si_pipe.c revision 993e1d59
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
44static 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
107	DEBUG_NAMED_VALUE_END /* must be last */
108};
109
110static void si_init_compiler(struct si_screen *sscreen,
111			     struct ac_llvm_compiler *compiler)
112{
113	/* Only create the less-optimizing version of the compiler on APUs
114	 * predating Ryzen (Raven). */
115	bool create_low_opt_compiler = !sscreen->info.has_dedicated_vram &&
116				       sscreen->info.chip_class <= VI;
117
118	enum ac_target_machine_options tm_options =
119		(sscreen->debug_flags & DBG(SI_SCHED) ? AC_TM_SISCHED : 0) |
120		(sscreen->debug_flags & DBG(GISEL) ? AC_TM_ENABLE_GLOBAL_ISEL : 0) |
121		(sscreen->info.chip_class >= GFX9 ? AC_TM_FORCE_ENABLE_XNACK : 0) |
122		(sscreen->info.chip_class < GFX9 ? AC_TM_FORCE_DISABLE_XNACK : 0) |
123		(!sscreen->llvm_has_working_vgpr_indexing ? AC_TM_PROMOTE_ALLOCA_TO_SCRATCH : 0) |
124		(sscreen->debug_flags & DBG(CHECK_IR) ? AC_TM_CHECK_IR : 0) |
125		(create_low_opt_compiler ? AC_TM_CREATE_LOW_OPT : 0);
126
127	ac_init_llvm_once();
128	ac_init_llvm_compiler(compiler, true, sscreen->info.family, tm_options);
129	compiler->passes = ac_create_llvm_passes(compiler->tm);
130
131	if (compiler->low_opt_tm)
132		compiler->low_opt_passes = ac_create_llvm_passes(compiler->low_opt_tm);
133}
134
135static void si_destroy_compiler(struct ac_llvm_compiler *compiler)
136{
137	ac_destroy_llvm_passes(compiler->passes);
138	ac_destroy_llvm_passes(compiler->low_opt_passes);
139	ac_destroy_llvm_compiler(compiler);
140}
141
142/*
143 * pipe_context
144 */
145static void si_destroy_context(struct pipe_context *context)
146{
147	struct si_context *sctx = (struct si_context *)context;
148	int i;
149
150	/* Unreference the framebuffer normally to disable related logic
151	 * properly.
152	 */
153	struct pipe_framebuffer_state fb = {};
154	if (context->set_framebuffer_state)
155		context->set_framebuffer_state(context, &fb);
156
157	si_release_all_descriptors(sctx);
158
159	pipe_resource_reference(&sctx->esgs_ring, NULL);
160	pipe_resource_reference(&sctx->gsvs_ring, NULL);
161	pipe_resource_reference(&sctx->tess_rings, NULL);
162	pipe_resource_reference(&sctx->null_const_buf.buffer, NULL);
163	pipe_resource_reference(&sctx->sample_pos_buffer, NULL);
164	r600_resource_reference(&sctx->border_color_buffer, NULL);
165	free(sctx->border_color_table);
166	r600_resource_reference(&sctx->scratch_buffer, NULL);
167	r600_resource_reference(&sctx->compute_scratch_buffer, NULL);
168	r600_resource_reference(&sctx->wait_mem_scratch, NULL);
169
170	si_pm4_free_state(sctx, sctx->init_config, ~0);
171	if (sctx->init_config_gs_rings)
172		si_pm4_free_state(sctx, sctx->init_config_gs_rings, ~0);
173	for (i = 0; i < ARRAY_SIZE(sctx->vgt_shader_config); i++)
174		si_pm4_delete_state(sctx, vgt_shader_config, sctx->vgt_shader_config[i]);
175
176	if (sctx->fixed_func_tcs_shader.cso)
177		sctx->b.delete_tcs_state(&sctx->b, sctx->fixed_func_tcs_shader.cso);
178	if (sctx->custom_dsa_flush)
179		sctx->b.delete_depth_stencil_alpha_state(&sctx->b, sctx->custom_dsa_flush);
180	if (sctx->custom_blend_resolve)
181		sctx->b.delete_blend_state(&sctx->b, sctx->custom_blend_resolve);
182	if (sctx->custom_blend_fmask_decompress)
183		sctx->b.delete_blend_state(&sctx->b, sctx->custom_blend_fmask_decompress);
184	if (sctx->custom_blend_eliminate_fastclear)
185		sctx->b.delete_blend_state(&sctx->b, sctx->custom_blend_eliminate_fastclear);
186	if (sctx->custom_blend_dcc_decompress)
187		sctx->b.delete_blend_state(&sctx->b, sctx->custom_blend_dcc_decompress);
188	if (sctx->vs_blit_pos)
189		sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_pos);
190	if (sctx->vs_blit_pos_layered)
191		sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_pos_layered);
192	if (sctx->vs_blit_color)
193		sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_color);
194	if (sctx->vs_blit_color_layered)
195		sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_color_layered);
196	if (sctx->vs_blit_texcoord)
197		sctx->b.delete_vs_state(&sctx->b, sctx->vs_blit_texcoord);
198	if (sctx->cs_clear_buffer)
199		sctx->b.delete_compute_state(&sctx->b, sctx->cs_clear_buffer);
200	if (sctx->cs_copy_buffer)
201		sctx->b.delete_compute_state(&sctx->b, sctx->cs_copy_buffer);
202
203	if (sctx->blitter)
204		util_blitter_destroy(sctx->blitter);
205
206	/* Release DCC stats. */
207	for (int i = 0; i < ARRAY_SIZE(sctx->dcc_stats); i++) {
208		assert(!sctx->dcc_stats[i].query_active);
209
210		for (int j = 0; j < ARRAY_SIZE(sctx->dcc_stats[i].ps_stats); j++)
211			if (sctx->dcc_stats[i].ps_stats[j])
212				sctx->b.destroy_query(&sctx->b,
213							sctx->dcc_stats[i].ps_stats[j]);
214
215		si_texture_reference(&sctx->dcc_stats[i].tex, NULL);
216	}
217
218	if (sctx->query_result_shader)
219		sctx->b.delete_compute_state(&sctx->b, sctx->query_result_shader);
220
221	if (sctx->gfx_cs)
222		sctx->ws->cs_destroy(sctx->gfx_cs);
223	if (sctx->dma_cs)
224		sctx->ws->cs_destroy(sctx->dma_cs);
225	if (sctx->ctx)
226		sctx->ws->ctx_destroy(sctx->ctx);
227
228	if (sctx->b.stream_uploader)
229		u_upload_destroy(sctx->b.stream_uploader);
230	if (sctx->b.const_uploader)
231		u_upload_destroy(sctx->b.const_uploader);
232	if (sctx->cached_gtt_allocator)
233		u_upload_destroy(sctx->cached_gtt_allocator);
234
235	slab_destroy_child(&sctx->pool_transfers);
236	slab_destroy_child(&sctx->pool_transfers_unsync);
237
238	if (sctx->allocator_zeroed_memory)
239		u_suballocator_destroy(sctx->allocator_zeroed_memory);
240
241	sctx->ws->fence_reference(&sctx->last_gfx_fence, NULL);
242	sctx->ws->fence_reference(&sctx->last_sdma_fence, NULL);
243	r600_resource_reference(&sctx->eop_bug_scratch, NULL);
244
245	si_destroy_compiler(&sctx->compiler);
246
247	si_saved_cs_reference(&sctx->current_saved_cs, NULL);
248
249	_mesa_hash_table_destroy(sctx->tex_handles, NULL);
250	_mesa_hash_table_destroy(sctx->img_handles, NULL);
251
252	util_dynarray_fini(&sctx->resident_tex_handles);
253	util_dynarray_fini(&sctx->resident_img_handles);
254	util_dynarray_fini(&sctx->resident_tex_needs_color_decompress);
255	util_dynarray_fini(&sctx->resident_img_needs_color_decompress);
256	util_dynarray_fini(&sctx->resident_tex_needs_depth_decompress);
257	FREE(sctx);
258}
259
260static enum pipe_reset_status si_get_reset_status(struct pipe_context *ctx)
261{
262	struct si_context *sctx = (struct si_context *)ctx;
263
264	if (sctx->screen->info.has_gpu_reset_status_query)
265		return sctx->ws->ctx_query_reset_status(sctx->ctx);
266
267	if (sctx->screen->info.has_gpu_reset_counter_query) {
268		unsigned latest = sctx->ws->query_value(sctx->ws,
269							RADEON_GPU_RESET_COUNTER);
270
271		if (sctx->gpu_reset_counter == latest)
272			return PIPE_NO_RESET;
273
274		sctx->gpu_reset_counter = latest;
275		return PIPE_UNKNOWN_CONTEXT_RESET;
276	}
277
278	return PIPE_NO_RESET;
279}
280
281static void si_set_device_reset_callback(struct pipe_context *ctx,
282					   const struct pipe_device_reset_callback *cb)
283{
284	struct si_context *sctx = (struct si_context *)ctx;
285
286	if (cb)
287		sctx->device_reset_callback = *cb;
288	else
289		memset(&sctx->device_reset_callback, 0,
290		       sizeof(sctx->device_reset_callback));
291}
292
293bool si_check_device_reset(struct si_context *sctx)
294{
295	enum pipe_reset_status status;
296
297	if (!sctx->device_reset_callback.reset)
298		return false;
299
300	if (!sctx->b.get_device_reset_status)
301		return false;
302
303	status = sctx->b.get_device_reset_status(&sctx->b);
304	if (status == PIPE_NO_RESET)
305		return false;
306
307	sctx->device_reset_callback.reset(sctx->device_reset_callback.data, status);
308	return true;
309}
310
311/* Apitrace profiling:
312 *   1) qapitrace : Tools -> Profile: Measure CPU & GPU times
313 *   2) In the middle panel, zoom in (mouse wheel) on some bad draw call
314 *      and remember its number.
315 *   3) In Mesa, enable queries and performance counters around that draw
316 *      call and print the results.
317 *   4) glretrace --benchmark --markers ..
318 */
319static void si_emit_string_marker(struct pipe_context *ctx,
320				  const char *string, int len)
321{
322	struct si_context *sctx = (struct si_context *)ctx;
323
324	dd_parse_apitrace_marker(string, len, &sctx->apitrace_call_number);
325
326	if (sctx->log)
327		u_log_printf(sctx->log, "\nString marker: %*s\n", len, string);
328}
329
330static void si_set_debug_callback(struct pipe_context *ctx,
331				  const struct pipe_debug_callback *cb)
332{
333	struct si_context *sctx = (struct si_context *)ctx;
334	struct si_screen *screen = sctx->screen;
335
336	util_queue_finish(&screen->shader_compiler_queue);
337	util_queue_finish(&screen->shader_compiler_queue_low_priority);
338
339	if (cb)
340		sctx->debug = *cb;
341	else
342		memset(&sctx->debug, 0, sizeof(sctx->debug));
343}
344
345static void si_set_log_context(struct pipe_context *ctx,
346			       struct u_log_context *log)
347{
348	struct si_context *sctx = (struct si_context *)ctx;
349	sctx->log = log;
350
351	if (log)
352		u_log_add_auto_logger(log, si_auto_log_cs, sctx);
353}
354
355static void si_set_context_param(struct pipe_context *ctx,
356				 enum pipe_context_param param,
357				 unsigned value)
358{
359	struct radeon_winsys *ws = ((struct si_context *)ctx)->ws;
360
361	switch (param) {
362	case PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE:
363		ws->pin_threads_to_L3_cache(ws, value);
364		break;
365	default:;
366	}
367}
368
369static struct pipe_context *si_create_context(struct pipe_screen *screen,
370                                              unsigned flags)
371{
372	struct si_context *sctx = CALLOC_STRUCT(si_context);
373	struct si_screen* sscreen = (struct si_screen *)screen;
374	struct radeon_winsys *ws = sscreen->ws;
375	int shader, i;
376
377	if (!sctx)
378		return NULL;
379
380	if (flags & PIPE_CONTEXT_DEBUG)
381		sscreen->record_llvm_ir = true; /* racy but not critical */
382
383	sctx->b.screen = screen; /* this must be set first */
384	sctx->b.priv = NULL;
385	sctx->b.destroy = si_destroy_context;
386	sctx->b.emit_string_marker = si_emit_string_marker;
387	sctx->b.set_debug_callback = si_set_debug_callback;
388	sctx->b.set_log_context = si_set_log_context;
389	sctx->b.set_context_param = si_set_context_param;
390	sctx->screen = sscreen; /* Easy accessing of screen/winsys. */
391	sctx->is_debug = (flags & PIPE_CONTEXT_DEBUG) != 0;
392
393	slab_create_child(&sctx->pool_transfers, &sscreen->pool_transfers);
394	slab_create_child(&sctx->pool_transfers_unsync, &sscreen->pool_transfers);
395
396	sctx->ws = sscreen->ws;
397	sctx->family = sscreen->info.family;
398	sctx->chip_class = sscreen->info.chip_class;
399
400	if (sscreen->info.has_gpu_reset_counter_query) {
401		sctx->gpu_reset_counter =
402			sctx->ws->query_value(sctx->ws, RADEON_GPU_RESET_COUNTER);
403	}
404
405	sctx->b.get_device_reset_status = si_get_reset_status;
406	sctx->b.set_device_reset_callback = si_set_device_reset_callback;
407
408	si_init_context_texture_functions(sctx);
409	si_init_query_functions(sctx);
410
411	if (sctx->chip_class == CIK ||
412	    sctx->chip_class == VI ||
413	    sctx->chip_class == GFX9) {
414		sctx->eop_bug_scratch = r600_resource(
415			pipe_buffer_create(&sscreen->b, 0, PIPE_USAGE_DEFAULT,
416					   16 * sscreen->info.num_render_backends));
417		if (!sctx->eop_bug_scratch)
418			goto fail;
419	}
420
421	sctx->allocator_zeroed_memory =
422			u_suballocator_create(&sctx->b, sscreen->info.gart_page_size,
423					      0, PIPE_USAGE_DEFAULT,
424					      SI_RESOURCE_FLAG_SO_FILLED_SIZE, true);
425	if (!sctx->allocator_zeroed_memory)
426		goto fail;
427
428	sctx->b.stream_uploader = u_upload_create(&sctx->b, 1024 * 1024,
429						    0, PIPE_USAGE_STREAM,
430						    SI_RESOURCE_FLAG_READ_ONLY);
431	if (!sctx->b.stream_uploader)
432		goto fail;
433
434	sctx->b.const_uploader = u_upload_create(&sctx->b, 128 * 1024,
435						   0, PIPE_USAGE_DEFAULT,
436						   SI_RESOURCE_FLAG_32BIT |
437						   (sscreen->cpdma_prefetch_writes_memory ?
438							    0 : SI_RESOURCE_FLAG_READ_ONLY));
439	if (!sctx->b.const_uploader)
440		goto fail;
441
442	sctx->cached_gtt_allocator = u_upload_create(&sctx->b, 16 * 1024,
443						       0, PIPE_USAGE_STAGING, 0);
444	if (!sctx->cached_gtt_allocator)
445		goto fail;
446
447	sctx->ctx = sctx->ws->ctx_create(sctx->ws);
448	if (!sctx->ctx)
449		goto fail;
450
451	if (sscreen->info.num_sdma_rings && !(sscreen->debug_flags & DBG(NO_ASYNC_DMA))) {
452		sctx->dma_cs = sctx->ws->cs_create(sctx->ctx, RING_DMA,
453						       (void*)si_flush_dma_cs,
454						       sctx);
455	}
456
457	si_init_buffer_functions(sctx);
458	si_init_clear_functions(sctx);
459	si_init_blit_functions(sctx);
460	si_init_compute_functions(sctx);
461	si_init_compute_blit_functions(sctx);
462	si_init_debug_functions(sctx);
463	si_init_msaa_functions(sctx);
464	si_init_streamout_functions(sctx);
465
466	if (sscreen->info.has_hw_decode) {
467		sctx->b.create_video_codec = si_uvd_create_decoder;
468		sctx->b.create_video_buffer = si_video_buffer_create;
469	} else {
470		sctx->b.create_video_codec = vl_create_decoder;
471		sctx->b.create_video_buffer = vl_video_buffer_create;
472	}
473
474	sctx->gfx_cs = ws->cs_create(sctx->ctx, RING_GFX,
475				       (void*)si_flush_gfx_cs, sctx);
476
477	/* Border colors. */
478	sctx->border_color_table = malloc(SI_MAX_BORDER_COLORS *
479					  sizeof(*sctx->border_color_table));
480	if (!sctx->border_color_table)
481		goto fail;
482
483	sctx->border_color_buffer = r600_resource(
484		pipe_buffer_create(screen, 0, PIPE_USAGE_DEFAULT,
485				   SI_MAX_BORDER_COLORS *
486				   sizeof(*sctx->border_color_table)));
487	if (!sctx->border_color_buffer)
488		goto fail;
489
490	sctx->border_color_map =
491		ws->buffer_map(sctx->border_color_buffer->buf,
492			       NULL, PIPE_TRANSFER_WRITE);
493	if (!sctx->border_color_map)
494		goto fail;
495
496	si_init_all_descriptors(sctx);
497	si_init_fence_functions(sctx);
498	si_init_state_functions(sctx);
499	si_init_shader_functions(sctx);
500	si_init_viewport_functions(sctx);
501	si_init_ia_multi_vgt_param_table(sctx);
502
503	if (sctx->chip_class >= CIK)
504		cik_init_sdma_functions(sctx);
505	else
506		si_init_dma_functions(sctx);
507
508	if (sscreen->debug_flags & DBG(FORCE_DMA))
509		sctx->b.resource_copy_region = sctx->dma_copy;
510
511	sctx->blitter = util_blitter_create(&sctx->b);
512	if (sctx->blitter == NULL)
513		goto fail;
514	sctx->blitter->draw_rectangle = si_draw_rectangle;
515	sctx->blitter->skip_viewport_restore = true;
516
517	sctx->sample_mask = 0xffff;
518
519	if (sctx->chip_class >= GFX9) {
520		sctx->wait_mem_scratch = r600_resource(
521			pipe_buffer_create(screen, 0, PIPE_USAGE_DEFAULT, 4));
522		if (!sctx->wait_mem_scratch)
523			goto fail;
524
525		/* Initialize the memory. */
526		struct radeon_cmdbuf *cs = sctx->gfx_cs;
527		radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 3, 0));
528		radeon_emit(cs, S_370_DST_SEL(V_370_MEMORY_SYNC) |
529			    S_370_WR_CONFIRM(1) |
530			    S_370_ENGINE_SEL(V_370_ME));
531		radeon_emit(cs, sctx->wait_mem_scratch->gpu_address);
532		radeon_emit(cs, sctx->wait_mem_scratch->gpu_address >> 32);
533		radeon_emit(cs, sctx->wait_mem_number);
534		radeon_add_to_buffer_list(sctx, cs, sctx->wait_mem_scratch,
535					  RADEON_USAGE_WRITE, RADEON_PRIO_FENCE);
536	}
537
538	/* CIK cannot unbind a constant buffer (S_BUFFER_LOAD doesn't skip loads
539	 * if NUM_RECORDS == 0). We need to use a dummy buffer instead. */
540	if (sctx->chip_class == CIK) {
541		sctx->null_const_buf.buffer =
542			pipe_aligned_buffer_create(screen,
543						   SI_RESOURCE_FLAG_32BIT,
544						   PIPE_USAGE_DEFAULT, 16,
545						   sctx->screen->info.tcc_cache_line_size);
546		if (!sctx->null_const_buf.buffer)
547			goto fail;
548		sctx->null_const_buf.buffer_size = sctx->null_const_buf.buffer->width0;
549
550		for (shader = 0; shader < SI_NUM_SHADERS; shader++) {
551			for (i = 0; i < SI_NUM_CONST_BUFFERS; i++) {
552				sctx->b.set_constant_buffer(&sctx->b, shader, i,
553							      &sctx->null_const_buf);
554			}
555		}
556
557		si_set_rw_buffer(sctx, SI_HS_CONST_DEFAULT_TESS_LEVELS,
558				 &sctx->null_const_buf);
559		si_set_rw_buffer(sctx, SI_VS_CONST_INSTANCE_DIVISORS,
560				 &sctx->null_const_buf);
561		si_set_rw_buffer(sctx, SI_VS_CONST_CLIP_PLANES,
562				 &sctx->null_const_buf);
563		si_set_rw_buffer(sctx, SI_PS_CONST_POLY_STIPPLE,
564				 &sctx->null_const_buf);
565		si_set_rw_buffer(sctx, SI_PS_CONST_SAMPLE_POSITIONS,
566				 &sctx->null_const_buf);
567	}
568
569	uint64_t max_threads_per_block;
570	screen->get_compute_param(screen, PIPE_SHADER_IR_TGSI,
571				  PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK,
572				  &max_threads_per_block);
573
574	/* The maximum number of scratch waves. Scratch space isn't divided
575	 * evenly between CUs. The number is only a function of the number of CUs.
576	 * We can decrease the constant to decrease the scratch buffer size.
577	 *
578	 * sctx->scratch_waves must be >= the maximum posible size of
579	 * 1 threadgroup, so that the hw doesn't hang from being unable
580	 * to start any.
581	 *
582	 * The recommended value is 4 per CU at most. Higher numbers don't
583	 * bring much benefit, but they still occupy chip resources (think
584	 * async compute). I've seen ~2% performance difference between 4 and 32.
585	 */
586	sctx->scratch_waves = MAX2(32 * sscreen->info.num_good_compute_units,
587				   max_threads_per_block / 64);
588
589	si_init_compiler(sscreen, &sctx->compiler);
590
591	/* Bindless handles. */
592	sctx->tex_handles = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
593						    _mesa_key_pointer_equal);
594	sctx->img_handles = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
595						    _mesa_key_pointer_equal);
596
597	util_dynarray_init(&sctx->resident_tex_handles, NULL);
598	util_dynarray_init(&sctx->resident_img_handles, NULL);
599	util_dynarray_init(&sctx->resident_tex_needs_color_decompress, NULL);
600	util_dynarray_init(&sctx->resident_img_needs_color_decompress, NULL);
601	util_dynarray_init(&sctx->resident_tex_needs_depth_decompress, NULL);
602
603	sctx->sample_pos_buffer =
604		pipe_buffer_create(sctx->b.screen, 0, PIPE_USAGE_DEFAULT,
605				   sizeof(sctx->sample_positions));
606	pipe_buffer_write(&sctx->b, sctx->sample_pos_buffer, 0,
607			  sizeof(sctx->sample_positions), &sctx->sample_positions);
608
609	/* this must be last */
610	si_begin_new_gfx_cs(sctx);
611
612	if (sctx->chip_class == CIK) {
613		/* Clear the NULL constant buffer, because loads should return zeros. */
614		uint32_t clear_value = 0;
615		si_clear_buffer(sctx, sctx->null_const_buf.buffer, 0,
616				sctx->null_const_buf.buffer->width0,
617				&clear_value, 4, SI_COHERENCY_SHADER);
618	}
619	return &sctx->b;
620fail:
621	fprintf(stderr, "radeonsi: Failed to create a context.\n");
622	si_destroy_context(&sctx->b);
623	return NULL;
624}
625
626static struct pipe_context *si_pipe_create_context(struct pipe_screen *screen,
627						   void *priv, unsigned flags)
628{
629	struct si_screen *sscreen = (struct si_screen *)screen;
630	struct pipe_context *ctx;
631
632	if (sscreen->debug_flags & DBG(CHECK_VM))
633		flags |= PIPE_CONTEXT_DEBUG;
634
635	ctx = si_create_context(screen, flags);
636
637	if (!(flags & PIPE_CONTEXT_PREFER_THREADED))
638		return ctx;
639
640	/* Clover (compute-only) is unsupported. */
641	if (flags & PIPE_CONTEXT_COMPUTE_ONLY)
642		return ctx;
643
644	/* When shaders are logged to stderr, asynchronous compilation is
645	 * disabled too. */
646	if (sscreen->debug_flags & DBG_ALL_SHADERS)
647		return ctx;
648
649	/* Use asynchronous flushes only on amdgpu, since the radeon
650	 * implementation for fence_server_sync is incomplete. */
651	return threaded_context_create(ctx, &sscreen->pool_transfers,
652				       si_replace_buffer_storage,
653				       sscreen->info.drm_major >= 3 ? si_create_fence : NULL,
654				       &((struct si_context*)ctx)->tc);
655}
656
657/*
658 * pipe_screen
659 */
660static void si_destroy_screen(struct pipe_screen* pscreen)
661{
662	struct si_screen *sscreen = (struct si_screen *)pscreen;
663	struct si_shader_part *parts[] = {
664		sscreen->vs_prologs,
665		sscreen->tcs_epilogs,
666		sscreen->gs_prologs,
667		sscreen->ps_prologs,
668		sscreen->ps_epilogs
669	};
670	unsigned i;
671
672	if (!sscreen->ws->unref(sscreen->ws))
673		return;
674
675	util_queue_destroy(&sscreen->shader_compiler_queue);
676	util_queue_destroy(&sscreen->shader_compiler_queue_low_priority);
677
678	for (i = 0; i < ARRAY_SIZE(sscreen->compiler); i++)
679		si_destroy_compiler(&sscreen->compiler[i]);
680
681	for (i = 0; i < ARRAY_SIZE(sscreen->compiler_lowp); i++)
682		si_destroy_compiler(&sscreen->compiler_lowp[i]);
683
684	/* Free shader parts. */
685	for (i = 0; i < ARRAY_SIZE(parts); i++) {
686		while (parts[i]) {
687			struct si_shader_part *part = parts[i];
688
689			parts[i] = part->next;
690			ac_shader_binary_clean(&part->binary);
691			FREE(part);
692		}
693	}
694	mtx_destroy(&sscreen->shader_parts_mutex);
695	si_destroy_shader_cache(sscreen);
696
697	si_perfcounters_destroy(sscreen);
698	si_gpu_load_kill_thread(sscreen);
699
700	mtx_destroy(&sscreen->gpu_load_mutex);
701	mtx_destroy(&sscreen->aux_context_lock);
702	sscreen->aux_context->destroy(sscreen->aux_context);
703
704	slab_destroy_parent(&sscreen->pool_transfers);
705
706	disk_cache_destroy(sscreen->disk_shader_cache);
707	sscreen->ws->destroy(sscreen->ws);
708	FREE(sscreen);
709}
710
711static void si_init_gs_info(struct si_screen *sscreen)
712{
713	sscreen->gs_table_depth = ac_get_gs_table_depth(sscreen->info.chip_class,
714							sscreen->info.family);
715}
716
717static void si_handle_env_var_force_family(struct si_screen *sscreen)
718{
719	const char *family = debug_get_option("SI_FORCE_FAMILY", NULL);
720	unsigned i;
721
722	if (!family)
723		return;
724
725	for (i = CHIP_TAHITI; i < CHIP_LAST; i++) {
726		if (!strcmp(family, ac_get_llvm_processor_name(i))) {
727			/* Override family and chip_class. */
728			sscreen->info.family = i;
729			sscreen->info.name = "GCN-NOOP";
730
731			if (i >= CHIP_VEGA10)
732				sscreen->info.chip_class = GFX9;
733			else if (i >= CHIP_TONGA)
734				sscreen->info.chip_class = VI;
735			else if (i >= CHIP_BONAIRE)
736				sscreen->info.chip_class = CIK;
737			else
738				sscreen->info.chip_class = SI;
739
740			/* Don't submit any IBs. */
741			setenv("RADEON_NOOP", "1", 1);
742			return;
743		}
744	}
745
746	fprintf(stderr, "radeonsi: Unknown family: %s\n", family);
747	exit(1);
748}
749
750static void si_test_vmfault(struct si_screen *sscreen)
751{
752	struct pipe_context *ctx = sscreen->aux_context;
753	struct si_context *sctx = (struct si_context *)ctx;
754	struct pipe_resource *buf =
755		pipe_buffer_create_const0(&sscreen->b, 0, PIPE_USAGE_DEFAULT, 64);
756
757	if (!buf) {
758		puts("Buffer allocation failed.");
759		exit(1);
760	}
761
762	r600_resource(buf)->gpu_address = 0; /* cause a VM fault */
763
764	if (sscreen->debug_flags & DBG(TEST_VMFAULT_CP)) {
765		si_cp_dma_copy_buffer(sctx, buf, buf, 0, 4, 4, 0,
766				      SI_COHERENCY_NONE, L2_BYPASS);
767		ctx->flush(ctx, NULL, 0);
768		puts("VM fault test: CP - done.");
769	}
770	if (sscreen->debug_flags & DBG(TEST_VMFAULT_SDMA)) {
771		si_sdma_clear_buffer(sctx, buf, 0, 4, 0);
772		ctx->flush(ctx, NULL, 0);
773		puts("VM fault test: SDMA - done.");
774	}
775	if (sscreen->debug_flags & DBG(TEST_VMFAULT_SHADER)) {
776		util_test_constant_buffer(ctx, buf);
777		puts("VM fault test: Shader - done.");
778	}
779	exit(0);
780}
781
782static void si_disk_cache_create(struct si_screen *sscreen)
783{
784	/* Don't use the cache if shader dumping is enabled. */
785	if (sscreen->debug_flags & DBG_ALL_SHADERS)
786		return;
787
788	struct mesa_sha1 ctx;
789	unsigned char sha1[20];
790	char cache_id[20 * 2 + 1];
791
792	_mesa_sha1_init(&ctx);
793
794	if (!disk_cache_get_function_identifier(si_disk_cache_create, &ctx) ||
795	    !disk_cache_get_function_identifier(LLVMInitializeAMDGPUTargetInfo,
796						&ctx))
797		return;
798
799	_mesa_sha1_final(&ctx, sha1);
800	disk_cache_format_hex_id(cache_id, sha1, 20 * 2);
801
802	/* These flags affect shader compilation. */
803	#define ALL_FLAGS (DBG(FS_CORRECT_DERIVS_AFTER_KILL) |	\
804			   DBG(SI_SCHED) |			\
805			   DBG(GISEL) |				\
806			   DBG(UNSAFE_MATH) |			\
807			   DBG(NIR))
808	uint64_t shader_debug_flags = sscreen->debug_flags &
809		ALL_FLAGS;
810
811	/* Add the high bits of 32-bit addresses, which affects
812	 * how 32-bit addresses are expanded to 64 bits.
813	 */
814	STATIC_ASSERT(ALL_FLAGS <= UINT_MAX);
815	shader_debug_flags |= (uint64_t)sscreen->info.address32_hi << 32;
816
817	sscreen->disk_shader_cache =
818		disk_cache_create(sscreen->info.name,
819				  cache_id,
820				  shader_debug_flags);
821}
822
823struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws,
824					   const struct pipe_screen_config *config)
825{
826	struct si_screen *sscreen = CALLOC_STRUCT(si_screen);
827	unsigned hw_threads, num_comp_hi_threads, num_comp_lo_threads, i;
828
829	if (!sscreen) {
830		return NULL;
831	}
832
833	sscreen->ws = ws;
834	ws->query_info(ws, &sscreen->info);
835	si_handle_env_var_force_family(sscreen);
836
837	if (sscreen->info.chip_class >= GFX9) {
838		sscreen->se_tile_repeat = 32 * sscreen->info.max_se;
839	} else {
840		ac_get_raster_config(&sscreen->info,
841				     &sscreen->pa_sc_raster_config,
842				     &sscreen->pa_sc_raster_config_1,
843				     &sscreen->se_tile_repeat);
844	}
845
846	sscreen->debug_flags = debug_get_flags_option("R600_DEBUG",
847							debug_options, 0);
848
849	/* Set functions first. */
850	sscreen->b.context_create = si_pipe_create_context;
851	sscreen->b.destroy = si_destroy_screen;
852
853	si_init_screen_get_functions(sscreen);
854	si_init_screen_buffer_functions(sscreen);
855	si_init_screen_fence_functions(sscreen);
856	si_init_screen_state_functions(sscreen);
857	si_init_screen_texture_functions(sscreen);
858	si_init_screen_query_functions(sscreen);
859
860	/* Set these flags in debug_flags early, so that the shader cache takes
861	 * them into account.
862	 */
863	if (driQueryOptionb(config->options,
864			    "glsl_correct_derivatives_after_discard"))
865		sscreen->debug_flags |= DBG(FS_CORRECT_DERIVS_AFTER_KILL);
866	if (driQueryOptionb(config->options, "radeonsi_enable_sisched"))
867		sscreen->debug_flags |= DBG(SI_SCHED);
868	if (driQueryOptionb(config->options, "radeonsi_enable_nir"))
869		sscreen->debug_flags |= DBG(NIR);
870
871	if (sscreen->debug_flags & DBG(INFO))
872		ac_print_gpu_info(&sscreen->info);
873
874	slab_create_parent(&sscreen->pool_transfers,
875			   sizeof(struct si_transfer), 64);
876
877	sscreen->force_aniso = MIN2(16, debug_get_num_option("R600_TEX_ANISO", -1));
878	if (sscreen->force_aniso >= 0) {
879		printf("radeonsi: Forcing anisotropy filter to %ix\n",
880		       /* round down to a power of two */
881		       1 << util_logbase2(sscreen->force_aniso));
882	}
883
884	(void) mtx_init(&sscreen->aux_context_lock, mtx_plain);
885	(void) mtx_init(&sscreen->gpu_load_mutex, mtx_plain);
886
887	si_init_gs_info(sscreen);
888	if (!si_init_shader_cache(sscreen)) {
889		FREE(sscreen);
890		return NULL;
891	}
892
893	si_disk_cache_create(sscreen);
894
895	/* Determine the number of shader compiler threads. */
896	hw_threads = sysconf(_SC_NPROCESSORS_ONLN);
897
898	if (hw_threads >= 12) {
899		num_comp_hi_threads = hw_threads * 3 / 4;
900		num_comp_lo_threads = hw_threads / 3;
901	} else if (hw_threads >= 6) {
902		num_comp_hi_threads = hw_threads - 2;
903		num_comp_lo_threads = hw_threads / 2;
904	} else if (hw_threads >= 2) {
905		num_comp_hi_threads = hw_threads - 1;
906		num_comp_lo_threads = hw_threads / 2;
907	} else {
908		num_comp_hi_threads = 1;
909		num_comp_lo_threads = 1;
910	}
911
912	num_comp_hi_threads = MIN2(num_comp_hi_threads,
913				   ARRAY_SIZE(sscreen->compiler));
914	num_comp_lo_threads = MIN2(num_comp_lo_threads,
915				   ARRAY_SIZE(sscreen->compiler_lowp));
916
917	if (!util_queue_init(&sscreen->shader_compiler_queue, "sh",
918			     64, num_comp_hi_threads,
919			     UTIL_QUEUE_INIT_RESIZE_IF_FULL |
920			     UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY)) {
921		si_destroy_shader_cache(sscreen);
922		FREE(sscreen);
923		return NULL;
924	}
925
926	if (!util_queue_init(&sscreen->shader_compiler_queue_low_priority,
927			     "shlo",
928			     64, num_comp_lo_threads,
929			     UTIL_QUEUE_INIT_RESIZE_IF_FULL |
930			     UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY |
931			     UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY)) {
932	       si_destroy_shader_cache(sscreen);
933	       FREE(sscreen);
934	       return NULL;
935	}
936
937	if (!debug_get_bool_option("RADEON_DISABLE_PERFCOUNTERS", false))
938		si_init_perfcounters(sscreen);
939
940	/* Determine tessellation ring info. */
941	bool double_offchip_buffers = sscreen->info.chip_class >= CIK &&
942				      sscreen->info.family != CHIP_CARRIZO &&
943				      sscreen->info.family != CHIP_STONEY;
944	/* This must be one less than the maximum number due to a hw limitation.
945	 * Various hardware bugs in SI, CIK, and GFX9 need this.
946	 */
947	unsigned max_offchip_buffers_per_se;
948
949	/* Only certain chips can use the maximum value. */
950	if (sscreen->info.family == CHIP_VEGA12 ||
951	    sscreen->info.family == CHIP_VEGA20)
952		max_offchip_buffers_per_se = double_offchip_buffers ? 128 : 64;
953	else
954		max_offchip_buffers_per_se = double_offchip_buffers ? 127 : 63;
955
956	unsigned max_offchip_buffers = max_offchip_buffers_per_se *
957				       sscreen->info.max_se;
958	unsigned offchip_granularity;
959
960	/* Hawaii has a bug with offchip buffers > 256 that can be worked
961	 * around by setting 4K granularity.
962	 */
963	if (sscreen->info.family == CHIP_HAWAII) {
964		sscreen->tess_offchip_block_dw_size = 4096;
965		offchip_granularity = V_03093C_X_4K_DWORDS;
966	} else {
967		sscreen->tess_offchip_block_dw_size = 8192;
968		offchip_granularity = V_03093C_X_8K_DWORDS;
969	}
970
971	sscreen->tess_factor_ring_size = 32768 * sscreen->info.max_se;
972	assert(((sscreen->tess_factor_ring_size / 4) & C_030938_SIZE) == 0);
973	sscreen->tess_offchip_ring_size = max_offchip_buffers *
974					  sscreen->tess_offchip_block_dw_size * 4;
975
976	if (sscreen->info.chip_class >= CIK) {
977		if (sscreen->info.chip_class >= VI)
978			--max_offchip_buffers;
979		sscreen->vgt_hs_offchip_param =
980			S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers) |
981			S_03093C_OFFCHIP_GRANULARITY(offchip_granularity);
982	} else {
983		assert(offchip_granularity == V_03093C_X_8K_DWORDS);
984		sscreen->vgt_hs_offchip_param =
985			S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers);
986	}
987
988	/* The mere presense of CLEAR_STATE in the IB causes random GPU hangs
989        * on SI. Some CLEAR_STATE cause asic hang on radeon kernel, etc.
990        * SPI_VS_OUT_CONFIG. So only enable CI CLEAR_STATE on amdgpu kernel.*/
991       sscreen->has_clear_state = sscreen->info.chip_class >= CIK &&
992                                  sscreen->info.drm_major == 3;
993
994	sscreen->has_distributed_tess =
995		sscreen->info.chip_class >= VI &&
996		sscreen->info.max_se >= 2;
997
998	sscreen->has_draw_indirect_multi =
999		(sscreen->info.family >= CHIP_POLARIS10) ||
1000		(sscreen->info.chip_class == VI &&
1001		 sscreen->info.pfp_fw_version >= 121 &&
1002		 sscreen->info.me_fw_version >= 87) ||
1003		(sscreen->info.chip_class == CIK &&
1004		 sscreen->info.pfp_fw_version >= 211 &&
1005		 sscreen->info.me_fw_version >= 173) ||
1006		(sscreen->info.chip_class == SI &&
1007		 sscreen->info.pfp_fw_version >= 79 &&
1008		 sscreen->info.me_fw_version >= 142);
1009
1010	sscreen->has_out_of_order_rast = sscreen->info.chip_class >= VI &&
1011					 sscreen->info.max_se >= 2 &&
1012					 !(sscreen->debug_flags & DBG(NO_OUT_OF_ORDER));
1013	sscreen->assume_no_z_fights =
1014		driQueryOptionb(config->options, "radeonsi_assume_no_z_fights");
1015	sscreen->commutative_blend_add =
1016		driQueryOptionb(config->options, "radeonsi_commutative_blend_add");
1017	sscreen->clear_db_cache_before_clear =
1018		driQueryOptionb(config->options, "radeonsi_clear_db_cache_before_clear");
1019	sscreen->has_msaa_sample_loc_bug = (sscreen->info.family >= CHIP_POLARIS10 &&
1020					    sscreen->info.family <= CHIP_POLARIS12) ||
1021					   sscreen->info.family == CHIP_VEGA10 ||
1022					   sscreen->info.family == CHIP_RAVEN;
1023	sscreen->has_ls_vgpr_init_bug = sscreen->info.family == CHIP_VEGA10 ||
1024					sscreen->info.family == CHIP_RAVEN;
1025
1026	if (sscreen->debug_flags & DBG(DPBB)) {
1027		sscreen->dpbb_allowed = true;
1028	} else {
1029		/* Only enable primitive binning on APUs by default. */
1030		/* TODO: Investigate if binning is profitable on Vega12. */
1031		sscreen->dpbb_allowed = !(sscreen->debug_flags & DBG(NO_DPBB)) &&
1032					(sscreen->info.family == CHIP_RAVEN ||
1033					 sscreen->info.family == CHIP_RAVEN2);
1034	}
1035
1036	if (sscreen->debug_flags & DBG(DFSM)) {
1037		sscreen->dfsm_allowed = sscreen->dpbb_allowed;
1038	} else {
1039		sscreen->dfsm_allowed = sscreen->dpbb_allowed &&
1040					!(sscreen->debug_flags & DBG(NO_DFSM));
1041	}
1042
1043	/* While it would be nice not to have this flag, we are constrained
1044	 * by the reality that LLVM 5.0 doesn't have working VGPR indexing
1045	 * on GFX9.
1046	 */
1047	sscreen->llvm_has_working_vgpr_indexing = sscreen->info.chip_class <= VI;
1048
1049	/* Some chips have RB+ registers, but don't support RB+. Those must
1050	 * always disable it.
1051	 */
1052	if (sscreen->info.family == CHIP_STONEY ||
1053	    sscreen->info.chip_class >= GFX9) {
1054		sscreen->has_rbplus = true;
1055
1056		sscreen->rbplus_allowed =
1057			!(sscreen->debug_flags & DBG(NO_RB_PLUS)) &&
1058			(sscreen->info.family == CHIP_STONEY ||
1059			 sscreen->info.family == CHIP_VEGA12 ||
1060			 sscreen->info.family == CHIP_RAVEN ||
1061			 sscreen->info.family == CHIP_RAVEN2);
1062	}
1063
1064	sscreen->dcc_msaa_allowed =
1065		!(sscreen->debug_flags & DBG(NO_DCC_MSAA));
1066
1067	sscreen->cpdma_prefetch_writes_memory = sscreen->info.chip_class <= VI;
1068
1069	(void) mtx_init(&sscreen->shader_parts_mutex, mtx_plain);
1070	sscreen->use_monolithic_shaders =
1071		(sscreen->debug_flags & DBG(MONOLITHIC_SHADERS)) != 0;
1072
1073	sscreen->barrier_flags.cp_to_L2 = SI_CONTEXT_INV_SMEM_L1 |
1074					    SI_CONTEXT_INV_VMEM_L1;
1075	if (sscreen->info.chip_class <= VI) {
1076		sscreen->barrier_flags.cp_to_L2 |= SI_CONTEXT_INV_GLOBAL_L2;
1077		sscreen->barrier_flags.L2_to_cp |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
1078	}
1079
1080	if (debug_get_bool_option("RADEON_DUMP_SHADERS", false))
1081		sscreen->debug_flags |= DBG_ALL_SHADERS;
1082
1083	/* Syntax:
1084	 *     EQAA=s,z,c
1085	 * Example:
1086	 *     EQAA=8,4,2
1087
1088	 * That means 8 coverage samples, 4 Z/S samples, and 2 color samples.
1089	 * Constraints:
1090	 *     s >= z >= c (ignoring this only wastes memory)
1091	 *     s = [2..16]
1092	 *     z = [2..8]
1093	 *     c = [2..8]
1094	 *
1095	 * Only MSAA color and depth buffers are overriden.
1096	 */
1097	if (sscreen->info.has_eqaa_surface_allocator) {
1098		const char *eqaa = debug_get_option("EQAA", NULL);
1099		unsigned s,z,f;
1100
1101		if (eqaa && sscanf(eqaa, "%u,%u,%u", &s, &z, &f) == 3 && s && z && f) {
1102			sscreen->eqaa_force_coverage_samples = s;
1103			sscreen->eqaa_force_z_samples = z;
1104			sscreen->eqaa_force_color_samples = f;
1105		}
1106	}
1107
1108	for (i = 0; i < num_comp_hi_threads; i++)
1109		si_init_compiler(sscreen, &sscreen->compiler[i]);
1110	for (i = 0; i < num_comp_lo_threads; i++)
1111		si_init_compiler(sscreen, &sscreen->compiler_lowp[i]);
1112
1113	/* Create the auxiliary context. This must be done last. */
1114	sscreen->aux_context = si_create_context(&sscreen->b, 0);
1115
1116	if (sscreen->debug_flags & DBG(TEST_DMA))
1117		si_test_dma(sscreen);
1118
1119	if (sscreen->debug_flags & DBG(TEST_DMA_PERF)) {
1120		si_test_dma_perf(sscreen);
1121	}
1122
1123	if (sscreen->debug_flags & (DBG(TEST_VMFAULT_CP) |
1124				      DBG(TEST_VMFAULT_SDMA) |
1125				      DBG(TEST_VMFAULT_SHADER)))
1126		si_test_vmfault(sscreen);
1127
1128	if (sscreen->debug_flags & DBG(TEST_GDS))
1129		si_test_gds((struct si_context*)sscreen->aux_context);
1130
1131	return &sscreen->b;
1132}
1133