projects / mesa.git / blob
? search:


6b89a1192d9edea775320b1e8bdbd14665b0ecbc
[mesa.git] / src / gallium / drivers / radeonsi / si_pipe.c
1 /*
2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
3 * Copyright 2018 Advanced Micro Devices, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * on the rights to use, copy, modify, merge, publish, distribute, sub
10 * license, and/or sell copies of the Software, and to permit persons to whom
11 * the Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
21 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
22 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
23 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 */
25
26 #include "si_pipe.h"
27 #include "si_public.h"
28 #include "si_shader_internal.h"
29 #include "sid.h"
30
31 #include "ac_llvm_util.h"
32 #include "radeon/radeon_uvd.h"
33 #include "gallivm/lp_bld_misc.h"
34 #include "util/disk_cache.h"
35 #include "util/u_log.h"
36 #include "util/u_memory.h"
37 #include "util/u_suballoc.h"
38 #include "util/u_tests.h"
39 #include "util/u_upload_mgr.h"
40 #include "util/xmlconfig.h"
41 #include "vl/vl_decoder.h"
42 #include "driver_ddebug/dd_util.h"
43
44 static const struct debug_named_value debug_options[] = {
45 /* Shader logging options: */
46 { "vs", DBG(VS), "Print vertex shaders" },
47 { "ps", DBG(PS), "Print pixel shaders" },
48 { "gs", DBG(GS), "Print geometry shaders" },
49 { "tcs", DBG(TCS), "Print tessellation control shaders" },
50 { "tes", DBG(TES), "Print tessellation evaluation shaders" },
51 { "cs", DBG(CS), "Print compute shaders" },
52 { "noir", DBG(NO_IR), "Don't print the LLVM IR"},
53 { "notgsi", DBG(NO_TGSI), "Don't print the TGSI"},
54 { "noasm", DBG(NO_ASM), "Don't print disassembled shaders"},
55 { "preoptir", DBG(PREOPT_IR), "Print the LLVM IR before initial optimizations" },
56
57 /* Shader compiler options the shader cache should be aware of: */
58 { "unsafemath", DBG(UNSAFE_MATH), "Enable unsafe math shader optimizations" },
59 { "sisched", DBG(SI_SCHED), "Enable LLVM SI Machine Instruction Scheduler." },
60 { "gisel", DBG(GISEL), "Enable LLVM global instruction selector." },
61
62 /* Shader compiler options (with no effect on the shader cache): */
63 { "checkir", DBG(CHECK_IR), "Enable additional sanity checks on shader IR" },
64 { "nir", DBG(NIR), "Enable experimental NIR shaders" },
65 { "mono", DBG(MONOLITHIC_SHADERS), "Use old-style monolithic shaders compiled on demand" },
66 { "nooptvariant", DBG(NO_OPT_VARIANT), "Disable compiling optimized shader variants." },
67
68 /* Information logging options: */
69 { "info", DBG(INFO), "Print driver information" },
70 { "tex", DBG(TEX), "Print texture info" },
71 { "compute", DBG(COMPUTE), "Print compute info" },
72 { "vm", DBG(VM), "Print virtual addresses when creating resources" },
73
74 /* Driver options: */
75 { "forcedma", DBG(FORCE_DMA), "Use asynchronous DMA for all operations when possible." },
76 { "nodma", DBG(NO_ASYNC_DMA), "Disable asynchronous DMA" },
77 { "nowc", DBG(NO_WC), "Disable GTT write combining" },
78 { "check_vm", DBG(CHECK_VM), "Check VM faults and dump debug info." },
79 { "reserve_vmid", DBG(RESERVE_VMID), "Force VMID reservation per context." },
80 { "zerovram", DBG(ZERO_VRAM), "Clear VRAM allocations." },
81
82 /* 3D engine options: */
83 { "switch_on_eop", DBG(SWITCH_ON_EOP), "Program WD/IA to switch on end-of-packet." },
84 { "nooutoforder", DBG(NO_OUT_OF_ORDER), "Disable out-of-order rasterization" },
85 { "nodpbb", DBG(NO_DPBB), "Disable DPBB." },
86 { "nodfsm", DBG(NO_DFSM), "Disable DFSM." },
87 { "dpbb", DBG(DPBB), "Enable DPBB." },
88 { "dfsm", DBG(DFSM), "Enable DFSM." },
89 { "nohyperz", DBG(NO_HYPERZ), "Disable Hyper-Z" },
90 { "norbplus", DBG(NO_RB_PLUS), "Disable RB+." },
91 { "no2d", DBG(NO_2D_TILING), "Disable 2D tiling" },
92 { "notiling", DBG(NO_TILING), "Disable tiling" },
93 { "nodcc", DBG(NO_DCC), "Disable DCC." },
94 { "nodccclear", DBG(NO_DCC_CLEAR), "Disable DCC fast clear." },
95 { "nodccfb", DBG(NO_DCC_FB), "Disable separate DCC on the main framebuffer" },
96 { "nodccmsaa", DBG(NO_DCC_MSAA), "Disable DCC for MSAA" },
97 { "nofmask", DBG(NO_FMASK), "Disable MSAA compression" },
98
99 /* Tests: */
100 { "testdma", DBG(TEST_DMA), "Invoke SDMA tests and exit." },
101 { "testvmfaultcp", DBG(TEST_VMFAULT_CP), "Invoke a CP VM fault test and exit." },
102 { "testvmfaultsdma", DBG(TEST_VMFAULT_SDMA), "Invoke a SDMA VM fault test and exit." },
103 { "testvmfaultshader", DBG(TEST_VMFAULT_SHADER), "Invoke a shader VM fault test and exit." },
104 { "testdmaperf", DBG(TEST_DMA_PERF), "Test DMA performance" },
105 { "testgds", DBG(TEST_GDS), "Test GDS." },
106 { "testgdsmm", DBG(TEST_GDS_MM), "Test GDS memory management." },
107 { "testgdsoamm", DBG(TEST_GDS_OA_MM), "Test GDS OA memory management." },
108
109 DEBUG_NAMED_VALUE_END /* must be last */
110 };
111
112 static void si_init_compiler(struct si_screen *sscreen,
113 struct ac_llvm_compiler *compiler)
114 {
115 /* Only create the less-optimizing version of the compiler on APUs
116 * predating Ryzen (Raven). */
117 bool create_low_opt_compiler = !sscreen->info.has_dedicated_vram &&
118 sscreen->info.chip_class <= VI;
119
120 enum ac_target_machine_options tm_options =
121 (sscreen->debug_flags & DBG(SI_SCHED) ? AC_TM_SISCHED : 0) |
122 (sscreen->debug_flags & DBG(GISEL) ? AC_TM_ENABLE_GLOBAL_ISEL : 0) |
123 (sscreen->info.chip_class >= GFX9 ? AC_TM_FORCE_ENABLE_XNACK : 0) |
124 (sscreen->info.chip_class < GFX9 ? AC_TM_FORCE_DISABLE_XNACK : 0) |
125 (!sscreen->llvm_has_working_vgpr_indexing ? AC_TM_PROMOTE_ALLOCA_TO_SCRATCH : 0) |
126 (sscreen->debug_flags & DBG(CHECK_IR) ? AC_TM_CHECK_IR : 0) |
127 (create_low_opt_compiler ? AC_TM_CREATE_LOW_OPT : 0);
128
129 ac_init_llvm_once();
130 ac_init_llvm_compiler(compiler, sscreen->info.family, tm_options);
131 compiler->passes = ac_create_llvm_passes(compiler->tm);
132
133 if (compiler->low_opt_tm)
134 compiler->low_opt_passes = ac_create_llvm_passes(compiler->low_opt_tm);
135 }
136
137 static void si_destroy_compiler(struct ac_llvm_compiler *compiler)
138 {
139 ac_destroy_llvm_passes(compiler->passes);
140 ac_destroy_llvm_passes(compiler->low_opt_passes);
141 ac_destroy_llvm_compiler(compiler);
142 }
143
144 /*
145 * pipe_context
146 */
147 static void si_destroy_context(struct pipe_context *context)
148 {
149 struct si_context *sctx = (struct si_context *)context;
150 int i;
151
152 /* Unreference the framebuffer normally to disable related logic
153 * properly.
154 */
155 struct pipe_framebuffer_state fb = {};
156 if (context->set_framebuffer_state)
157 context->set_framebuffer_state(context, &fb);
158
159 si_release_all_descriptors(sctx);
160
161 pipe_resource_reference(&sctx->esgs_ring, NULL);
162 pipe_resource_reference(&sctx->gsvs_ring, NULL);
163 pipe_resource_reference(&sctx->tess_rings, NULL);
164 pipe_resource_reference(&sctx->null_const_buf.buffer, NULL);
165 pipe_resource_reference(&sctx->sample_pos_buffer, NULL);
166 r600_resource_reference(&sctx->border_color_buffer, NULL);
167 free(sctx->border_color_table);
168 r600_resource_reference(&sctx->scratch_buffer, NULL);
169 r600_resource_reference(&sctx->compute_scratch_buffer, NULL);
170 r600_resource_reference(&sctx->wait_mem_scratch, NULL);
171
172 si_pm4_free_state(sctx, sctx->init_config, ~0);
173 if (sctx->init_config_gs_rings)
174 si_pm4_free_state(sctx, sctx->init_config_gs_rings, ~0);
175 for (i = 0; i < ARRAY_SIZE(sctx->vgt_shader_config); i++)
176 si_pm4_delete_state(sctx, vgt_shader_config, sctx->vgt_shader_config[i]);
177
178 if (sctx->fixed_func_tcs_shader.cso)
179 sctx->b.delete_tcs_state(&sctx->b, sctx->fixed_func_tcs_shader.cso);
180 if (sctx->custom_dsa_flush)
181 sctx->b.delete_depth_stencil_alpha_state(&sctx->b, sctx->custom_dsa_flush);
182 if (sctx->custom_blend_resolve)
183 sctx->b.delete_blend_state(&sctx->b, sctx->custom_blend_resolve);
184 if (sctx->custom_blend_fmask_decompress)
185 sctx->b.delete_blend_state(&sctx->b, sctx->custom_blend_fmask_decompress);
186 if (sctx->custom_blend_eliminate_fastclear)
187 sctx->b.delete_blend_state(&sctx->b, sctx->custom_blend_eliminate_fastclear);
188 if (sctx->custom_blend_dcc_decompress)
189 sctx->b.delete_blend_state(&sctx->b, sctx->custom_blend_dcc_decompress);
190 if (sctx->vs_blit_pos)
191 sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_pos);
192 if (sctx->vs_blit_pos_layered)
193 sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_pos_layered);
194 if (sctx->vs_blit_color)
195 sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_color);
196 if (sctx->vs_blit_color_layered)
197 sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_color_layered);
198 if (sctx->vs_blit_texcoord)
199 sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_texcoord);
200 if (sctx->cs_clear_buffer)
201 sctx->b.delete_compute_state(&sctx->b, sctx->cs_clear_buffer);
202 if (sctx->cs_copy_buffer)
203 sctx->b.delete_compute_state(&sctx->b, sctx->cs_copy_buffer);
204
205 if (sctx->blitter)
206 util_blitter_destroy(sctx->blitter);
207
208 /* Release DCC stats. */
209 for (int i = 0; i < ARRAY_SIZE(sctx->dcc_stats); i++) {
210 assert(!sctx->dcc_stats[i].query_active);
211
212 for (int j = 0; j < ARRAY_SIZE(sctx->dcc_stats[i].ps_stats); j++)
213 if (sctx->dcc_stats[i].ps_stats[j])
214 sctx->b.destroy_query(&sctx->b,
215 sctx->dcc_stats[i].ps_stats[j]);
216
217 si_texture_reference(&sctx->dcc_stats[i].tex, NULL);
218 }
219
220 if (sctx->query_result_shader)
221 sctx->b.delete_compute_state(&sctx->b, sctx->query_result_shader);
222
223 if (sctx->gfx_cs)
224 sctx->ws->cs_destroy(sctx->gfx_cs);
225 if (sctx->dma_cs)
226 sctx->ws->cs_destroy(sctx->dma_cs);
227 if (sctx->ctx)
228 sctx->ws->ctx_destroy(sctx->ctx);
229
230 if (sctx->b.stream_uploader)
231 u_upload_destroy(sctx->b.stream_uploader);
232 if (sctx->b.const_uploader)
233 u_upload_destroy(sctx->b.const_uploader);
234 if (sctx->cached_gtt_allocator)
235 u_upload_destroy(sctx->cached_gtt_allocator);
236
237 slab_destroy_child(&sctx->pool_transfers);
238 slab_destroy_child(&sctx->pool_transfers_unsync);
239
240 if (sctx->allocator_zeroed_memory)
241 u_suballocator_destroy(sctx->allocator_zeroed_memory);
242
243 sctx->ws->fence_reference(&sctx->last_gfx_fence, NULL);
244 sctx->ws->fence_reference(&sctx->last_sdma_fence, NULL);
245 r600_resource_reference(&sctx->eop_bug_scratch, NULL);
246
247 si_destroy_compiler(&sctx->compiler);
248
249 si_saved_cs_reference(&sctx->current_saved_cs, NULL);
250
251 _mesa_hash_table_destroy(sctx->tex_handles, NULL);
252 _mesa_hash_table_destroy(sctx->img_handles, NULL);
253
254 util_dynarray_fini(&sctx->resident_tex_handles);
255 util_dynarray_fini(&sctx->resident_img_handles);
256 util_dynarray_fini(&sctx->resident_tex_needs_color_decompress);
257 util_dynarray_fini(&sctx->resident_img_needs_color_decompress);
258 util_dynarray_fini(&sctx->resident_tex_needs_depth_decompress);
259 FREE(sctx);
260 }
261
262 static enum pipe_reset_status si_get_reset_status(struct pipe_context *ctx)
263 {
264 struct si_context *sctx = (struct si_context *)ctx;
265
266 if (sctx->screen->info.has_gpu_reset_status_query)
267 return sctx->ws->ctx_query_reset_status(sctx->ctx);
268
269 if (sctx->screen->info.has_gpu_reset_counter_query) {
270 unsigned latest = sctx->ws->query_value(sctx->ws,
271 RADEON_GPU_RESET_COUNTER);
272
273 if (sctx->gpu_reset_counter == latest)
274 return PIPE_NO_RESET;
275
276 sctx->gpu_reset_counter = latest;
277 return PIPE_UNKNOWN_CONTEXT_RESET;
278 }
279
280 return PIPE_NO_RESET;
281 }
282
283 static void si_set_device_reset_callback(struct pipe_context *ctx,
284 const struct pipe_device_reset_callback *cb)
285 {
286 struct si_context *sctx = (struct si_context *)ctx;
287
288 if (cb)
289 sctx->device_reset_callback = *cb;
290 else
291 memset(&sctx->device_reset_callback, 0,
292 sizeof(sctx->device_reset_callback));
293 }
294
295 bool si_check_device_reset(struct si_context *sctx)
296 {
297 enum pipe_reset_status status;
298
299 if (!sctx->device_reset_callback.reset)
300 return false;
301
302 if (!sctx->b.get_device_reset_status)
303 return false;
304
305 status = sctx->b.get_device_reset_status(&sctx->b);
306 if (status == PIPE_NO_RESET)
307 return false;
308
309 sctx->device_reset_callback.reset(sctx->device_reset_callback.data, status);
310 return true;
311 }
312
313 /* Apitrace profiling:
314 * 1) qapitrace : Tools -> Profile: Measure CPU & GPU times
315 * 2) In the middle panel, zoom in (mouse wheel) on some bad draw call
316 * and remember its number.
317 * 3) In Mesa, enable queries and performance counters around that draw
318 * call and print the results.
319 * 4) glretrace --benchmark --markers ..
320 */
321 static void si_emit_string_marker(struct pipe_context *ctx,
322 const char *string, int len)
323 {
324 struct si_context *sctx = (struct si_context *)ctx;
325
326 dd_parse_apitrace_marker(string, len, &sctx->apitrace_call_number);
327
328 if (sctx->log)
329 u_log_printf(sctx->log, "\nString marker: %*s\n", len, string);
330 }
331
332 static void si_set_debug_callback(struct pipe_context *ctx,
333 const struct pipe_debug_callback *cb)
334 {
335 struct si_context *sctx = (struct si_context *)ctx;
336 struct si_screen *screen = sctx->screen;
337
338 util_queue_finish(&screen->shader_compiler_queue);
339 util_queue_finish(&screen->shader_compiler_queue_low_priority);
340
341 if (cb)
342 sctx->debug = *cb;
343 else
344 memset(&sctx->debug, 0, sizeof(sctx->debug));
345 }
346
347 static void si_set_log_context(struct pipe_context *ctx,
348 struct u_log_context *log)
349 {
350 struct si_context *sctx = (struct si_context *)ctx;
351 sctx->log = log;
352
353 if (log)
354 u_log_add_auto_logger(log, si_auto_log_cs, sctx);
355 }
356
357 static void si_set_context_param(struct pipe_context *ctx,
358 enum pipe_context_param param,
359 unsigned value)
360 {
361 struct radeon_winsys *ws = ((struct si_context *)ctx)->ws;
362
363 switch (param) {
364 case PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE:
365 ws->pin_threads_to_L3_cache(ws, value);
366 break;
367 default:;
368 }
369 }
370
371 static struct pipe_context *si_create_context(struct pipe_screen *screen,
372 unsigned flags)
373 {
374 struct si_context *sctx = CALLOC_STRUCT(si_context);
375 struct si_screen* sscreen = (struct si_screen *)screen;
376 struct radeon_winsys *ws = sscreen->ws;
377 int shader, i;
378 bool stop_exec_on_failure = (flags & PIPE_CONTEXT_LOSE_CONTEXT_ON_RESET) != 0;
379
380 if (!sctx)
381 return NULL;
382
383 if (flags & PIPE_CONTEXT_DEBUG)
384 sscreen->record_llvm_ir = true; /* racy but not critical */
385
386 sctx->b.screen = screen; /* this must be set first */
387 sctx->b.priv = NULL;
388 sctx->b.destroy = si_destroy_context;
389 sctx->b.emit_string_marker = si_emit_string_marker;
390 sctx->b.set_debug_callback = si_set_debug_callback;
391 sctx->b.set_log_context = si_set_log_context;
392 sctx->b.set_context_param = si_set_context_param;
393 sctx->screen = sscreen; /* Easy accessing of screen/winsys. */
394 sctx->is_debug = (flags & PIPE_CONTEXT_DEBUG) != 0;
395
396 slab_create_child(&sctx->pool_transfers, &sscreen->pool_transfers);
397 slab_create_child(&sctx->pool_transfers_unsync, &sscreen->pool_transfers);
398
399 sctx->ws = sscreen->ws;
400 sctx->family = sscreen->info.family;
401 sctx->chip_class = sscreen->info.chip_class;
402
403 if (sscreen->info.has_gpu_reset_counter_query) {
404 sctx->gpu_reset_counter =
405 sctx->ws->query_value(sctx->ws, RADEON_GPU_RESET_COUNTER);
406 }
407
408 sctx->b.get_device_reset_status = si_get_reset_status;
409 sctx->b.set_device_reset_callback = si_set_device_reset_callback;
410
411 si_init_context_texture_functions(sctx);
412 si_init_query_functions(sctx);
413
414 if (sctx->chip_class == CIK ||
415 sctx->chip_class == VI ||
416 sctx->chip_class == GFX9) {
417 sctx->eop_bug_scratch = r600_resource(
418 pipe_buffer_create(&sscreen->b, 0, PIPE_USAGE_DEFAULT,
419 16 * sscreen->info.num_render_backends));
420 if (!sctx->eop_bug_scratch)
421 goto fail;
422 }
423
424 sctx->allocator_zeroed_memory =
425 u_suballocator_create(&sctx->b, sscreen->info.gart_page_size,
426 0, PIPE_USAGE_DEFAULT,
427 SI_RESOURCE_FLAG_SO_FILLED_SIZE, true);
428 if (!sctx->allocator_zeroed_memory)
429 goto fail;
430
431 sctx->b.stream_uploader = u_upload_create(&sctx->b, 1024 * 1024,
432 0, PIPE_USAGE_STREAM,
433 SI_RESOURCE_FLAG_READ_ONLY);
434 if (!sctx->b.stream_uploader)
435 goto fail;
436
437 sctx->b.const_uploader = u_upload_create(&sctx->b, 128 * 1024,
438 0, PIPE_USAGE_DEFAULT,
439 SI_RESOURCE_FLAG_32BIT |
440 (sscreen->cpdma_prefetch_writes_memory ?
441 0 : SI_RESOURCE_FLAG_READ_ONLY));
442 if (!sctx->b.const_uploader)
443 goto fail;
444
445 sctx->cached_gtt_allocator = u_upload_create(&sctx->b, 16 * 1024,
446 0, PIPE_USAGE_STAGING, 0);
447 if (!sctx->cached_gtt_allocator)
448 goto fail;
449
450 sctx->ctx = sctx->ws->ctx_create(sctx->ws);
451 if (!sctx->ctx)
452 goto fail;
453
454 if (sscreen->info.num_sdma_rings && !(sscreen->debug_flags & DBG(NO_ASYNC_DMA))) {
455 sctx->dma_cs = sctx->ws->cs_create(sctx->ctx, RING_DMA,
456 (void*)si_flush_dma_cs,
457 sctx, stop_exec_on_failure);
458 }
459
460 si_init_buffer_functions(sctx);
461 si_init_clear_functions(sctx);
462 si_init_blit_functions(sctx);
463 si_init_compute_functions(sctx);
464 si_init_compute_blit_functions(sctx);
465 si_init_debug_functions(sctx);
466 si_init_msaa_functions(sctx);
467 si_init_streamout_functions(sctx);
468
469 if (sscreen->info.has_hw_decode) {
470 sctx->b.create_video_codec = si_uvd_create_decoder;
471 sctx->b.create_video_buffer = si_video_buffer_create;
472 } else {
473 sctx->b.create_video_codec = vl_create_decoder;
474 sctx->b.create_video_buffer = vl_video_buffer_create;
475 }
476
477 sctx->gfx_cs = ws->cs_create(sctx->ctx, RING_GFX,
478 (void*)si_flush_gfx_cs, sctx, stop_exec_on_failure);
479
480 /* Border colors. */
481 sctx->border_color_table = malloc(SI_MAX_BORDER_COLORS *
482 sizeof(*sctx->border_color_table));
483 if (!sctx->border_color_table)
484 goto fail;
485
486 sctx->border_color_buffer = r600_resource(
487 pipe_buffer_create(screen, 0, PIPE_USAGE_DEFAULT,
488 SI_MAX_BORDER_COLORS *
489 sizeof(*sctx->border_color_table)));
490 if (!sctx->border_color_buffer)
491 goto fail;
492
493 sctx->border_color_map =
494 ws->buffer_map(sctx->border_color_buffer->buf,
495 NULL, PIPE_TRANSFER_WRITE);
496 if (!sctx->border_color_map)
497 goto fail;
498
499 si_init_all_descriptors(sctx);
500 si_init_fence_functions(sctx);
501 si_init_state_functions(sctx);
502 si_init_shader_functions(sctx);
503 si_init_viewport_functions(sctx);
504
505 if (sctx->chip_class >= CIK)
506 cik_init_sdma_functions(sctx);
507 else
508 si_init_dma_functions(sctx);
509
510 if (sscreen->debug_flags & DBG(FORCE_DMA))
511 sctx->b.resource_copy_region = sctx->dma_copy;
512
513 sctx->blitter = util_blitter_create(&sctx->b);
514 if (sctx->blitter == NULL)
515 goto fail;
516 sctx->blitter->skip_viewport_restore = true;
517
518 si_init_draw_functions(sctx);
519
520 sctx->sample_mask = 0xffff;
521
522 if (sctx->chip_class >= GFX9) {
523 sctx->wait_mem_scratch = r600_resource(
524 pipe_buffer_create(screen, 0, PIPE_USAGE_DEFAULT, 4));
525 if (!sctx->wait_mem_scratch)
526 goto fail;
527
528 /* Initialize the memory. */
529 struct radeon_cmdbuf *cs = sctx->gfx_cs;
530 radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 3, 0));
531 radeon_emit(cs, S_370_DST_SEL(V_370_MEMORY_SYNC) |
532 S_370_WR_CONFIRM(1) |
533 S_370_ENGINE_SEL(V_370_ME));
534 radeon_emit(cs, sctx->wait_mem_scratch->gpu_address);
535 radeon_emit(cs, sctx->wait_mem_scratch->gpu_address >> 32);
536 radeon_emit(cs, sctx->wait_mem_number);
537 radeon_add_to_buffer_list(sctx, cs, sctx->wait_mem_scratch,
538 RADEON_USAGE_WRITE, RADEON_PRIO_FENCE);
539 }
540
541 /* CIK cannot unbind a constant buffer (S_BUFFER_LOAD doesn't skip loads
542 * if NUM_RECORDS == 0). We need to use a dummy buffer instead. */
543 if (sctx->chip_class == CIK) {
544 sctx->null_const_buf.buffer =
545 pipe_aligned_buffer_create(screen,
546 SI_RESOURCE_FLAG_32BIT,
547 PIPE_USAGE_DEFAULT, 16,
548 sctx->screen->info.tcc_cache_line_size);
549 if (!sctx->null_const_buf.buffer)
550 goto fail;
551 sctx->null_const_buf.buffer_size = sctx->null_const_buf.buffer->width0;
552
553 for (shader = 0; shader < SI_NUM_SHADERS; shader++) {
554 for (i = 0; i < SI_NUM_CONST_BUFFERS; i++) {
555 sctx->b.set_constant_buffer(&sctx->b, shader, i,
556 &sctx->null_const_buf);
557 }
558 }
559
560 si_set_rw_buffer(sctx, SI_HS_CONST_DEFAULT_TESS_LEVELS,
561 &sctx->null_const_buf);
562 si_set_rw_buffer(sctx, SI_VS_CONST_INSTANCE_DIVISORS,
563 &sctx->null_const_buf);
564 si_set_rw_buffer(sctx, SI_VS_CONST_CLIP_PLANES,
565 &sctx->null_const_buf);
566 si_set_rw_buffer(sctx, SI_PS_CONST_POLY_STIPPLE,
567 &sctx->null_const_buf);
568 si_set_rw_buffer(sctx, SI_PS_CONST_SAMPLE_POSITIONS,
569 &sctx->null_const_buf);
570 }
571
572 uint64_t max_threads_per_block;
573 screen->get_compute_param(screen, PIPE_SHADER_IR_TGSI,
574 PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK,
575 &max_threads_per_block);
576
577 /* The maximum number of scratch waves. Scratch space isn't divided
578 * evenly between CUs. The number is only a function of the number of CUs.
579 * We can decrease the constant to decrease the scratch buffer size.
580 *
581 * sctx->scratch_waves must be >= the maximum posible size of
582 * 1 threadgroup, so that the hw doesn't hang from being unable
583 * to start any.
584 *
585 * The recommended value is 4 per CU at most. Higher numbers don't
586 * bring much benefit, but they still occupy chip resources (think
587 * async compute). I've seen ~2% performance difference between 4 and 32.
588 */
589 sctx->scratch_waves = MAX2(32 * sscreen->info.num_good_compute_units,
590 max_threads_per_block / 64);
591
592 si_init_compiler(sscreen, &sctx->compiler);
593
594 /* Bindless handles. */
595 sctx->tex_handles = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
596 _mesa_key_pointer_equal);
597 sctx->img_handles = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
598 _mesa_key_pointer_equal);
599
600 util_dynarray_init(&sctx->resident_tex_handles, NULL);
601 util_dynarray_init(&sctx->resident_img_handles, NULL);
602 util_dynarray_init(&sctx->resident_tex_needs_color_decompress, NULL);
603 util_dynarray_init(&sctx->resident_img_needs_color_decompress, NULL);
604 util_dynarray_init(&sctx->resident_tex_needs_depth_decompress, NULL);
605
606 sctx->sample_pos_buffer =
607 pipe_buffer_create(sctx->b.screen, 0, PIPE_USAGE_DEFAULT,
608 sizeof(sctx->sample_positions));
609 pipe_buffer_write(&sctx->b, sctx->sample_pos_buffer, 0,
610 sizeof(sctx->sample_positions), &sctx->sample_positions);
611
612 /* this must be last */
613 si_begin_new_gfx_cs(sctx);
614
615 if (sctx->chip_class == CIK) {
616 /* Clear the NULL constant buffer, because loads should return zeros. */
617 uint32_t clear_value = 0;
618 si_clear_buffer(sctx, sctx->null_const_buf.buffer, 0,
619 sctx->null_const_buf.buffer->width0,
620 &clear_value, 4, SI_COHERENCY_SHADER);
621 }
622 return &sctx->b;
623 fail:
624 fprintf(stderr, "radeonsi: Failed to create a context.\n");
625 si_destroy_context(&sctx->b);
626 return NULL;
627 }
628
629 static struct pipe_context *si_pipe_create_context(struct pipe_screen *screen,
630 void *priv, unsigned flags)
631 {
632 struct si_screen *sscreen = (struct si_screen *)screen;
633 struct pipe_context *ctx;
634
635 if (sscreen->debug_flags & DBG(CHECK_VM))
636 flags |= PIPE_CONTEXT_DEBUG;
637
638 ctx = si_create_context(screen, flags);
639
640 if (!(flags & PIPE_CONTEXT_PREFER_THREADED))
641 return ctx;
642
643 /* Clover (compute-only) is unsupported. */
644 if (flags & PIPE_CONTEXT_COMPUTE_ONLY)
645 return ctx;
646
647 /* When shaders are logged to stderr, asynchronous compilation is
648 * disabled too. */
649 if (sscreen->debug_flags & DBG_ALL_SHADERS)
650 return ctx;
651
652 /* Use asynchronous flushes only on amdgpu, since the radeon
653 * implementation for fence_server_sync is incomplete. */
654 return threaded_context_create(ctx, &sscreen->pool_transfers,
655 si_replace_buffer_storage,
656 sscreen->info.drm_major >= 3 ? si_create_fence : NULL,
657 &((struct si_context*)ctx)->tc);
658 }
659
660 /*
661 * pipe_screen
662 */
663 static void si_destroy_screen(struct pipe_screen* pscreen)
664 {
665 struct si_screen *sscreen = (struct si_screen *)pscreen;
666 struct si_shader_part *parts[] = {
667 sscreen->vs_prologs,
668 sscreen->tcs_epilogs,
669 sscreen->gs_prologs,
670 sscreen->ps_prologs,
671 sscreen->ps_epilogs
672 };
673 unsigned i;
674
675 if (!sscreen->ws->unref(sscreen->ws))
676 return;
677
678 util_queue_destroy(&sscreen->shader_compiler_queue);
679 util_queue_destroy(&sscreen->shader_compiler_queue_low_priority);
680
681 for (i = 0; i < ARRAY_SIZE(sscreen->compiler); i++)
682 si_destroy_compiler(&sscreen->compiler[i]);
683
684 for (i = 0; i < ARRAY_SIZE(sscreen->compiler_lowp); i++)
685 si_destroy_compiler(&sscreen->compiler_lowp[i]);
686
687 /* Free shader parts. */
688 for (i = 0; i < ARRAY_SIZE(parts); i++) {
689 while (parts[i]) {
690 struct si_shader_part *part = parts[i];
691
692 parts[i] = part->next;
693 ac_shader_binary_clean(&part->binary);
694 FREE(part);
695 }
696 }
697 mtx_destroy(&sscreen->shader_parts_mutex);
698 si_destroy_shader_cache(sscreen);
699
700 si_destroy_perfcounters(sscreen);
701 si_gpu_load_kill_thread(sscreen);
702
703 mtx_destroy(&sscreen->gpu_load_mutex);
704 mtx_destroy(&sscreen->aux_context_lock);
705 sscreen->aux_context->destroy(sscreen->aux_context);
706
707 slab_destroy_parent(&sscreen->pool_transfers);
708
709 disk_cache_destroy(sscreen->disk_shader_cache);
710 sscreen->ws->destroy(sscreen->ws);
711 FREE(sscreen);
712 }
713
714 static void si_init_gs_info(struct si_screen *sscreen)
715 {
716 sscreen->gs_table_depth = ac_get_gs_table_depth(sscreen->info.chip_class,
717 sscreen->info.family);
718 }
719
720 static void si_test_vmfault(struct si_screen *sscreen)
721 {
722 struct pipe_context *ctx = sscreen->aux_context;
723 struct si_context *sctx = (struct si_context *)ctx;
724 struct pipe_resource *buf =
725 pipe_buffer_create_const0(&sscreen->b, 0, PIPE_USAGE_DEFAULT, 64);
726
727 if (!buf) {
728 puts("Buffer allocation failed.");
729 exit(1);
730 }
731
732 r600_resource(buf)->gpu_address = 0; /* cause a VM fault */
733
734 if (sscreen->debug_flags & DBG(TEST_VMFAULT_CP)) {
735 si_cp_dma_copy_buffer(sctx, buf, buf, 0, 4, 4, 0,
736 SI_COHERENCY_NONE, L2_BYPASS);
737 ctx->flush(ctx, NULL, 0);
738 puts("VM fault test: CP - done.");
739 }
740 if (sscreen->debug_flags & DBG(TEST_VMFAULT_SDMA)) {
741 si_sdma_clear_buffer(sctx, buf, 0, 4, 0);
742 ctx->flush(ctx, NULL, 0);
743 puts("VM fault test: SDMA - done.");
744 }
745 if (sscreen->debug_flags & DBG(TEST_VMFAULT_SHADER)) {
746 util_test_constant_buffer(ctx, buf);
747 puts("VM fault test: Shader - done.");
748 }
749 exit(0);
750 }
751
752 static void si_test_gds_memory_management(struct si_context *sctx,
753 unsigned alloc_size, unsigned alignment,
754 enum radeon_bo_domain domain)
755 {
756 struct radeon_winsys *ws = sctx->ws;
757 struct radeon_cmdbuf *cs[8];
758 struct pb_buffer *gds_bo[ARRAY_SIZE(cs)];
759
760 for (unsigned i = 0; i < ARRAY_SIZE(cs); i++) {
761 cs[i] = ws->cs_create(sctx->ctx, RING_COMPUTE,
762 NULL, NULL, false);
763 gds_bo[i] = ws->buffer_create(ws, alloc_size, alignment, domain, 0);
764 assert(gds_bo[i]);
765 }
766
767 for (unsigned iterations = 0; iterations < 20000; iterations++) {
768 for (unsigned i = 0; i < ARRAY_SIZE(cs); i++) {
769 /* This clears GDS with CP DMA.
770 *
771 * We don't care if GDS is present. Just add some packet
772 * to make the GPU busy for a moment.
773 */
774 si_cp_dma_clear_buffer(sctx, cs[i], NULL, 0, alloc_size, 0,
775 SI_CPDMA_SKIP_BO_LIST_UPDATE |
776 SI_CPDMA_SKIP_CHECK_CS_SPACE |
777 SI_CPDMA_SKIP_GFX_SYNC, 0, 0);
778
779 ws->cs_add_buffer(cs[i], gds_bo[i], domain,
780 RADEON_USAGE_READWRITE, 0);
781 ws->cs_flush(cs[i], PIPE_FLUSH_ASYNC, NULL);
782 }
783 }
784 exit(0);
785 }
786
787 static void si_disk_cache_create(struct si_screen *sscreen)
788 {
789 /* Don't use the cache if shader dumping is enabled. */
790 if (sscreen->debug_flags & DBG_ALL_SHADERS)
791 return;
792
793 struct mesa_sha1 ctx;
794 unsigned char sha1[20];
795 char cache_id[20 * 2 + 1];
796
797 _mesa_sha1_init(&ctx);
798
799 if (!disk_cache_get_function_identifier(si_disk_cache_create, &ctx) ||
800 !disk_cache_get_function_identifier(LLVMInitializeAMDGPUTargetInfo,
801 &ctx))
802 return;
803
804 _mesa_sha1_final(&ctx, sha1);
805 disk_cache_format_hex_id(cache_id, sha1, 20 * 2);
806
807 /* These flags affect shader compilation. */
808 #define ALL_FLAGS (DBG(FS_CORRECT_DERIVS_AFTER_KILL) | \
809 DBG(SI_SCHED) | \
810 DBG(GISEL) | \
811 DBG(UNSAFE_MATH) | \
812 DBG(NIR))
813 uint64_t shader_debug_flags = sscreen->debug_flags &
814 ALL_FLAGS;
815
816 /* Add the high bits of 32-bit addresses, which affects
817 * how 32-bit addresses are expanded to 64 bits.
818 */
819 STATIC_ASSERT(ALL_FLAGS <= UINT_MAX);
820 shader_debug_flags |= (uint64_t)sscreen->info.address32_hi << 32;
821
822 sscreen->disk_shader_cache =
823 disk_cache_create(sscreen->info.name,
824 cache_id,
825 shader_debug_flags);
826 }
827
828 struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws,
829 const struct pipe_screen_config *config)
830 {
831 struct si_screen *sscreen = CALLOC_STRUCT(si_screen);
832 unsigned hw_threads, num_comp_hi_threads, num_comp_lo_threads, i;
833
834 if (!sscreen) {
835 return NULL;
836 }
837
838 sscreen->ws = ws;
839 ws->query_info(ws, &sscreen->info);
840
841 if (sscreen->info.chip_class >= GFX9) {
842 sscreen->se_tile_repeat = 32 * sscreen->info.max_se;
843 } else {
844 ac_get_raster_config(&sscreen->info,
845 &sscreen->pa_sc_raster_config,
846 &sscreen->pa_sc_raster_config_1,
847 &sscreen->se_tile_repeat);
848 }
849
850 sscreen->debug_flags = debug_get_flags_option("R600_DEBUG",
851 debug_options, 0);
852
853 /* Set functions first. */
854 sscreen->b.context_create = si_pipe_create_context;
855 sscreen->b.destroy = si_destroy_screen;
856
857 si_init_screen_get_functions(sscreen);
858 si_init_screen_buffer_functions(sscreen);
859 si_init_screen_fence_functions(sscreen);
860 si_init_screen_state_functions(sscreen);
861 si_init_screen_texture_functions(sscreen);
862 si_init_screen_query_functions(sscreen);
863
864 /* Set these flags in debug_flags early, so that the shader cache takes
865 * them into account.
866 */
867 if (driQueryOptionb(config->options,
868 "glsl_correct_derivatives_after_discard"))
869 sscreen->debug_flags |= DBG(FS_CORRECT_DERIVS_AFTER_KILL);
870 if (driQueryOptionb(config->options, "radeonsi_enable_sisched"))
871 sscreen->debug_flags |= DBG(SI_SCHED);
872
873
874 if (sscreen->debug_flags & DBG(INFO))
875 ac_print_gpu_info(&sscreen->info);
876
877 slab_create_parent(&sscreen->pool_transfers,
878 sizeof(struct si_transfer), 64);
879
880 sscreen->force_aniso = MIN2(16, debug_get_num_option("R600_TEX_ANISO", -1));
881 if (sscreen->force_aniso >= 0) {
882 printf("radeonsi: Forcing anisotropy filter to %ix\n",
883 /* round down to a power of two */
884 1 << util_logbase2(sscreen->force_aniso));
885 }
886
887 (void) mtx_init(&sscreen->aux_context_lock, mtx_plain);
888 (void) mtx_init(&sscreen->gpu_load_mutex, mtx_plain);
889
890 si_init_gs_info(sscreen);
891 if (!si_init_shader_cache(sscreen)) {
892 FREE(sscreen);
893 return NULL;
894 }
895
896 si_disk_cache_create(sscreen);
897
898 /* Determine the number of shader compiler threads. */
899 hw_threads = sysconf(_SC_NPROCESSORS_ONLN);
900
901 if (hw_threads >= 12) {
902 num_comp_hi_threads = hw_threads * 3 / 4;
903 num_comp_lo_threads = hw_threads / 3;
904 } else if (hw_threads >= 6) {
905 num_comp_hi_threads = hw_threads - 2;
906 num_comp_lo_threads = hw_threads / 2;
907 } else if (hw_threads >= 2) {
908 num_comp_hi_threads = hw_threads - 1;
909 num_comp_lo_threads = hw_threads / 2;
910 } else {
911 num_comp_hi_threads = 1;
912 num_comp_lo_threads = 1;
913 }
914
915 num_comp_hi_threads = MIN2(num_comp_hi_threads,
916 ARRAY_SIZE(sscreen->compiler));
917 num_comp_lo_threads = MIN2(num_comp_lo_threads,
918 ARRAY_SIZE(sscreen->compiler_lowp));
919
920 if (!util_queue_init(&sscreen->shader_compiler_queue, "sh",
921 64, num_comp_hi_threads,
922 UTIL_QUEUE_INIT_RESIZE_IF_FULL |
923 UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY)) {
924 si_destroy_shader_cache(sscreen);
925 FREE(sscreen);
926 return NULL;
927 }
928
929 if (!util_queue_init(&sscreen->shader_compiler_queue_low_priority,
930 "shlo",
931 64, num_comp_lo_threads,
932 UTIL_QUEUE_INIT_RESIZE_IF_FULL |
933 UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY |
934 UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY)) {
935 si_destroy_shader_cache(sscreen);
936 FREE(sscreen);
937 return NULL;
938 }
939
940 if (!debug_get_bool_option("RADEON_DISABLE_PERFCOUNTERS", false))
941 si_init_perfcounters(sscreen);
942
943 /* Determine tessellation ring info. */
944 bool double_offchip_buffers = sscreen->info.chip_class >= CIK &&
945 sscreen->info.family != CHIP_CARRIZO &&
946 sscreen->info.family != CHIP_STONEY;
947 /* This must be one less than the maximum number due to a hw limitation.
948 * Various hardware bugs in SI, CIK, and GFX9 need this.
949 */
950 unsigned max_offchip_buffers_per_se;
951
952 /* Only certain chips can use the maximum value. */
953 if (sscreen->info.family == CHIP_VEGA12 ||
954 sscreen->info.family == CHIP_VEGA20)
955 max_offchip_buffers_per_se = double_offchip_buffers ? 128 : 64;
956 else
957 max_offchip_buffers_per_se = double_offchip_buffers ? 127 : 63;
958
959 unsigned max_offchip_buffers = max_offchip_buffers_per_se *
960 sscreen->info.max_se;
961 unsigned offchip_granularity;
962
963 /* Hawaii has a bug with offchip buffers > 256 that can be worked
964 * around by setting 4K granularity.
965 */
966 if (sscreen->info.family == CHIP_HAWAII) {
967 sscreen->tess_offchip_block_dw_size = 4096;
968 offchip_granularity = V_03093C_X_4K_DWORDS;
969 } else {
970 sscreen->tess_offchip_block_dw_size = 8192;
971 offchip_granularity = V_03093C_X_8K_DWORDS;
972 }
973
974 sscreen->tess_factor_ring_size = 32768 * sscreen->info.max_se;
975 assert(((sscreen->tess_factor_ring_size / 4) & C_030938_SIZE) == 0);
976 sscreen->tess_offchip_ring_size = max_offchip_buffers *
977 sscreen->tess_offchip_block_dw_size * 4;
978
979 if (sscreen->info.chip_class >= CIK) {
980 if (sscreen->info.chip_class >= VI)
981 --max_offchip_buffers;
982 sscreen->vgt_hs_offchip_param =
983 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers) |
984 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity);
985 } else {
986 assert(offchip_granularity == V_03093C_X_8K_DWORDS);
987 sscreen->vgt_hs_offchip_param =
988 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers);
989 }
990
991 /* The mere presense of CLEAR_STATE in the IB causes random GPU hangs
992 * on SI. Some CLEAR_STATE cause asic hang on radeon kernel, etc.
993 * SPI_VS_OUT_CONFIG. So only enable CI CLEAR_STATE on amdgpu kernel.*/
994 sscreen->has_clear_state = sscreen->info.chip_class >= CIK &&
995 sscreen->info.drm_major == 3;
996
997 sscreen->has_distributed_tess =
998 sscreen->info.chip_class >= VI &&
999 sscreen->info.max_se >= 2;
1000
1001 sscreen->has_draw_indirect_multi =
1002 (sscreen->info.family >= CHIP_POLARIS10) ||
1003 (sscreen->info.chip_class == VI &&
1004 sscreen->info.pfp_fw_version >= 121 &&
1005 sscreen->info.me_fw_version >= 87) ||
1006 (sscreen->info.chip_class == CIK &&
1007 sscreen->info.pfp_fw_version >= 211 &&
1008 sscreen->info.me_fw_version >= 173) ||
1009 (sscreen->info.chip_class == SI &&
1010 sscreen->info.pfp_fw_version >= 79 &&
1011 sscreen->info.me_fw_version >= 142);
1012
1013 sscreen->has_out_of_order_rast = sscreen->info.chip_class >= VI &&
1014 sscreen->info.max_se >= 2 &&
1015 !(sscreen->debug_flags & DBG(NO_OUT_OF_ORDER));
1016 sscreen->assume_no_z_fights =
1017 driQueryOptionb(config->options, "radeonsi_assume_no_z_fights");
1018 sscreen->commutative_blend_add =
1019 driQueryOptionb(config->options, "radeonsi_commutative_blend_add");
1020 sscreen->clear_db_cache_before_clear =
1021 driQueryOptionb(config->options, "radeonsi_clear_db_cache_before_clear");
1022 sscreen->has_msaa_sample_loc_bug = (sscreen->info.family >= CHIP_POLARIS10 &&
1023 sscreen->info.family <= CHIP_POLARIS12) ||
1024 sscreen->info.family == CHIP_VEGA10 ||
1025 sscreen->info.family == CHIP_RAVEN;
1026 sscreen->has_ls_vgpr_init_bug = sscreen->info.family == CHIP_VEGA10 ||
1027 sscreen->info.family == CHIP_RAVEN;
1028 sscreen->has_dcc_constant_encode = sscreen->info.family == CHIP_RAVEN2;
1029
1030 /* Only enable primitive binning on APUs by default. */
1031 sscreen->dpbb_allowed = sscreen->info.family == CHIP_RAVEN ||
1032 sscreen->info.family == CHIP_RAVEN2;
1033
1034 sscreen->dfsm_allowed = sscreen->info.family == CHIP_RAVEN ||
1035 sscreen->info.family == CHIP_RAVEN2;
1036
1037 /* Process DPBB enable flags. */
1038 if (sscreen->debug_flags & DBG(DPBB)) {
1039 sscreen->dpbb_allowed = true;
1040 if (sscreen->debug_flags & DBG(DFSM))
1041 sscreen->dfsm_allowed = true;
1042 }
1043
1044 /* Process DPBB disable flags. */
1045 if (sscreen->debug_flags & DBG(NO_DPBB)) {
1046 sscreen->dpbb_allowed = false;
1047 sscreen->dfsm_allowed = false;
1048 } else if (sscreen->debug_flags & DBG(NO_DFSM)) {
1049 sscreen->dfsm_allowed = false;
1050 }
1051
1052 /* While it would be nice not to have this flag, we are constrained
1053 * by the reality that LLVM 5.0 doesn't have working VGPR indexing
1054 * on GFX9.
1055 */
1056 sscreen->llvm_has_working_vgpr_indexing = sscreen->info.chip_class <= VI;
1057
1058 /* Some chips have RB+ registers, but don't support RB+. Those must
1059 * always disable it.
1060 */
1061 if (sscreen->info.family == CHIP_STONEY ||
1062 sscreen->info.chip_class >= GFX9) {
1063 sscreen->has_rbplus = true;
1064
1065 sscreen->rbplus_allowed =
1066 !(sscreen->debug_flags & DBG(NO_RB_PLUS)) &&
1067 (sscreen->info.family == CHIP_STONEY ||
1068 sscreen->info.family == CHIP_VEGA12 ||
1069 sscreen->info.family == CHIP_RAVEN ||
1070 sscreen->info.family == CHIP_RAVEN2);
1071 }
1072
1073 sscreen->dcc_msaa_allowed =
1074 !(sscreen->debug_flags & DBG(NO_DCC_MSAA));
1075
1076 sscreen->cpdma_prefetch_writes_memory = sscreen->info.chip_class <= VI;
1077
1078 (void) mtx_init(&sscreen->shader_parts_mutex, mtx_plain);
1079 sscreen->use_monolithic_shaders =
1080 (sscreen->debug_flags & DBG(MONOLITHIC_SHADERS)) != 0;
1081
1082 sscreen->barrier_flags.cp_to_L2 = SI_CONTEXT_INV_SMEM_L1 |
1083 SI_CONTEXT_INV_VMEM_L1;
1084 if (sscreen->info.chip_class <= VI) {
1085 sscreen->barrier_flags.cp_to_L2 |= SI_CONTEXT_INV_GLOBAL_L2;
1086 sscreen->barrier_flags.L2_to_cp |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
1087 }
1088
1089 if (debug_get_bool_option("RADEON_DUMP_SHADERS", false))
1090 sscreen->debug_flags |= DBG_ALL_SHADERS;
1091
1092 /* Syntax:
1093 * EQAA=s,z,c
1094 * Example:
1095 * EQAA=8,4,2
1096
1097 * That means 8 coverage samples, 4 Z/S samples, and 2 color samples.
1098 * Constraints:
1099 * s >= z >= c (ignoring this only wastes memory)
1100 * s = [2..16]
1101 * z = [2..8]
1102 * c = [2..8]
1103 *
1104 * Only MSAA color and depth buffers are overriden.
1105 */
1106 if (sscreen->info.has_eqaa_surface_allocator) {
1107 const char *eqaa = debug_get_option("EQAA", NULL);
1108 unsigned s,z,f;
1109
1110 if (eqaa && sscanf(eqaa, "%u,%u,%u", &s, &z, &f) == 3 && s && z && f) {
1111 sscreen->eqaa_force_coverage_samples = s;
1112 sscreen->eqaa_force_z_samples = z;
1113 sscreen->eqaa_force_color_samples = f;
1114 }
1115 }
1116
1117 for (i = 0; i < num_comp_hi_threads; i++)
1118 si_init_compiler(sscreen, &sscreen->compiler[i]);
1119 for (i = 0; i < num_comp_lo_threads; i++)
1120 si_init_compiler(sscreen, &sscreen->compiler_lowp[i]);
1121
1122 /* Create the auxiliary context. This must be done last. */
1123 sscreen->aux_context = si_create_context(&sscreen->b, 0);
1124
1125 if (sscreen->debug_flags & DBG(TEST_DMA))
1126 si_test_dma(sscreen);
1127
1128 if (sscreen->debug_flags & DBG(TEST_DMA_PERF)) {
1129 si_test_dma_perf(sscreen);
1130 }
1131
1132 if (sscreen->debug_flags & (DBG(TEST_VMFAULT_CP) |
1133 DBG(TEST_VMFAULT_SDMA) |
1134 DBG(TEST_VMFAULT_SHADER)))
1135 si_test_vmfault(sscreen);
1136
1137 if (sscreen->debug_flags & DBG(TEST_GDS))
1138 si_test_gds((struct si_context*)sscreen->aux_context);
1139
1140 if (sscreen->debug_flags & DBG(TEST_GDS_MM)) {
1141 si_test_gds_memory_management((struct si_context*)sscreen->aux_context,
1142 32 * 1024, 4, RADEON_DOMAIN_GDS);
1143 }
1144 if (sscreen->debug_flags & DBG(TEST_GDS_OA_MM)) {
1145 si_test_gds_memory_management((struct si_context*)sscreen->aux_context,
1146 4, 1, RADEON_DOMAIN_OA);
1147 }
1148
1149 return &sscreen->b;
1150 }