2 * Copyright © 2017 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included
12 * in all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
15 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
20 * DEALINGS IN THE SOFTWARE.
24 * @file iris_program_cache.c
26 * The in-memory program cache. This is basically a hash table mapping
27 * API-specified shaders and a state key to a compiled variant. It also
28 * takes care of uploading shader assembly into a BO for use on the GPU.
33 #include "pipe/p_defines.h"
34 #include "pipe/p_state.h"
35 #include "pipe/p_context.h"
36 #include "pipe/p_screen.h"
37 #include "util/u_atomic.h"
38 #include "util/u_upload_mgr.h"
39 #include "compiler/nir/nir.h"
40 #include "compiler/nir/nir_builder.h"
41 #include "intel/compiler/brw_compiler.h"
42 #include "intel/compiler/brw_eu.h"
43 #include "intel/compiler/brw_nir.h"
44 #include "iris_context.h"
45 #include "iris_resource.h"
49 enum iris_program_cache_id cache_id
;
54 key_size_for_cache(enum iris_program_cache_id cache_id
)
56 static const unsigned key_sizes
[] = {
57 [IRIS_CACHE_VS
] = sizeof(struct brw_vs_prog_key
),
58 [IRIS_CACHE_TCS
] = sizeof(struct brw_tcs_prog_key
),
59 [IRIS_CACHE_TES
] = sizeof(struct brw_tes_prog_key
),
60 [IRIS_CACHE_GS
] = sizeof(struct brw_gs_prog_key
),
61 [IRIS_CACHE_FS
] = sizeof(struct brw_wm_prog_key
),
62 [IRIS_CACHE_CS
] = sizeof(struct brw_cs_prog_key
),
65 /* BLORP keys aren't all the same size. */
66 assert(cache_id
!= IRIS_CACHE_BLORP
);
68 return key_sizes
[cache_id
];
71 static struct keybox
*
72 make_keybox(void *mem_ctx
,
73 enum iris_program_cache_id cache_id
,
77 struct keybox
*keybox
=
78 ralloc_size(mem_ctx
, sizeof(struct keybox
) + key_size
);
80 keybox
->cache_id
= cache_id
;
81 keybox
->size
= key_size
;
82 memcpy(keybox
->data
, key
, key_size
);
88 keybox_hash(const void *void_key
)
90 const struct keybox
*key
= void_key
;
91 return _mesa_hash_data(&key
->cache_id
, key
->size
+ sizeof(key
->cache_id
));
95 keybox_equals(const void *void_a
, const void *void_b
)
97 const struct keybox
*a
= void_a
, *b
= void_b
;
98 if (a
->size
!= b
->size
)
101 return memcmp(a
->data
, b
->data
, a
->size
) == 0;
105 dirty_flag_for_cache(enum iris_program_cache_id cache_id
)
107 assert(cache_id
<= MESA_SHADER_STAGES
);
110 // XXX: move this flagging out to a higher level, allow comparison of
111 // XXX: new and old programs to decide what bits to twiddle
112 // XXX: CLIP: toggle if barycentric modes has any NONPERSPECTIVE or not
113 if (cache_id
== IRIS_CACHE_FS
)
114 return IRIS_DIRTY_WM
| IRIS_DIRTY_FS
| IRIS_DIRTY_CLIP
| IRIS_DIRTY_SBE
;
115 if (cache_id
== IRIS_CACHE_VS
)
116 return IRIS_DIRTY_VS
| IRIS_DIRTY_VF_SGVS
;
118 /* For compute, prog_data->threads needs to be uploaded as constants. */
119 if (cache_id
== IRIS_CACHE_CS
)
120 return IRIS_DIRTY_CS
| IRIS_DIRTY_CONSTANTS_CS
;
122 return IRIS_DIRTY_VS
<< cache_id
| IRIS_DIRTY_BINDINGS_VS
<< cache_id
;
126 get_program_string_id(enum iris_program_cache_id cache_id
, const void *key
)
130 return ((struct brw_vs_prog_key
*) key
)->program_string_id
;
132 return ((struct brw_tcs_prog_key
*) key
)->program_string_id
;
134 return ((struct brw_tes_prog_key
*) key
)->program_string_id
;
136 return ((struct brw_gs_prog_key
*) key
)->program_string_id
;
138 return ((struct brw_cs_prog_key
*) key
)->program_string_id
;
140 return ((struct brw_wm_prog_key
*) key
)->program_string_id
;
142 unreachable("no program string id for this kind of program");
146 static struct iris_compiled_shader
*
147 iris_find_cached_shader(struct iris_context
*ice
,
148 enum iris_program_cache_id cache_id
,
152 struct keybox
*keybox
=
153 make_keybox(ice
->shaders
.cache
, cache_id
, key
, key_size
);
154 struct hash_entry
*entry
=
155 _mesa_hash_table_search(ice
->shaders
.cache
, keybox
);
159 return entry
? entry
->data
: NULL
;
163 * Looks for a program in the cache and binds it.
165 * If no program was found, returns false and leaves the binding alone.
168 iris_bind_cached_shader(struct iris_context
*ice
,
169 enum iris_program_cache_id cache_id
,
172 unsigned key_size
= key_size_for_cache(cache_id
);
173 struct iris_compiled_shader
*shader
=
174 iris_find_cached_shader(ice
, cache_id
, key
, key_size
);
180 if (!ice
->shaders
.prog
[cache_id
] ||
181 memcmp(shader
, ice
->shaders
.prog
[cache_id
], sizeof(*shader
)) != 0) {
182 ice
->shaders
.prog
[cache_id
] = shader
;
183 ice
->state
.dirty
|= dirty_flag_for_cache(cache_id
);
190 iris_unbind_shader(struct iris_context
*ice
,
191 enum iris_program_cache_id cache_id
)
193 if (ice
->shaders
.prog
[cache_id
]) {
194 ice
->shaders
.prog
[cache_id
] = NULL
;
195 ice
->state
.dirty
|= dirty_flag_for_cache(cache_id
);
200 iris_find_previous_compile(const struct iris_context
*ice
,
201 enum iris_program_cache_id cache_id
,
202 unsigned program_string_id
)
204 hash_table_foreach(ice
->shaders
.cache
, entry
) {
205 const struct keybox
*keybox
= entry
->key
;
206 if (keybox
->cache_id
== cache_id
&&
207 get_program_string_id(cache_id
, keybox
->data
) == program_string_id
) {
216 * Look for an existing entry in the cache that has identical assembly code.
218 * This is useful for programs generating shaders at runtime, where multiple
219 * distinct shaders (from an API perspective) may compile to the same assembly
220 * in our backend. This saves space in the program cache buffer.
222 static const struct iris_compiled_shader
*
223 find_existing_assembly(struct hash_table
*cache
,
224 const void *assembly
,
225 unsigned assembly_size
)
227 hash_table_foreach(cache
, entry
) {
228 const struct iris_compiled_shader
*existing
= entry
->data
;
229 if (existing
->prog_data
->program_size
== assembly_size
&&
230 memcmp(existing
->map
, assembly
, assembly_size
) == 0)
236 static struct iris_compiled_shader
*
237 iris_upload_shader(struct iris_context
*ice
,
238 enum iris_program_cache_id cache_id
,
241 const void *assembly
,
242 struct brw_stage_prog_data
*prog_data
,
245 struct iris_screen
*screen
= (void *) ice
->ctx
.screen
;
246 struct gen_device_info
*devinfo
= &screen
->devinfo
;
247 struct hash_table
*cache
= ice
->shaders
.cache
;
248 struct iris_compiled_shader
*shader
=
249 rzalloc_size(cache
, sizeof(struct iris_compiled_shader
) +
250 ice
->vtbl
.derived_program_state_size(cache_id
));
251 const struct iris_compiled_shader
*existing
=
252 find_existing_assembly(cache
, assembly
, prog_data
->program_size
);
254 /* If we can find a matching prog in the cache already, then reuse the
255 * existing stuff without creating new copy into the underlying buffer
256 * object. This is notably useful for programs generating shaders at
257 * runtime, where multiple shaders may compile to the same thing in our
261 pipe_resource_reference(&shader
->assembly
.res
, existing
->assembly
.res
);
262 shader
->assembly
.offset
= existing
->assembly
.offset
;
263 shader
->map
= existing
->map
;
265 shader
->assembly
.res
= NULL
;
266 u_upload_alloc(ice
->shaders
.uploader
, 0, prog_data
->program_size
, 64,
267 &shader
->assembly
.offset
, &shader
->assembly
.res
,
269 memcpy(shader
->map
, assembly
, prog_data
->program_size
);
272 shader
->prog_data
= prog_data
;
273 shader
->streamout
= streamout
;
275 ralloc_steal(shader
, shader
->prog_data
);
276 ralloc_steal(shader
->prog_data
, prog_data
->param
);
277 ralloc_steal(shader
->prog_data
, prog_data
->pull_param
);
278 ralloc_steal(shader
, shader
->streamout
);
280 /* Store the 3DSTATE shader packets and other derived state. */
281 ice
->vtbl
.store_derived_program_state(devinfo
, cache_id
, shader
);
283 struct keybox
*keybox
= make_keybox(cache
, cache_id
, key
, key_size
);
284 _mesa_hash_table_insert(ice
->shaders
.cache
, keybox
, shader
);
290 * Upload a new shader to the program cache, and bind it for use.
292 * \param prog_data must be ralloc'd and will be stolen.
295 iris_upload_and_bind_shader(struct iris_context
*ice
,
296 enum iris_program_cache_id cache_id
,
298 const void *assembly
,
299 struct brw_stage_prog_data
*prog_data
,
302 assert(cache_id
!= IRIS_CACHE_BLORP
);
304 struct iris_compiled_shader
*shader
=
305 iris_upload_shader(ice
, cache_id
, key_size_for_cache(cache_id
), key
,
306 assembly
, prog_data
, streamout
);
308 ice
->shaders
.prog
[cache_id
] = shader
;
309 ice
->state
.dirty
|= dirty_flag_for_cache(cache_id
);
313 iris_blorp_lookup_shader(struct blorp_batch
*blorp_batch
,
314 const void *key
, uint32_t key_size
,
315 uint32_t *kernel_out
, void *prog_data_out
)
317 struct blorp_context
*blorp
= blorp_batch
->blorp
;
318 struct iris_context
*ice
= blorp
->driver_ctx
;
319 struct iris_batch
*batch
= blorp_batch
->driver_batch
;
320 struct iris_compiled_shader
*shader
=
321 iris_find_cached_shader(ice
, IRIS_CACHE_BLORP
, key
, key_size
);
326 struct iris_bo
*bo
= iris_resource_bo(shader
->assembly
.res
);
328 iris_bo_offset_from_base_address(bo
) + shader
->assembly
.offset
;
329 *((void **) prog_data_out
) = shader
->prog_data
;
331 iris_use_pinned_bo(batch
, bo
, false);
337 iris_blorp_upload_shader(struct blorp_batch
*blorp_batch
,
338 const void *key
, uint32_t key_size
,
339 const void *kernel
, UNUSED
uint32_t kernel_size
,
340 const struct brw_stage_prog_data
*prog_data_templ
,
341 UNUSED
uint32_t prog_data_size
,
342 uint32_t *kernel_out
, void *prog_data_out
)
344 struct blorp_context
*blorp
= blorp_batch
->blorp
;
345 struct iris_context
*ice
= blorp
->driver_ctx
;
346 struct iris_batch
*batch
= blorp_batch
->driver_batch
;
348 void *prog_data
= ralloc_size(NULL
, prog_data_size
);
349 memcpy(prog_data
, prog_data_templ
, prog_data_size
);
351 struct iris_compiled_shader
*shader
=
352 iris_upload_shader(ice
, IRIS_CACHE_BLORP
, key_size
, key
, kernel
,
355 struct iris_bo
*bo
= iris_resource_bo(shader
->assembly
.res
);
357 iris_bo_offset_from_base_address(bo
) + shader
->assembly
.offset
;
358 *((void **) prog_data_out
) = shader
->prog_data
;
360 iris_use_pinned_bo(batch
, bo
, false);
366 iris_init_program_cache(struct iris_context
*ice
)
369 _mesa_hash_table_create(ice
, keybox_hash
, keybox_equals
);
371 ice
->shaders
.uploader
=
372 u_upload_create(&ice
->ctx
, 16384, PIPE_BIND_CUSTOM
, PIPE_USAGE_IMMUTABLE
,
373 IRIS_RESOURCE_FLAG_SHADER_MEMZONE
);
377 iris_destroy_program_cache(struct iris_context
*ice
)
379 for (int i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
380 ice
->shaders
.prog
[i
] = NULL
;
383 hash_table_foreach(ice
->shaders
.cache
, entry
) {
384 struct iris_compiled_shader
*shader
= entry
->data
;
385 pipe_resource_reference(&shader
->assembly
.res
, NULL
);
388 u_upload_destroy(ice
->shaders
.uploader
);
390 ralloc_free(ice
->shaders
.cache
);
394 cache_name(enum iris_program_cache_id cache_id
)
396 if (cache_id
== IRIS_CACHE_BLORP
)
399 return _mesa_shader_stage_to_string(cache_id
);
403 iris_print_program_cache(struct iris_context
*ice
)
405 struct iris_screen
*screen
= (struct iris_screen
*)ice
->ctx
.screen
;
406 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
408 hash_table_foreach(ice
->shaders
.cache
, entry
) {
409 const struct keybox
*keybox
= entry
->key
;
410 struct iris_compiled_shader
*shader
= entry
->data
;
411 fprintf(stderr
, "%s:\n", cache_name(keybox
->cache_id
));
412 brw_disassemble(devinfo
, shader
->map
, 0,
413 shader
->prog_data
->program_size
, stderr
);