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 "radeon/radeon_uvd.h"
33 #include "util/disk_cache.h"
34 #include "util/u_log.h"
35 #include "util/u_memory.h"
36 #include "util/u_suballoc.h"
37 #include "util/u_tests.h"
38 #include "util/u_upload_mgr.h"
39 #include "util/xmlconfig.h"
40 #include "vl/vl_decoder.h"
41 #include "driver_ddebug/dd_util.h"
43 #include "gallium/winsys/radeon/drm/radeon_drm_public.h"
44 #include "gallium/winsys/amdgpu/drm/amdgpu_public.h"
47 static struct pipe_context
*si_create_context(struct pipe_screen
*screen
,
50 static const struct debug_named_value debug_options
[] = {
51 /* Shader logging options: */
52 { "vs", DBG(VS
), "Print vertex shaders" },
53 { "ps", DBG(PS
), "Print pixel shaders" },
54 { "gs", DBG(GS
), "Print geometry shaders" },
55 { "tcs", DBG(TCS
), "Print tessellation control shaders" },
56 { "tes", DBG(TES
), "Print tessellation evaluation shaders" },
57 { "cs", DBG(CS
), "Print compute shaders" },
58 { "noir", DBG(NO_IR
), "Don't print the LLVM IR"},
59 { "nonir", DBG(NO_NIR
), "Don't print NIR when printing shaders"},
60 { "noasm", DBG(NO_ASM
), "Don't print disassembled shaders"},
61 { "preoptir", DBG(PREOPT_IR
), "Print the LLVM IR before initial optimizations" },
63 /* Shader compiler options the shader cache should be aware of: */
64 { "sisched", DBG(SI_SCHED
), "Enable LLVM SI Machine Instruction Scheduler." },
65 { "gisel", DBG(GISEL
), "Enable LLVM global instruction selector." },
66 { "w32ge", DBG(W32_GE
), "Use Wave32 for vertex, tessellation, and geometry shaders." },
67 { "w32ps", DBG(W32_PS
), "Use Wave32 for pixel shaders." },
68 { "w32cs", DBG(W32_CS
), "Use Wave32 for computes shaders." },
69 { "w64ge", DBG(W64_GE
), "Use Wave64 for vertex, tessellation, and geometry shaders." },
70 { "w64ps", DBG(W64_PS
), "Use Wave64 for pixel shaders." },
71 { "w64cs", DBG(W64_CS
), "Use Wave64 for computes shaders." },
73 /* Shader compiler options (with no effect on the shader cache): */
74 { "checkir", DBG(CHECK_IR
), "Enable additional sanity checks on shader IR" },
75 { "mono", DBG(MONOLITHIC_SHADERS
), "Use old-style monolithic shaders compiled on demand" },
76 { "nooptvariant", DBG(NO_OPT_VARIANT
), "Disable compiling optimized shader variants." },
78 /* Information logging options: */
79 { "info", DBG(INFO
), "Print driver information" },
80 { "tex", DBG(TEX
), "Print texture info" },
81 { "compute", DBG(COMPUTE
), "Print compute info" },
82 { "vm", DBG(VM
), "Print virtual addresses when creating resources" },
83 { "cache_stats", DBG(CACHE_STATS
), "Print shader cache statistics." },
86 { "forcedma", DBG(FORCE_SDMA
), "Use SDMA for all operations when possible." },
87 { "nodma", DBG(NO_SDMA
), "Disable SDMA" },
88 { "nodmaclear", DBG(NO_SDMA_CLEARS
), "Disable SDMA clears" },
89 { "nodmacopyimage", DBG(NO_SDMA_COPY_IMAGE
), "Disable SDMA image copies" },
90 { "nowc", DBG(NO_WC
), "Disable GTT write combining" },
91 { "check_vm", DBG(CHECK_VM
), "Check VM faults and dump debug info." },
92 { "reserve_vmid", DBG(RESERVE_VMID
), "Force VMID reservation per context." },
93 { "zerovram", DBG(ZERO_VRAM
), "Clear VRAM allocations." },
95 /* 3D engine options: */
96 { "nogfx", DBG(NO_GFX
), "Disable graphics. Only multimedia compute paths can be used." },
97 { "nongg", DBG(NO_NGG
), "Disable NGG and use the legacy pipeline." },
98 { "nggc", DBG(ALWAYS_NGG_CULLING
), "Always use NGG culling even when it can hurt." },
99 { "nonggc", DBG(NO_NGG_CULLING
), "Disable NGG culling." },
100 { "alwayspd", DBG(ALWAYS_PD
), "Always enable the primitive discard compute shader." },
101 { "pd", DBG(PD
), "Enable the primitive discard compute shader for large draw calls." },
102 { "nopd", DBG(NO_PD
), "Disable the primitive discard compute shader." },
103 { "switch_on_eop", DBG(SWITCH_ON_EOP
), "Program WD/IA to switch on end-of-packet." },
104 { "nooutoforder", DBG(NO_OUT_OF_ORDER
), "Disable out-of-order rasterization" },
105 { "nodpbb", DBG(NO_DPBB
), "Disable DPBB." },
106 { "nodfsm", DBG(NO_DFSM
), "Disable DFSM." },
107 { "dpbb", DBG(DPBB
), "Enable DPBB." },
108 { "dfsm", DBG(DFSM
), "Enable DFSM." },
109 { "nohyperz", DBG(NO_HYPERZ
), "Disable Hyper-Z" },
110 { "norbplus", DBG(NO_RB_PLUS
), "Disable RB+." },
111 { "no2d", DBG(NO_2D_TILING
), "Disable 2D tiling" },
112 { "notiling", DBG(NO_TILING
), "Disable tiling" },
113 { "nodcc", DBG(NO_DCC
), "Disable DCC." },
114 { "nodccclear", DBG(NO_DCC_CLEAR
), "Disable DCC fast clear." },
115 { "nodccfb", DBG(NO_DCC_FB
), "Disable separate DCC on the main framebuffer" },
116 { "nodccmsaa", DBG(NO_DCC_MSAA
), "Disable DCC for MSAA" },
117 { "nofmask", DBG(NO_FMASK
), "Disable MSAA compression" },
119 DEBUG_NAMED_VALUE_END
/* must be last */
122 static const struct debug_named_value test_options
[] = {
124 { "testdma", DBG(TEST_DMA
), "Invoke SDMA tests and exit." },
125 { "testvmfaultcp", DBG(TEST_VMFAULT_CP
), "Invoke a CP VM fault test and exit." },
126 { "testvmfaultsdma", DBG(TEST_VMFAULT_SDMA
), "Invoke a SDMA VM fault test and exit." },
127 { "testvmfaultshader", DBG(TEST_VMFAULT_SHADER
), "Invoke a shader VM fault test and exit." },
128 { "testdmaperf", DBG(TEST_DMA_PERF
), "Test DMA performance" },
129 { "testgds", DBG(TEST_GDS
), "Test GDS." },
130 { "testgdsmm", DBG(TEST_GDS_MM
), "Test GDS memory management." },
131 { "testgdsoamm", DBG(TEST_GDS_OA_MM
), "Test GDS OA memory management." },
133 DEBUG_NAMED_VALUE_END
/* must be last */
136 void si_init_compiler(struct si_screen
*sscreen
, struct ac_llvm_compiler
*compiler
)
138 /* Only create the less-optimizing version of the compiler on APUs
139 * predating Ryzen (Raven). */
140 bool create_low_opt_compiler
= !sscreen
->info
.has_dedicated_vram
&&
141 sscreen
->info
.chip_class
<= GFX8
;
143 enum ac_target_machine_options tm_options
=
144 (sscreen
->debug_flags
& DBG(SI_SCHED
) ? AC_TM_SISCHED
: 0) |
145 (sscreen
->debug_flags
& DBG(GISEL
) ? AC_TM_ENABLE_GLOBAL_ISEL
: 0) |
146 (sscreen
->info
.chip_class
>= GFX9
? AC_TM_FORCE_ENABLE_XNACK
: 0) |
147 (sscreen
->info
.chip_class
< GFX9
? AC_TM_FORCE_DISABLE_XNACK
: 0) |
148 (!sscreen
->llvm_has_working_vgpr_indexing
? AC_TM_PROMOTE_ALLOCA_TO_SCRATCH
: 0) |
149 (sscreen
->debug_flags
& DBG(CHECK_IR
) ? AC_TM_CHECK_IR
: 0) |
150 (create_low_opt_compiler
? AC_TM_CREATE_LOW_OPT
: 0);
153 ac_init_llvm_compiler(compiler
, sscreen
->info
.family
, tm_options
);
154 compiler
->passes
= ac_create_llvm_passes(compiler
->tm
);
156 if (compiler
->tm_wave32
)
157 compiler
->passes_wave32
= ac_create_llvm_passes(compiler
->tm_wave32
);
158 if (compiler
->low_opt_tm
)
159 compiler
->low_opt_passes
= ac_create_llvm_passes(compiler
->low_opt_tm
);
162 static void si_destroy_compiler(struct ac_llvm_compiler
*compiler
)
164 ac_destroy_llvm_compiler(compiler
);
170 static void si_destroy_context(struct pipe_context
*context
)
172 struct si_context
*sctx
= (struct si_context
*)context
;
175 /* Unreference the framebuffer normally to disable related logic
178 struct pipe_framebuffer_state fb
= {};
179 if (context
->set_framebuffer_state
)
180 context
->set_framebuffer_state(context
, &fb
);
182 si_release_all_descriptors(sctx
);
184 if (sctx
->chip_class
>= GFX10
&& sctx
->has_graphics
)
185 gfx10_destroy_query(sctx
);
187 pipe_resource_reference(&sctx
->esgs_ring
, NULL
);
188 pipe_resource_reference(&sctx
->gsvs_ring
, NULL
);
189 pipe_resource_reference(&sctx
->tess_rings
, NULL
);
190 pipe_resource_reference(&sctx
->null_const_buf
.buffer
, NULL
);
191 pipe_resource_reference(&sctx
->sample_pos_buffer
, NULL
);
192 si_resource_reference(&sctx
->border_color_buffer
, NULL
);
193 free(sctx
->border_color_table
);
194 si_resource_reference(&sctx
->scratch_buffer
, NULL
);
195 si_resource_reference(&sctx
->compute_scratch_buffer
, NULL
);
196 si_resource_reference(&sctx
->wait_mem_scratch
, NULL
);
197 si_resource_reference(&sctx
->small_prim_cull_info_buf
, NULL
);
199 si_pm4_free_state(sctx
, sctx
->init_config
, ~0);
200 if (sctx
->init_config_gs_rings
)
201 si_pm4_free_state(sctx
, sctx
->init_config_gs_rings
, ~0);
202 for (i
= 0; i
< ARRAY_SIZE(sctx
->vgt_shader_config
); i
++)
203 si_pm4_delete_state(sctx
, vgt_shader_config
, sctx
->vgt_shader_config
[i
]);
205 if (sctx
->fixed_func_tcs_shader
.cso
)
206 sctx
->b
.delete_tcs_state(&sctx
->b
, sctx
->fixed_func_tcs_shader
.cso
);
207 if (sctx
->custom_dsa_flush
)
208 sctx
->b
.delete_depth_stencil_alpha_state(&sctx
->b
, sctx
->custom_dsa_flush
);
209 if (sctx
->custom_blend_resolve
)
210 sctx
->b
.delete_blend_state(&sctx
->b
, sctx
->custom_blend_resolve
);
211 if (sctx
->custom_blend_fmask_decompress
)
212 sctx
->b
.delete_blend_state(&sctx
->b
, sctx
->custom_blend_fmask_decompress
);
213 if (sctx
->custom_blend_eliminate_fastclear
)
214 sctx
->b
.delete_blend_state(&sctx
->b
, sctx
->custom_blend_eliminate_fastclear
);
215 if (sctx
->custom_blend_dcc_decompress
)
216 sctx
->b
.delete_blend_state(&sctx
->b
, sctx
->custom_blend_dcc_decompress
);
217 if (sctx
->vs_blit_pos
)
218 sctx
->b
.delete_vs_state(&sctx
->b
, sctx
->vs_blit_pos
);
219 if (sctx
->vs_blit_pos_layered
)
220 sctx
->b
.delete_vs_state(&sctx
->b
, sctx
->vs_blit_pos_layered
);
221 if (sctx
->vs_blit_color
)
222 sctx
->b
.delete_vs_state(&sctx
->b
, sctx
->vs_blit_color
);
223 if (sctx
->vs_blit_color_layered
)
224 sctx
->b
.delete_vs_state(&sctx
->b
, sctx
->vs_blit_color_layered
);
225 if (sctx
->vs_blit_texcoord
)
226 sctx
->b
.delete_vs_state(&sctx
->b
, sctx
->vs_blit_texcoord
);
227 if (sctx
->cs_clear_buffer
)
228 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_clear_buffer
);
229 if (sctx
->cs_copy_buffer
)
230 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_copy_buffer
);
231 if (sctx
->cs_copy_image
)
232 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_copy_image
);
233 if (sctx
->cs_copy_image_1d_array
)
234 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_copy_image_1d_array
);
235 if (sctx
->cs_clear_render_target
)
236 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_clear_render_target
);
237 if (sctx
->cs_clear_render_target_1d_array
)
238 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_clear_render_target_1d_array
);
239 if (sctx
->cs_clear_12bytes_buffer
)
240 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_clear_12bytes_buffer
);
241 if (sctx
->cs_dcc_retile
)
242 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->cs_dcc_retile
);
244 for (unsigned i
= 0; i
< ARRAY_SIZE(sctx
->cs_fmask_expand
); i
++) {
245 for (unsigned j
= 0; j
< ARRAY_SIZE(sctx
->cs_fmask_expand
[i
]); j
++) {
246 if (sctx
->cs_fmask_expand
[i
][j
]) {
247 sctx
->b
.delete_compute_state(&sctx
->b
,
248 sctx
->cs_fmask_expand
[i
][j
]);
254 util_blitter_destroy(sctx
->blitter
);
256 /* Release DCC stats. */
257 for (int i
= 0; i
< ARRAY_SIZE(sctx
->dcc_stats
); i
++) {
258 assert(!sctx
->dcc_stats
[i
].query_active
);
260 for (int j
= 0; j
< ARRAY_SIZE(sctx
->dcc_stats
[i
].ps_stats
); j
++)
261 if (sctx
->dcc_stats
[i
].ps_stats
[j
])
262 sctx
->b
.destroy_query(&sctx
->b
,
263 sctx
->dcc_stats
[i
].ps_stats
[j
]);
265 si_texture_reference(&sctx
->dcc_stats
[i
].tex
, NULL
);
268 if (sctx
->query_result_shader
)
269 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->query_result_shader
);
270 if (sctx
->sh_query_result_shader
)
271 sctx
->b
.delete_compute_state(&sctx
->b
, sctx
->sh_query_result_shader
);
274 sctx
->ws
->cs_destroy(sctx
->gfx_cs
);
276 sctx
->ws
->cs_destroy(sctx
->sdma_cs
);
278 sctx
->ws
->ctx_destroy(sctx
->ctx
);
280 if (sctx
->b
.stream_uploader
)
281 u_upload_destroy(sctx
->b
.stream_uploader
);
282 if (sctx
->b
.const_uploader
)
283 u_upload_destroy(sctx
->b
.const_uploader
);
284 if (sctx
->cached_gtt_allocator
)
285 u_upload_destroy(sctx
->cached_gtt_allocator
);
287 slab_destroy_child(&sctx
->pool_transfers
);
288 slab_destroy_child(&sctx
->pool_transfers_unsync
);
290 if (sctx
->allocator_zeroed_memory
)
291 u_suballocator_destroy(sctx
->allocator_zeroed_memory
);
293 sctx
->ws
->fence_reference(&sctx
->last_gfx_fence
, NULL
);
294 sctx
->ws
->fence_reference(&sctx
->last_sdma_fence
, NULL
);
295 sctx
->ws
->fence_reference(&sctx
->last_ib_barrier_fence
, NULL
);
296 si_resource_reference(&sctx
->eop_bug_scratch
, NULL
);
297 si_resource_reference(&sctx
->index_ring
, NULL
);
298 si_resource_reference(&sctx
->barrier_buf
, NULL
);
299 si_resource_reference(&sctx
->last_ib_barrier_buf
, NULL
);
300 pb_reference(&sctx
->gds
, NULL
);
301 pb_reference(&sctx
->gds_oa
, NULL
);
303 si_destroy_compiler(&sctx
->compiler
);
305 si_saved_cs_reference(&sctx
->current_saved_cs
, NULL
);
307 _mesa_hash_table_destroy(sctx
->tex_handles
, NULL
);
308 _mesa_hash_table_destroy(sctx
->img_handles
, NULL
);
310 util_dynarray_fini(&sctx
->resident_tex_handles
);
311 util_dynarray_fini(&sctx
->resident_img_handles
);
312 util_dynarray_fini(&sctx
->resident_tex_needs_color_decompress
);
313 util_dynarray_fini(&sctx
->resident_img_needs_color_decompress
);
314 util_dynarray_fini(&sctx
->resident_tex_needs_depth_decompress
);
315 si_unref_sdma_uploads(sctx
);
316 free(sctx
->sdma_uploads
);
320 static enum pipe_reset_status
si_get_reset_status(struct pipe_context
*ctx
)
322 struct si_context
*sctx
= (struct si_context
*)ctx
;
323 struct si_screen
*sscreen
= sctx
->screen
;
324 enum pipe_reset_status status
= sctx
->ws
->ctx_query_reset_status(sctx
->ctx
);
326 if (status
!= PIPE_NO_RESET
) {
327 /* Call the state tracker to set a no-op API dispatch. */
328 if (sctx
->device_reset_callback
.reset
) {
329 sctx
->device_reset_callback
.reset(sctx
->device_reset_callback
.data
,
333 /* Re-create the auxiliary context, because it won't submit
334 * any new IBs due to a GPU reset.
336 simple_mtx_lock(&sscreen
->aux_context_lock
);
338 struct u_log_context
*aux_log
= ((struct si_context
*)sscreen
->aux_context
)->log
;
339 sscreen
->aux_context
->set_log_context(sscreen
->aux_context
, NULL
);
340 sscreen
->aux_context
->destroy(sscreen
->aux_context
);
342 sscreen
->aux_context
= si_create_context(&sscreen
->b
,
343 (sscreen
->options
.aux_debug
? PIPE_CONTEXT_DEBUG
: 0) |
344 (sscreen
->info
.has_graphics
? 0 : PIPE_CONTEXT_COMPUTE_ONLY
));
345 sscreen
->aux_context
->set_log_context(sscreen
->aux_context
, aux_log
);
346 simple_mtx_unlock(&sscreen
->aux_context_lock
);
351 static void si_set_device_reset_callback(struct pipe_context
*ctx
,
352 const struct pipe_device_reset_callback
*cb
)
354 struct si_context
*sctx
= (struct si_context
*)ctx
;
357 sctx
->device_reset_callback
= *cb
;
359 memset(&sctx
->device_reset_callback
, 0,
360 sizeof(sctx
->device_reset_callback
));
363 /* Apitrace profiling:
364 * 1) qapitrace : Tools -> Profile: Measure CPU & GPU times
365 * 2) In the middle panel, zoom in (mouse wheel) on some bad draw call
366 * and remember its number.
367 * 3) In Mesa, enable queries and performance counters around that draw
368 * call and print the results.
369 * 4) glretrace --benchmark --markers ..
371 static void si_emit_string_marker(struct pipe_context
*ctx
,
372 const char *string
, int len
)
374 struct si_context
*sctx
= (struct si_context
*)ctx
;
376 dd_parse_apitrace_marker(string
, len
, &sctx
->apitrace_call_number
);
379 u_log_printf(sctx
->log
, "\nString marker: %*s\n", len
, string
);
382 static void si_set_debug_callback(struct pipe_context
*ctx
,
383 const struct pipe_debug_callback
*cb
)
385 struct si_context
*sctx
= (struct si_context
*)ctx
;
386 struct si_screen
*screen
= sctx
->screen
;
388 util_queue_finish(&screen
->shader_compiler_queue
);
389 util_queue_finish(&screen
->shader_compiler_queue_low_priority
);
394 memset(&sctx
->debug
, 0, sizeof(sctx
->debug
));
397 static void si_set_log_context(struct pipe_context
*ctx
,
398 struct u_log_context
*log
)
400 struct si_context
*sctx
= (struct si_context
*)ctx
;
404 u_log_add_auto_logger(log
, si_auto_log_cs
, sctx
);
407 static void si_set_context_param(struct pipe_context
*ctx
,
408 enum pipe_context_param param
,
411 struct radeon_winsys
*ws
= ((struct si_context
*)ctx
)->ws
;
414 case PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE
:
415 ws
->pin_threads_to_L3_cache(ws
, value
);
421 static struct pipe_context
*si_create_context(struct pipe_screen
*screen
,
424 struct si_screen
* sscreen
= (struct si_screen
*)screen
;
425 STATIC_ASSERT(DBG_COUNT
<= 64);
427 /* Don't create a context if it's not compute-only and hw is compute-only. */
428 if (!sscreen
->info
.has_graphics
&&
429 !(flags
& PIPE_CONTEXT_COMPUTE_ONLY
))
432 struct si_context
*sctx
= CALLOC_STRUCT(si_context
);
433 struct radeon_winsys
*ws
= sscreen
->ws
;
435 bool stop_exec_on_failure
= (flags
& PIPE_CONTEXT_LOSE_CONTEXT_ON_RESET
) != 0;
440 sctx
->has_graphics
= sscreen
->info
.chip_class
== GFX6
||
441 !(flags
& PIPE_CONTEXT_COMPUTE_ONLY
);
443 if (flags
& PIPE_CONTEXT_DEBUG
)
444 sscreen
->record_llvm_ir
= true; /* racy but not critical */
446 sctx
->b
.screen
= screen
; /* this must be set first */
448 sctx
->b
.destroy
= si_destroy_context
;
449 sctx
->screen
= sscreen
; /* Easy accessing of screen/winsys. */
450 sctx
->is_debug
= (flags
& PIPE_CONTEXT_DEBUG
) != 0;
452 slab_create_child(&sctx
->pool_transfers
, &sscreen
->pool_transfers
);
453 slab_create_child(&sctx
->pool_transfers_unsync
, &sscreen
->pool_transfers
);
455 sctx
->ws
= sscreen
->ws
;
456 sctx
->family
= sscreen
->info
.family
;
457 sctx
->chip_class
= sscreen
->info
.chip_class
;
459 if (sctx
->chip_class
== GFX7
||
460 sctx
->chip_class
== GFX8
||
461 sctx
->chip_class
== GFX9
) {
462 sctx
->eop_bug_scratch
= si_resource(
463 pipe_buffer_create(&sscreen
->b
, 0, PIPE_USAGE_DEFAULT
,
464 16 * sscreen
->info
.num_render_backends
));
465 if (!sctx
->eop_bug_scratch
)
469 /* Initialize context allocators. */
470 sctx
->allocator_zeroed_memory
=
471 u_suballocator_create(&sctx
->b
, 128 * 1024,
472 0, PIPE_USAGE_DEFAULT
,
473 SI_RESOURCE_FLAG_UNMAPPABLE
|
474 SI_RESOURCE_FLAG_CLEAR
, false);
475 if (!sctx
->allocator_zeroed_memory
)
478 sctx
->b
.stream_uploader
= u_upload_create(&sctx
->b
, 1024 * 1024,
479 0, PIPE_USAGE_STREAM
,
480 SI_RESOURCE_FLAG_READ_ONLY
);
481 if (!sctx
->b
.stream_uploader
)
484 sctx
->cached_gtt_allocator
= u_upload_create(&sctx
->b
, 16 * 1024,
485 0, PIPE_USAGE_STAGING
, 0);
486 if (!sctx
->cached_gtt_allocator
)
489 sctx
->ctx
= sctx
->ws
->ctx_create(sctx
->ws
);
493 if (sscreen
->info
.num_rings
[RING_DMA
] &&
494 !(sscreen
->debug_flags
& DBG(NO_SDMA
)) &&
495 /* SDMA causes corruption on RX 580:
496 * https://gitlab.freedesktop.org/mesa/mesa/issues/1399
497 * https://gitlab.freedesktop.org/mesa/mesa/issues/1889
499 (sctx
->chip_class
!= GFX8
|| sscreen
->debug_flags
& DBG(FORCE_SDMA
)) &&
500 /* SDMA timeouts sometimes on gfx10 so disable it for now. See:
501 * https://bugs.freedesktop.org/show_bug.cgi?id=111481
502 * https://gitlab.freedesktop.org/mesa/mesa/issues/1907
504 (sctx
->chip_class
!= GFX10
|| sscreen
->debug_flags
& DBG(FORCE_SDMA
))) {
505 sctx
->sdma_cs
= sctx
->ws
->cs_create(sctx
->ctx
, RING_DMA
,
506 (void*)si_flush_dma_cs
,
507 sctx
, stop_exec_on_failure
);
510 bool use_sdma_upload
= sscreen
->info
.has_dedicated_vram
&& sctx
->sdma_cs
;
511 sctx
->b
.const_uploader
= u_upload_create(&sctx
->b
, 256 * 1024,
512 0, PIPE_USAGE_DEFAULT
,
513 SI_RESOURCE_FLAG_32BIT
|
515 SI_RESOURCE_FLAG_UPLOAD_FLUSH_EXPLICIT_VIA_SDMA
: 0));
516 if (!sctx
->b
.const_uploader
)
520 u_upload_enable_flush_explicit(sctx
->b
.const_uploader
);
522 sctx
->gfx_cs
= ws
->cs_create(sctx
->ctx
,
523 sctx
->has_graphics
? RING_GFX
: RING_COMPUTE
,
524 (void*)si_flush_gfx_cs
, sctx
, stop_exec_on_failure
);
527 sctx
->border_color_table
= malloc(SI_MAX_BORDER_COLORS
*
528 sizeof(*sctx
->border_color_table
));
529 if (!sctx
->border_color_table
)
532 sctx
->border_color_buffer
= si_resource(
533 pipe_buffer_create(screen
, 0, PIPE_USAGE_DEFAULT
,
534 SI_MAX_BORDER_COLORS
*
535 sizeof(*sctx
->border_color_table
)));
536 if (!sctx
->border_color_buffer
)
539 sctx
->border_color_map
=
540 ws
->buffer_map(sctx
->border_color_buffer
->buf
,
541 NULL
, PIPE_TRANSFER_WRITE
);
542 if (!sctx
->border_color_map
)
545 sctx
->ngg
= sscreen
->use_ngg
;
547 /* Initialize context functions used by graphics and compute. */
548 if (sctx
->chip_class
>= GFX10
)
549 sctx
->emit_cache_flush
= gfx10_emit_cache_flush
;
551 sctx
->emit_cache_flush
= si_emit_cache_flush
;
553 sctx
->b
.emit_string_marker
= si_emit_string_marker
;
554 sctx
->b
.set_debug_callback
= si_set_debug_callback
;
555 sctx
->b
.set_log_context
= si_set_log_context
;
556 sctx
->b
.set_context_param
= si_set_context_param
;
557 sctx
->b
.get_device_reset_status
= si_get_reset_status
;
558 sctx
->b
.set_device_reset_callback
= si_set_device_reset_callback
;
560 si_init_all_descriptors(sctx
);
561 si_init_buffer_functions(sctx
);
562 si_init_clear_functions(sctx
);
563 si_init_blit_functions(sctx
);
564 si_init_compute_functions(sctx
);
565 si_init_compute_blit_functions(sctx
);
566 si_init_debug_functions(sctx
);
567 si_init_fence_functions(sctx
);
568 si_init_query_functions(sctx
);
569 si_init_state_compute_functions(sctx
);
570 si_init_context_texture_functions(sctx
);
572 /* Initialize graphics-only context functions. */
573 if (sctx
->has_graphics
) {
574 if (sctx
->chip_class
>= GFX10
)
575 gfx10_init_query(sctx
);
576 si_init_msaa_functions(sctx
);
577 si_init_shader_functions(sctx
);
578 si_init_state_functions(sctx
);
579 si_init_streamout_functions(sctx
);
580 si_init_viewport_functions(sctx
);
582 sctx
->blitter
= util_blitter_create(&sctx
->b
);
583 if (sctx
->blitter
== NULL
)
585 sctx
->blitter
->skip_viewport_restore
= true;
587 /* Some states are expected to be always non-NULL. */
588 sctx
->noop_blend
= util_blitter_get_noop_blend_state(sctx
->blitter
);
589 sctx
->queued
.named
.blend
= sctx
->noop_blend
;
591 sctx
->noop_dsa
= util_blitter_get_noop_dsa_state(sctx
->blitter
);
592 sctx
->queued
.named
.dsa
= sctx
->noop_dsa
;
594 sctx
->discard_rasterizer_state
=
595 util_blitter_get_discard_rasterizer_state(sctx
->blitter
);
596 sctx
->queued
.named
.rasterizer
= sctx
->discard_rasterizer_state
;
598 si_init_draw_functions(sctx
);
600 /* If aux_context == NULL, we are initializing aux_context right now. */
601 bool is_aux_context
= !sscreen
->aux_context
;
602 si_initialize_prim_discard_tunables(sscreen
, is_aux_context
,
603 &sctx
->prim_discard_vertex_count_threshold
,
604 &sctx
->index_ring_size_per_ib
);
607 /* Initialize SDMA functions. */
608 if (sctx
->chip_class
>= GFX7
)
609 cik_init_sdma_functions(sctx
);
611 sctx
->dma_copy
= si_resource_copy_region
;
613 if (sscreen
->debug_flags
& DBG(FORCE_SDMA
))
614 sctx
->b
.resource_copy_region
= sctx
->dma_copy
;
616 sctx
->sample_mask
= 0xffff;
618 /* Initialize multimedia functions. */
619 if (sscreen
->info
.has_hw_decode
) {
620 sctx
->b
.create_video_codec
= si_uvd_create_decoder
;
621 sctx
->b
.create_video_buffer
= si_video_buffer_create
;
623 sctx
->b
.create_video_codec
= vl_create_decoder
;
624 sctx
->b
.create_video_buffer
= vl_video_buffer_create
;
627 if (sctx
->chip_class
>= GFX9
) {
628 sctx
->wait_mem_scratch
= si_resource(
629 pipe_buffer_create(screen
, 0, PIPE_USAGE_DEFAULT
, 8));
630 if (!sctx
->wait_mem_scratch
)
633 /* Initialize the memory. */
634 si_cp_write_data(sctx
, sctx
->wait_mem_scratch
, 0, 4,
635 V_370_MEM
, V_370_ME
, &sctx
->wait_mem_number
);
638 /* GFX7 cannot unbind a constant buffer (S_BUFFER_LOAD doesn't skip loads
639 * if NUM_RECORDS == 0). We need to use a dummy buffer instead. */
640 if (sctx
->chip_class
== GFX7
) {
641 sctx
->null_const_buf
.buffer
=
642 pipe_aligned_buffer_create(screen
,
643 SI_RESOURCE_FLAG_32BIT
,
644 PIPE_USAGE_DEFAULT
, 16,
645 sctx
->screen
->info
.tcc_cache_line_size
);
646 if (!sctx
->null_const_buf
.buffer
)
648 sctx
->null_const_buf
.buffer_size
= sctx
->null_const_buf
.buffer
->width0
;
650 unsigned start_shader
= sctx
->has_graphics
? 0 : PIPE_SHADER_COMPUTE
;
651 for (shader
= start_shader
; shader
< SI_NUM_SHADERS
; shader
++) {
652 for (i
= 0; i
< SI_NUM_CONST_BUFFERS
; i
++) {
653 sctx
->b
.set_constant_buffer(&sctx
->b
, shader
, i
,
654 &sctx
->null_const_buf
);
658 si_set_rw_buffer(sctx
, SI_HS_CONST_DEFAULT_TESS_LEVELS
,
659 &sctx
->null_const_buf
);
660 si_set_rw_buffer(sctx
, SI_VS_CONST_INSTANCE_DIVISORS
,
661 &sctx
->null_const_buf
);
662 si_set_rw_buffer(sctx
, SI_VS_CONST_CLIP_PLANES
,
663 &sctx
->null_const_buf
);
664 si_set_rw_buffer(sctx
, SI_PS_CONST_POLY_STIPPLE
,
665 &sctx
->null_const_buf
);
666 si_set_rw_buffer(sctx
, SI_PS_CONST_SAMPLE_POSITIONS
,
667 &sctx
->null_const_buf
);
670 uint64_t max_threads_per_block
;
671 screen
->get_compute_param(screen
, PIPE_SHADER_IR_NIR
,
672 PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK
,
673 &max_threads_per_block
);
675 /* The maximum number of scratch waves. Scratch space isn't divided
676 * evenly between CUs. The number is only a function of the number of CUs.
677 * We can decrease the constant to decrease the scratch buffer size.
679 * sctx->scratch_waves must be >= the maximum posible size of
680 * 1 threadgroup, so that the hw doesn't hang from being unable
683 * The recommended value is 4 per CU at most. Higher numbers don't
684 * bring much benefit, but they still occupy chip resources (think
685 * async compute). I've seen ~2% performance difference between 4 and 32.
687 sctx
->scratch_waves
= MAX2(32 * sscreen
->info
.num_good_compute_units
,
688 max_threads_per_block
/ 64);
690 /* Bindless handles. */
691 sctx
->tex_handles
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
692 _mesa_key_pointer_equal
);
693 sctx
->img_handles
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
694 _mesa_key_pointer_equal
);
696 util_dynarray_init(&sctx
->resident_tex_handles
, NULL
);
697 util_dynarray_init(&sctx
->resident_img_handles
, NULL
);
698 util_dynarray_init(&sctx
->resident_tex_needs_color_decompress
, NULL
);
699 util_dynarray_init(&sctx
->resident_img_needs_color_decompress
, NULL
);
700 util_dynarray_init(&sctx
->resident_tex_needs_depth_decompress
, NULL
);
702 sctx
->sample_pos_buffer
=
703 pipe_buffer_create(sctx
->b
.screen
, 0, PIPE_USAGE_DEFAULT
,
704 sizeof(sctx
->sample_positions
));
705 pipe_buffer_write(&sctx
->b
, sctx
->sample_pos_buffer
, 0,
706 sizeof(sctx
->sample_positions
), &sctx
->sample_positions
);
708 /* this must be last */
709 si_begin_new_gfx_cs(sctx
);
711 if (sctx
->chip_class
== GFX7
) {
712 /* Clear the NULL constant buffer, because loads should return zeros.
713 * Note that this forces CP DMA to be used, because clover deadlocks
714 * for some reason when the compute codepath is used.
716 uint32_t clear_value
= 0;
717 si_clear_buffer(sctx
, sctx
->null_const_buf
.buffer
, 0,
718 sctx
->null_const_buf
.buffer
->width0
,
719 &clear_value
, 4, SI_COHERENCY_SHADER
, true);
723 fprintf(stderr
, "radeonsi: Failed to create a context.\n");
724 si_destroy_context(&sctx
->b
);
728 static struct pipe_context
*si_pipe_create_context(struct pipe_screen
*screen
,
729 void *priv
, unsigned flags
)
731 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
732 struct pipe_context
*ctx
;
734 if (sscreen
->debug_flags
& DBG(CHECK_VM
))
735 flags
|= PIPE_CONTEXT_DEBUG
;
737 ctx
= si_create_context(screen
, flags
);
739 if (!(flags
& PIPE_CONTEXT_PREFER_THREADED
))
742 /* Clover (compute-only) is unsupported. */
743 if (flags
& PIPE_CONTEXT_COMPUTE_ONLY
)
746 /* When shaders are logged to stderr, asynchronous compilation is
748 if (sscreen
->debug_flags
& DBG_ALL_SHADERS
)
751 /* Use asynchronous flushes only on amdgpu, since the radeon
752 * implementation for fence_server_sync is incomplete. */
753 return threaded_context_create(ctx
, &sscreen
->pool_transfers
,
754 si_replace_buffer_storage
,
755 sscreen
->info
.is_amdgpu
? si_create_fence
: NULL
,
756 &((struct si_context
*)ctx
)->tc
);
762 static void si_destroy_screen(struct pipe_screen
* pscreen
)
764 struct si_screen
*sscreen
= (struct si_screen
*)pscreen
;
765 struct si_shader_part
*parts
[] = {
767 sscreen
->tcs_epilogs
,
774 if (!sscreen
->ws
->unref(sscreen
->ws
))
777 if (sscreen
->debug_flags
& DBG(CACHE_STATS
)) {
778 printf("live shader cache: hits = %u, misses = %u\n",
779 sscreen
->live_shader_cache
.hits
,
780 sscreen
->live_shader_cache
.misses
);
781 printf("memory shader cache: hits = %u, misses = %u\n",
782 sscreen
->num_memory_shader_cache_hits
,
783 sscreen
->num_memory_shader_cache_misses
);
784 printf("disk shader cache: hits = %u, misses = %u\n",
785 sscreen
->num_disk_shader_cache_hits
,
786 sscreen
->num_disk_shader_cache_misses
);
789 simple_mtx_destroy(&sscreen
->aux_context_lock
);
791 struct u_log_context
*aux_log
= ((struct si_context
*)sscreen
->aux_context
)->log
;
793 sscreen
->aux_context
->set_log_context(sscreen
->aux_context
, NULL
);
794 u_log_context_destroy(aux_log
);
798 sscreen
->aux_context
->destroy(sscreen
->aux_context
);
800 util_queue_destroy(&sscreen
->shader_compiler_queue
);
801 util_queue_destroy(&sscreen
->shader_compiler_queue_low_priority
);
803 /* Release the reference on glsl types of the compiler threads. */
804 glsl_type_singleton_decref();
806 for (i
= 0; i
< ARRAY_SIZE(sscreen
->compiler
); i
++)
807 si_destroy_compiler(&sscreen
->compiler
[i
]);
809 for (i
= 0; i
< ARRAY_SIZE(sscreen
->compiler_lowp
); i
++)
810 si_destroy_compiler(&sscreen
->compiler_lowp
[i
]);
812 /* Free shader parts. */
813 for (i
= 0; i
< ARRAY_SIZE(parts
); i
++) {
815 struct si_shader_part
*part
= parts
[i
];
817 parts
[i
] = part
->next
;
818 si_shader_binary_clean(&part
->binary
);
822 simple_mtx_destroy(&sscreen
->shader_parts_mutex
);
823 si_destroy_shader_cache(sscreen
);
825 si_destroy_perfcounters(sscreen
);
826 si_gpu_load_kill_thread(sscreen
);
828 simple_mtx_destroy(&sscreen
->gpu_load_mutex
);
830 slab_destroy_parent(&sscreen
->pool_transfers
);
832 disk_cache_destroy(sscreen
->disk_shader_cache
);
833 util_live_shader_cache_deinit(&sscreen
->live_shader_cache
);
834 sscreen
->ws
->destroy(sscreen
->ws
);
838 static void si_init_gs_info(struct si_screen
*sscreen
)
840 sscreen
->gs_table_depth
= ac_get_gs_table_depth(sscreen
->info
.chip_class
,
841 sscreen
->info
.family
);
844 static void si_test_vmfault(struct si_screen
*sscreen
, uint64_t test_flags
)
846 struct pipe_context
*ctx
= sscreen
->aux_context
;
847 struct si_context
*sctx
= (struct si_context
*)ctx
;
848 struct pipe_resource
*buf
=
849 pipe_buffer_create_const0(&sscreen
->b
, 0, PIPE_USAGE_DEFAULT
, 64);
852 puts("Buffer allocation failed.");
856 si_resource(buf
)->gpu_address
= 0; /* cause a VM fault */
858 if (test_flags
& DBG(TEST_VMFAULT_CP
)) {
859 si_cp_dma_copy_buffer(sctx
, buf
, buf
, 0, 4, 4, 0,
860 SI_COHERENCY_NONE
, L2_BYPASS
);
861 ctx
->flush(ctx
, NULL
, 0);
862 puts("VM fault test: CP - done.");
864 if (test_flags
& DBG(TEST_VMFAULT_SDMA
)) {
865 si_sdma_clear_buffer(sctx
, buf
, 0, 4, 0);
866 ctx
->flush(ctx
, NULL
, 0);
867 puts("VM fault test: SDMA - done.");
869 if (test_flags
& DBG(TEST_VMFAULT_SHADER
)) {
870 util_test_constant_buffer(ctx
, buf
);
871 puts("VM fault test: Shader - done.");
876 static void si_test_gds_memory_management(struct si_context
*sctx
,
877 unsigned alloc_size
, unsigned alignment
,
878 enum radeon_bo_domain domain
)
880 struct radeon_winsys
*ws
= sctx
->ws
;
881 struct radeon_cmdbuf
*cs
[8];
882 struct pb_buffer
*gds_bo
[ARRAY_SIZE(cs
)];
884 for (unsigned i
= 0; i
< ARRAY_SIZE(cs
); i
++) {
885 cs
[i
] = ws
->cs_create(sctx
->ctx
, RING_COMPUTE
,
887 gds_bo
[i
] = ws
->buffer_create(ws
, alloc_size
, alignment
, domain
, 0);
891 for (unsigned iterations
= 0; iterations
< 20000; iterations
++) {
892 for (unsigned i
= 0; i
< ARRAY_SIZE(cs
); i
++) {
893 /* This clears GDS with CP DMA.
895 * We don't care if GDS is present. Just add some packet
896 * to make the GPU busy for a moment.
898 si_cp_dma_clear_buffer(sctx
, cs
[i
], NULL
, 0, alloc_size
, 0,
899 SI_CPDMA_SKIP_BO_LIST_UPDATE
|
900 SI_CPDMA_SKIP_CHECK_CS_SPACE
|
901 SI_CPDMA_SKIP_GFX_SYNC
, 0, 0);
903 ws
->cs_add_buffer(cs
[i
], gds_bo
[i
], domain
,
904 RADEON_USAGE_READWRITE
, 0);
905 ws
->cs_flush(cs
[i
], PIPE_FLUSH_ASYNC
, NULL
);
911 static void si_disk_cache_create(struct si_screen
*sscreen
)
913 /* Don't use the cache if shader dumping is enabled. */
914 if (sscreen
->debug_flags
& DBG_ALL_SHADERS
)
917 struct mesa_sha1 ctx
;
918 unsigned char sha1
[20];
919 char cache_id
[20 * 2 + 1];
921 _mesa_sha1_init(&ctx
);
923 if (!disk_cache_get_function_identifier(si_disk_cache_create
, &ctx
) ||
924 !disk_cache_get_function_identifier(LLVMInitializeAMDGPUTargetInfo
,
928 _mesa_sha1_final(&ctx
, sha1
);
929 disk_cache_format_hex_id(cache_id
, sha1
, 20 * 2);
931 /* These flags affect shader compilation. */
932 #define ALL_FLAGS (DBG(SI_SCHED) | DBG(GISEL))
933 uint64_t shader_debug_flags
= sscreen
->debug_flags
& ALL_FLAGS
;
935 /* Add the high bits of 32-bit addresses, which affects
936 * how 32-bit addresses are expanded to 64 bits.
938 STATIC_ASSERT(ALL_FLAGS
<= UINT_MAX
);
939 assert((int16_t)sscreen
->info
.address32_hi
== (int32_t)sscreen
->info
.address32_hi
);
940 shader_debug_flags
|= (uint64_t)(sscreen
->info
.address32_hi
& 0xffff) << 32;
942 sscreen
->disk_shader_cache
=
943 disk_cache_create(sscreen
->info
.name
,
948 static void si_set_max_shader_compiler_threads(struct pipe_screen
*screen
,
949 unsigned max_threads
)
951 struct si_screen
*sscreen
= (struct si_screen
*)screen
;
953 /* This function doesn't allow a greater number of threads than
954 * the queue had at its creation. */
955 util_queue_adjust_num_threads(&sscreen
->shader_compiler_queue
,
957 /* Don't change the number of threads on the low priority queue. */
960 static bool si_is_parallel_shader_compilation_finished(struct pipe_screen
*screen
,
962 enum pipe_shader_type shader_type
)
964 struct si_shader_selector
*sel
= (struct si_shader_selector
*)shader
;
966 return util_queue_fence_is_signalled(&sel
->ready
);
969 static struct pipe_screen
*
970 radeonsi_screen_create_impl(struct radeon_winsys
*ws
,
971 const struct pipe_screen_config
*config
)
973 struct si_screen
*sscreen
= CALLOC_STRUCT(si_screen
);
974 unsigned hw_threads
, num_comp_hi_threads
, num_comp_lo_threads
;
982 ws
->query_info(ws
, &sscreen
->info
);
984 if (sscreen
->info
.chip_class
== GFX10
&& LLVM_VERSION_MAJOR
< 9) {
985 fprintf(stderr
, "radeonsi: Navi family support requires LLVM 9 or higher\n");
990 if (sscreen
->info
.chip_class
>= GFX9
) {
991 sscreen
->se_tile_repeat
= 32 * sscreen
->info
.max_se
;
993 ac_get_raster_config(&sscreen
->info
,
994 &sscreen
->pa_sc_raster_config
,
995 &sscreen
->pa_sc_raster_config_1
,
996 &sscreen
->se_tile_repeat
);
999 sscreen
->debug_flags
= debug_get_flags_option("R600_DEBUG",
1001 sscreen
->debug_flags
|= debug_get_flags_option("AMD_DEBUG",
1003 test_flags
= debug_get_flags_option("AMD_TEST",
1006 if (sscreen
->debug_flags
& DBG(NO_GFX
))
1007 sscreen
->info
.has_graphics
= false;
1009 /* Set functions first. */
1010 sscreen
->b
.context_create
= si_pipe_create_context
;
1011 sscreen
->b
.destroy
= si_destroy_screen
;
1012 sscreen
->b
.set_max_shader_compiler_threads
=
1013 si_set_max_shader_compiler_threads
;
1014 sscreen
->b
.is_parallel_shader_compilation_finished
=
1015 si_is_parallel_shader_compilation_finished
;
1016 sscreen
->b
.finalize_nir
= si_finalize_nir
;
1018 si_init_screen_get_functions(sscreen
);
1019 si_init_screen_buffer_functions(sscreen
);
1020 si_init_screen_fence_functions(sscreen
);
1021 si_init_screen_state_functions(sscreen
);
1022 si_init_screen_texture_functions(sscreen
);
1023 si_init_screen_query_functions(sscreen
);
1024 si_init_screen_live_shader_cache(sscreen
);
1026 /* Set these flags in debug_flags early, so that the shader cache takes
1027 * them into account.
1029 if (driQueryOptionb(config
->options
,
1030 "glsl_correct_derivatives_after_discard"))
1031 sscreen
->debug_flags
|= DBG(FS_CORRECT_DERIVS_AFTER_KILL
);
1032 if (driQueryOptionb(config
->options
, "radeonsi_enable_sisched"))
1033 sscreen
->debug_flags
|= DBG(SI_SCHED
);
1035 if (sscreen
->debug_flags
& DBG(INFO
))
1036 ac_print_gpu_info(&sscreen
->info
);
1038 slab_create_parent(&sscreen
->pool_transfers
,
1039 sizeof(struct si_transfer
), 64);
1041 sscreen
->force_aniso
= MIN2(16, debug_get_num_option("R600_TEX_ANISO", -1));
1042 if (sscreen
->force_aniso
== -1) {
1043 sscreen
->force_aniso
= MIN2(16, debug_get_num_option("AMD_TEX_ANISO", -1));
1046 if (sscreen
->force_aniso
>= 0) {
1047 printf("radeonsi: Forcing anisotropy filter to %ix\n",
1048 /* round down to a power of two */
1049 1 << util_logbase2(sscreen
->force_aniso
));
1052 (void) simple_mtx_init(&sscreen
->aux_context_lock
, mtx_plain
);
1053 (void) simple_mtx_init(&sscreen
->gpu_load_mutex
, mtx_plain
);
1055 si_init_gs_info(sscreen
);
1056 if (!si_init_shader_cache(sscreen
)) {
1062 #define OPT_BOOL(name, dflt, description) \
1063 sscreen->options.name = \
1064 driQueryOptionb(config->options, "radeonsi_"#name);
1065 #include "si_debug_options.h"
1068 si_disk_cache_create(sscreen
);
1070 /* Determine the number of shader compiler threads. */
1071 hw_threads
= sysconf(_SC_NPROCESSORS_ONLN
);
1073 if (hw_threads
>= 12) {
1074 num_comp_hi_threads
= hw_threads
* 3 / 4;
1075 num_comp_lo_threads
= hw_threads
/ 3;
1076 } else if (hw_threads
>= 6) {
1077 num_comp_hi_threads
= hw_threads
- 2;
1078 num_comp_lo_threads
= hw_threads
/ 2;
1079 } else if (hw_threads
>= 2) {
1080 num_comp_hi_threads
= hw_threads
- 1;
1081 num_comp_lo_threads
= hw_threads
/ 2;
1083 num_comp_hi_threads
= 1;
1084 num_comp_lo_threads
= 1;
1087 num_comp_hi_threads
= MIN2(num_comp_hi_threads
,
1088 ARRAY_SIZE(sscreen
->compiler
));
1089 num_comp_lo_threads
= MIN2(num_comp_lo_threads
,
1090 ARRAY_SIZE(sscreen
->compiler_lowp
));
1092 /* Take a reference on the glsl types for the compiler threads. */
1093 glsl_type_singleton_init_or_ref();
1095 if (!util_queue_init(&sscreen
->shader_compiler_queue
, "sh",
1096 64, num_comp_hi_threads
,
1097 UTIL_QUEUE_INIT_RESIZE_IF_FULL
|
1098 UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY
)) {
1099 si_destroy_shader_cache(sscreen
);
1101 glsl_type_singleton_decref();
1105 if (!util_queue_init(&sscreen
->shader_compiler_queue_low_priority
,
1107 64, num_comp_lo_threads
,
1108 UTIL_QUEUE_INIT_RESIZE_IF_FULL
|
1109 UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY
|
1110 UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY
)) {
1111 si_destroy_shader_cache(sscreen
);
1113 glsl_type_singleton_decref();
1117 if (!debug_get_bool_option("RADEON_DISABLE_PERFCOUNTERS", false))
1118 si_init_perfcounters(sscreen
);
1120 unsigned prim_discard_vertex_count_threshold
, tmp
;
1121 si_initialize_prim_discard_tunables(sscreen
, false,
1122 &prim_discard_vertex_count_threshold
,
1124 /* Compute-shader-based culling doesn't support VBOs in user SGPRs. */
1125 if (prim_discard_vertex_count_threshold
== UINT_MAX
)
1126 sscreen
->num_vbos_in_user_sgprs
= sscreen
->info
.chip_class
>= GFX9
? 5 : 1;
1128 /* Determine tessellation ring info. */
1129 bool double_offchip_buffers
= sscreen
->info
.chip_class
>= GFX7
&&
1130 sscreen
->info
.family
!= CHIP_CARRIZO
&&
1131 sscreen
->info
.family
!= CHIP_STONEY
;
1132 /* This must be one less than the maximum number due to a hw limitation.
1133 * Various hardware bugs need this.
1135 unsigned max_offchip_buffers_per_se
;
1137 if (sscreen
->info
.chip_class
>= GFX10
)
1138 max_offchip_buffers_per_se
= 256;
1139 /* Only certain chips can use the maximum value. */
1140 else if (sscreen
->info
.family
== CHIP_VEGA12
||
1141 sscreen
->info
.family
== CHIP_VEGA20
)
1142 max_offchip_buffers_per_se
= double_offchip_buffers
? 128 : 64;
1144 max_offchip_buffers_per_se
= double_offchip_buffers
? 127 : 63;
1146 unsigned max_offchip_buffers
= max_offchip_buffers_per_se
*
1147 sscreen
->info
.max_se
;
1148 unsigned offchip_granularity
;
1150 /* Hawaii has a bug with offchip buffers > 256 that can be worked
1151 * around by setting 4K granularity.
1153 if (sscreen
->info
.family
== CHIP_HAWAII
) {
1154 sscreen
->tess_offchip_block_dw_size
= 4096;
1155 offchip_granularity
= V_03093C_X_4K_DWORDS
;
1157 sscreen
->tess_offchip_block_dw_size
= 8192;
1158 offchip_granularity
= V_03093C_X_8K_DWORDS
;
1161 sscreen
->tess_factor_ring_size
= 32768 * sscreen
->info
.max_se
;
1162 sscreen
->tess_offchip_ring_size
= max_offchip_buffers
*
1163 sscreen
->tess_offchip_block_dw_size
* 4;
1165 if (sscreen
->info
.chip_class
>= GFX7
) {
1166 if (sscreen
->info
.chip_class
>= GFX8
)
1167 --max_offchip_buffers
;
1168 sscreen
->vgt_hs_offchip_param
=
1169 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers
) |
1170 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity
);
1172 assert(offchip_granularity
== V_03093C_X_8K_DWORDS
);
1173 sscreen
->vgt_hs_offchip_param
=
1174 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers
);
1177 sscreen
->has_draw_indirect_multi
=
1178 (sscreen
->info
.family
>= CHIP_POLARIS10
) ||
1179 (sscreen
->info
.chip_class
== GFX8
&&
1180 sscreen
->info
.pfp_fw_version
>= 121 &&
1181 sscreen
->info
.me_fw_version
>= 87) ||
1182 (sscreen
->info
.chip_class
== GFX7
&&
1183 sscreen
->info
.pfp_fw_version
>= 211 &&
1184 sscreen
->info
.me_fw_version
>= 173) ||
1185 (sscreen
->info
.chip_class
== GFX6
&&
1186 sscreen
->info
.pfp_fw_version
>= 79 &&
1187 sscreen
->info
.me_fw_version
>= 142);
1189 sscreen
->has_out_of_order_rast
= sscreen
->info
.has_out_of_order_rast
&&
1190 !(sscreen
->debug_flags
& DBG(NO_OUT_OF_ORDER
));
1191 sscreen
->assume_no_z_fights
=
1192 driQueryOptionb(config
->options
, "radeonsi_assume_no_z_fights");
1193 sscreen
->commutative_blend_add
=
1194 driQueryOptionb(config
->options
, "radeonsi_commutative_blend_add");
1196 sscreen
->use_ngg
= sscreen
->info
.chip_class
>= GFX10
&&
1197 sscreen
->info
.family
!= CHIP_NAVI14
&&
1198 !(sscreen
->debug_flags
& DBG(NO_NGG
));
1199 sscreen
->use_ngg_culling
= sscreen
->use_ngg
&&
1200 !(sscreen
->debug_flags
& DBG(NO_NGG_CULLING
));
1201 sscreen
->always_use_ngg_culling
= sscreen
->use_ngg_culling
&&
1202 sscreen
->debug_flags
& DBG(ALWAYS_NGG_CULLING
);
1203 sscreen
->use_ngg_streamout
= false;
1205 /* Only enable primitive binning on APUs by default. */
1206 if (sscreen
->info
.chip_class
>= GFX10
) {
1207 sscreen
->dpbb_allowed
= true;
1208 sscreen
->dfsm_allowed
= !sscreen
->info
.has_dedicated_vram
;
1209 } else if (sscreen
->info
.chip_class
== GFX9
) {
1210 sscreen
->dpbb_allowed
= !sscreen
->info
.has_dedicated_vram
;
1211 sscreen
->dfsm_allowed
= !sscreen
->info
.has_dedicated_vram
;
1214 /* Process DPBB enable flags. */
1215 if (sscreen
->debug_flags
& DBG(DPBB
)) {
1216 sscreen
->dpbb_allowed
= true;
1217 if (sscreen
->debug_flags
& DBG(DFSM
))
1218 sscreen
->dfsm_allowed
= true;
1221 /* Process DPBB disable flags. */
1222 if (sscreen
->debug_flags
& DBG(NO_DPBB
)) {
1223 sscreen
->dpbb_allowed
= false;
1224 sscreen
->dfsm_allowed
= false;
1225 } else if (sscreen
->debug_flags
& DBG(NO_DFSM
)) {
1226 sscreen
->dfsm_allowed
= false;
1229 /* While it would be nice not to have this flag, we are constrained
1230 * by the reality that LLVM 9.0 has buggy VGPR indexing on GFX9.
1232 sscreen
->llvm_has_working_vgpr_indexing
= sscreen
->info
.chip_class
!= GFX9
;
1234 sscreen
->dcc_msaa_allowed
=
1235 !(sscreen
->debug_flags
& DBG(NO_DCC_MSAA
));
1237 (void) simple_mtx_init(&sscreen
->shader_parts_mutex
, mtx_plain
);
1238 sscreen
->use_monolithic_shaders
=
1239 (sscreen
->debug_flags
& DBG(MONOLITHIC_SHADERS
)) != 0;
1241 sscreen
->barrier_flags
.cp_to_L2
= SI_CONTEXT_INV_SCACHE
|
1242 SI_CONTEXT_INV_VCACHE
;
1243 if (sscreen
->info
.chip_class
<= GFX8
) {
1244 sscreen
->barrier_flags
.cp_to_L2
|= SI_CONTEXT_INV_L2
;
1245 sscreen
->barrier_flags
.L2_to_cp
|= SI_CONTEXT_WB_L2
;
1248 if (debug_get_bool_option("RADEON_DUMP_SHADERS", false))
1249 sscreen
->debug_flags
|= DBG_ALL_SHADERS
;
1256 * That means 8 coverage samples, 4 Z/S samples, and 2 color samples.
1258 * s >= z >= c (ignoring this only wastes memory)
1263 * Only MSAA color and depth buffers are overriden.
1265 if (sscreen
->info
.has_eqaa_surface_allocator
) {
1266 const char *eqaa
= debug_get_option("EQAA", NULL
);
1269 if (eqaa
&& sscanf(eqaa
, "%u,%u,%u", &s
, &z
, &f
) == 3 && s
&& z
&& f
) {
1270 sscreen
->eqaa_force_coverage_samples
= s
;
1271 sscreen
->eqaa_force_z_samples
= z
;
1272 sscreen
->eqaa_force_color_samples
= f
;
1276 sscreen
->ge_wave_size
= 64;
1277 sscreen
->ps_wave_size
= 64;
1278 sscreen
->compute_wave_size
= 64;
1280 if (sscreen
->info
.chip_class
>= GFX10
) {
1281 /* Pixels shaders: Wave64 is recommended.
1282 * Compute shaders: There are piglit failures with Wave32.
1284 sscreen
->ge_wave_size
= 32;
1286 if (sscreen
->debug_flags
& DBG(W32_GE
))
1287 sscreen
->ge_wave_size
= 32;
1288 if (sscreen
->debug_flags
& DBG(W32_PS
))
1289 sscreen
->ps_wave_size
= 32;
1290 if (sscreen
->debug_flags
& DBG(W32_CS
))
1291 sscreen
->compute_wave_size
= 32;
1293 if (sscreen
->debug_flags
& DBG(W64_GE
))
1294 sscreen
->ge_wave_size
= 64;
1295 if (sscreen
->debug_flags
& DBG(W64_PS
))
1296 sscreen
->ps_wave_size
= 64;
1297 if (sscreen
->debug_flags
& DBG(W64_CS
))
1298 sscreen
->compute_wave_size
= 64;
1301 /* Create the auxiliary context. This must be done last. */
1302 sscreen
->aux_context
= si_create_context(&sscreen
->b
,
1303 (sscreen
->options
.aux_debug
? PIPE_CONTEXT_DEBUG
: 0) |
1304 (sscreen
->info
.has_graphics
? 0 : PIPE_CONTEXT_COMPUTE_ONLY
));
1305 if (sscreen
->options
.aux_debug
) {
1306 struct u_log_context
*log
= CALLOC_STRUCT(u_log_context
);
1307 u_log_context_init(log
);
1308 sscreen
->aux_context
->set_log_context(sscreen
->aux_context
, log
);
1311 if (test_flags
& DBG(TEST_DMA
))
1312 si_test_dma(sscreen
);
1314 if (test_flags
& DBG(TEST_DMA_PERF
)) {
1315 si_test_dma_perf(sscreen
);
1318 if (test_flags
& (DBG(TEST_VMFAULT_CP
) |
1319 DBG(TEST_VMFAULT_SDMA
) |
1320 DBG(TEST_VMFAULT_SHADER
)))
1321 si_test_vmfault(sscreen
, test_flags
);
1323 if (test_flags
& DBG(TEST_GDS
))
1324 si_test_gds((struct si_context
*)sscreen
->aux_context
);
1326 if (test_flags
& DBG(TEST_GDS_MM
)) {
1327 si_test_gds_memory_management((struct si_context
*)sscreen
->aux_context
,
1328 32 * 1024, 4, RADEON_DOMAIN_GDS
);
1330 if (test_flags
& DBG(TEST_GDS_OA_MM
)) {
1331 si_test_gds_memory_management((struct si_context
*)sscreen
->aux_context
,
1332 4, 1, RADEON_DOMAIN_OA
);
1335 STATIC_ASSERT(sizeof(union si_vgt_stages_key
) == 4);
1339 struct pipe_screen
*radeonsi_screen_create(int fd
, const struct pipe_screen_config
*config
)
1341 drmVersionPtr version
= drmGetVersion(fd
);
1342 struct radeon_winsys
*rw
= NULL
;
1344 switch (version
->version_major
) {
1346 rw
= radeon_drm_winsys_create(fd
, config
, radeonsi_screen_create_impl
);
1349 rw
= amdgpu_winsys_create(fd
, config
, radeonsi_screen_create_impl
);
1353 drmFreeVersion(version
);
1354 return rw
? rw
->screen
: NULL
;