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"
29 #include "si_compute.h"
32 #include "ac_llvm_util.h"
33 #include "radeon/radeon_uvd.h"
34 #include "gallivm/lp_bld_misc.h"
35 #include "util/disk_cache.h"
36 #include "util/u_log.h"
37 #include "util/u_memory.h"
38 #include "util/u_suballoc.h"
39 #include "util/u_tests.h"
40 #include "util/u_upload_mgr.h"
41 #include "util/xmlconfig.h"
42 #include "vl/vl_decoder.h"
43 #include "driver_ddebug/dd_util.h"
45 #include "gallium/winsys/radeon/drm/radeon_drm_public.h"
46 #include "gallium/winsys/amdgpu/drm/amdgpu_public.h"
49 static const struct debug_named_value debug_options
[] = {
50 /* Shader logging options: */
51 { "vs", DBG(VS
), "Print vertex shaders" },
52 { "ps", DBG(PS
), "Print pixel shaders" },
53 { "gs", DBG(GS
), "Print geometry shaders" },
54 { "tcs", DBG(TCS
), "Print tessellation control shaders" },
55 { "tes", DBG(TES
), "Print tessellation evaluation shaders" },
56 { "cs", DBG(CS
), "Print compute shaders" },
57 { "noir", DBG(NO_IR
), "Don't print the LLVM IR"},
58 { "notgsi", DBG(NO_TGSI
), "Don't print the TGSI"},
59 { "noasm", DBG(NO_ASM
), "Don't print disassembled shaders"},
60 { "preoptir", DBG(PREOPT_IR
), "Print the LLVM IR before initial optimizations" },
62 /* Shader compiler options the shader cache should be aware of: */
63 { "unsafemath", DBG(UNSAFE_MATH
), "Enable unsafe math shader optimizations" },
64 { "sisched", DBG(SI_SCHED
), "Enable LLVM SI Machine Instruction Scheduler." },
65 { "gisel", DBG(GISEL
), "Enable LLVM global instruction selector." },
67 /* Shader compiler options (with no effect on the shader cache): */
68 { "checkir", DBG(CHECK_IR
), "Enable additional sanity checks on shader IR" },
69 { "mono", DBG(MONOLITHIC_SHADERS
), "Use old-style monolithic shaders compiled on demand" },
70 { "nooptvariant", DBG(NO_OPT_VARIANT
), "Disable compiling optimized shader variants." },
72 /* Information logging options: */
73 { "info", DBG(INFO
), "Print driver information" },
74 { "tex", DBG(TEX
), "Print texture info" },
75 { "compute", DBG(COMPUTE
), "Print compute info" },
76 { "vm", DBG(VM
), "Print virtual addresses when creating resources" },
79 { "forcedma", DBG(FORCE_DMA
), "Use asynchronous DMA for all operations when possible." },
80 { "nodma", DBG(NO_ASYNC_DMA
), "Disable asynchronous DMA" },
81 { "nowc", DBG(NO_WC
), "Disable GTT write combining" },
82 { "check_vm", DBG(CHECK_VM
), "Check VM faults and dump debug info." },
83 { "reserve_vmid", DBG(RESERVE_VMID
), "Force VMID reservation per context." },
84 { "zerovram", DBG(ZERO_VRAM
), "Clear VRAM allocations." },
86 /* 3D engine options: */
87 { "alwayspd", DBG(ALWAYS_PD
), "Always enable the primitive discard compute shader." },
88 { "pd", DBG(PD
), "Enable the primitive discard compute shader for large draw calls." },
89 { "nopd", DBG(NO_PD
), "Disable the primitive discard compute shader." },
90 { "switch_on_eop", DBG(SWITCH_ON_EOP
), "Program WD/IA to switch on end-of-packet." },
91 { "nooutoforder", DBG(NO_OUT_OF_ORDER
), "Disable out-of-order rasterization" },
92 { "nodpbb", DBG(NO_DPBB
), "Disable DPBB." },
93 { "nodfsm", DBG(NO_DFSM
), "Disable DFSM." },
94 { "dpbb", DBG(DPBB
), "Enable DPBB." },
95 { "dfsm", DBG(DFSM
), "Enable DFSM." },
96 { "nohyperz", DBG(NO_HYPERZ
), "Disable Hyper-Z" },
97 { "norbplus", DBG(NO_RB_PLUS
), "Disable RB+." },
98 { "no2d", DBG(NO_2D_TILING
), "Disable 2D tiling" },
99 { "notiling", DBG(NO_TILING
), "Disable tiling" },
100 { "nodcc", DBG(NO_DCC
), "Disable DCC." },
101 { "nodccclear", DBG(NO_DCC_CLEAR
), "Disable DCC fast clear." },
102 { "nodccfb", DBG(NO_DCC_FB
), "Disable separate DCC on the main framebuffer" },
103 { "nodccmsaa", DBG(NO_DCC_MSAA
), "Disable DCC for MSAA" },
104 { "nofmask", DBG(NO_FMASK
), "Disable MSAA compression" },
107 { "testdma", DBG(TEST_DMA
), "Invoke SDMA tests and exit." },
108 { "testvmfaultcp", DBG(TEST_VMFAULT_CP
), "Invoke a CP VM fault test and exit." },
109 { "testvmfaultsdma", DBG(TEST_VMFAULT_SDMA
), "Invoke a SDMA VM fault test and exit." },
110 { "testvmfaultshader", DBG(TEST_VMFAULT_SHADER
), "Invoke a shader VM fault test and exit." },
111 { "testdmaperf", DBG(TEST_DMA_PERF
), "Test DMA performance" },
112 { "testgds", DBG(TEST_GDS
), "Test GDS." },
113 { "testgdsmm", DBG(TEST_GDS_MM
), "Test GDS memory management." },
114 { "testgdsoamm", DBG(TEST_GDS_OA_MM
), "Test GDS OA memory management." },
116 DEBUG_NAMED_VALUE_END
/* must be last */
119 static void si_init_compiler(struct si_screen
*sscreen
,
120 struct ac_llvm_compiler
*compiler
)
122 /* Only create the less-optimizing version of the compiler on APUs
123 * predating Ryzen (Raven). */
124 bool create_low_opt_compiler
= !sscreen
->info
.has_dedicated_vram
&&
125 sscreen
->info
.chip_class
<= GFX8
;
127 enum ac_target_machine_options tm_options
=
128 (sscreen
->debug_flags
& DBG(SI_SCHED
) ? AC_TM_SISCHED
: 0) |
129 (sscreen
->debug_flags
& DBG(GISEL
) ? AC_TM_ENABLE_GLOBAL_ISEL
: 0) |
130 (sscreen
->info
.chip_class
>= GFX9
? AC_TM_FORCE_ENABLE_XNACK
: 0) |
131 (sscreen
->info
.chip_class
< GFX9
? AC_TM_FORCE_DISABLE_XNACK
: 0) |
132 (!sscreen
->llvm_has_working_vgpr_indexing
? AC_TM_PROMOTE_ALLOCA_TO_SCRATCH
: 0) |
133 (sscreen
->debug_flags
& DBG(CHECK_IR
) ? AC_TM_CHECK_IR
: 0) |
134 (create_low_opt_compiler
? AC_TM_CREATE_LOW_OPT
: 0);
137 ac_init_llvm_compiler(compiler
, sscreen
->info
.family
, tm_options
);
138 compiler
->passes
= ac_create_llvm_passes(compiler
->tm
);
140 if (compiler
->low_opt_tm
)
141 compiler
->low_opt_passes
= ac_create_llvm_passes(compiler
->low_opt_tm
);
144 static void si_destroy_compiler(struct ac_llvm_compiler
*compiler
)
146 ac_destroy_llvm_passes(compiler
->passes
);
147 ac_destroy_llvm_passes(compiler
->low_opt_passes
);
148 ac_destroy_llvm_compiler(compiler
);
154 static void si_destroy_context(struct pipe_context
*context
)
156 struct si_context
*sctx
= (struct si_context
*)context
;
159 util_queue_finish(&sctx
->screen
->shader_compiler_queue
);
160 util_queue_finish(&sctx
->screen
->shader_compiler_queue_low_priority
);
162 /* Unreference the framebuffer normally to disable related logic
165 struct pipe_framebuffer_state fb
= {};
166 if (context
->set_framebuffer_state
)
167 context
->set_framebuffer_state(context
, &fb
);
169 si_release_all_descriptors(sctx
);
171 if (sctx
->chip_class
>= GFX10
)
172 gfx10_destroy_query(sctx
);
174 pipe_resource_reference(&sctx
->esgs_ring
, NULL
);
175 pipe_resource_reference(&sctx
->gsvs_ring
, NULL
);
176 pipe_resource_reference(&sctx
->tess_rings
, NULL
);
177 pipe_resource_reference(&sctx
->null_const_buf
.buffer
, NULL
);
178 pipe_resource_reference(&sctx
->sample_pos_buffer
, NULL
);
179 si_resource_reference(&sctx
->border_color_buffer
, NULL
);
180 free(sctx
->border_color_table
);
181 si_resource_reference(&sctx
->scratch_buffer
, NULL
);
182 si_resource_reference(&sctx
->compute_scratch_buffer
, NULL
);
183 si_resource_reference(&sctx
->wait_mem_scratch
, NULL
);
185 si_pm4_free_state(sctx
, sctx
->init_config
, ~0);
186 if (sctx
->init_config_gs_rings
)
187 si_pm4_free_state(sctx
, sctx
->init_config_gs_rings
, ~0);
188 for (i
= 0; i
< ARRAY_SIZE(sctx
->vgt_shader_config
); i
++)
189 si_pm4_delete_state(sctx
, vgt_shader_config
, sctx
->vgt_shader_config
[i
]);
191 if (sctx
->fixed_func_tcs_shader
.cso
)
192 sctx
->b
.delete_tcs_state(&sctx
->b
, sctx
->fixed_func_tcs_shader
.cso
);
193 if (sctx
->custom_dsa_flush
)
194 sctx
->b
.delete_depth_stencil_alpha_state(&sctx
->b
, sctx
->custom_dsa_flush
);
195 if (sctx
->custom_blend_resolve
)
196 sctx
->b
.delete_blend_state(&sctx
->b
, sctx
->custom_blend_resolve
);
197 if (sctx
->custom_blend_fmask_decompress
)
198 sctx
->b
.delete_blend_state(&sctx
->b
, sctx
->custom_blend_fmask_decompress
);
199 if (sctx
->custom_blend_eliminate_fastclear
)
200 sctx
->b
.delete_blend_state(&sctx
->b
, sctx
->custom_blend_eliminate_fastclear
);
201 if (sctx
->custom_blend_dcc_decompress
)
202 sctx
->b
.delete_blend_state(&sctx
->b
, sctx
->custom_blend_dcc_decompress
);
203 if (sctx
->vs_blit_pos
)
204 sctx
->b
.delete_vs_state(&sctx
->b
, sctx
->vs_blit_pos
);
205 if (sctx
->vs_blit_pos_layered
)
206 sctx
->b
.delete_vs_state(&sctx
->b
, sctx
->vs_blit_pos_layered
);
207 if (sctx
->vs_blit_color
)
208 sctx
->b
.delete_vs_state(&sctx
->b
, sctx
->vs_blit_color
);
209 if (sctx
->vs_blit_color_layered
)
210 sctx
->b
.delete_vs_state(&sctx
->b
, sctx
->vs_blit_color_layered
);
211 if (sctx
->vs_blit_texcoord
)
212 sctx
->b
.delete_vs_state(&sctx
->b
, sctx
->vs_blit_texcoord
);
213 if (sctx
->cs_clear_buffer
)
214 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_clear_buffer
);
215 if (sctx
->cs_copy_buffer
)
216 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_copy_buffer
);
217 if (sctx
->cs_copy_image
)
218 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_copy_image
);
219 if (sctx
->cs_copy_image_1d_array
)
220 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_copy_image_1d_array
);
221 if (sctx
->cs_clear_render_target
)
222 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_clear_render_target
);
223 if (sctx
->cs_clear_render_target_1d_array
)
224 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_clear_render_target_1d_array
);
225 if (sctx
->cs_dcc_retile
)
226 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_dcc_retile
);
229 util_blitter_destroy(sctx
->blitter
);
231 /* Release DCC stats. */
232 for (int i
= 0; i
< ARRAY_SIZE(sctx
->dcc_stats
); i
++) {
233 assert(!sctx
->dcc_stats
[i
].query_active
);
235 for (int j
= 0; j
< ARRAY_SIZE(sctx
->dcc_stats
[i
].ps_stats
); j
++)
236 if (sctx
->dcc_stats
[i
].ps_stats
[j
])
237 sctx
->b
.destroy_query(&sctx
->b
,
238 sctx
->dcc_stats
[i
].ps_stats
[j
]);
240 si_texture_reference(&sctx
->dcc_stats
[i
].tex
, NULL
);
243 if (sctx
->query_result_shader
)
244 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->query_result_shader
);
245 if (sctx
->sh_query_result_shader
)
246 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->sh_query_result_shader
);
249 sctx
->ws
->cs_destroy(sctx
->gfx_cs
);
251 sctx
->ws
->cs_destroy(sctx
->dma_cs
);
253 sctx
->ws
->ctx_destroy(sctx
->ctx
);
255 if (sctx
->b
.stream_uploader
)
256 u_upload_destroy(sctx
->b
.stream_uploader
);
257 if (sctx
->b
.const_uploader
)
258 u_upload_destroy(sctx
->b
.const_uploader
);
259 if (sctx
->cached_gtt_allocator
)
260 u_upload_destroy(sctx
->cached_gtt_allocator
);
262 slab_destroy_child(&sctx
->pool_transfers
);
263 slab_destroy_child(&sctx
->pool_transfers_unsync
);
265 if (sctx
->allocator_zeroed_memory
)
266 u_suballocator_destroy(sctx
->allocator_zeroed_memory
);
268 sctx
->ws
->fence_reference(&sctx
->last_gfx_fence
, NULL
);
269 sctx
->ws
->fence_reference(&sctx
->last_sdma_fence
, NULL
);
270 sctx
->ws
->fence_reference(&sctx
->last_ib_barrier_fence
, NULL
);
271 si_resource_reference(&sctx
->eop_bug_scratch
, NULL
);
272 si_resource_reference(&sctx
->index_ring
, NULL
);
273 si_resource_reference(&sctx
->barrier_buf
, NULL
);
274 si_resource_reference(&sctx
->last_ib_barrier_buf
, NULL
);
275 pb_reference(&sctx
->gds
, NULL
);
276 pb_reference(&sctx
->gds_oa
, NULL
);
278 si_destroy_compiler(&sctx
->compiler
);
280 si_saved_cs_reference(&sctx
->current_saved_cs
, NULL
);
282 _mesa_hash_table_destroy(sctx
->tex_handles
, NULL
);
283 _mesa_hash_table_destroy(sctx
->img_handles
, NULL
);
285 util_dynarray_fini(&sctx
->resident_tex_handles
);
286 util_dynarray_fini(&sctx
->resident_img_handles
);
287 util_dynarray_fini(&sctx
->resident_tex_needs_color_decompress
);
288 util_dynarray_fini(&sctx
->resident_img_needs_color_decompress
);
289 util_dynarray_fini(&sctx
->resident_tex_needs_depth_decompress
);
290 si_unref_sdma_uploads(sctx
);
294 static enum pipe_reset_status
si_get_reset_status(struct pipe_context
*ctx
)
296 struct si_context
*sctx
= (struct si_context
*)ctx
;
298 return sctx
->ws
->ctx_query_reset_status(sctx
->ctx
);
301 static void si_set_device_reset_callback(struct pipe_context
*ctx
,
302 const struct pipe_device_reset_callback
*cb
)
304 struct si_context
*sctx
= (struct si_context
*)ctx
;
307 sctx
->device_reset_callback
= *cb
;
309 memset(&sctx
->device_reset_callback
, 0,
310 sizeof(sctx
->device_reset_callback
));
313 bool si_check_device_reset(struct si_context
*sctx
)
315 enum pipe_reset_status status
;
317 if (!sctx
->device_reset_callback
.reset
)
320 status
= sctx
->ws
->ctx_query_reset_status(sctx
->ctx
);
321 if (status
== PIPE_NO_RESET
)
324 sctx
->device_reset_callback
.reset(sctx
->device_reset_callback
.data
, status
);
328 /* Apitrace profiling:
329 * 1) qapitrace : Tools -> Profile: Measure CPU & GPU times
330 * 2) In the middle panel, zoom in (mouse wheel) on some bad draw call
331 * and remember its number.
332 * 3) In Mesa, enable queries and performance counters around that draw
333 * call and print the results.
334 * 4) glretrace --benchmark --markers ..
336 static void si_emit_string_marker(struct pipe_context
*ctx
,
337 const char *string
, int len
)
339 struct si_context
*sctx
= (struct si_context
*)ctx
;
341 dd_parse_apitrace_marker(string
, len
, &sctx
->apitrace_call_number
);
344 u_log_printf(sctx
->log
, "\nString marker: %*s\n", len
, string
);
347 static void si_set_debug_callback(struct pipe_context
*ctx
,
348 const struct pipe_debug_callback
*cb
)
350 struct si_context
*sctx
= (struct si_context
*)ctx
;
351 struct si_screen
*screen
= sctx
->screen
;
353 util_queue_finish(&screen
->shader_compiler_queue
);
354 util_queue_finish(&screen
->shader_compiler_queue_low_priority
);
359 memset(&sctx
->debug
, 0, sizeof(sctx
->debug
));
362 static void si_set_log_context(struct pipe_context
*ctx
,
363 struct u_log_context
*log
)
365 struct si_context
*sctx
= (struct si_context
*)ctx
;
369 u_log_add_auto_logger(log
, si_auto_log_cs
, sctx
);
372 static void si_set_context_param(struct pipe_context
*ctx
,
373 enum pipe_context_param param
,
376 struct radeon_winsys
*ws
= ((struct si_context
*)ctx
)->ws
;
379 case PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE
:
380 ws
->pin_threads_to_L3_cache(ws
, value
);
386 static struct pipe_context
*si_create_context(struct pipe_screen
*screen
,
389 struct si_context
*sctx
= CALLOC_STRUCT(si_context
);
390 struct si_screen
* sscreen
= (struct si_screen
*)screen
;
391 struct radeon_winsys
*ws
= sscreen
->ws
;
393 bool stop_exec_on_failure
= (flags
& PIPE_CONTEXT_LOSE_CONTEXT_ON_RESET
) != 0;
398 sctx
->has_graphics
= sscreen
->info
.chip_class
== GFX6
||
399 !(flags
& PIPE_CONTEXT_COMPUTE_ONLY
);
401 if (flags
& PIPE_CONTEXT_DEBUG
)
402 sscreen
->record_llvm_ir
= true; /* racy but not critical */
404 sctx
->b
.screen
= screen
; /* this must be set first */
406 sctx
->b
.destroy
= si_destroy_context
;
407 sctx
->screen
= sscreen
; /* Easy accessing of screen/winsys. */
408 sctx
->is_debug
= (flags
& PIPE_CONTEXT_DEBUG
) != 0;
410 slab_create_child(&sctx
->pool_transfers
, &sscreen
->pool_transfers
);
411 slab_create_child(&sctx
->pool_transfers_unsync
, &sscreen
->pool_transfers
);
413 sctx
->ws
= sscreen
->ws
;
414 sctx
->family
= sscreen
->info
.family
;
415 sctx
->chip_class
= sscreen
->info
.chip_class
;
417 if (sctx
->chip_class
== GFX7
||
418 sctx
->chip_class
== GFX8
||
419 sctx
->chip_class
== GFX9
) {
420 sctx
->eop_bug_scratch
= si_resource(
421 pipe_buffer_create(&sscreen
->b
, 0, PIPE_USAGE_DEFAULT
,
422 16 * sscreen
->info
.num_render_backends
));
423 if (!sctx
->eop_bug_scratch
)
427 /* Initialize context allocators. */
428 sctx
->allocator_zeroed_memory
=
429 u_suballocator_create(&sctx
->b
, 128 * 1024,
430 0, PIPE_USAGE_DEFAULT
,
431 SI_RESOURCE_FLAG_UNMAPPABLE
|
432 SI_RESOURCE_FLAG_CLEAR
, false);
433 if (!sctx
->allocator_zeroed_memory
)
436 sctx
->b
.stream_uploader
= u_upload_create(&sctx
->b
, 1024 * 1024,
437 0, PIPE_USAGE_STREAM
,
438 SI_RESOURCE_FLAG_READ_ONLY
);
439 if (!sctx
->b
.stream_uploader
)
442 sctx
->cached_gtt_allocator
= u_upload_create(&sctx
->b
, 16 * 1024,
443 0, PIPE_USAGE_STAGING
, 0);
444 if (!sctx
->cached_gtt_allocator
)
447 sctx
->ctx
= sctx
->ws
->ctx_create(sctx
->ws
);
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
, stop_exec_on_failure
);
457 bool use_sdma_upload
= sscreen
->info
.has_dedicated_vram
&& sctx
->dma_cs
;
458 sctx
->b
.const_uploader
= u_upload_create(&sctx
->b
, 256 * 1024,
459 0, PIPE_USAGE_DEFAULT
,
460 SI_RESOURCE_FLAG_32BIT
|
462 SI_RESOURCE_FLAG_UPLOAD_FLUSH_EXPLICIT_VIA_SDMA
:
463 (sscreen
->cpdma_prefetch_writes_memory
?
464 0 : SI_RESOURCE_FLAG_READ_ONLY
)));
465 if (!sctx
->b
.const_uploader
)
469 u_upload_enable_flush_explicit(sctx
->b
.const_uploader
);
471 sctx
->gfx_cs
= ws
->cs_create(sctx
->ctx
,
472 sctx
->has_graphics
? RING_GFX
: RING_COMPUTE
,
473 (void*)si_flush_gfx_cs
, sctx
, stop_exec_on_failure
);
476 sctx
->border_color_table
= malloc(SI_MAX_BORDER_COLORS
*
477 sizeof(*sctx
->border_color_table
));
478 if (!sctx
->border_color_table
)
481 sctx
->border_color_buffer
= si_resource(
482 pipe_buffer_create(screen
, 0, PIPE_USAGE_DEFAULT
,
483 SI_MAX_BORDER_COLORS
*
484 sizeof(*sctx
->border_color_table
)));
485 if (!sctx
->border_color_buffer
)
488 sctx
->border_color_map
=
489 ws
->buffer_map(sctx
->border_color_buffer
->buf
,
490 NULL
, PIPE_TRANSFER_WRITE
);
491 if (!sctx
->border_color_map
)
494 if (sctx
->chip_class
>= GFX10
)
495 sctx
->ngg
= !sscreen
->options
.disable_ngg
;
497 /* Initialize context functions used by graphics and compute. */
498 if (sctx
->chip_class
>= GFX10
)
499 sctx
->emit_cache_flush
= gfx10_emit_cache_flush
;
501 sctx
->emit_cache_flush
= si_emit_cache_flush
;
503 sctx
->b
.emit_string_marker
= si_emit_string_marker
;
504 sctx
->b
.set_debug_callback
= si_set_debug_callback
;
505 sctx
->b
.set_log_context
= si_set_log_context
;
506 sctx
->b
.set_context_param
= si_set_context_param
;
507 sctx
->b
.get_device_reset_status
= si_get_reset_status
;
508 sctx
->b
.set_device_reset_callback
= si_set_device_reset_callback
;
510 si_init_all_descriptors(sctx
);
511 si_init_buffer_functions(sctx
);
512 si_init_clear_functions(sctx
);
513 si_init_blit_functions(sctx
);
514 si_init_compute_functions(sctx
);
515 si_init_compute_blit_functions(sctx
);
516 si_init_debug_functions(sctx
);
517 si_init_fence_functions(sctx
);
518 si_init_query_functions(sctx
);
519 si_init_state_compute_functions(sctx
);
521 /* Initialize graphics-only context functions. */
522 if (sctx
->has_graphics
) {
523 si_init_context_texture_functions(sctx
);
524 if (sctx
->chip_class
>= GFX10
)
525 gfx10_init_query(sctx
);
526 si_init_msaa_functions(sctx
);
527 si_init_shader_functions(sctx
);
528 si_init_state_functions(sctx
);
529 si_init_streamout_functions(sctx
);
530 si_init_viewport_functions(sctx
);
532 sctx
->blitter
= util_blitter_create(&sctx
->b
);
533 if (sctx
->blitter
== NULL
)
535 sctx
->blitter
->skip_viewport_restore
= true;
537 si_init_draw_functions(sctx
);
538 si_initialize_prim_discard_tunables(sctx
);
541 /* Initialize SDMA functions. */
542 if (sctx
->chip_class
>= GFX7
)
543 cik_init_sdma_functions(sctx
);
545 si_init_dma_functions(sctx
);
547 if (sscreen
->debug_flags
& DBG(FORCE_DMA
))
548 sctx
->b
.resource_copy_region
= sctx
->dma_copy
;
550 sctx
->sample_mask
= 0xffff;
552 /* Initialize multimedia functions. */
553 if (sscreen
->info
.has_hw_decode
) {
554 sctx
->b
.create_video_codec
= si_uvd_create_decoder
;
555 sctx
->b
.create_video_buffer
= si_video_buffer_create
;
557 sctx
->b
.create_video_codec
= vl_create_decoder
;
558 sctx
->b
.create_video_buffer
= vl_video_buffer_create
;
561 if (sctx
->chip_class
>= GFX9
) {
562 sctx
->wait_mem_scratch
= si_resource(
563 pipe_buffer_create(screen
, 0, PIPE_USAGE_DEFAULT
, 8));
564 if (!sctx
->wait_mem_scratch
)
567 /* Initialize the memory. */
568 si_cp_write_data(sctx
, sctx
->wait_mem_scratch
, 0, 4,
569 V_370_MEM
, V_370_ME
, &sctx
->wait_mem_number
);
572 /* GFX7 cannot unbind a constant buffer (S_BUFFER_LOAD doesn't skip loads
573 * if NUM_RECORDS == 0). We need to use a dummy buffer instead. */
574 if (sctx
->chip_class
== GFX7
) {
575 sctx
->null_const_buf
.buffer
=
576 pipe_aligned_buffer_create(screen
,
577 SI_RESOURCE_FLAG_32BIT
,
578 PIPE_USAGE_DEFAULT
, 16,
579 sctx
->screen
->info
.tcc_cache_line_size
);
580 if (!sctx
->null_const_buf
.buffer
)
582 sctx
->null_const_buf
.buffer_size
= sctx
->null_const_buf
.buffer
->width0
;
584 unsigned start_shader
= sctx
->has_graphics
? 0 : PIPE_SHADER_COMPUTE
;
585 for (shader
= start_shader
; shader
< SI_NUM_SHADERS
; shader
++) {
586 for (i
= 0; i
< SI_NUM_CONST_BUFFERS
; i
++) {
587 sctx
->b
.set_constant_buffer(&sctx
->b
, shader
, i
,
588 &sctx
->null_const_buf
);
592 si_set_rw_buffer(sctx
, SI_HS_CONST_DEFAULT_TESS_LEVELS
,
593 &sctx
->null_const_buf
);
594 si_set_rw_buffer(sctx
, SI_VS_CONST_INSTANCE_DIVISORS
,
595 &sctx
->null_const_buf
);
596 si_set_rw_buffer(sctx
, SI_VS_CONST_CLIP_PLANES
,
597 &sctx
->null_const_buf
);
598 si_set_rw_buffer(sctx
, SI_PS_CONST_POLY_STIPPLE
,
599 &sctx
->null_const_buf
);
600 si_set_rw_buffer(sctx
, SI_PS_CONST_SAMPLE_POSITIONS
,
601 &sctx
->null_const_buf
);
604 uint64_t max_threads_per_block
;
605 screen
->get_compute_param(screen
, PIPE_SHADER_IR_TGSI
,
606 PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK
,
607 &max_threads_per_block
);
609 /* The maximum number of scratch waves. Scratch space isn't divided
610 * evenly between CUs. The number is only a function of the number of CUs.
611 * We can decrease the constant to decrease the scratch buffer size.
613 * sctx->scratch_waves must be >= the maximum posible size of
614 * 1 threadgroup, so that the hw doesn't hang from being unable
617 * The recommended value is 4 per CU at most. Higher numbers don't
618 * bring much benefit, but they still occupy chip resources (think
619 * async compute). I've seen ~2% performance difference between 4 and 32.
621 sctx
->scratch_waves
= MAX2(32 * sscreen
->info
.num_good_compute_units
,
622 max_threads_per_block
/ 64);
624 si_init_compiler(sscreen
, &sctx
->compiler
);
626 /* Bindless handles. */
627 sctx
->tex_handles
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
628 _mesa_key_pointer_equal
);
629 sctx
->img_handles
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
630 _mesa_key_pointer_equal
);
632 util_dynarray_init(&sctx
->resident_tex_handles
, NULL
);
633 util_dynarray_init(&sctx
->resident_img_handles
, NULL
);
634 util_dynarray_init(&sctx
->resident_tex_needs_color_decompress
, NULL
);
635 util_dynarray_init(&sctx
->resident_img_needs_color_decompress
, NULL
);
636 util_dynarray_init(&sctx
->resident_tex_needs_depth_decompress
, NULL
);
638 sctx
->sample_pos_buffer
=
639 pipe_buffer_create(sctx
->b
.screen
, 0, PIPE_USAGE_DEFAULT
,
640 sizeof(sctx
->sample_positions
));
641 pipe_buffer_write(&sctx
->b
, sctx
->sample_pos_buffer
, 0,
642 sizeof(sctx
->sample_positions
), &sctx
->sample_positions
);
644 /* this must be last */
645 si_begin_new_gfx_cs(sctx
);
647 if (sctx
->chip_class
== GFX7
) {
648 /* Clear the NULL constant buffer, because loads should return zeros.
649 * Note that this forces CP DMA to be used, because clover deadlocks
650 * for some reason when the compute codepath is used.
652 uint32_t clear_value
= 0;
653 si_clear_buffer(sctx
, sctx
->null_const_buf
.buffer
, 0,
654 sctx
->null_const_buf
.buffer
->width0
,
655 &clear_value
, 4, SI_COHERENCY_SHADER
, true);
659 fprintf(stderr
, "radeonsi: Failed to create a context.\n");
660 si_destroy_context(&sctx
->b
);
664 static struct pipe_context
*si_pipe_create_context(struct pipe_screen
*screen
,
665 void *priv
, unsigned flags
)
667 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
668 struct pipe_context
*ctx
;
670 if (sscreen
->debug_flags
& DBG(CHECK_VM
))
671 flags
|= PIPE_CONTEXT_DEBUG
;
673 ctx
= si_create_context(screen
, flags
);
675 if (!(flags
& PIPE_CONTEXT_PREFER_THREADED
))
678 /* Clover (compute-only) is unsupported. */
679 if (flags
& PIPE_CONTEXT_COMPUTE_ONLY
)
682 /* When shaders are logged to stderr, asynchronous compilation is
684 if (sscreen
->debug_flags
& DBG_ALL_SHADERS
)
687 /* Use asynchronous flushes only on amdgpu, since the radeon
688 * implementation for fence_server_sync is incomplete. */
689 return threaded_context_create(ctx
, &sscreen
->pool_transfers
,
690 si_replace_buffer_storage
,
691 sscreen
->info
.is_amdgpu
? si_create_fence
: NULL
,
692 &((struct si_context
*)ctx
)->tc
);
698 static void si_destroy_screen(struct pipe_screen
* pscreen
)
700 struct si_screen
*sscreen
= (struct si_screen
*)pscreen
;
701 struct si_shader_part
*parts
[] = {
703 sscreen
->tcs_epilogs
,
710 if (!sscreen
->ws
->unref(sscreen
->ws
))
713 mtx_destroy(&sscreen
->aux_context_lock
);
715 struct u_log_context
*aux_log
= ((struct si_context
*)sscreen
->aux_context
)->log
;
717 sscreen
->aux_context
->set_log_context(sscreen
->aux_context
, NULL
);
718 u_log_context_destroy(aux_log
);
722 sscreen
->aux_context
->destroy(sscreen
->aux_context
);
724 util_queue_destroy(&sscreen
->shader_compiler_queue
);
725 util_queue_destroy(&sscreen
->shader_compiler_queue_low_priority
);
727 for (i
= 0; i
< ARRAY_SIZE(sscreen
->compiler
); i
++)
728 si_destroy_compiler(&sscreen
->compiler
[i
]);
730 for (i
= 0; i
< ARRAY_SIZE(sscreen
->compiler_lowp
); i
++)
731 si_destroy_compiler(&sscreen
->compiler_lowp
[i
]);
733 /* Free shader parts. */
734 for (i
= 0; i
< ARRAY_SIZE(parts
); i
++) {
736 struct si_shader_part
*part
= parts
[i
];
738 parts
[i
] = part
->next
;
739 si_shader_binary_clean(&part
->binary
);
743 mtx_destroy(&sscreen
->shader_parts_mutex
);
744 si_destroy_shader_cache(sscreen
);
746 si_destroy_perfcounters(sscreen
);
747 si_gpu_load_kill_thread(sscreen
);
749 mtx_destroy(&sscreen
->gpu_load_mutex
);
751 slab_destroy_parent(&sscreen
->pool_transfers
);
753 disk_cache_destroy(sscreen
->disk_shader_cache
);
754 sscreen
->ws
->destroy(sscreen
->ws
);
758 static void si_init_gs_info(struct si_screen
*sscreen
)
760 sscreen
->gs_table_depth
= ac_get_gs_table_depth(sscreen
->info
.chip_class
,
761 sscreen
->info
.family
);
764 static void si_test_vmfault(struct si_screen
*sscreen
)
766 struct pipe_context
*ctx
= sscreen
->aux_context
;
767 struct si_context
*sctx
= (struct si_context
*)ctx
;
768 struct pipe_resource
*buf
=
769 pipe_buffer_create_const0(&sscreen
->b
, 0, PIPE_USAGE_DEFAULT
, 64);
772 puts("Buffer allocation failed.");
776 si_resource(buf
)->gpu_address
= 0; /* cause a VM fault */
778 if (sscreen
->debug_flags
& DBG(TEST_VMFAULT_CP
)) {
779 si_cp_dma_copy_buffer(sctx
, buf
, buf
, 0, 4, 4, 0,
780 SI_COHERENCY_NONE
, L2_BYPASS
);
781 ctx
->flush(ctx
, NULL
, 0);
782 puts("VM fault test: CP - done.");
784 if (sscreen
->debug_flags
& DBG(TEST_VMFAULT_SDMA
)) {
785 si_sdma_clear_buffer(sctx
, buf
, 0, 4, 0);
786 ctx
->flush(ctx
, NULL
, 0);
787 puts("VM fault test: SDMA - done.");
789 if (sscreen
->debug_flags
& DBG(TEST_VMFAULT_SHADER
)) {
790 util_test_constant_buffer(ctx
, buf
);
791 puts("VM fault test: Shader - done.");
796 static void si_test_gds_memory_management(struct si_context
*sctx
,
797 unsigned alloc_size
, unsigned alignment
,
798 enum radeon_bo_domain domain
)
800 struct radeon_winsys
*ws
= sctx
->ws
;
801 struct radeon_cmdbuf
*cs
[8];
802 struct pb_buffer
*gds_bo
[ARRAY_SIZE(cs
)];
804 for (unsigned i
= 0; i
< ARRAY_SIZE(cs
); i
++) {
805 cs
[i
] = ws
->cs_create(sctx
->ctx
, RING_COMPUTE
,
807 gds_bo
[i
] = ws
->buffer_create(ws
, alloc_size
, alignment
, domain
, 0);
811 for (unsigned iterations
= 0; iterations
< 20000; iterations
++) {
812 for (unsigned i
= 0; i
< ARRAY_SIZE(cs
); i
++) {
813 /* This clears GDS with CP DMA.
815 * We don't care if GDS is present. Just add some packet
816 * to make the GPU busy for a moment.
818 si_cp_dma_clear_buffer(sctx
, cs
[i
], NULL
, 0, alloc_size
, 0,
819 SI_CPDMA_SKIP_BO_LIST_UPDATE
|
820 SI_CPDMA_SKIP_CHECK_CS_SPACE
|
821 SI_CPDMA_SKIP_GFX_SYNC
, 0, 0);
823 ws
->cs_add_buffer(cs
[i
], gds_bo
[i
], domain
,
824 RADEON_USAGE_READWRITE
, 0);
825 ws
->cs_flush(cs
[i
], PIPE_FLUSH_ASYNC
, NULL
);
831 static void si_disk_cache_create(struct si_screen
*sscreen
)
833 /* Don't use the cache if shader dumping is enabled. */
834 if (sscreen
->debug_flags
& DBG_ALL_SHADERS
)
837 struct mesa_sha1 ctx
;
838 unsigned char sha1
[20];
839 char cache_id
[20 * 2 + 1];
841 _mesa_sha1_init(&ctx
);
843 if (!disk_cache_get_function_identifier(si_disk_cache_create
, &ctx
) ||
844 !disk_cache_get_function_identifier(LLVMInitializeAMDGPUTargetInfo
,
848 _mesa_sha1_final(&ctx
, sha1
);
849 disk_cache_format_hex_id(cache_id
, sha1
, 20 * 2);
851 /* These flags affect shader compilation. */
852 #define ALL_FLAGS (DBG(FS_CORRECT_DERIVS_AFTER_KILL) | \
856 uint64_t shader_debug_flags
= sscreen
->debug_flags
&
859 /* Add the high bits of 32-bit addresses, which affects
860 * how 32-bit addresses are expanded to 64 bits.
862 STATIC_ASSERT(ALL_FLAGS
<= UINT_MAX
);
863 assert((int16_t)sscreen
->info
.address32_hi
== (int32_t)sscreen
->info
.address32_hi
);
864 shader_debug_flags
|= (uint64_t)(sscreen
->info
.address32_hi
& 0xffff) << 32;
866 if (sscreen
->options
.enable_nir
)
867 shader_debug_flags
|= 1ull << 48;
869 sscreen
->disk_shader_cache
=
870 disk_cache_create(sscreen
->info
.name
,
875 static void si_set_max_shader_compiler_threads(struct pipe_screen
*screen
,
876 unsigned max_threads
)
878 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
880 /* This function doesn't allow a greater number of threads than
881 * the queue had at its creation. */
882 util_queue_adjust_num_threads(&sscreen
->shader_compiler_queue
,
884 /* Don't change the number of threads on the low priority queue. */
887 static bool si_is_parallel_shader_compilation_finished(struct pipe_screen
*screen
,
889 unsigned shader_type
)
891 if (shader_type
== PIPE_SHADER_COMPUTE
) {
892 struct si_compute
*cs
= (struct si_compute
*)shader
;
894 return util_queue_fence_is_signalled(&cs
->ready
);
896 struct si_shader_selector
*sel
= (struct si_shader_selector
*)shader
;
898 return util_queue_fence_is_signalled(&sel
->ready
);
901 static struct pipe_screen
*
902 radeonsi_screen_create_impl(struct radeon_winsys
*ws
,
903 const struct pipe_screen_config
*config
)
905 struct si_screen
*sscreen
= CALLOC_STRUCT(si_screen
);
906 unsigned hw_threads
, num_comp_hi_threads
, num_comp_lo_threads
, i
;
913 ws
->query_info(ws
, &sscreen
->info
);
915 if (sscreen
->info
.chip_class
== GFX10
&& HAVE_LLVM
< 0x0900) {
916 fprintf(stderr
, "radeonsi: Navi family support requires LLVM 9 or higher\n");
921 if (sscreen
->info
.chip_class
>= GFX9
) {
922 sscreen
->se_tile_repeat
= 32 * sscreen
->info
.max_se
;
924 ac_get_raster_config(&sscreen
->info
,
925 &sscreen
->pa_sc_raster_config
,
926 &sscreen
->pa_sc_raster_config_1
,
927 &sscreen
->se_tile_repeat
);
930 sscreen
->debug_flags
= debug_get_flags_option("R600_DEBUG",
932 sscreen
->debug_flags
|= debug_get_flags_option("AMD_DEBUG",
935 /* Set functions first. */
936 sscreen
->b
.context_create
= si_pipe_create_context
;
937 sscreen
->b
.destroy
= si_destroy_screen
;
938 sscreen
->b
.set_max_shader_compiler_threads
=
939 si_set_max_shader_compiler_threads
;
940 sscreen
->b
.is_parallel_shader_compilation_finished
=
941 si_is_parallel_shader_compilation_finished
;
943 si_init_screen_get_functions(sscreen
);
944 si_init_screen_buffer_functions(sscreen
);
945 si_init_screen_fence_functions(sscreen
);
946 si_init_screen_state_functions(sscreen
);
947 si_init_screen_texture_functions(sscreen
);
948 si_init_screen_query_functions(sscreen
);
950 /* Set these flags in debug_flags early, so that the shader cache takes
953 if (driQueryOptionb(config
->options
,
954 "glsl_correct_derivatives_after_discard"))
955 sscreen
->debug_flags
|= DBG(FS_CORRECT_DERIVS_AFTER_KILL
);
956 if (driQueryOptionb(config
->options
, "radeonsi_enable_sisched"))
957 sscreen
->debug_flags
|= DBG(SI_SCHED
);
959 if (sscreen
->debug_flags
& DBG(INFO
))
960 ac_print_gpu_info(&sscreen
->info
);
962 slab_create_parent(&sscreen
->pool_transfers
,
963 sizeof(struct si_transfer
), 64);
965 sscreen
->force_aniso
= MIN2(16, debug_get_num_option("R600_TEX_ANISO", -1));
966 if (sscreen
->force_aniso
== -1) {
967 sscreen
->force_aniso
= MIN2(16, debug_get_num_option("AMD_TEX_ANISO", -1));
970 if (sscreen
->force_aniso
>= 0) {
971 printf("radeonsi: Forcing anisotropy filter to %ix\n",
972 /* round down to a power of two */
973 1 << util_logbase2(sscreen
->force_aniso
));
976 (void) mtx_init(&sscreen
->aux_context_lock
, mtx_plain
);
977 (void) mtx_init(&sscreen
->gpu_load_mutex
, mtx_plain
);
979 si_init_gs_info(sscreen
);
980 if (!si_init_shader_cache(sscreen
)) {
985 si_disk_cache_create(sscreen
);
987 /* Determine the number of shader compiler threads. */
988 hw_threads
= sysconf(_SC_NPROCESSORS_ONLN
);
990 if (hw_threads
>= 12) {
991 num_comp_hi_threads
= hw_threads
* 3 / 4;
992 num_comp_lo_threads
= hw_threads
/ 3;
993 } else if (hw_threads
>= 6) {
994 num_comp_hi_threads
= hw_threads
- 2;
995 num_comp_lo_threads
= hw_threads
/ 2;
996 } else if (hw_threads
>= 2) {
997 num_comp_hi_threads
= hw_threads
- 1;
998 num_comp_lo_threads
= hw_threads
/ 2;
1000 num_comp_hi_threads
= 1;
1001 num_comp_lo_threads
= 1;
1004 num_comp_hi_threads
= MIN2(num_comp_hi_threads
,
1005 ARRAY_SIZE(sscreen
->compiler
));
1006 num_comp_lo_threads
= MIN2(num_comp_lo_threads
,
1007 ARRAY_SIZE(sscreen
->compiler_lowp
));
1009 if (!util_queue_init(&sscreen
->shader_compiler_queue
, "sh",
1010 64, num_comp_hi_threads
,
1011 UTIL_QUEUE_INIT_RESIZE_IF_FULL
|
1012 UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY
)) {
1013 si_destroy_shader_cache(sscreen
);
1018 if (!util_queue_init(&sscreen
->shader_compiler_queue_low_priority
,
1020 64, num_comp_lo_threads
,
1021 UTIL_QUEUE_INIT_RESIZE_IF_FULL
|
1022 UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY
|
1023 UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY
)) {
1024 si_destroy_shader_cache(sscreen
);
1029 if (!debug_get_bool_option("RADEON_DISABLE_PERFCOUNTERS", false))
1030 si_init_perfcounters(sscreen
);
1032 /* Determine tessellation ring info. */
1033 bool double_offchip_buffers
= sscreen
->info
.chip_class
>= GFX7
&&
1034 sscreen
->info
.family
!= CHIP_CARRIZO
&&
1035 sscreen
->info
.family
!= CHIP_STONEY
;
1036 /* This must be one less than the maximum number due to a hw limitation.
1037 * Various hardware bugs need this.
1039 unsigned max_offchip_buffers_per_se
;
1041 if (sscreen
->info
.chip_class
>= GFX10
)
1042 max_offchip_buffers_per_se
= 256;
1043 /* Only certain chips can use the maximum value. */
1044 else if (sscreen
->info
.family
== CHIP_VEGA12
||
1045 sscreen
->info
.family
== CHIP_VEGA20
)
1046 max_offchip_buffers_per_se
= double_offchip_buffers
? 128 : 64;
1048 max_offchip_buffers_per_se
= double_offchip_buffers
? 127 : 63;
1050 unsigned max_offchip_buffers
= max_offchip_buffers_per_se
*
1051 sscreen
->info
.max_se
;
1052 unsigned offchip_granularity
;
1054 /* Hawaii has a bug with offchip buffers > 256 that can be worked
1055 * around by setting 4K granularity.
1057 if (sscreen
->info
.family
== CHIP_HAWAII
) {
1058 sscreen
->tess_offchip_block_dw_size
= 4096;
1059 offchip_granularity
= V_03093C_X_4K_DWORDS
;
1061 sscreen
->tess_offchip_block_dw_size
= 8192;
1062 offchip_granularity
= V_03093C_X_8K_DWORDS
;
1065 sscreen
->tess_factor_ring_size
= 32768 * sscreen
->info
.max_se
;
1066 assert(((sscreen
->tess_factor_ring_size
/ 4) & C_030938_SIZE
) == 0);
1067 sscreen
->tess_offchip_ring_size
= max_offchip_buffers
*
1068 sscreen
->tess_offchip_block_dw_size
* 4;
1070 if (sscreen
->info
.chip_class
>= GFX7
) {
1071 if (sscreen
->info
.chip_class
>= GFX8
)
1072 --max_offchip_buffers
;
1073 sscreen
->vgt_hs_offchip_param
=
1074 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers
) |
1075 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity
);
1077 assert(offchip_granularity
== V_03093C_X_8K_DWORDS
);
1078 sscreen
->vgt_hs_offchip_param
=
1079 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers
);
1082 /* The mere presense of CLEAR_STATE in the IB causes random GPU hangs
1083 * on GFX6. Some CLEAR_STATE cause asic hang on radeon kernel, etc.
1084 * SPI_VS_OUT_CONFIG. So only enable GFX7 CLEAR_STATE on amdgpu kernel.*/
1085 sscreen
->has_clear_state
= sscreen
->info
.chip_class
>= GFX7
&&
1086 sscreen
->info
.is_amdgpu
;
1088 sscreen
->has_distributed_tess
=
1089 sscreen
->info
.chip_class
>= GFX8
&&
1090 sscreen
->info
.max_se
>= 2;
1092 sscreen
->has_draw_indirect_multi
=
1093 (sscreen
->info
.family
>= CHIP_POLARIS10
) ||
1094 (sscreen
->info
.chip_class
== GFX8
&&
1095 sscreen
->info
.pfp_fw_version
>= 121 &&
1096 sscreen
->info
.me_fw_version
>= 87) ||
1097 (sscreen
->info
.chip_class
== GFX7
&&
1098 sscreen
->info
.pfp_fw_version
>= 211 &&
1099 sscreen
->info
.me_fw_version
>= 173) ||
1100 (sscreen
->info
.chip_class
== GFX6
&&
1101 sscreen
->info
.pfp_fw_version
>= 79 &&
1102 sscreen
->info
.me_fw_version
>= 142);
1104 sscreen
->has_out_of_order_rast
= sscreen
->info
.chip_class
>= GFX8
&&
1105 sscreen
->info
.max_se
>= 2 &&
1106 !(sscreen
->debug_flags
& DBG(NO_OUT_OF_ORDER
));
1107 sscreen
->assume_no_z_fights
=
1108 driQueryOptionb(config
->options
, "radeonsi_assume_no_z_fights");
1109 sscreen
->commutative_blend_add
=
1110 driQueryOptionb(config
->options
, "radeonsi_commutative_blend_add");
1113 #define OPT_BOOL(name, dflt, description) \
1114 sscreen->options.name = \
1115 driQueryOptionb(config->options, "radeonsi_"#name);
1116 #include "si_debug_options.h"
1119 sscreen
->has_gfx9_scissor_bug
= sscreen
->info
.family
== CHIP_VEGA10
||
1120 sscreen
->info
.family
== CHIP_RAVEN
;
1121 sscreen
->has_msaa_sample_loc_bug
= (sscreen
->info
.family
>= CHIP_POLARIS10
&&
1122 sscreen
->info
.family
<= CHIP_POLARIS12
) ||
1123 sscreen
->info
.family
== CHIP_VEGA10
||
1124 sscreen
->info
.family
== CHIP_RAVEN
;
1125 sscreen
->has_ls_vgpr_init_bug
= sscreen
->info
.family
== CHIP_VEGA10
||
1126 sscreen
->info
.family
== CHIP_RAVEN
;
1127 sscreen
->has_dcc_constant_encode
= sscreen
->info
.family
== CHIP_RAVEN2
;
1129 /* Only enable primitive binning on APUs by default. */
1130 sscreen
->dpbb_allowed
= sscreen
->info
.family
== CHIP_RAVEN
||
1131 sscreen
->info
.family
== CHIP_RAVEN2
;
1133 sscreen
->dfsm_allowed
= sscreen
->info
.family
== CHIP_RAVEN
||
1134 sscreen
->info
.family
== CHIP_RAVEN2
;
1136 /* Process DPBB enable flags. */
1137 if (sscreen
->debug_flags
& DBG(DPBB
)) {
1138 sscreen
->dpbb_allowed
= true;
1139 if (sscreen
->debug_flags
& DBG(DFSM
))
1140 sscreen
->dfsm_allowed
= true;
1143 /* Process DPBB disable flags. */
1144 if (sscreen
->debug_flags
& DBG(NO_DPBB
)) {
1145 sscreen
->dpbb_allowed
= false;
1146 sscreen
->dfsm_allowed
= false;
1147 } else if (sscreen
->debug_flags
& DBG(NO_DFSM
)) {
1148 sscreen
->dfsm_allowed
= false;
1151 /* While it would be nice not to have this flag, we are constrained
1152 * by the reality that LLVM 9.0 has buggy VGPR indexing on GFX9.
1154 sscreen
->llvm_has_working_vgpr_indexing
= sscreen
->info
.chip_class
!= GFX9
;
1156 /* Some chips have RB+ registers, but don't support RB+. Those must
1157 * always disable it.
1159 if (sscreen
->info
.family
== CHIP_STONEY
||
1160 sscreen
->info
.chip_class
>= GFX9
) {
1161 sscreen
->has_rbplus
= true;
1163 sscreen
->rbplus_allowed
=
1164 !(sscreen
->debug_flags
& DBG(NO_RB_PLUS
)) &&
1165 (sscreen
->info
.family
== CHIP_STONEY
||
1166 sscreen
->info
.family
== CHIP_VEGA12
||
1167 sscreen
->info
.family
== CHIP_RAVEN
||
1168 sscreen
->info
.family
== CHIP_RAVEN2
);
1171 sscreen
->dcc_msaa_allowed
=
1172 !(sscreen
->debug_flags
& DBG(NO_DCC_MSAA
));
1174 sscreen
->cpdma_prefetch_writes_memory
= sscreen
->info
.chip_class
<= GFX8
;
1176 (void) mtx_init(&sscreen
->shader_parts_mutex
, mtx_plain
);
1177 sscreen
->use_monolithic_shaders
=
1178 (sscreen
->debug_flags
& DBG(MONOLITHIC_SHADERS
)) != 0;
1180 sscreen
->barrier_flags
.cp_to_L2
= SI_CONTEXT_INV_SCACHE
|
1181 SI_CONTEXT_INV_VCACHE
;
1182 if (sscreen
->info
.chip_class
<= GFX8
) {
1183 sscreen
->barrier_flags
.cp_to_L2
|= SI_CONTEXT_INV_L2
;
1184 sscreen
->barrier_flags
.L2_to_cp
|= SI_CONTEXT_WB_L2
;
1187 if (debug_get_bool_option("RADEON_DUMP_SHADERS", false))
1188 sscreen
->debug_flags
|= DBG_ALL_SHADERS
;
1195 * That means 8 coverage samples, 4 Z/S samples, and 2 color samples.
1197 * s >= z >= c (ignoring this only wastes memory)
1202 * Only MSAA color and depth buffers are overriden.
1204 if (sscreen
->info
.has_eqaa_surface_allocator
) {
1205 const char *eqaa
= debug_get_option("EQAA", NULL
);
1208 if (eqaa
&& sscanf(eqaa
, "%u,%u,%u", &s
, &z
, &f
) == 3 && s
&& z
&& f
) {
1209 sscreen
->eqaa_force_coverage_samples
= s
;
1210 sscreen
->eqaa_force_z_samples
= z
;
1211 sscreen
->eqaa_force_color_samples
= f
;
1215 for (i
= 0; i
< num_comp_hi_threads
; i
++)
1216 si_init_compiler(sscreen
, &sscreen
->compiler
[i
]);
1217 for (i
= 0; i
< num_comp_lo_threads
; i
++)
1218 si_init_compiler(sscreen
, &sscreen
->compiler_lowp
[i
]);
1220 /* Create the auxiliary context. This must be done last. */
1221 sscreen
->aux_context
= si_create_context(
1222 &sscreen
->b
, sscreen
->options
.aux_debug
? PIPE_CONTEXT_DEBUG
: 0);
1223 if (sscreen
->options
.aux_debug
) {
1224 struct u_log_context
*log
= CALLOC_STRUCT(u_log_context
);
1225 u_log_context_init(log
);
1226 sscreen
->aux_context
->set_log_context(sscreen
->aux_context
, log
);
1229 if (sscreen
->debug_flags
& DBG(TEST_DMA
))
1230 si_test_dma(sscreen
);
1232 if (sscreen
->debug_flags
& DBG(TEST_DMA_PERF
)) {
1233 si_test_dma_perf(sscreen
);
1236 if (sscreen
->debug_flags
& (DBG(TEST_VMFAULT_CP
) |
1237 DBG(TEST_VMFAULT_SDMA
) |
1238 DBG(TEST_VMFAULT_SHADER
)))
1239 si_test_vmfault(sscreen
);
1241 if (sscreen
->debug_flags
& DBG(TEST_GDS
))
1242 si_test_gds((struct si_context
*)sscreen
->aux_context
);
1244 if (sscreen
->debug_flags
& DBG(TEST_GDS_MM
)) {
1245 si_test_gds_memory_management((struct si_context
*)sscreen
->aux_context
,
1246 32 * 1024, 4, RADEON_DOMAIN_GDS
);
1248 if (sscreen
->debug_flags
& DBG(TEST_GDS_OA_MM
)) {
1249 si_test_gds_memory_management((struct si_context
*)sscreen
->aux_context
,
1250 4, 1, RADEON_DOMAIN_OA
);
1256 struct pipe_screen
*radeonsi_screen_create(int fd
, const struct pipe_screen_config
*config
)
1258 drmVersionPtr version
= drmGetVersion(fd
);
1259 struct radeon_winsys
*rw
= NULL
;
1261 switch (version
->version_major
) {
1263 rw
= radeon_drm_winsys_create(fd
, config
, radeonsi_screen_create_impl
);
1266 rw
= amdgpu_winsys_create(fd
, config
, radeonsi_screen_create_impl
);
1270 drmFreeVersion(version
);
1271 return rw
? rw
->screen
: NULL
;