2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
3 * Copyright 2018 Advanced Micro Devices, Inc.
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:
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
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.
27 #include "si_public.h"
28 #include "si_shader_internal.h"
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"
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" },
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." },
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." },
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" },
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." },
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" },
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." },
109 DEBUG_NAMED_VALUE_END
/* must be last */
112 static void si_init_compiler(struct si_screen
*sscreen
,
113 struct ac_llvm_compiler
*compiler
)
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
;
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);
130 ac_init_llvm_compiler(compiler
, sscreen
->info
.family
, tm_options
);
131 compiler
->passes
= ac_create_llvm_passes(compiler
->tm
);
133 if (compiler
->low_opt_tm
)
134 compiler
->low_opt_passes
= ac_create_llvm_passes(compiler
->low_opt_tm
);
137 static void si_destroy_compiler(struct ac_llvm_compiler
*compiler
)
139 ac_destroy_llvm_passes(compiler
->passes
);
140 ac_destroy_llvm_passes(compiler
->low_opt_passes
);
141 ac_destroy_llvm_compiler(compiler
);
147 static void si_destroy_context(struct pipe_context
*context
)
149 struct si_context
*sctx
= (struct si_context
*)context
;
152 /* Unreference the framebuffer normally to disable related logic
155 struct pipe_framebuffer_state fb
= {};
156 if (context
->set_framebuffer_state
)
157 context
->set_framebuffer_state(context
, &fb
);
159 si_release_all_descriptors(sctx
);
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 si_resource_reference(&sctx
->border_color_buffer
, NULL
);
167 free(sctx
->border_color_table
);
168 si_resource_reference(&sctx
->scratch_buffer
, NULL
);
169 si_resource_reference(&sctx
->compute_scratch_buffer
, NULL
);
170 si_resource_reference(&sctx
->wait_mem_scratch
, NULL
);
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
]);
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 if (sctx
->cs_copy_image
)
205 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_copy_image
);
206 if (sctx
->cs_copy_image_1d_array
)
207 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_copy_image_1d_array
);
208 if (sctx
->cs_clear_render_target
)
209 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_clear_render_target
);
210 if (sctx
->cs_clear_render_target_1d_array
)
211 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_clear_render_target_1d_array
);
214 util_blitter_destroy(sctx
->blitter
);
216 /* Release DCC stats. */
217 for (int i
= 0; i
< ARRAY_SIZE(sctx
->dcc_stats
); i
++) {
218 assert(!sctx
->dcc_stats
[i
].query_active
);
220 for (int j
= 0; j
< ARRAY_SIZE(sctx
->dcc_stats
[i
].ps_stats
); j
++)
221 if (sctx
->dcc_stats
[i
].ps_stats
[j
])
222 sctx
->b
.destroy_query(&sctx
->b
,
223 sctx
->dcc_stats
[i
].ps_stats
[j
]);
225 si_texture_reference(&sctx
->dcc_stats
[i
].tex
, NULL
);
228 if (sctx
->query_result_shader
)
229 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->query_result_shader
);
232 sctx
->ws
->cs_destroy(sctx
->gfx_cs
);
234 sctx
->ws
->cs_destroy(sctx
->dma_cs
);
236 sctx
->ws
->ctx_destroy(sctx
->ctx
);
238 if (sctx
->b
.stream_uploader
)
239 u_upload_destroy(sctx
->b
.stream_uploader
);
240 if (sctx
->b
.const_uploader
)
241 u_upload_destroy(sctx
->b
.const_uploader
);
242 if (sctx
->cached_gtt_allocator
)
243 u_upload_destroy(sctx
->cached_gtt_allocator
);
245 slab_destroy_child(&sctx
->pool_transfers
);
246 slab_destroy_child(&sctx
->pool_transfers_unsync
);
248 if (sctx
->allocator_zeroed_memory
)
249 u_suballocator_destroy(sctx
->allocator_zeroed_memory
);
251 sctx
->ws
->fence_reference(&sctx
->last_gfx_fence
, NULL
);
252 sctx
->ws
->fence_reference(&sctx
->last_sdma_fence
, NULL
);
253 si_resource_reference(&sctx
->eop_bug_scratch
, NULL
);
255 si_destroy_compiler(&sctx
->compiler
);
257 si_saved_cs_reference(&sctx
->current_saved_cs
, NULL
);
259 _mesa_hash_table_destroy(sctx
->tex_handles
, NULL
);
260 _mesa_hash_table_destroy(sctx
->img_handles
, NULL
);
262 util_dynarray_fini(&sctx
->resident_tex_handles
);
263 util_dynarray_fini(&sctx
->resident_img_handles
);
264 util_dynarray_fini(&sctx
->resident_tex_needs_color_decompress
);
265 util_dynarray_fini(&sctx
->resident_img_needs_color_decompress
);
266 util_dynarray_fini(&sctx
->resident_tex_needs_depth_decompress
);
270 static enum pipe_reset_status
si_get_reset_status(struct pipe_context
*ctx
)
272 struct si_context
*sctx
= (struct si_context
*)ctx
;
274 if (sctx
->screen
->info
.has_gpu_reset_status_query
)
275 return sctx
->ws
->ctx_query_reset_status(sctx
->ctx
);
277 if (sctx
->screen
->info
.has_gpu_reset_counter_query
) {
278 unsigned latest
= sctx
->ws
->query_value(sctx
->ws
,
279 RADEON_GPU_RESET_COUNTER
);
281 if (sctx
->gpu_reset_counter
== latest
)
282 return PIPE_NO_RESET
;
284 sctx
->gpu_reset_counter
= latest
;
285 return PIPE_UNKNOWN_CONTEXT_RESET
;
288 return PIPE_NO_RESET
;
291 static void si_set_device_reset_callback(struct pipe_context
*ctx
,
292 const struct pipe_device_reset_callback
*cb
)
294 struct si_context
*sctx
= (struct si_context
*)ctx
;
297 sctx
->device_reset_callback
= *cb
;
299 memset(&sctx
->device_reset_callback
, 0,
300 sizeof(sctx
->device_reset_callback
));
303 bool si_check_device_reset(struct si_context
*sctx
)
305 enum pipe_reset_status status
;
307 if (!sctx
->device_reset_callback
.reset
)
310 if (!sctx
->b
.get_device_reset_status
)
313 status
= sctx
->b
.get_device_reset_status(&sctx
->b
);
314 if (status
== PIPE_NO_RESET
)
317 sctx
->device_reset_callback
.reset(sctx
->device_reset_callback
.data
, status
);
321 /* Apitrace profiling:
322 * 1) qapitrace : Tools -> Profile: Measure CPU & GPU times
323 * 2) In the middle panel, zoom in (mouse wheel) on some bad draw call
324 * and remember its number.
325 * 3) In Mesa, enable queries and performance counters around that draw
326 * call and print the results.
327 * 4) glretrace --benchmark --markers ..
329 static void si_emit_string_marker(struct pipe_context
*ctx
,
330 const char *string
, int len
)
332 struct si_context
*sctx
= (struct si_context
*)ctx
;
334 dd_parse_apitrace_marker(string
, len
, &sctx
->apitrace_call_number
);
337 u_log_printf(sctx
->log
, "\nString marker: %*s\n", len
, string
);
340 static void si_set_debug_callback(struct pipe_context
*ctx
,
341 const struct pipe_debug_callback
*cb
)
343 struct si_context
*sctx
= (struct si_context
*)ctx
;
344 struct si_screen
*screen
= sctx
->screen
;
346 util_queue_finish(&screen
->shader_compiler_queue
);
347 util_queue_finish(&screen
->shader_compiler_queue_low_priority
);
352 memset(&sctx
->debug
, 0, sizeof(sctx
->debug
));
355 static void si_set_log_context(struct pipe_context
*ctx
,
356 struct u_log_context
*log
)
358 struct si_context
*sctx
= (struct si_context
*)ctx
;
362 u_log_add_auto_logger(log
, si_auto_log_cs
, sctx
);
365 static void si_set_context_param(struct pipe_context
*ctx
,
366 enum pipe_context_param param
,
369 struct radeon_winsys
*ws
= ((struct si_context
*)ctx
)->ws
;
372 case PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE
:
373 ws
->pin_threads_to_L3_cache(ws
, value
);
379 static struct pipe_context
*si_create_context(struct pipe_screen
*screen
,
382 struct si_context
*sctx
= CALLOC_STRUCT(si_context
);
383 struct si_screen
* sscreen
= (struct si_screen
*)screen
;
384 struct radeon_winsys
*ws
= sscreen
->ws
;
386 bool stop_exec_on_failure
= (flags
& PIPE_CONTEXT_LOSE_CONTEXT_ON_RESET
) != 0;
391 if (flags
& PIPE_CONTEXT_DEBUG
)
392 sscreen
->record_llvm_ir
= true; /* racy but not critical */
394 sctx
->b
.screen
= screen
; /* this must be set first */
396 sctx
->b
.destroy
= si_destroy_context
;
397 sctx
->b
.emit_string_marker
= si_emit_string_marker
;
398 sctx
->b
.set_debug_callback
= si_set_debug_callback
;
399 sctx
->b
.set_log_context
= si_set_log_context
;
400 sctx
->b
.set_context_param
= si_set_context_param
;
401 sctx
->screen
= sscreen
; /* Easy accessing of screen/winsys. */
402 sctx
->is_debug
= (flags
& PIPE_CONTEXT_DEBUG
) != 0;
404 slab_create_child(&sctx
->pool_transfers
, &sscreen
->pool_transfers
);
405 slab_create_child(&sctx
->pool_transfers_unsync
, &sscreen
->pool_transfers
);
407 sctx
->ws
= sscreen
->ws
;
408 sctx
->family
= sscreen
->info
.family
;
409 sctx
->chip_class
= sscreen
->info
.chip_class
;
411 if (sscreen
->info
.has_gpu_reset_counter_query
) {
412 sctx
->gpu_reset_counter
=
413 sctx
->ws
->query_value(sctx
->ws
, RADEON_GPU_RESET_COUNTER
);
416 sctx
->b
.get_device_reset_status
= si_get_reset_status
;
417 sctx
->b
.set_device_reset_callback
= si_set_device_reset_callback
;
419 si_init_context_texture_functions(sctx
);
420 si_init_query_functions(sctx
);
422 if (sctx
->chip_class
== CIK
||
423 sctx
->chip_class
== VI
||
424 sctx
->chip_class
== GFX9
) {
425 sctx
->eop_bug_scratch
= si_resource(
426 pipe_buffer_create(&sscreen
->b
, 0, PIPE_USAGE_DEFAULT
,
427 16 * sscreen
->info
.num_render_backends
));
428 if (!sctx
->eop_bug_scratch
)
432 sctx
->allocator_zeroed_memory
=
433 u_suballocator_create(&sctx
->b
, 128 * 1024,
434 0, PIPE_USAGE_DEFAULT
,
435 SI_RESOURCE_FLAG_UNMAPPABLE
|
436 SI_RESOURCE_FLAG_CLEAR
, false);
437 if (!sctx
->allocator_zeroed_memory
)
440 sctx
->b
.stream_uploader
= u_upload_create(&sctx
->b
, 1024 * 1024,
441 0, PIPE_USAGE_STREAM
,
442 SI_RESOURCE_FLAG_READ_ONLY
);
443 if (!sctx
->b
.stream_uploader
)
446 sctx
->b
.const_uploader
= u_upload_create(&sctx
->b
, 128 * 1024,
447 0, PIPE_USAGE_DEFAULT
,
448 SI_RESOURCE_FLAG_32BIT
|
449 (sscreen
->cpdma_prefetch_writes_memory
?
450 0 : SI_RESOURCE_FLAG_READ_ONLY
));
451 if (!sctx
->b
.const_uploader
)
454 sctx
->cached_gtt_allocator
= u_upload_create(&sctx
->b
, 16 * 1024,
455 0, PIPE_USAGE_STAGING
, 0);
456 if (!sctx
->cached_gtt_allocator
)
459 sctx
->ctx
= sctx
->ws
->ctx_create(sctx
->ws
);
463 if (sscreen
->info
.num_sdma_rings
&& !(sscreen
->debug_flags
& DBG(NO_ASYNC_DMA
))) {
464 sctx
->dma_cs
= sctx
->ws
->cs_create(sctx
->ctx
, RING_DMA
,
465 (void*)si_flush_dma_cs
,
466 sctx
, stop_exec_on_failure
);
469 si_init_buffer_functions(sctx
);
470 si_init_clear_functions(sctx
);
471 si_init_blit_functions(sctx
);
472 si_init_compute_functions(sctx
);
473 si_init_compute_blit_functions(sctx
);
474 si_init_debug_functions(sctx
);
475 si_init_msaa_functions(sctx
);
476 si_init_streamout_functions(sctx
);
478 if (sscreen
->info
.has_hw_decode
) {
479 sctx
->b
.create_video_codec
= si_uvd_create_decoder
;
480 sctx
->b
.create_video_buffer
= si_video_buffer_create
;
482 sctx
->b
.create_video_codec
= vl_create_decoder
;
483 sctx
->b
.create_video_buffer
= vl_video_buffer_create
;
486 sctx
->gfx_cs
= ws
->cs_create(sctx
->ctx
, RING_GFX
,
487 (void*)si_flush_gfx_cs
, sctx
, stop_exec_on_failure
);
490 sctx
->border_color_table
= malloc(SI_MAX_BORDER_COLORS
*
491 sizeof(*sctx
->border_color_table
));
492 if (!sctx
->border_color_table
)
495 sctx
->border_color_buffer
= si_resource(
496 pipe_buffer_create(screen
, 0, PIPE_USAGE_DEFAULT
,
497 SI_MAX_BORDER_COLORS
*
498 sizeof(*sctx
->border_color_table
)));
499 if (!sctx
->border_color_buffer
)
502 sctx
->border_color_map
=
503 ws
->buffer_map(sctx
->border_color_buffer
->buf
,
504 NULL
, PIPE_TRANSFER_WRITE
);
505 if (!sctx
->border_color_map
)
508 si_init_all_descriptors(sctx
);
509 si_init_fence_functions(sctx
);
510 si_init_state_functions(sctx
);
511 si_init_shader_functions(sctx
);
512 si_init_viewport_functions(sctx
);
514 if (sctx
->chip_class
>= CIK
)
515 cik_init_sdma_functions(sctx
);
517 si_init_dma_functions(sctx
);
519 if (sscreen
->debug_flags
& DBG(FORCE_DMA
))
520 sctx
->b
.resource_copy_region
= sctx
->dma_copy
;
522 sctx
->blitter
= util_blitter_create(&sctx
->b
);
523 if (sctx
->blitter
== NULL
)
525 sctx
->blitter
->skip_viewport_restore
= true;
527 si_init_draw_functions(sctx
);
529 sctx
->sample_mask
= 0xffff;
531 if (sctx
->chip_class
>= GFX9
) {
532 sctx
->wait_mem_scratch
= si_resource(
533 pipe_buffer_create(screen
, 0, PIPE_USAGE_DEFAULT
, 4));
534 if (!sctx
->wait_mem_scratch
)
537 /* Initialize the memory. */
538 si_cp_write_data(sctx
, sctx
->wait_mem_scratch
, 0, 4,
539 V_370_MEM
, V_370_ME
, &sctx
->wait_mem_number
);
542 /* CIK cannot unbind a constant buffer (S_BUFFER_LOAD doesn't skip loads
543 * if NUM_RECORDS == 0). We need to use a dummy buffer instead. */
544 if (sctx
->chip_class
== CIK
) {
545 sctx
->null_const_buf
.buffer
=
546 pipe_aligned_buffer_create(screen
,
547 SI_RESOURCE_FLAG_32BIT
,
548 PIPE_USAGE_DEFAULT
, 16,
549 sctx
->screen
->info
.tcc_cache_line_size
);
550 if (!sctx
->null_const_buf
.buffer
)
552 sctx
->null_const_buf
.buffer_size
= sctx
->null_const_buf
.buffer
->width0
;
554 for (shader
= 0; shader
< SI_NUM_SHADERS
; shader
++) {
555 for (i
= 0; i
< SI_NUM_CONST_BUFFERS
; i
++) {
556 sctx
->b
.set_constant_buffer(&sctx
->b
, shader
, i
,
557 &sctx
->null_const_buf
);
561 si_set_rw_buffer(sctx
, SI_HS_CONST_DEFAULT_TESS_LEVELS
,
562 &sctx
->null_const_buf
);
563 si_set_rw_buffer(sctx
, SI_VS_CONST_INSTANCE_DIVISORS
,
564 &sctx
->null_const_buf
);
565 si_set_rw_buffer(sctx
, SI_VS_CONST_CLIP_PLANES
,
566 &sctx
->null_const_buf
);
567 si_set_rw_buffer(sctx
, SI_PS_CONST_POLY_STIPPLE
,
568 &sctx
->null_const_buf
);
569 si_set_rw_buffer(sctx
, SI_PS_CONST_SAMPLE_POSITIONS
,
570 &sctx
->null_const_buf
);
573 uint64_t max_threads_per_block
;
574 screen
->get_compute_param(screen
, PIPE_SHADER_IR_TGSI
,
575 PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK
,
576 &max_threads_per_block
);
578 /* The maximum number of scratch waves. Scratch space isn't divided
579 * evenly between CUs. The number is only a function of the number of CUs.
580 * We can decrease the constant to decrease the scratch buffer size.
582 * sctx->scratch_waves must be >= the maximum posible size of
583 * 1 threadgroup, so that the hw doesn't hang from being unable
586 * The recommended value is 4 per CU at most. Higher numbers don't
587 * bring much benefit, but they still occupy chip resources (think
588 * async compute). I've seen ~2% performance difference between 4 and 32.
590 sctx
->scratch_waves
= MAX2(32 * sscreen
->info
.num_good_compute_units
,
591 max_threads_per_block
/ 64);
593 si_init_compiler(sscreen
, &sctx
->compiler
);
595 /* Bindless handles. */
596 sctx
->tex_handles
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
597 _mesa_key_pointer_equal
);
598 sctx
->img_handles
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
599 _mesa_key_pointer_equal
);
601 util_dynarray_init(&sctx
->resident_tex_handles
, NULL
);
602 util_dynarray_init(&sctx
->resident_img_handles
, NULL
);
603 util_dynarray_init(&sctx
->resident_tex_needs_color_decompress
, NULL
);
604 util_dynarray_init(&sctx
->resident_img_needs_color_decompress
, NULL
);
605 util_dynarray_init(&sctx
->resident_tex_needs_depth_decompress
, NULL
);
607 sctx
->sample_pos_buffer
=
608 pipe_buffer_create(sctx
->b
.screen
, 0, PIPE_USAGE_DEFAULT
,
609 sizeof(sctx
->sample_positions
));
610 pipe_buffer_write(&sctx
->b
, sctx
->sample_pos_buffer
, 0,
611 sizeof(sctx
->sample_positions
), &sctx
->sample_positions
);
613 /* this must be last */
614 si_begin_new_gfx_cs(sctx
);
616 if (sctx
->chip_class
== CIK
) {
617 /* Clear the NULL constant buffer, because loads should return zeros. */
618 uint32_t clear_value
= 0;
619 si_clear_buffer(sctx
, sctx
->null_const_buf
.buffer
, 0,
620 sctx
->null_const_buf
.buffer
->width0
,
621 &clear_value
, 4, SI_COHERENCY_SHADER
);
625 fprintf(stderr
, "radeonsi: Failed to create a context.\n");
626 si_destroy_context(&sctx
->b
);
630 static struct pipe_context
*si_pipe_create_context(struct pipe_screen
*screen
,
631 void *priv
, unsigned flags
)
633 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
634 struct pipe_context
*ctx
;
636 if (sscreen
->debug_flags
& DBG(CHECK_VM
))
637 flags
|= PIPE_CONTEXT_DEBUG
;
639 ctx
= si_create_context(screen
, flags
);
641 if (!(flags
& PIPE_CONTEXT_PREFER_THREADED
))
644 /* Clover (compute-only) is unsupported. */
645 if (flags
& PIPE_CONTEXT_COMPUTE_ONLY
)
648 /* When shaders are logged to stderr, asynchronous compilation is
650 if (sscreen
->debug_flags
& DBG_ALL_SHADERS
)
653 /* Use asynchronous flushes only on amdgpu, since the radeon
654 * implementation for fence_server_sync is incomplete. */
655 return threaded_context_create(ctx
, &sscreen
->pool_transfers
,
656 si_replace_buffer_storage
,
657 sscreen
->info
.drm_major
>= 3 ? si_create_fence
: NULL
,
658 &((struct si_context
*)ctx
)->tc
);
664 static void si_destroy_screen(struct pipe_screen
* pscreen
)
666 struct si_screen
*sscreen
= (struct si_screen
*)pscreen
;
667 struct si_shader_part
*parts
[] = {
669 sscreen
->tcs_epilogs
,
676 if (!sscreen
->ws
->unref(sscreen
->ws
))
679 util_queue_destroy(&sscreen
->shader_compiler_queue
);
680 util_queue_destroy(&sscreen
->shader_compiler_queue_low_priority
);
682 for (i
= 0; i
< ARRAY_SIZE(sscreen
->compiler
); i
++)
683 si_destroy_compiler(&sscreen
->compiler
[i
]);
685 for (i
= 0; i
< ARRAY_SIZE(sscreen
->compiler_lowp
); i
++)
686 si_destroy_compiler(&sscreen
->compiler_lowp
[i
]);
688 /* Free shader parts. */
689 for (i
= 0; i
< ARRAY_SIZE(parts
); i
++) {
691 struct si_shader_part
*part
= parts
[i
];
693 parts
[i
] = part
->next
;
694 ac_shader_binary_clean(&part
->binary
);
698 mtx_destroy(&sscreen
->shader_parts_mutex
);
699 si_destroy_shader_cache(sscreen
);
701 si_destroy_perfcounters(sscreen
);
702 si_gpu_load_kill_thread(sscreen
);
704 mtx_destroy(&sscreen
->gpu_load_mutex
);
705 mtx_destroy(&sscreen
->aux_context_lock
);
706 sscreen
->aux_context
->destroy(sscreen
->aux_context
);
708 slab_destroy_parent(&sscreen
->pool_transfers
);
710 disk_cache_destroy(sscreen
->disk_shader_cache
);
711 sscreen
->ws
->destroy(sscreen
->ws
);
715 static void si_init_gs_info(struct si_screen
*sscreen
)
717 sscreen
->gs_table_depth
= ac_get_gs_table_depth(sscreen
->info
.chip_class
,
718 sscreen
->info
.family
);
721 static void si_test_vmfault(struct si_screen
*sscreen
)
723 struct pipe_context
*ctx
= sscreen
->aux_context
;
724 struct si_context
*sctx
= (struct si_context
*)ctx
;
725 struct pipe_resource
*buf
=
726 pipe_buffer_create_const0(&sscreen
->b
, 0, PIPE_USAGE_DEFAULT
, 64);
729 puts("Buffer allocation failed.");
733 si_resource(buf
)->gpu_address
= 0; /* cause a VM fault */
735 if (sscreen
->debug_flags
& DBG(TEST_VMFAULT_CP
)) {
736 si_cp_dma_copy_buffer(sctx
, buf
, buf
, 0, 4, 4, 0,
737 SI_COHERENCY_NONE
, L2_BYPASS
);
738 ctx
->flush(ctx
, NULL
, 0);
739 puts("VM fault test: CP - done.");
741 if (sscreen
->debug_flags
& DBG(TEST_VMFAULT_SDMA
)) {
742 si_sdma_clear_buffer(sctx
, buf
, 0, 4, 0);
743 ctx
->flush(ctx
, NULL
, 0);
744 puts("VM fault test: SDMA - done.");
746 if (sscreen
->debug_flags
& DBG(TEST_VMFAULT_SHADER
)) {
747 util_test_constant_buffer(ctx
, buf
);
748 puts("VM fault test: Shader - done.");
753 static void si_test_gds_memory_management(struct si_context
*sctx
,
754 unsigned alloc_size
, unsigned alignment
,
755 enum radeon_bo_domain domain
)
757 struct radeon_winsys
*ws
= sctx
->ws
;
758 struct radeon_cmdbuf
*cs
[8];
759 struct pb_buffer
*gds_bo
[ARRAY_SIZE(cs
)];
761 for (unsigned i
= 0; i
< ARRAY_SIZE(cs
); i
++) {
762 cs
[i
] = ws
->cs_create(sctx
->ctx
, RING_COMPUTE
,
764 gds_bo
[i
] = ws
->buffer_create(ws
, alloc_size
, alignment
, domain
, 0);
768 for (unsigned iterations
= 0; iterations
< 20000; iterations
++) {
769 for (unsigned i
= 0; i
< ARRAY_SIZE(cs
); i
++) {
770 /* This clears GDS with CP DMA.
772 * We don't care if GDS is present. Just add some packet
773 * to make the GPU busy for a moment.
775 si_cp_dma_clear_buffer(sctx
, cs
[i
], NULL
, 0, alloc_size
, 0,
776 SI_CPDMA_SKIP_BO_LIST_UPDATE
|
777 SI_CPDMA_SKIP_CHECK_CS_SPACE
|
778 SI_CPDMA_SKIP_GFX_SYNC
, 0, 0);
780 ws
->cs_add_buffer(cs
[i
], gds_bo
[i
], domain
,
781 RADEON_USAGE_READWRITE
, 0);
782 ws
->cs_flush(cs
[i
], PIPE_FLUSH_ASYNC
, NULL
);
788 static void si_disk_cache_create(struct si_screen
*sscreen
)
790 /* Don't use the cache if shader dumping is enabled. */
791 if (sscreen
->debug_flags
& DBG_ALL_SHADERS
)
794 struct mesa_sha1 ctx
;
795 unsigned char sha1
[20];
796 char cache_id
[20 * 2 + 1];
798 _mesa_sha1_init(&ctx
);
800 if (!disk_cache_get_function_identifier(si_disk_cache_create
, &ctx
) ||
801 !disk_cache_get_function_identifier(LLVMInitializeAMDGPUTargetInfo
,
805 _mesa_sha1_final(&ctx
, sha1
);
806 disk_cache_format_hex_id(cache_id
, sha1
, 20 * 2);
808 /* These flags affect shader compilation. */
809 #define ALL_FLAGS (DBG(FS_CORRECT_DERIVS_AFTER_KILL) | \
814 uint64_t shader_debug_flags
= sscreen
->debug_flags
&
817 /* Add the high bits of 32-bit addresses, which affects
818 * how 32-bit addresses are expanded to 64 bits.
820 STATIC_ASSERT(ALL_FLAGS
<= UINT_MAX
);
821 shader_debug_flags
|= (uint64_t)sscreen
->info
.address32_hi
<< 32;
823 sscreen
->disk_shader_cache
=
824 disk_cache_create(sscreen
->info
.name
,
829 struct pipe_screen
*radeonsi_screen_create(struct radeon_winsys
*ws
,
830 const struct pipe_screen_config
*config
)
832 struct si_screen
*sscreen
= CALLOC_STRUCT(si_screen
);
833 unsigned hw_threads
, num_comp_hi_threads
, num_comp_lo_threads
, i
;
840 ws
->query_info(ws
, &sscreen
->info
);
842 if (sscreen
->info
.chip_class
>= GFX9
) {
843 sscreen
->se_tile_repeat
= 32 * sscreen
->info
.max_se
;
845 ac_get_raster_config(&sscreen
->info
,
846 &sscreen
->pa_sc_raster_config
,
847 &sscreen
->pa_sc_raster_config_1
,
848 &sscreen
->se_tile_repeat
);
851 sscreen
->debug_flags
= debug_get_flags_option("R600_DEBUG",
853 sscreen
->debug_flags
|= debug_get_flags_option("AMD_DEBUG",
856 /* Set functions first. */
857 sscreen
->b
.context_create
= si_pipe_create_context
;
858 sscreen
->b
.destroy
= si_destroy_screen
;
860 si_init_screen_get_functions(sscreen
);
861 si_init_screen_buffer_functions(sscreen
);
862 si_init_screen_fence_functions(sscreen
);
863 si_init_screen_state_functions(sscreen
);
864 si_init_screen_texture_functions(sscreen
);
865 si_init_screen_query_functions(sscreen
);
867 /* Set these flags in debug_flags early, so that the shader cache takes
870 if (driQueryOptionb(config
->options
,
871 "glsl_correct_derivatives_after_discard"))
872 sscreen
->debug_flags
|= DBG(FS_CORRECT_DERIVS_AFTER_KILL
);
873 if (driQueryOptionb(config
->options
, "radeonsi_enable_sisched"))
874 sscreen
->debug_flags
|= DBG(SI_SCHED
);
875 if (driQueryOptionb(config
->options
, "radeonsi_enable_nir"))
876 sscreen
->debug_flags
|= DBG(NIR
);
878 if (sscreen
->debug_flags
& DBG(INFO
))
879 ac_print_gpu_info(&sscreen
->info
);
881 slab_create_parent(&sscreen
->pool_transfers
,
882 sizeof(struct si_transfer
), 64);
884 sscreen
->force_aniso
= MIN2(16, debug_get_num_option("R600_TEX_ANISO", -1));
885 if (sscreen
->force_aniso
>= 0) {
886 printf("radeonsi: Forcing anisotropy filter to %ix\n",
887 /* round down to a power of two */
888 1 << util_logbase2(sscreen
->force_aniso
));
891 (void) mtx_init(&sscreen
->aux_context_lock
, mtx_plain
);
892 (void) mtx_init(&sscreen
->gpu_load_mutex
, mtx_plain
);
894 si_init_gs_info(sscreen
);
895 if (!si_init_shader_cache(sscreen
)) {
900 si_disk_cache_create(sscreen
);
902 /* Determine the number of shader compiler threads. */
903 hw_threads
= sysconf(_SC_NPROCESSORS_ONLN
);
905 if (hw_threads
>= 12) {
906 num_comp_hi_threads
= hw_threads
* 3 / 4;
907 num_comp_lo_threads
= hw_threads
/ 3;
908 } else if (hw_threads
>= 6) {
909 num_comp_hi_threads
= hw_threads
- 2;
910 num_comp_lo_threads
= hw_threads
/ 2;
911 } else if (hw_threads
>= 2) {
912 num_comp_hi_threads
= hw_threads
- 1;
913 num_comp_lo_threads
= hw_threads
/ 2;
915 num_comp_hi_threads
= 1;
916 num_comp_lo_threads
= 1;
919 num_comp_hi_threads
= MIN2(num_comp_hi_threads
,
920 ARRAY_SIZE(sscreen
->compiler
));
921 num_comp_lo_threads
= MIN2(num_comp_lo_threads
,
922 ARRAY_SIZE(sscreen
->compiler_lowp
));
924 if (!util_queue_init(&sscreen
->shader_compiler_queue
, "sh",
925 64, num_comp_hi_threads
,
926 UTIL_QUEUE_INIT_RESIZE_IF_FULL
|
927 UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY
)) {
928 si_destroy_shader_cache(sscreen
);
933 if (!util_queue_init(&sscreen
->shader_compiler_queue_low_priority
,
935 64, num_comp_lo_threads
,
936 UTIL_QUEUE_INIT_RESIZE_IF_FULL
|
937 UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY
|
938 UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY
)) {
939 si_destroy_shader_cache(sscreen
);
944 if (!debug_get_bool_option("RADEON_DISABLE_PERFCOUNTERS", false))
945 si_init_perfcounters(sscreen
);
947 /* Determine tessellation ring info. */
948 bool double_offchip_buffers
= sscreen
->info
.chip_class
>= CIK
&&
949 sscreen
->info
.family
!= CHIP_CARRIZO
&&
950 sscreen
->info
.family
!= CHIP_STONEY
;
951 /* This must be one less than the maximum number due to a hw limitation.
952 * Various hardware bugs in SI, CIK, and GFX9 need this.
954 unsigned max_offchip_buffers_per_se
;
956 /* Only certain chips can use the maximum value. */
957 if (sscreen
->info
.family
== CHIP_VEGA12
||
958 sscreen
->info
.family
== CHIP_VEGA20
)
959 max_offchip_buffers_per_se
= double_offchip_buffers
? 128 : 64;
961 max_offchip_buffers_per_se
= double_offchip_buffers
? 127 : 63;
963 unsigned max_offchip_buffers
= max_offchip_buffers_per_se
*
964 sscreen
->info
.max_se
;
965 unsigned offchip_granularity
;
967 /* Hawaii has a bug with offchip buffers > 256 that can be worked
968 * around by setting 4K granularity.
970 if (sscreen
->info
.family
== CHIP_HAWAII
) {
971 sscreen
->tess_offchip_block_dw_size
= 4096;
972 offchip_granularity
= V_03093C_X_4K_DWORDS
;
974 sscreen
->tess_offchip_block_dw_size
= 8192;
975 offchip_granularity
= V_03093C_X_8K_DWORDS
;
978 sscreen
->tess_factor_ring_size
= 32768 * sscreen
->info
.max_se
;
979 assert(((sscreen
->tess_factor_ring_size
/ 4) & C_030938_SIZE
) == 0);
980 sscreen
->tess_offchip_ring_size
= max_offchip_buffers
*
981 sscreen
->tess_offchip_block_dw_size
* 4;
983 if (sscreen
->info
.chip_class
>= CIK
) {
984 if (sscreen
->info
.chip_class
>= VI
)
985 --max_offchip_buffers
;
986 sscreen
->vgt_hs_offchip_param
=
987 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers
) |
988 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity
);
990 assert(offchip_granularity
== V_03093C_X_8K_DWORDS
);
991 sscreen
->vgt_hs_offchip_param
=
992 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers
);
995 /* The mere presense of CLEAR_STATE in the IB causes random GPU hangs
996 * on SI. Some CLEAR_STATE cause asic hang on radeon kernel, etc.
997 * SPI_VS_OUT_CONFIG. So only enable CI CLEAR_STATE on amdgpu kernel.*/
998 sscreen
->has_clear_state
= sscreen
->info
.chip_class
>= CIK
&&
999 sscreen
->info
.drm_major
== 3;
1001 sscreen
->has_distributed_tess
=
1002 sscreen
->info
.chip_class
>= VI
&&
1003 sscreen
->info
.max_se
>= 2;
1005 sscreen
->has_draw_indirect_multi
=
1006 (sscreen
->info
.family
>= CHIP_POLARIS10
) ||
1007 (sscreen
->info
.chip_class
== VI
&&
1008 sscreen
->info
.pfp_fw_version
>= 121 &&
1009 sscreen
->info
.me_fw_version
>= 87) ||
1010 (sscreen
->info
.chip_class
== CIK
&&
1011 sscreen
->info
.pfp_fw_version
>= 211 &&
1012 sscreen
->info
.me_fw_version
>= 173) ||
1013 (sscreen
->info
.chip_class
== SI
&&
1014 sscreen
->info
.pfp_fw_version
>= 79 &&
1015 sscreen
->info
.me_fw_version
>= 142);
1017 sscreen
->has_out_of_order_rast
= sscreen
->info
.chip_class
>= VI
&&
1018 sscreen
->info
.max_se
>= 2 &&
1019 !(sscreen
->debug_flags
& DBG(NO_OUT_OF_ORDER
));
1020 sscreen
->assume_no_z_fights
=
1021 driQueryOptionb(config
->options
, "radeonsi_assume_no_z_fights");
1022 sscreen
->commutative_blend_add
=
1023 driQueryOptionb(config
->options
, "radeonsi_commutative_blend_add");
1024 sscreen
->clear_db_cache_before_clear
=
1025 driQueryOptionb(config
->options
, "radeonsi_clear_db_cache_before_clear");
1026 sscreen
->has_msaa_sample_loc_bug
= (sscreen
->info
.family
>= CHIP_POLARIS10
&&
1027 sscreen
->info
.family
<= CHIP_POLARIS12
) ||
1028 sscreen
->info
.family
== CHIP_VEGA10
||
1029 sscreen
->info
.family
== CHIP_RAVEN
;
1030 sscreen
->has_ls_vgpr_init_bug
= sscreen
->info
.family
== CHIP_VEGA10
||
1031 sscreen
->info
.family
== CHIP_RAVEN
;
1032 sscreen
->has_dcc_constant_encode
= sscreen
->info
.family
== CHIP_RAVEN2
;
1034 /* Only enable primitive binning on APUs by default. */
1035 sscreen
->dpbb_allowed
= sscreen
->info
.family
== CHIP_RAVEN
||
1036 sscreen
->info
.family
== CHIP_RAVEN2
;
1038 sscreen
->dfsm_allowed
= sscreen
->info
.family
== CHIP_RAVEN
||
1039 sscreen
->info
.family
== CHIP_RAVEN2
;
1041 /* Process DPBB enable flags. */
1042 if (sscreen
->debug_flags
& DBG(DPBB
)) {
1043 sscreen
->dpbb_allowed
= true;
1044 if (sscreen
->debug_flags
& DBG(DFSM
))
1045 sscreen
->dfsm_allowed
= true;
1048 /* Process DPBB disable flags. */
1049 if (sscreen
->debug_flags
& DBG(NO_DPBB
)) {
1050 sscreen
->dpbb_allowed
= false;
1051 sscreen
->dfsm_allowed
= false;
1052 } else if (sscreen
->debug_flags
& DBG(NO_DFSM
)) {
1053 sscreen
->dfsm_allowed
= false;
1056 /* While it would be nice not to have this flag, we are constrained
1057 * by the reality that LLVM 5.0 doesn't have working VGPR indexing
1060 sscreen
->llvm_has_working_vgpr_indexing
= sscreen
->info
.chip_class
<= VI
;
1062 /* Some chips have RB+ registers, but don't support RB+. Those must
1063 * always disable it.
1065 if (sscreen
->info
.family
== CHIP_STONEY
||
1066 sscreen
->info
.chip_class
>= GFX9
) {
1067 sscreen
->has_rbplus
= true;
1069 sscreen
->rbplus_allowed
=
1070 !(sscreen
->debug_flags
& DBG(NO_RB_PLUS
)) &&
1071 (sscreen
->info
.family
== CHIP_STONEY
||
1072 sscreen
->info
.family
== CHIP_VEGA12
||
1073 sscreen
->info
.family
== CHIP_RAVEN
||
1074 sscreen
->info
.family
== CHIP_RAVEN2
);
1077 sscreen
->dcc_msaa_allowed
=
1078 !(sscreen
->debug_flags
& DBG(NO_DCC_MSAA
));
1080 sscreen
->cpdma_prefetch_writes_memory
= sscreen
->info
.chip_class
<= VI
;
1082 (void) mtx_init(&sscreen
->shader_parts_mutex
, mtx_plain
);
1083 sscreen
->use_monolithic_shaders
=
1084 (sscreen
->debug_flags
& DBG(MONOLITHIC_SHADERS
)) != 0;
1086 sscreen
->barrier_flags
.cp_to_L2
= SI_CONTEXT_INV_SMEM_L1
|
1087 SI_CONTEXT_INV_VMEM_L1
;
1088 if (sscreen
->info
.chip_class
<= VI
) {
1089 sscreen
->barrier_flags
.cp_to_L2
|= SI_CONTEXT_INV_GLOBAL_L2
;
1090 sscreen
->barrier_flags
.L2_to_cp
|= SI_CONTEXT_WRITEBACK_GLOBAL_L2
;
1093 if (debug_get_bool_option("RADEON_DUMP_SHADERS", false))
1094 sscreen
->debug_flags
|= DBG_ALL_SHADERS
;
1101 * That means 8 coverage samples, 4 Z/S samples, and 2 color samples.
1103 * s >= z >= c (ignoring this only wastes memory)
1108 * Only MSAA color and depth buffers are overriden.
1110 if (sscreen
->info
.has_eqaa_surface_allocator
) {
1111 const char *eqaa
= debug_get_option("EQAA", NULL
);
1114 if (eqaa
&& sscanf(eqaa
, "%u,%u,%u", &s
, &z
, &f
) == 3 && s
&& z
&& f
) {
1115 sscreen
->eqaa_force_coverage_samples
= s
;
1116 sscreen
->eqaa_force_z_samples
= z
;
1117 sscreen
->eqaa_force_color_samples
= f
;
1121 for (i
= 0; i
< num_comp_hi_threads
; i
++)
1122 si_init_compiler(sscreen
, &sscreen
->compiler
[i
]);
1123 for (i
= 0; i
< num_comp_lo_threads
; i
++)
1124 si_init_compiler(sscreen
, &sscreen
->compiler_lowp
[i
]);
1126 /* Create the auxiliary context. This must be done last. */
1127 sscreen
->aux_context
= si_create_context(&sscreen
->b
, 0);
1129 if (sscreen
->debug_flags
& DBG(TEST_DMA
))
1130 si_test_dma(sscreen
);
1132 if (sscreen
->debug_flags
& DBG(TEST_DMA_PERF
)) {
1133 si_test_dma_perf(sscreen
);
1136 if (sscreen
->debug_flags
& (DBG(TEST_VMFAULT_CP
) |
1137 DBG(TEST_VMFAULT_SDMA
) |
1138 DBG(TEST_VMFAULT_SHADER
)))
1139 si_test_vmfault(sscreen
);
1141 if (sscreen
->debug_flags
& DBG(TEST_GDS
))
1142 si_test_gds((struct si_context
*)sscreen
->aux_context
);
1144 if (sscreen
->debug_flags
& DBG(TEST_GDS_MM
)) {
1145 si_test_gds_memory_management((struct si_context
*)sscreen
->aux_context
,
1146 32 * 1024, 4, RADEON_DOMAIN_GDS
);
1148 if (sscreen
->debug_flags
& DBG(TEST_GDS_OA_MM
)) {
1149 si_test_gds_memory_management((struct si_context
*)sscreen
->aux_context
,
1150 4, 1, RADEON_DOMAIN_OA
);