; 550 551 COMPUTE_DBG(ctx->screen, "*** evergreen_launch_grid: pc = %u\n", pc); 552 553#ifdef HAVE_OPENCL 554 555 if (!kernel->code_bo) { 556 void *p; 557 struct r600_bytecode *bc = &kernel->bc; 558 LLVMModuleRef mod = kernel->llvm_module; 559 boolean use_kill = false; 560 bool dump = (ctx->screen->b.debug_flags & DBG_CS) != 0; 561 unsigned use_sb = ctx->screen->b.debug_flags & DBG_SB_CS; 562 unsigned sb_disasm = use_sb || 563 (ctx->screen->b.debug_flags & DBG_SB_DISASM); 564 565 r600_bytecode_init(bc, ctx->b.chip_class, ctx->b.family, 566 ctx->screen->has_compressed_msaa_texturing); 567 bc->type = TGSI_PROCESSOR_COMPUTE; 568 bc->isa = ctx->isa; 569 r600_llvm_compile(mod, ctx->b.family, bc, &use_kill, dump); 570 571 if (dump && !sb_disasm) { 572 r600_bytecode_disasm(bc); 573 } else if ((dump && sb_disasm) || use_sb) { 574 if (r600_sb_bytecode_process(ctx, bc, NULL, dump, use_sb)) 575 R600_ERR("r600_sb_bytecode_process failed!\n"); 576 } 577 578 kernel->code_bo = r600_compute_buffer_alloc_vram(ctx->screen, 579 kernel->bc.ndw * 4); 580 p = r600_buffer_map_sync_with_rings(&ctx->b, kernel->code_bo, PIPE_TRANSFER_WRITE); 581 memcpy(p, kernel->bc.bytecode, kernel->bc.ndw * 4); 582 ctx->b.ws->buffer_unmap(kernel->code_bo->cs_buf); 583 } 584#endif 585 shader->active_kernel = kernel; 586 ctx->cs_shader_state.kernel_index = pc; 587 evergreen_compute_upload_input(ctx_, block_layout, grid_layout, input); 588 compute_emit_cs(ctx, block_layout, grid_layout); 589} 590 591static void evergreen_set_compute_resources(struct pipe_context * ctx_, 592 unsigned start, unsigned count, 593 struct pipe_surface ** surfaces) 594{ 595 struct r600_context *ctx = (struct r600_context *)ctx_; 596 struct r600_surface **resources = (struct r600_surface **)surfaces; 597 598 COMPUTE_DBG(ctx->screen, "*** evergreen_set_compute_resources: start = %u count = %u\n", 599 start, count); 600 601 for (int i = 0; i < count; i++) { 602 /* The First two vertex buffers are reserved for parameters and 603 * global buffers. */ 604 unsigned vtx_id = 2 + i; 605 if (resources[i]) { 606 struct r600_resource_global *buffer = 607 (struct r600_resource_global*) 608 resources[i]->base.texture; 609 if (resources[i]->base.writable) { 610 assert(i+1 < 12); 611 612 evergreen_set_rat(ctx->cs_shader_state.shader, i+1, 613 (struct r600_resource *)resources[i]->base.texture, 614 buffer->chunk->start_in_dw*4, 615 resources[i]->base.texture->width0); 616 } 617 618 evergreen_cs_set_vertex_buffer(ctx, vtx_id, 619 buffer->chunk->start_in_dw * 4, 620 resources[i]->base.texture); 621 } 622 } 623} 624 625void evergreen_set_cs_sampler_view(struct pipe_context *ctx_, 626 unsigned start_slot, unsigned count, 627 struct pipe_sampler_view **views) 628{ 629 struct r600_pipe_sampler_view **resource = 630 (struct r600_pipe_sampler_view **)views; 631 632 for (int i = 0; i < count; i++) { 633 if (resource[i]) { 634 assert(i+1 < 12); 635 /* XXX: Implement */ 636 assert(!"Compute samplers not implemented."); 637 ///FETCH0 = VTX0 (param buffer), 638 //FETCH1 = VTX1 (global buffer pool), FETCH2... = TEX 639 } 640 } 641} 642 643 644static void evergreen_set_global_binding( 645 struct pipe_context *ctx_, unsigned first, unsigned n, 646 struct pipe_resource **resources, 647 uint32_t **handles) 648{ 649 struct r600_context *ctx = (struct r600_context *)ctx_; 650 struct compute_memory_pool *pool = ctx->screen->global_pool; 651 struct r600_resource_global **buffers = 652 (struct r600_resource_global **)resources; 653 654 COMPUTE_DBG(ctx->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 (int i = 0; i < 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 compute_memory_finalize_pending(pool, ctx_); 672 673 for (int i = 0; i < n; i++) 674 { 675 uint32_t buffer_offset; 676 uint32_t handle; 677 assert(resources[i]->target == PIPE_BUFFER); 678 assert(resources[i]->bind & PIPE_BIND_GLOBAL); 679 680 buffer_offset = util_le32_to_cpu(*(handles[i])); 681 handle = buffer_offset + buffers[i]->chunk->start_in_dw * 4; 682 683 *(handles[i]) = util_cpu_to_le32(handle); 684 } 685 686 evergreen_set_rat(ctx->cs_shader_state.shader, 0, pool->bo, 0, pool->size_in_dw * 4); 687 evergreen_cs_set_vertex_buffer(ctx, 1, 0, 688 (struct pipe_resource*)pool->bo); 689} 690 691/** 692 * This function initializes all the compute specific registers that need to 693 * be initialized for each compute command stream. Registers that are common 694 * to both compute and 3D will be initialized at the beginning of each compute 695 * command stream by the start_cs_cmd atom. However, since the SET_CONTEXT_REG 696 * packet requires that the shader type bit be set, we must initialize all 697 * context registers needed for compute in this function. The registers 698 * intialized by the start_cs_cmd atom can be found in evereen_state.c in the 699 * functions evergreen_init_atom_start_cs or cayman_init_atom_start_cs depending 700 * on the GPU family. 701 */ 702void evergreen_init_atom_start_compute_cs(struct r600_context *ctx) 703{ 704 struct r600_command_buffer *cb = &ctx->start_compute_cs_cmd; 705 int num_threads; 706 int num_stack_entries; 707 708 /* since all required registers are initialised in the 709 * start_compute_cs_cmd atom, we can EMIT_EARLY here. 710 */ 711 r600_init_command_buffer(cb, 256); 712 cb->pkt_flags = RADEON_CP_PACKET3_COMPUTE_MODE; 713 714 /* This must be first. */ 715 r600_store_value(cb, PKT3(PKT3_CONTEXT_CONTROL, 1, 0)); 716 r600_store_value(cb, 0x80000000); 717 r600_store_value(cb, 0x80000000); 718 719 /* We're setting config registers here. */ 720 r600_store_value(cb, PKT3(PKT3_EVENT_WRITE, 0, 0)); 721 r600_store_value(cb, EVENT_TYPE(EVENT_TYPE_CS_PARTIAL_FLUSH) | EVENT_INDEX(4)); 722 723 switch (ctx->b.family) { 724 case CHIP_CEDAR: 725 default: 726 num_threads = 128; 727 num_stack_entries = 256; 728 break; 729 case CHIP_REDWOOD: 730 num_threads = 128; 731 num_stack_entries = 256; 732 break; 733 case CHIP_JUNIPER: 734 num_threads = 128; 735 num_stack_entries = 512; 736 break; 737 case CHIP_CYPRESS: 738 case CHIP_HEMLOCK: 739 num_threads = 128; 740 num_stack_entries = 512; 741 break; 742 case CHIP_PALM: 743 num_threads = 128; 744 num_stack_entries = 256; 745 break; 746 case CHIP_SUMO: 747 num_threads = 128; 748 num_stack_entries = 256; 749 break; 750 case CHIP_SUMO2: 751 num_threads = 128; 752 num_stack_entries = 512; 753 break; 754 case CHIP_BARTS: 755 num_threads = 128; 756 num_stack_entries = 512; 757 break; 758 case CHIP_TURKS: 759 num_threads = 128; 760 num_stack_entries = 256; 761 break; 762 case CHIP_CAICOS: 763 num_threads = 128; 764 num_stack_entries = 256; 765 break; 766 } 767 768 /* Config Registers */ 769 if (ctx->b.chip_class < CAYMAN) 770 evergreen_init_common_regs(cb, ctx->b.chip_class, ctx->b.family, 771 ctx->screen->b.info.drm_minor); 772 else 773 cayman_init_common_regs(cb, ctx->b.chip_class, ctx->b.family, 774 ctx->screen->b.info.drm_minor); 775 776 /* The primitive type always needs to be POINTLIST for compute. */ 777 r600_store_config_reg(cb, R_008958_VGT_PRIMITIVE_TYPE, 778 V_008958_DI_PT_POINTLIST); 779 780 if (ctx->b.chip_class < CAYMAN) { 781 782 /* These registers control which simds can be used by each stage. 783 * The default for these registers is 0xffffffff, which means 784 * all simds are available for each stage. It's possible we may 785 * want to play around with these in the future, but for now 786 * the default value is fine. 787 * 788 * R_008E20_SQ_STATIC_THREAD_MGMT1 789 * R_008E24_SQ_STATIC_THREAD_MGMT2 790 * R_008E28_SQ_STATIC_THREAD_MGMT3 791 */ 792 793 /* XXX: We may need to adjust the thread and stack resouce 794 * values for 3D/compute interop */ 795 796 r600_store_config_reg_seq(cb, R_008C18_SQ_THREAD_RESOURCE_MGMT_1, 5); 797 798 /* R_008C18_SQ_THREAD_RESOURCE_MGMT_1 799 * Set the number of threads used by the PS/VS/GS/ES stage to 800 * 0. 801 */ 802 r600_store_value(cb, 0); 803 804 /* R_008C1C_SQ_THREAD_RESOURCE_MGMT_2 805 * Set the number of threads used by the CS (aka LS) stage to 806 * the maximum number of threads and set the number of threads 807 * for the HS stage to 0. */ 808 r600_store_value(cb, S_008C1C_NUM_LS_THREADS(num_threads)); 809 810 /* R_008C20_SQ_STACK_RESOURCE_MGMT_1 811 * Set the Control Flow stack entries to 0 for PS/VS stages */ 812 r600_store_value(cb, 0); 813 814 /* R_008C24_SQ_STACK_RESOURCE_MGMT_2 815 * Set the Control Flow stack entries to 0 for GS/ES stages */ 816 r600_store_value(cb, 0); 817 818 /* R_008C28_SQ_STACK_RESOURCE_MGMT_3 819 * Set the Contol Flow stack entries to 0 for the HS stage, and 820 * set it to the maximum value for the CS (aka LS) stage. */ 821 r600_store_value(cb, 822 S_008C28_NUM_LS_STACK_ENTRIES(num_stack_entries)); 823 } 824 /* Give the compute shader all the available LDS space. 825 * NOTE: This only sets the maximum number of dwords that a compute 826 * shader can allocate. When a shader is executed, we still need to 827 * allocate the appropriate amount of LDS dwords using the 828 * CM_R_0288E8_SQ_LDS_ALLOC register. 829 */ 830 if (ctx->b.chip_class < CAYMAN) { 831 r600_store_config_reg(cb, R_008E2C_SQ_LDS_RESOURCE_MGMT, 832 S_008E2C_NUM_PS_LDS(0x0000) | S_008E2C_NUM_LS_LDS(8192)); 833 } else { 834 r600_store_context_reg(cb, CM_R_0286FC_SPI_LDS_MGMT, 835 S_0286FC_NUM_PS_LDS(0) | 836 S_0286FC_NUM_LS_LDS(255)); /* 255 * 32 = 8160 dwords */ 837 } 838 839 /* Context Registers */ 840 841 if (ctx->b.chip_class < CAYMAN) { 842 /* workaround for hw issues with dyn gpr - must set all limits 843 * to 240 instead of 0, 0x1e == 240 / 8 844 */ 845 r600_store_context_reg(cb, R_028838_SQ_DYN_GPR_RESOURCE_LIMIT_1, 846 S_028838_PS_GPRS(0x1e) | 847 S_028838_VS_GPRS(0x1e) | 848 S_028838_GS_GPRS(0x1e) | 849 S_028838_ES_GPRS(0x1e) | 850 S_028838_HS_GPRS(0x1e) | 851 S_028838_LS_GPRS(0x1e)); 852 } 853 854 /* XXX: Investigate setting bit 15, which is FAST_COMPUTE_MODE */ 855 r600_store_context_reg(cb, R_028A40_VGT_GS_MODE, 856 S_028A40_COMPUTE_MODE(1) | S_028A40_PARTIAL_THD_AT_EOI(1)); 857 858 r600_store_context_reg(cb, R_028B54_VGT_SHADER_STAGES_EN, 2/*CS_ON*/); 859 860 r600_store_context_reg(cb, R_0286E8_SPI_COMPUTE_INPUT_CNTL, 861 S_0286E8_TID_IN_GROUP_ENA 862 | S_0286E8_TGID_ENA 863 | S_0286E8_DISABLE_INDEX_PACK) 864 ; 865 866 /* The LOOP_CONST registers are an optimizations for loops that allows 867 * you to store the initial counter, increment value, and maximum 868 * counter value in a register so that hardware can calculate the 869 * correct number of iterations for the loop, so that you don't need 870 * to have the loop counter in your shader code. We don't currently use 871 * this optimization, so we must keep track of the counter in the 872 * shader and use a break instruction to exit loops. However, the 873 * hardware will still uses this register to determine when to exit a 874 * loop, so we need to initialize the counter to 0, set the increment 875 * value to 1 and the maximum counter value to the 4095 (0xfff) which 876 * is the maximum value allowed. This gives us a maximum of 4096 877 * iterations for our loops, but hopefully our break instruction will 878 * execute before some time before the 4096th iteration. 879 */ 880 eg_store_loop_const(cb, R_03A200_SQ_LOOP_CONST_0 + (160 * 4), 0x1000FFF); 881} 882 883void evergreen_init_compute_state_functions(struct r600_context *ctx) 884{ 885 ctx->b.b.create_compute_state = evergreen_create_compute_state; 886 ctx->b.b.delete_compute_state = evergreen_delete_compute_state; 887 ctx->b.b.bind_compute_state = evergreen_bind_compute_state; 888// ctx->context.create_sampler_view = evergreen_compute_create_sampler_view; 889 ctx->b.b.set_compute_resources = evergreen_set_compute_resources; 890 ctx->b.b.set_global_binding = evergreen_set_global_binding; 891 ctx->b.b.launch_grid = evergreen_launch_grid; 892 893 /* We always use at least one vertex buffer for parameters (id = 1)*/ 894 ctx->cs_vertex_buffer_state.enabled_mask = 895 ctx->cs_vertex_buffer_state.dirty_mask = 0x2; 896} 897 898struct pipe_resource *r600_compute_global_buffer_create( 899 struct pipe_screen *screen, 900 const struct pipe_resource *templ) 901{ 902 struct r600_resource_global* result = NULL; 903 struct r600_screen* rscreen = NULL; 904 int size_in_dw = 0; 905 906 assert(templ->target == PIPE_BUFFER); 907 assert(templ->bind & PIPE_BIND_GLOBAL); 908 assert(templ->array_size == 1 || templ->array_size == 0); 909 assert(templ->depth0 == 1 || templ->depth0 == 0); 910 assert(templ->height0 == 1 || templ->height0 == 0); 911 912 result = (struct r600_resource_global*) 913 CALLOC(sizeof(struct r600_resource_global), 1); 914 rscreen = (struct r600_screen*)screen; 915 916 COMPUTE_DBG(rscreen, "*** r600_compute_global_buffer_create\n"); 917 COMPUTE_DBG(rscreen, "width = %u array_size = %u\n", templ->width0, 918 templ->array_size); 919 920 result->base.b.vtbl = &r600_global_buffer_vtbl; 921 result->base.b.b.screen = screen; 922 result->base.b.b = *templ; 923 pipe_reference_init(&result->base.b.b.reference, 1); 924 925 size_in_dw = (templ->width0+3) / 4; 926 927 result->chunk = compute_memory_alloc(rscreen->global_pool, size_in_dw); 928 929 if (result->chunk == NULL) 930 { 931 free(result); 932 return NULL; 933 } 934 935 return &result->base.b.b; 936} 937 938void r600_compute_global_buffer_destroy( 939 struct pipe_screen *screen, 940 struct pipe_resource *res) 941{ 942 struct r600_resource_global* buffer = NULL; 943 struct r600_screen* rscreen = NULL; 944 945 assert(res->target == PIPE_BUFFER); 946 assert(res->bind & PIPE_BIND_GLOBAL); 947 948 buffer = (struct r600_resource_global*)res; 949 rscreen = (struct r600_screen*)screen; 950 951 compute_memory_free(rscreen->global_pool, buffer->chunk->id); 952 953 buffer->chunk = NULL; 954 free(res); 955} 956 957void *r600_compute_global_transfer_map( 958 struct pipe_context *ctx_, 959 struct pipe_resource *resource, 960 unsigned level, 961 unsigned usage, 962 const struct pipe_box *box, 963 struct pipe_transfer **ptransfer) 964{ 965 struct r600_context *rctx = (struct r600_context*)ctx_; 966 struct compute_memory_pool *pool = rctx->screen->global_pool; 967 struct r600_resource_global* buffer = 968 (struct r600_resource_global*)resource; 969 970 struct pipe_resource *dst; 971 unsigned offset = box->x; 972 973 if (is_item_in_pool(buffer->chunk)) { 974 compute_memory_demote_item(pool, buffer->chunk, ctx_); 975 } 976 977 dst = (struct pipe_resource*)buffer->chunk->real_buffer; 978 979 if (usage & PIPE_TRANSFER_READ) 980 buffer->chunk->status |= ITEM_MAPPED_FOR_READING; 981 982 COMPUTE_DBG(rctx->screen, "* r600_compute_global_transfer_map()\n" 983 "level = %u, usage = %u, box(x = %u, y = %u, z = %u " 984 "width = %u, height = %u, depth = %u)\n", level, usage, 985 box->x, box->y, box->z, box->width, box->height, 986 box->depth); 987 COMPUTE_DBG(rctx->screen, "Buffer id = %u offset = " 988 "%u (box.x)\n", buffer->chunk->id, box->x); 989 990 991 assert(resource->target == PIPE_BUFFER); 992 assert(resource->bind & PIPE_BIND_GLOBAL); 993 assert(box->x >= 0); 994 assert(box->y == 0); 995 assert(box->z == 0); 996 997 ///TODO: do it better, mapping is not possible if the pool is too big 998 return pipe_buffer_map_range(ctx_, dst, 999 offset, box->width, usage, ptransfer); 1000} 1001 1002void r600_compute_global_transfer_unmap( 1003 struct pipe_context *ctx_, 1004 struct pipe_transfer* transfer) 1005{ 1006 /* struct r600_resource_global are not real resources, they just map 1007 * to an offset within the compute memory pool. The function 1008 * r600_compute_global_transfer_map() maps the memory pool 1009 * resource rather than the struct r600_resource_global passed to 1010 * it as an argument and then initalizes ptransfer->resource with 1011 * the memory pool resource (via pipe_buffer_map_range). 1012 * When transfer_unmap is called it uses the memory pool's 1013 * vtable which calls r600_buffer_transfer_map() rather than 1014 * this function. 1015 */ 1016 assert (!"This function should not be called"); 1017} 1018 1019void r600_compute_global_transfer_flush_region( 1020 struct pipe_context *ctx_, 1021 struct pipe_transfer *transfer, 1022 const struct pipe_box *box) 1023{ 1024 assert(0 && "TODO"); 1025} 1026 1027void r600_compute_global_transfer_inline_write( 1028 struct pipe_context *pipe, 1029 struct pipe_resource *resource, 1030 unsigned level, 1031 unsigned usage, 1032 const struct pipe_box *box, 1033 const void *data, 1034 unsigned stride, 1035 unsigned layer_stride) 1036{ 1037 assert(0 && "TODO"); 1038} 1039