2 * Copyright 2012 Advanced Micro Devices, Inc.
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 * Christian König <christian.koenig@amd.com>
25 * Marek Olšák <maraeo@gmail.com>
30 #include "radeon/r600_cs.h"
32 #include "tgsi/tgsi_parse.h"
33 #include "tgsi/tgsi_ureg.h"
34 #include "util/hash_table.h"
35 #include "util/crc32.h"
36 #include "util/u_memory.h"
37 #include "util/u_prim.h"
42 * Return the TGSI binary in a buffer. The first 4 bytes contain its size as
45 static void *si_get_tgsi_binary(struct si_shader_selector
*sel
)
47 unsigned tgsi_size
= tgsi_num_tokens(sel
->tokens
) *
48 sizeof(struct tgsi_token
);
49 unsigned size
= 4 + tgsi_size
+ sizeof(sel
->so
);
50 char *result
= (char*)MALLOC(size
);
55 *((uint32_t*)result
) = size
;
56 memcpy(result
+ 4, sel
->tokens
, tgsi_size
);
57 memcpy(result
+ 4 + tgsi_size
, &sel
->so
, sizeof(sel
->so
));
61 /** Copy "data" to "ptr" and return the next dword following copied data. */
62 static uint32_t *write_data(uint32_t *ptr
, const void *data
, unsigned size
)
64 /* data may be NULL if size == 0 */
66 memcpy(ptr
, data
, size
);
67 ptr
+= DIV_ROUND_UP(size
, 4);
71 /** Read data from "ptr". Return the next dword following the data. */
72 static uint32_t *read_data(uint32_t *ptr
, void *data
, unsigned size
)
74 memcpy(data
, ptr
, size
);
75 ptr
+= DIV_ROUND_UP(size
, 4);
80 * Write the size as uint followed by the data. Return the next dword
81 * following the copied data.
83 static uint32_t *write_chunk(uint32_t *ptr
, const void *data
, unsigned size
)
86 return write_data(ptr
, data
, size
);
90 * Read the size as uint followed by the data. Return both via parameters.
91 * Return the next dword following the data.
93 static uint32_t *read_chunk(uint32_t *ptr
, void **data
, unsigned *size
)
96 assert(*data
== NULL
);
99 *data
= malloc(*size
);
100 return read_data(ptr
, *data
, *size
);
104 * Return the shader binary in a buffer. The first 4 bytes contain its size
107 static void *si_get_shader_binary(struct si_shader
*shader
)
109 /* There is always a size of data followed by the data itself. */
110 unsigned relocs_size
= shader
->binary
.reloc_count
*
111 sizeof(shader
->binary
.relocs
[0]);
112 unsigned disasm_size
= strlen(shader
->binary
.disasm_string
) + 1;
113 unsigned llvm_ir_size
= shader
->binary
.llvm_ir_string
?
114 strlen(shader
->binary
.llvm_ir_string
) + 1 : 0;
117 4 + /* CRC32 of the data below */
118 align(sizeof(shader
->config
), 4) +
119 align(sizeof(shader
->info
), 4) +
120 4 + align(shader
->binary
.code_size
, 4) +
121 4 + align(shader
->binary
.rodata_size
, 4) +
122 4 + align(relocs_size
, 4) +
123 4 + align(disasm_size
, 4) +
124 4 + align(llvm_ir_size
, 4);
125 void *buffer
= CALLOC(1, size
);
126 uint32_t *ptr
= (uint32_t*)buffer
;
132 ptr
++; /* CRC32 is calculated at the end. */
134 ptr
= write_data(ptr
, &shader
->config
, sizeof(shader
->config
));
135 ptr
= write_data(ptr
, &shader
->info
, sizeof(shader
->info
));
136 ptr
= write_chunk(ptr
, shader
->binary
.code
, shader
->binary
.code_size
);
137 ptr
= write_chunk(ptr
, shader
->binary
.rodata
, shader
->binary
.rodata_size
);
138 ptr
= write_chunk(ptr
, shader
->binary
.relocs
, relocs_size
);
139 ptr
= write_chunk(ptr
, shader
->binary
.disasm_string
, disasm_size
);
140 ptr
= write_chunk(ptr
, shader
->binary
.llvm_ir_string
, llvm_ir_size
);
141 assert((char *)ptr
- (char *)buffer
== size
);
144 ptr
= (uint32_t*)buffer
;
146 *ptr
= util_hash_crc32(ptr
+ 1, size
- 8);
151 static bool si_load_shader_binary(struct si_shader
*shader
, void *binary
)
153 uint32_t *ptr
= (uint32_t*)binary
;
154 uint32_t size
= *ptr
++;
155 uint32_t crc32
= *ptr
++;
158 if (util_hash_crc32(ptr
, size
- 8) != crc32
) {
159 fprintf(stderr
, "radeonsi: binary shader has invalid CRC32\n");
163 ptr
= read_data(ptr
, &shader
->config
, sizeof(shader
->config
));
164 ptr
= read_data(ptr
, &shader
->info
, sizeof(shader
->info
));
165 ptr
= read_chunk(ptr
, (void**)&shader
->binary
.code
,
166 &shader
->binary
.code_size
);
167 ptr
= read_chunk(ptr
, (void**)&shader
->binary
.rodata
,
168 &shader
->binary
.rodata_size
);
169 ptr
= read_chunk(ptr
, (void**)&shader
->binary
.relocs
, &chunk_size
);
170 shader
->binary
.reloc_count
= chunk_size
/ sizeof(shader
->binary
.relocs
[0]);
171 ptr
= read_chunk(ptr
, (void**)&shader
->binary
.disasm_string
, &chunk_size
);
172 ptr
= read_chunk(ptr
, (void**)&shader
->binary
.llvm_ir_string
, &chunk_size
);
178 * Insert a shader into the cache. It's assumed the shader is not in the cache.
179 * Use si_shader_cache_load_shader before calling this.
181 * Returns false on failure, in which case the tgsi_binary should be freed.
183 static bool si_shader_cache_insert_shader(struct si_screen
*sscreen
,
185 struct si_shader
*shader
)
188 struct hash_entry
*entry
;
190 entry
= _mesa_hash_table_search(sscreen
->shader_cache
, tgsi_binary
);
192 return false; /* already added */
194 hw_binary
= si_get_shader_binary(shader
);
198 if (_mesa_hash_table_insert(sscreen
->shader_cache
, tgsi_binary
,
199 hw_binary
) == NULL
) {
207 static bool si_shader_cache_load_shader(struct si_screen
*sscreen
,
209 struct si_shader
*shader
)
211 struct hash_entry
*entry
=
212 _mesa_hash_table_search(sscreen
->shader_cache
, tgsi_binary
);
216 if (!si_load_shader_binary(shader
, entry
->data
))
219 p_atomic_inc(&sscreen
->b
.num_shader_cache_hits
);
223 static uint32_t si_shader_cache_key_hash(const void *key
)
225 /* The first dword is the key size. */
226 return util_hash_crc32(key
, *(uint32_t*)key
);
229 static bool si_shader_cache_key_equals(const void *a
, const void *b
)
231 uint32_t *keya
= (uint32_t*)a
;
232 uint32_t *keyb
= (uint32_t*)b
;
234 /* The first dword is the key size. */
238 return memcmp(keya
, keyb
, *keya
) == 0;
241 static void si_destroy_shader_cache_entry(struct hash_entry
*entry
)
243 FREE((void*)entry
->key
);
247 bool si_init_shader_cache(struct si_screen
*sscreen
)
249 pipe_mutex_init(sscreen
->shader_cache_mutex
);
250 sscreen
->shader_cache
=
251 _mesa_hash_table_create(NULL
,
252 si_shader_cache_key_hash
,
253 si_shader_cache_key_equals
);
254 return sscreen
->shader_cache
!= NULL
;
257 void si_destroy_shader_cache(struct si_screen
*sscreen
)
259 if (sscreen
->shader_cache
)
260 _mesa_hash_table_destroy(sscreen
->shader_cache
,
261 si_destroy_shader_cache_entry
);
262 pipe_mutex_destroy(sscreen
->shader_cache_mutex
);
267 static void si_set_tesseval_regs(struct si_screen
*sscreen
,
268 struct si_shader
*shader
,
269 struct si_pm4_state
*pm4
)
271 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
272 unsigned tes_prim_mode
= info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
273 unsigned tes_spacing
= info
->properties
[TGSI_PROPERTY_TES_SPACING
];
274 bool tes_vertex_order_cw
= info
->properties
[TGSI_PROPERTY_TES_VERTEX_ORDER_CW
];
275 bool tes_point_mode
= info
->properties
[TGSI_PROPERTY_TES_POINT_MODE
];
276 unsigned type
, partitioning
, topology
, distribution_mode
;
278 switch (tes_prim_mode
) {
279 case PIPE_PRIM_LINES
:
280 type
= V_028B6C_TESS_ISOLINE
;
282 case PIPE_PRIM_TRIANGLES
:
283 type
= V_028B6C_TESS_TRIANGLE
;
285 case PIPE_PRIM_QUADS
:
286 type
= V_028B6C_TESS_QUAD
;
293 switch (tes_spacing
) {
294 case PIPE_TESS_SPACING_FRACTIONAL_ODD
:
295 partitioning
= V_028B6C_PART_FRAC_ODD
;
297 case PIPE_TESS_SPACING_FRACTIONAL_EVEN
:
298 partitioning
= V_028B6C_PART_FRAC_EVEN
;
300 case PIPE_TESS_SPACING_EQUAL
:
301 partitioning
= V_028B6C_PART_INTEGER
;
309 topology
= V_028B6C_OUTPUT_POINT
;
310 else if (tes_prim_mode
== PIPE_PRIM_LINES
)
311 topology
= V_028B6C_OUTPUT_LINE
;
312 else if (tes_vertex_order_cw
)
313 /* for some reason, this must be the other way around */
314 topology
= V_028B6C_OUTPUT_TRIANGLE_CCW
;
316 topology
= V_028B6C_OUTPUT_TRIANGLE_CW
;
318 if (sscreen
->has_distributed_tess
) {
319 if (sscreen
->b
.family
== CHIP_FIJI
||
320 sscreen
->b
.family
>= CHIP_POLARIS10
)
321 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS
;
323 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_DONUTS
;
325 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_NO_DIST
;
327 si_pm4_set_reg(pm4
, R_028B6C_VGT_TF_PARAM
,
328 S_028B6C_TYPE(type
) |
329 S_028B6C_PARTITIONING(partitioning
) |
330 S_028B6C_TOPOLOGY(topology
) |
331 S_028B6C_DISTRIBUTION_MODE(distribution_mode
));
334 /* Polaris needs different VTX_REUSE_DEPTH settings depending on
335 * whether the "fractional odd" tessellation spacing is used.
337 * Possible VGT configurations and which state should set the register:
339 * Reg set in | VGT shader configuration | Value
340 * ------------------------------------------------------
342 * VS as ES | ES -> GS -> VS | 30
343 * TES as VS | LS -> HS -> VS | 14 or 30
344 * TES as ES | LS -> HS -> ES -> GS -> VS | 14 or 30
346 static void polaris_set_vgt_vertex_reuse(struct si_screen
*sscreen
,
347 struct si_shader
*shader
,
348 struct si_pm4_state
*pm4
)
350 unsigned type
= shader
->selector
->type
;
352 if (sscreen
->b
.family
< CHIP_POLARIS10
)
355 /* VS as VS, or VS as ES: */
356 if ((type
== PIPE_SHADER_VERTEX
&&
357 !shader
->key
.as_ls
&&
358 !shader
->is_gs_copy_shader
) ||
359 /* TES as VS, or TES as ES: */
360 type
== PIPE_SHADER_TESS_EVAL
) {
361 unsigned vtx_reuse_depth
= 30;
363 if (type
== PIPE_SHADER_TESS_EVAL
&&
364 shader
->selector
->info
.properties
[TGSI_PROPERTY_TES_SPACING
] ==
365 PIPE_TESS_SPACING_FRACTIONAL_ODD
)
366 vtx_reuse_depth
= 14;
368 si_pm4_set_reg(pm4
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
373 static struct si_pm4_state
*si_get_shader_pm4_state(struct si_shader
*shader
)
376 si_pm4_clear_state(shader
->pm4
);
378 shader
->pm4
= CALLOC_STRUCT(si_pm4_state
);
383 static void si_shader_ls(struct si_shader
*shader
)
385 struct si_pm4_state
*pm4
;
386 unsigned vgpr_comp_cnt
;
389 pm4
= si_get_shader_pm4_state(shader
);
393 va
= shader
->bo
->gpu_address
;
394 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
396 /* We need at least 2 components for LS.
397 * VGPR0-3: (VertexID, RelAutoindex, ???, InstanceID). */
398 vgpr_comp_cnt
= shader
->info
.uses_instanceid
? 3 : 1;
400 si_pm4_set_reg(pm4
, R_00B520_SPI_SHADER_PGM_LO_LS
, va
>> 8);
401 si_pm4_set_reg(pm4
, R_00B524_SPI_SHADER_PGM_HI_LS
, va
>> 40);
403 shader
->config
.rsrc1
= S_00B528_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
404 S_00B528_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
405 S_00B528_VGPR_COMP_CNT(vgpr_comp_cnt
) |
406 S_00B528_DX10_CLAMP(1) |
407 S_00B528_FLOAT_MODE(shader
->config
.float_mode
);
408 shader
->config
.rsrc2
= S_00B52C_USER_SGPR(SI_LS_NUM_USER_SGPR
) |
409 S_00B52C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
412 static void si_shader_hs(struct si_shader
*shader
)
414 struct si_pm4_state
*pm4
;
417 pm4
= si_get_shader_pm4_state(shader
);
421 va
= shader
->bo
->gpu_address
;
422 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
424 si_pm4_set_reg(pm4
, R_00B420_SPI_SHADER_PGM_LO_HS
, va
>> 8);
425 si_pm4_set_reg(pm4
, R_00B424_SPI_SHADER_PGM_HI_HS
, va
>> 40);
426 si_pm4_set_reg(pm4
, R_00B428_SPI_SHADER_PGM_RSRC1_HS
,
427 S_00B428_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
428 S_00B428_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
429 S_00B428_DX10_CLAMP(1) |
430 S_00B428_FLOAT_MODE(shader
->config
.float_mode
));
431 si_pm4_set_reg(pm4
, R_00B42C_SPI_SHADER_PGM_RSRC2_HS
,
432 S_00B42C_USER_SGPR(SI_TCS_NUM_USER_SGPR
) |
433 S_00B42C_OC_LDS_EN(1) |
434 S_00B42C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
437 static void si_shader_es(struct si_screen
*sscreen
, struct si_shader
*shader
)
439 struct si_pm4_state
*pm4
;
440 unsigned num_user_sgprs
;
441 unsigned vgpr_comp_cnt
;
445 pm4
= si_get_shader_pm4_state(shader
);
449 va
= shader
->bo
->gpu_address
;
450 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
452 if (shader
->selector
->type
== PIPE_SHADER_VERTEX
) {
453 vgpr_comp_cnt
= shader
->info
.uses_instanceid
? 3 : 0;
454 num_user_sgprs
= SI_ES_NUM_USER_SGPR
;
455 } else if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
456 vgpr_comp_cnt
= 3; /* all components are needed for TES */
457 num_user_sgprs
= SI_TES_NUM_USER_SGPR
;
459 unreachable("invalid shader selector type");
461 oc_lds_en
= shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
? 1 : 0;
463 si_pm4_set_reg(pm4
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
464 shader
->selector
->esgs_itemsize
/ 4);
465 si_pm4_set_reg(pm4
, R_00B320_SPI_SHADER_PGM_LO_ES
, va
>> 8);
466 si_pm4_set_reg(pm4
, R_00B324_SPI_SHADER_PGM_HI_ES
, va
>> 40);
467 si_pm4_set_reg(pm4
, R_00B328_SPI_SHADER_PGM_RSRC1_ES
,
468 S_00B328_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
469 S_00B328_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
470 S_00B328_VGPR_COMP_CNT(vgpr_comp_cnt
) |
471 S_00B328_DX10_CLAMP(1) |
472 S_00B328_FLOAT_MODE(shader
->config
.float_mode
));
473 si_pm4_set_reg(pm4
, R_00B32C_SPI_SHADER_PGM_RSRC2_ES
,
474 S_00B32C_USER_SGPR(num_user_sgprs
) |
475 S_00B32C_OC_LDS_EN(oc_lds_en
) |
476 S_00B32C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
478 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
479 si_set_tesseval_regs(sscreen
, shader
, pm4
);
481 polaris_set_vgt_vertex_reuse(sscreen
, shader
, pm4
);
485 * Calculate the appropriate setting of VGT_GS_MODE when \p shader is a
488 static uint32_t si_vgt_gs_mode(struct si_shader_selector
*sel
)
490 unsigned gs_max_vert_out
= sel
->gs_max_out_vertices
;
493 if (gs_max_vert_out
<= 128) {
494 cut_mode
= V_028A40_GS_CUT_128
;
495 } else if (gs_max_vert_out
<= 256) {
496 cut_mode
= V_028A40_GS_CUT_256
;
497 } else if (gs_max_vert_out
<= 512) {
498 cut_mode
= V_028A40_GS_CUT_512
;
500 assert(gs_max_vert_out
<= 1024);
501 cut_mode
= V_028A40_GS_CUT_1024
;
504 return S_028A40_MODE(V_028A40_GS_SCENARIO_G
) |
505 S_028A40_CUT_MODE(cut_mode
)|
506 S_028A40_ES_WRITE_OPTIMIZE(1) |
507 S_028A40_GS_WRITE_OPTIMIZE(1);
510 static void si_shader_gs(struct si_shader
*shader
)
512 struct si_shader_selector
*sel
= shader
->selector
;
513 const ubyte
*num_components
= sel
->info
.num_stream_output_components
;
514 unsigned gs_num_invocations
= sel
->gs_num_invocations
;
515 struct si_pm4_state
*pm4
;
517 unsigned max_stream
= sel
->max_gs_stream
;
520 pm4
= si_get_shader_pm4_state(shader
);
524 si_pm4_set_reg(pm4
, R_028A40_VGT_GS_MODE
, si_vgt_gs_mode(shader
->selector
));
526 offset
= num_components
[0] * sel
->gs_max_out_vertices
;
527 si_pm4_set_reg(pm4
, R_028A60_VGT_GSVS_RING_OFFSET_1
, offset
);
529 offset
+= num_components
[1] * sel
->gs_max_out_vertices
;
530 si_pm4_set_reg(pm4
, R_028A64_VGT_GSVS_RING_OFFSET_2
, offset
);
532 offset
+= num_components
[2] * sel
->gs_max_out_vertices
;
533 si_pm4_set_reg(pm4
, R_028A68_VGT_GSVS_RING_OFFSET_3
, offset
);
535 offset
+= num_components
[3] * sel
->gs_max_out_vertices
;
536 si_pm4_set_reg(pm4
, R_028AB0_VGT_GSVS_RING_ITEMSIZE
, offset
);
538 /* The GSVS_RING_ITEMSIZE register takes 15 bits */
539 assert(offset
< (1 << 15));
541 si_pm4_set_reg(pm4
, R_028B38_VGT_GS_MAX_VERT_OUT
, shader
->selector
->gs_max_out_vertices
);
543 si_pm4_set_reg(pm4
, R_028B5C_VGT_GS_VERT_ITEMSIZE
, num_components
[0]);
544 si_pm4_set_reg(pm4
, R_028B60_VGT_GS_VERT_ITEMSIZE_1
, (max_stream
>= 1) ? num_components
[1] : 0);
545 si_pm4_set_reg(pm4
, R_028B64_VGT_GS_VERT_ITEMSIZE_2
, (max_stream
>= 2) ? num_components
[2] : 0);
546 si_pm4_set_reg(pm4
, R_028B68_VGT_GS_VERT_ITEMSIZE_3
, (max_stream
>= 3) ? num_components
[3] : 0);
548 si_pm4_set_reg(pm4
, R_028B90_VGT_GS_INSTANCE_CNT
,
549 S_028B90_CNT(MIN2(gs_num_invocations
, 127)) |
550 S_028B90_ENABLE(gs_num_invocations
> 0));
552 va
= shader
->bo
->gpu_address
;
553 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
554 si_pm4_set_reg(pm4
, R_00B220_SPI_SHADER_PGM_LO_GS
, va
>> 8);
555 si_pm4_set_reg(pm4
, R_00B224_SPI_SHADER_PGM_HI_GS
, va
>> 40);
557 si_pm4_set_reg(pm4
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
,
558 S_00B228_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
559 S_00B228_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
560 S_00B228_DX10_CLAMP(1) |
561 S_00B228_FLOAT_MODE(shader
->config
.float_mode
));
562 si_pm4_set_reg(pm4
, R_00B22C_SPI_SHADER_PGM_RSRC2_GS
,
563 S_00B22C_USER_SGPR(SI_GS_NUM_USER_SGPR
) |
564 S_00B22C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
568 * Compute the state for \p shader, which will run as a vertex shader on the
571 * If \p gs is non-NULL, it points to the geometry shader for which this shader
572 * is the copy shader.
574 static void si_shader_vs(struct si_screen
*sscreen
, struct si_shader
*shader
,
575 struct si_shader_selector
*gs
)
577 struct si_pm4_state
*pm4
;
578 unsigned num_user_sgprs
;
579 unsigned nparams
, vgpr_comp_cnt
;
582 unsigned window_space
=
583 shader
->selector
->info
.properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
];
584 bool enable_prim_id
= si_vs_exports_prim_id(shader
);
586 pm4
= si_get_shader_pm4_state(shader
);
590 /* We always write VGT_GS_MODE in the VS state, because every switch
591 * between different shader pipelines involving a different GS or no
592 * GS at all involves a switch of the VS (different GS use different
593 * copy shaders). On the other hand, when the API switches from a GS to
594 * no GS and then back to the same GS used originally, the GS state is
598 si_pm4_set_reg(pm4
, R_028A40_VGT_GS_MODE
,
599 S_028A40_MODE(enable_prim_id
? V_028A40_GS_SCENARIO_A
: 0));
600 si_pm4_set_reg(pm4
, R_028A84_VGT_PRIMITIVEID_EN
, enable_prim_id
);
602 si_pm4_set_reg(pm4
, R_028A40_VGT_GS_MODE
, si_vgt_gs_mode(gs
));
603 si_pm4_set_reg(pm4
, R_028A84_VGT_PRIMITIVEID_EN
, 0);
606 va
= shader
->bo
->gpu_address
;
607 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
610 vgpr_comp_cnt
= 0; /* only VertexID is needed for GS-COPY. */
611 num_user_sgprs
= SI_GSCOPY_NUM_USER_SGPR
;
612 } else if (shader
->selector
->type
== PIPE_SHADER_VERTEX
) {
613 vgpr_comp_cnt
= shader
->info
.uses_instanceid
? 3 : (enable_prim_id
? 2 : 0);
614 num_user_sgprs
= SI_VS_NUM_USER_SGPR
;
615 } else if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
616 vgpr_comp_cnt
= 3; /* all components are needed for TES */
617 num_user_sgprs
= SI_TES_NUM_USER_SGPR
;
619 unreachable("invalid shader selector type");
621 /* VS is required to export at least one param. */
622 nparams
= MAX2(shader
->info
.nr_param_exports
, 1);
623 si_pm4_set_reg(pm4
, R_0286C4_SPI_VS_OUT_CONFIG
,
624 S_0286C4_VS_EXPORT_COUNT(nparams
- 1));
626 si_pm4_set_reg(pm4
, R_02870C_SPI_SHADER_POS_FORMAT
,
627 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP
) |
628 S_02870C_POS1_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 1 ?
629 V_02870C_SPI_SHADER_4COMP
:
630 V_02870C_SPI_SHADER_NONE
) |
631 S_02870C_POS2_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 2 ?
632 V_02870C_SPI_SHADER_4COMP
:
633 V_02870C_SPI_SHADER_NONE
) |
634 S_02870C_POS3_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 3 ?
635 V_02870C_SPI_SHADER_4COMP
:
636 V_02870C_SPI_SHADER_NONE
));
638 oc_lds_en
= shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
? 1 : 0;
640 si_pm4_set_reg(pm4
, R_00B120_SPI_SHADER_PGM_LO_VS
, va
>> 8);
641 si_pm4_set_reg(pm4
, R_00B124_SPI_SHADER_PGM_HI_VS
, va
>> 40);
642 si_pm4_set_reg(pm4
, R_00B128_SPI_SHADER_PGM_RSRC1_VS
,
643 S_00B128_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
644 S_00B128_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
645 S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt
) |
646 S_00B128_DX10_CLAMP(1) |
647 S_00B128_FLOAT_MODE(shader
->config
.float_mode
));
648 si_pm4_set_reg(pm4
, R_00B12C_SPI_SHADER_PGM_RSRC2_VS
,
649 S_00B12C_USER_SGPR(num_user_sgprs
) |
650 S_00B12C_OC_LDS_EN(oc_lds_en
) |
651 S_00B12C_SO_BASE0_EN(!!shader
->selector
->so
.stride
[0]) |
652 S_00B12C_SO_BASE1_EN(!!shader
->selector
->so
.stride
[1]) |
653 S_00B12C_SO_BASE2_EN(!!shader
->selector
->so
.stride
[2]) |
654 S_00B12C_SO_BASE3_EN(!!shader
->selector
->so
.stride
[3]) |
655 S_00B12C_SO_EN(!!shader
->selector
->so
.num_outputs
) |
656 S_00B12C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
658 si_pm4_set_reg(pm4
, R_028818_PA_CL_VTE_CNTL
,
659 S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1));
661 si_pm4_set_reg(pm4
, R_028818_PA_CL_VTE_CNTL
,
662 S_028818_VTX_W0_FMT(1) |
663 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
664 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
665 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
667 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
668 si_set_tesseval_regs(sscreen
, shader
, pm4
);
670 polaris_set_vgt_vertex_reuse(sscreen
, shader
, pm4
);
673 static unsigned si_get_ps_num_interp(struct si_shader
*ps
)
675 struct tgsi_shader_info
*info
= &ps
->selector
->info
;
676 unsigned num_colors
= !!(info
->colors_read
& 0x0f) +
677 !!(info
->colors_read
& 0xf0);
678 unsigned num_interp
= ps
->selector
->info
.num_inputs
+
679 (ps
->key
.part
.ps
.prolog
.color_two_side
? num_colors
: 0);
681 assert(num_interp
<= 32);
682 return MIN2(num_interp
, 32);
685 static unsigned si_get_spi_shader_col_format(struct si_shader
*shader
)
687 unsigned value
= shader
->key
.part
.ps
.epilog
.spi_shader_col_format
;
688 unsigned i
, num_targets
= (util_last_bit(value
) + 3) / 4;
690 /* If the i-th target format is set, all previous target formats must
691 * be non-zero to avoid hangs.
693 for (i
= 0; i
< num_targets
; i
++)
694 if (!(value
& (0xf << (i
* 4))))
695 value
|= V_028714_SPI_SHADER_32_R
<< (i
* 4);
700 static unsigned si_get_cb_shader_mask(unsigned spi_shader_col_format
)
702 unsigned i
, cb_shader_mask
= 0;
704 for (i
= 0; i
< 8; i
++) {
705 switch ((spi_shader_col_format
>> (i
* 4)) & 0xf) {
706 case V_028714_SPI_SHADER_ZERO
:
708 case V_028714_SPI_SHADER_32_R
:
709 cb_shader_mask
|= 0x1 << (i
* 4);
711 case V_028714_SPI_SHADER_32_GR
:
712 cb_shader_mask
|= 0x3 << (i
* 4);
714 case V_028714_SPI_SHADER_32_AR
:
715 cb_shader_mask
|= 0x9 << (i
* 4);
717 case V_028714_SPI_SHADER_FP16_ABGR
:
718 case V_028714_SPI_SHADER_UNORM16_ABGR
:
719 case V_028714_SPI_SHADER_SNORM16_ABGR
:
720 case V_028714_SPI_SHADER_UINT16_ABGR
:
721 case V_028714_SPI_SHADER_SINT16_ABGR
:
722 case V_028714_SPI_SHADER_32_ABGR
:
723 cb_shader_mask
|= 0xf << (i
* 4);
729 return cb_shader_mask
;
732 static void si_shader_ps(struct si_shader
*shader
)
734 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
735 struct si_pm4_state
*pm4
;
736 unsigned spi_ps_in_control
, spi_shader_col_format
, cb_shader_mask
;
737 unsigned spi_baryc_cntl
= S_0286E0_FRONT_FACE_ALL_BITS(1);
739 unsigned input_ena
= shader
->config
.spi_ps_input_ena
;
741 /* we need to enable at least one of them, otherwise we hang the GPU */
742 assert(G_0286CC_PERSP_SAMPLE_ENA(input_ena
) ||
743 G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
744 G_0286CC_PERSP_CENTROID_ENA(input_ena
) ||
745 G_0286CC_PERSP_PULL_MODEL_ENA(input_ena
) ||
746 G_0286CC_LINEAR_SAMPLE_ENA(input_ena
) ||
747 G_0286CC_LINEAR_CENTER_ENA(input_ena
) ||
748 G_0286CC_LINEAR_CENTROID_ENA(input_ena
) ||
749 G_0286CC_LINE_STIPPLE_TEX_ENA(input_ena
));
750 /* POS_W_FLOAT_ENA requires one of the perspective weights. */
751 assert(!G_0286CC_POS_W_FLOAT_ENA(input_ena
) ||
752 G_0286CC_PERSP_SAMPLE_ENA(input_ena
) ||
753 G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
754 G_0286CC_PERSP_CENTROID_ENA(input_ena
) ||
755 G_0286CC_PERSP_PULL_MODEL_ENA(input_ena
));
757 /* Validate interpolation optimization flags (read as implications). */
758 assert(!shader
->key
.part
.ps
.prolog
.bc_optimize_for_persp
||
759 (G_0286CC_PERSP_CENTER_ENA(input_ena
) &&
760 G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
761 assert(!shader
->key
.part
.ps
.prolog
.bc_optimize_for_linear
||
762 (G_0286CC_LINEAR_CENTER_ENA(input_ena
) &&
763 G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
764 assert(!shader
->key
.part
.ps
.prolog
.force_persp_center_interp
||
765 (!G_0286CC_PERSP_SAMPLE_ENA(input_ena
) &&
766 !G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
767 assert(!shader
->key
.part
.ps
.prolog
.force_linear_center_interp
||
768 (!G_0286CC_LINEAR_SAMPLE_ENA(input_ena
) &&
769 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
770 assert(!shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
||
771 (!G_0286CC_PERSP_CENTER_ENA(input_ena
) &&
772 !G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
773 assert(!shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
||
774 (!G_0286CC_LINEAR_CENTER_ENA(input_ena
) &&
775 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
777 /* Validate cases when the optimizations are off (read as implications). */
778 assert(shader
->key
.part
.ps
.prolog
.bc_optimize_for_persp
||
779 !G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
780 !G_0286CC_PERSP_CENTROID_ENA(input_ena
));
781 assert(shader
->key
.part
.ps
.prolog
.bc_optimize_for_linear
||
782 !G_0286CC_LINEAR_CENTER_ENA(input_ena
) ||
783 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
));
785 pm4
= si_get_shader_pm4_state(shader
);
789 /* SPI_BARYC_CNTL.POS_FLOAT_LOCATION
791 * 0 -> Position = pixel center
792 * 1 -> Position = pixel centroid
793 * 2 -> Position = at sample position
795 * From GLSL 4.5 specification, section 7.1:
796 * "The variable gl_FragCoord is available as an input variable from
797 * within fragment shaders and it holds the window relative coordinates
798 * (x, y, z, 1/w) values for the fragment. If multi-sampling, this
799 * value can be for any location within the pixel, or one of the
800 * fragment samples. The use of centroid does not further restrict
801 * this value to be inside the current primitive."
803 * Meaning that centroid has no effect and we can return anything within
804 * the pixel. Thus, return the value at sample position, because that's
805 * the most accurate one shaders can get.
807 spi_baryc_cntl
|= S_0286E0_POS_FLOAT_LOCATION(2);
809 if (info
->properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
] ==
810 TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)
811 spi_baryc_cntl
|= S_0286E0_POS_FLOAT_ULC(1);
813 spi_shader_col_format
= si_get_spi_shader_col_format(shader
);
814 cb_shader_mask
= si_get_cb_shader_mask(spi_shader_col_format
);
816 /* Ensure that some export memory is always allocated, for two reasons:
818 * 1) Correctness: The hardware ignores the EXEC mask if no export
819 * memory is allocated, so KILL and alpha test do not work correctly
821 * 2) Performance: Every shader needs at least a NULL export, even when
822 * it writes no color/depth output. The NULL export instruction
823 * stalls without this setting.
825 * Don't add this to CB_SHADER_MASK.
827 if (!spi_shader_col_format
&&
828 !info
->writes_z
&& !info
->writes_stencil
&& !info
->writes_samplemask
)
829 spi_shader_col_format
= V_028714_SPI_SHADER_32_R
;
831 si_pm4_set_reg(pm4
, R_0286CC_SPI_PS_INPUT_ENA
, input_ena
);
832 si_pm4_set_reg(pm4
, R_0286D0_SPI_PS_INPUT_ADDR
,
833 shader
->config
.spi_ps_input_addr
);
835 /* Set interpolation controls. */
836 spi_ps_in_control
= S_0286D8_NUM_INTERP(si_get_ps_num_interp(shader
));
839 si_pm4_set_reg(pm4
, R_0286E0_SPI_BARYC_CNTL
, spi_baryc_cntl
);
840 si_pm4_set_reg(pm4
, R_0286D8_SPI_PS_IN_CONTROL
, spi_ps_in_control
);
842 si_pm4_set_reg(pm4
, R_028710_SPI_SHADER_Z_FORMAT
,
843 si_get_spi_shader_z_format(info
->writes_z
,
844 info
->writes_stencil
,
845 info
->writes_samplemask
));
847 si_pm4_set_reg(pm4
, R_028714_SPI_SHADER_COL_FORMAT
, spi_shader_col_format
);
848 si_pm4_set_reg(pm4
, R_02823C_CB_SHADER_MASK
, cb_shader_mask
);
850 va
= shader
->bo
->gpu_address
;
851 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
852 si_pm4_set_reg(pm4
, R_00B020_SPI_SHADER_PGM_LO_PS
, va
>> 8);
853 si_pm4_set_reg(pm4
, R_00B024_SPI_SHADER_PGM_HI_PS
, va
>> 40);
855 si_pm4_set_reg(pm4
, R_00B028_SPI_SHADER_PGM_RSRC1_PS
,
856 S_00B028_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
857 S_00B028_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
858 S_00B028_DX10_CLAMP(1) |
859 S_00B028_FLOAT_MODE(shader
->config
.float_mode
));
860 si_pm4_set_reg(pm4
, R_00B02C_SPI_SHADER_PGM_RSRC2_PS
,
861 S_00B02C_EXTRA_LDS_SIZE(shader
->config
.lds_size
) |
862 S_00B02C_USER_SGPR(SI_PS_NUM_USER_SGPR
) |
863 S_00B32C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
866 static void si_shader_init_pm4_state(struct si_screen
*sscreen
,
867 struct si_shader
*shader
)
869 switch (shader
->selector
->type
) {
870 case PIPE_SHADER_VERTEX
:
871 if (shader
->key
.as_ls
)
872 si_shader_ls(shader
);
873 else if (shader
->key
.as_es
)
874 si_shader_es(sscreen
, shader
);
876 si_shader_vs(sscreen
, shader
, NULL
);
878 case PIPE_SHADER_TESS_CTRL
:
879 si_shader_hs(shader
);
881 case PIPE_SHADER_TESS_EVAL
:
882 if (shader
->key
.as_es
)
883 si_shader_es(sscreen
, shader
);
885 si_shader_vs(sscreen
, shader
, NULL
);
887 case PIPE_SHADER_GEOMETRY
:
888 si_shader_gs(shader
);
890 case PIPE_SHADER_FRAGMENT
:
891 si_shader_ps(shader
);
898 static unsigned si_get_alpha_test_func(struct si_context
*sctx
)
900 /* Alpha-test should be disabled if colorbuffer 0 is integer. */
901 if (sctx
->queued
.named
.dsa
)
902 return sctx
->queued
.named
.dsa
->alpha_func
;
904 return PIPE_FUNC_ALWAYS
;
907 static void si_shader_selector_key_hw_vs(struct si_context
*sctx
,
908 struct si_shader_selector
*vs
,
909 struct si_shader_key
*key
)
911 struct si_shader_selector
*ps
= sctx
->ps_shader
.cso
;
913 key
->opt
.hw_vs
.clip_disable
=
914 sctx
->queued
.named
.rasterizer
->clip_plane_enable
== 0 &&
915 (vs
->info
.clipdist_writemask
||
916 vs
->info
.writes_clipvertex
) &&
917 !vs
->info
.culldist_writemask
;
919 /* Find out if PS is disabled. */
920 bool ps_disabled
= true;
922 bool ps_modifies_zs
= ps
->info
.uses_kill
||
924 ps
->info
.writes_stencil
||
925 ps
->info
.writes_samplemask
||
926 si_get_alpha_test_func(sctx
) != PIPE_FUNC_ALWAYS
;
928 unsigned ps_colormask
= sctx
->framebuffer
.colorbuf_enabled_4bit
&
929 sctx
->queued
.named
.blend
->cb_target_mask
;
930 if (!ps
->info
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
])
931 ps_colormask
&= ps
->colors_written_4bit
;
933 ps_disabled
= sctx
->queued
.named
.rasterizer
->rasterizer_discard
||
936 !ps
->info
.writes_memory
);
939 /* Find out which VS outputs aren't used by the PS. */
940 uint64_t outputs_written
= vs
->outputs_written
;
941 uint32_t outputs_written2
= vs
->outputs_written2
;
942 uint64_t inputs_read
= 0;
943 uint32_t inputs_read2
= 0;
945 outputs_written
&= ~0x3; /* ignore POSITION, PSIZE */
948 inputs_read
= ps
->inputs_read
;
949 inputs_read2
= ps
->inputs_read2
;
952 uint64_t linked
= outputs_written
& inputs_read
;
953 uint32_t linked2
= outputs_written2
& inputs_read2
;
955 key
->opt
.hw_vs
.kill_outputs
= ~linked
& outputs_written
;
956 key
->opt
.hw_vs
.kill_outputs2
= ~linked2
& outputs_written2
;
959 /* Compute the key for the hw shader variant */
960 static inline void si_shader_selector_key(struct pipe_context
*ctx
,
961 struct si_shader_selector
*sel
,
962 struct si_shader_key
*key
)
964 struct si_context
*sctx
= (struct si_context
*)ctx
;
967 memset(key
, 0, sizeof(*key
));
970 case PIPE_SHADER_VERTEX
:
971 if (sctx
->vertex_elements
) {
972 unsigned count
= MIN2(sel
->info
.num_inputs
,
973 sctx
->vertex_elements
->count
);
974 for (i
= 0; i
< count
; ++i
)
975 key
->part
.vs
.prolog
.instance_divisors
[i
] =
976 sctx
->vertex_elements
->elements
[i
].instance_divisor
;
978 memcpy(key
->mono
.vs
.fix_fetch
,
979 sctx
->vertex_elements
->fix_fetch
, count
);
981 if (sctx
->tes_shader
.cso
)
983 else if (sctx
->gs_shader
.cso
)
986 si_shader_selector_key_hw_vs(sctx
, sel
, key
);
988 if (sctx
->ps_shader
.cso
&& sctx
->ps_shader
.cso
->info
.uses_primid
)
989 key
->part
.vs
.epilog
.export_prim_id
= 1;
992 case PIPE_SHADER_TESS_CTRL
:
993 key
->part
.tcs
.epilog
.prim_mode
=
994 sctx
->tes_shader
.cso
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
996 if (sel
== sctx
->fixed_func_tcs_shader
.cso
)
997 key
->mono
.tcs
.inputs_to_copy
= sctx
->vs_shader
.cso
->outputs_written
;
999 case PIPE_SHADER_TESS_EVAL
:
1000 if (sctx
->gs_shader
.cso
)
1003 si_shader_selector_key_hw_vs(sctx
, sel
, key
);
1005 if (sctx
->ps_shader
.cso
&& sctx
->ps_shader
.cso
->info
.uses_primid
)
1006 key
->part
.tes
.epilog
.export_prim_id
= 1;
1009 case PIPE_SHADER_GEOMETRY
:
1010 key
->part
.gs
.prolog
.tri_strip_adj_fix
= sctx
->gs_tri_strip_adj_fix
;
1012 case PIPE_SHADER_FRAGMENT
: {
1013 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
1014 struct si_state_blend
*blend
= sctx
->queued
.named
.blend
;
1016 if (sel
->info
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
] &&
1017 sel
->info
.colors_written
== 0x1)
1018 key
->part
.ps
.epilog
.last_cbuf
= MAX2(sctx
->framebuffer
.state
.nr_cbufs
, 1) - 1;
1021 /* Select the shader color format based on whether
1022 * blending or alpha are needed.
1024 key
->part
.ps
.epilog
.spi_shader_col_format
=
1025 (blend
->blend_enable_4bit
& blend
->need_src_alpha_4bit
&
1026 sctx
->framebuffer
.spi_shader_col_format_blend_alpha
) |
1027 (blend
->blend_enable_4bit
& ~blend
->need_src_alpha_4bit
&
1028 sctx
->framebuffer
.spi_shader_col_format_blend
) |
1029 (~blend
->blend_enable_4bit
& blend
->need_src_alpha_4bit
&
1030 sctx
->framebuffer
.spi_shader_col_format_alpha
) |
1031 (~blend
->blend_enable_4bit
& ~blend
->need_src_alpha_4bit
&
1032 sctx
->framebuffer
.spi_shader_col_format
);
1034 /* The output for dual source blending should have
1035 * the same format as the first output.
1037 if (blend
->dual_src_blend
)
1038 key
->part
.ps
.epilog
.spi_shader_col_format
|=
1039 (key
->part
.ps
.epilog
.spi_shader_col_format
& 0xf) << 4;
1041 key
->part
.ps
.epilog
.spi_shader_col_format
= sctx
->framebuffer
.spi_shader_col_format
;
1043 /* If alpha-to-coverage is enabled, we have to export alpha
1044 * even if there is no color buffer.
1046 if (!(key
->part
.ps
.epilog
.spi_shader_col_format
& 0xf) &&
1047 blend
&& blend
->alpha_to_coverage
)
1048 key
->part
.ps
.epilog
.spi_shader_col_format
|= V_028710_SPI_SHADER_32_AR
;
1050 /* On SI and CIK except Hawaii, the CB doesn't clamp outputs
1051 * to the range supported by the type if a channel has less
1052 * than 16 bits and the export format is 16_ABGR.
1054 if (sctx
->b
.chip_class
<= CIK
&& sctx
->b
.family
!= CHIP_HAWAII
)
1055 key
->part
.ps
.epilog
.color_is_int8
= sctx
->framebuffer
.color_is_int8
;
1057 /* Disable unwritten outputs (if WRITE_ALL_CBUFS isn't enabled). */
1058 if (!key
->part
.ps
.epilog
.last_cbuf
) {
1059 key
->part
.ps
.epilog
.spi_shader_col_format
&= sel
->colors_written_4bit
;
1060 key
->part
.ps
.epilog
.color_is_int8
&= sel
->info
.colors_written
;
1064 bool is_poly
= (sctx
->current_rast_prim
>= PIPE_PRIM_TRIANGLES
&&
1065 sctx
->current_rast_prim
<= PIPE_PRIM_POLYGON
) ||
1066 sctx
->current_rast_prim
>= PIPE_PRIM_TRIANGLES_ADJACENCY
;
1067 bool is_line
= !is_poly
&& sctx
->current_rast_prim
!= PIPE_PRIM_POINTS
;
1069 key
->part
.ps
.prolog
.color_two_side
= rs
->two_side
&& sel
->info
.colors_read
;
1070 key
->part
.ps
.prolog
.flatshade_colors
= rs
->flatshade
&& sel
->info
.colors_read
;
1072 if (sctx
->queued
.named
.blend
) {
1073 key
->part
.ps
.epilog
.alpha_to_one
= sctx
->queued
.named
.blend
->alpha_to_one
&&
1074 rs
->multisample_enable
;
1077 key
->part
.ps
.prolog
.poly_stipple
= rs
->poly_stipple_enable
&& is_poly
;
1078 key
->part
.ps
.epilog
.poly_line_smoothing
= ((is_poly
&& rs
->poly_smooth
) ||
1079 (is_line
&& rs
->line_smooth
)) &&
1080 sctx
->framebuffer
.nr_samples
<= 1;
1081 key
->part
.ps
.epilog
.clamp_color
= rs
->clamp_fragment_color
;
1083 if (rs
->force_persample_interp
&&
1084 rs
->multisample_enable
&&
1085 sctx
->framebuffer
.nr_samples
> 1 &&
1086 sctx
->ps_iter_samples
> 1) {
1087 key
->part
.ps
.prolog
.force_persp_sample_interp
=
1088 sel
->info
.uses_persp_center
||
1089 sel
->info
.uses_persp_centroid
;
1091 key
->part
.ps
.prolog
.force_linear_sample_interp
=
1092 sel
->info
.uses_linear_center
||
1093 sel
->info
.uses_linear_centroid
;
1094 } else if (rs
->multisample_enable
&&
1095 sctx
->framebuffer
.nr_samples
> 1) {
1096 key
->part
.ps
.prolog
.bc_optimize_for_persp
=
1097 sel
->info
.uses_persp_center
&&
1098 sel
->info
.uses_persp_centroid
;
1099 key
->part
.ps
.prolog
.bc_optimize_for_linear
=
1100 sel
->info
.uses_linear_center
&&
1101 sel
->info
.uses_linear_centroid
;
1103 /* Make sure SPI doesn't compute more than 1 pair
1104 * of (i,j), which is the optimization here. */
1105 key
->part
.ps
.prolog
.force_persp_center_interp
=
1106 sel
->info
.uses_persp_center
+
1107 sel
->info
.uses_persp_centroid
+
1108 sel
->info
.uses_persp_sample
> 1;
1110 key
->part
.ps
.prolog
.force_linear_center_interp
=
1111 sel
->info
.uses_linear_center
+
1112 sel
->info
.uses_linear_centroid
+
1113 sel
->info
.uses_linear_sample
> 1;
1117 key
->part
.ps
.epilog
.alpha_func
= si_get_alpha_test_func(sctx
);
1125 static void si_build_shader_variant(void *job
, int thread_index
)
1127 struct si_shader
*shader
= (struct si_shader
*)job
;
1128 struct si_shader_selector
*sel
= shader
->selector
;
1129 struct si_screen
*sscreen
= sel
->screen
;
1130 LLVMTargetMachineRef tm
;
1131 struct pipe_debug_callback
*debug
= &shader
->compiler_ctx_state
.debug
;
1134 if (thread_index
>= 0) {
1135 assert(thread_index
< ARRAY_SIZE(sscreen
->tm
));
1136 tm
= sscreen
->tm
[thread_index
];
1140 tm
= shader
->compiler_ctx_state
.tm
;
1143 r
= si_shader_create(sscreen
, tm
, shader
, debug
);
1145 R600_ERR("Failed to build shader variant (type=%u) %d\n",
1147 shader
->compilation_failed
= true;
1151 if (shader
->compiler_ctx_state
.is_debug_context
) {
1152 FILE *f
= open_memstream(&shader
->shader_log
,
1153 &shader
->shader_log_size
);
1155 si_shader_dump(sscreen
, shader
, NULL
, sel
->type
, f
, false);
1160 si_shader_init_pm4_state(sscreen
, shader
);
1163 /* Select the hw shader variant depending on the current state. */
1164 static int si_shader_select_with_key(struct si_screen
*sscreen
,
1165 struct si_shader_ctx_state
*state
,
1166 struct si_compiler_ctx_state
*compiler_state
,
1167 struct si_shader_key
*key
,
1170 static const struct si_shader_key zeroed
;
1171 struct si_shader_selector
*sel
= state
->cso
;
1172 struct si_shader
*current
= state
->current
;
1173 struct si_shader
*iter
, *shader
= NULL
;
1175 if (unlikely(sscreen
->b
.debug_flags
& DBG_NO_OPT_VARIANT
)) {
1176 memset(&key
->opt
, 0, sizeof(key
->opt
));
1180 /* Check if we don't need to change anything.
1181 * This path is also used for most shaders that don't need multiple
1182 * variants, it will cost just a computation of the key and this
1184 if (likely(current
&&
1185 memcmp(¤t
->key
, key
, sizeof(*key
)) == 0 &&
1186 (!current
->is_optimized
||
1187 util_queue_fence_is_signalled(¤t
->optimized_ready
))))
1190 /* This must be done before the mutex is locked, because async GS
1191 * compilation calls this function too, and therefore must enter
1194 * Only wait if we are in a draw call. Don't wait if we are
1195 * in a compiler thread.
1197 if (thread_index
< 0)
1198 util_queue_job_wait(&sel
->ready
);
1200 pipe_mutex_lock(sel
->mutex
);
1202 /* Find the shader variant. */
1203 for (iter
= sel
->first_variant
; iter
; iter
= iter
->next_variant
) {
1204 /* Don't check the "current" shader. We checked it above. */
1205 if (current
!= iter
&&
1206 memcmp(&iter
->key
, key
, sizeof(*key
)) == 0) {
1207 /* If it's an optimized shader and its compilation has
1208 * been started but isn't done, use the unoptimized
1209 * shader so as not to cause a stall due to compilation.
1211 if (iter
->is_optimized
&&
1212 !util_queue_fence_is_signalled(&iter
->optimized_ready
)) {
1213 memset(&key
->opt
, 0, sizeof(key
->opt
));
1214 pipe_mutex_unlock(sel
->mutex
);
1218 if (iter
->compilation_failed
) {
1219 pipe_mutex_unlock(sel
->mutex
);
1220 return -1; /* skip the draw call */
1223 state
->current
= iter
;
1224 pipe_mutex_unlock(sel
->mutex
);
1229 /* Build a new shader. */
1230 shader
= CALLOC_STRUCT(si_shader
);
1232 pipe_mutex_unlock(sel
->mutex
);
1235 shader
->selector
= sel
;
1237 shader
->compiler_ctx_state
= *compiler_state
;
1239 bool is_pure_monolithic
=
1240 memcmp(&key
->mono
, &zeroed
.mono
, sizeof(key
->mono
)) != 0;
1242 /* Monolithic-only shaders don't make a distinction between optimized
1243 * and unoptimized. */
1244 shader
->is_monolithic
=
1245 !sel
->main_shader_part
||
1246 sel
->main_shader_part
->key
.as_ls
!= key
->as_ls
||
1247 sel
->main_shader_part
->key
.as_es
!= key
->as_es
||
1248 is_pure_monolithic
||
1249 memcmp(&key
->opt
, &zeroed
.opt
, sizeof(key
->opt
)) != 0;
1251 shader
->is_optimized
=
1252 !sscreen
->use_monolithic_shaders
&&
1253 memcmp(&key
->opt
, &zeroed
.opt
, sizeof(key
->opt
)) != 0;
1254 if (shader
->is_optimized
)
1255 util_queue_fence_init(&shader
->optimized_ready
);
1257 if (!sel
->last_variant
) {
1258 sel
->first_variant
= shader
;
1259 sel
->last_variant
= shader
;
1261 sel
->last_variant
->next_variant
= shader
;
1262 sel
->last_variant
= shader
;
1265 /* If it's an optimized shader, compile it asynchronously. */
1266 if (shader
->is_optimized
&&
1267 !is_pure_monolithic
&&
1269 /* Compile it asynchronously. */
1270 util_queue_add_job(&sscreen
->shader_compiler_queue
,
1271 shader
, &shader
->optimized_ready
,
1272 si_build_shader_variant
, NULL
);
1274 /* Use the default (unoptimized) shader for now. */
1275 memset(&key
->opt
, 0, sizeof(key
->opt
));
1276 pipe_mutex_unlock(sel
->mutex
);
1280 assert(!shader
->is_optimized
);
1281 si_build_shader_variant(shader
, thread_index
);
1283 if (!shader
->compilation_failed
)
1284 state
->current
= shader
;
1286 pipe_mutex_unlock(sel
->mutex
);
1287 return shader
->compilation_failed
? -1 : 0;
1290 static int si_shader_select(struct pipe_context
*ctx
,
1291 struct si_shader_ctx_state
*state
,
1292 struct si_compiler_ctx_state
*compiler_state
)
1294 struct si_context
*sctx
= (struct si_context
*)ctx
;
1295 struct si_shader_key key
;
1297 si_shader_selector_key(ctx
, state
->cso
, &key
);
1298 return si_shader_select_with_key(sctx
->screen
, state
, compiler_state
,
1302 static void si_parse_next_shader_property(const struct tgsi_shader_info
*info
,
1303 struct si_shader_key
*key
)
1305 unsigned next_shader
= info
->properties
[TGSI_PROPERTY_NEXT_SHADER
];
1307 switch (info
->processor
) {
1308 case PIPE_SHADER_VERTEX
:
1309 switch (next_shader
) {
1310 case PIPE_SHADER_GEOMETRY
:
1313 case PIPE_SHADER_TESS_CTRL
:
1314 case PIPE_SHADER_TESS_EVAL
:
1318 /* If POSITION isn't written, it can't be a HW VS.
1319 * Assume that it's a HW LS. (the next shader is TCS)
1320 * This heuristic is needed for separate shader objects.
1322 if (!info
->writes_position
)
1327 case PIPE_SHADER_TESS_EVAL
:
1328 if (next_shader
== PIPE_SHADER_GEOMETRY
||
1329 !info
->writes_position
)
1336 * Compile the main shader part or the monolithic shader as part of
1337 * si_shader_selector initialization. Since it can be done asynchronously,
1338 * there is no way to report compile failures to applications.
1340 void si_init_shader_selector_async(void *job
, int thread_index
)
1342 struct si_shader_selector
*sel
= (struct si_shader_selector
*)job
;
1343 struct si_screen
*sscreen
= sel
->screen
;
1344 LLVMTargetMachineRef tm
;
1345 struct pipe_debug_callback
*debug
= &sel
->compiler_ctx_state
.debug
;
1348 if (thread_index
>= 0) {
1349 assert(thread_index
< ARRAY_SIZE(sscreen
->tm
));
1350 tm
= sscreen
->tm
[thread_index
];
1354 tm
= sel
->compiler_ctx_state
.tm
;
1357 /* Compile the main shader part for use with a prolog and/or epilog.
1358 * If this fails, the driver will try to compile a monolithic shader
1361 if (!sscreen
->use_monolithic_shaders
) {
1362 struct si_shader
*shader
= CALLOC_STRUCT(si_shader
);
1366 fprintf(stderr
, "radeonsi: can't allocate a main shader part\n");
1370 shader
->selector
= sel
;
1371 si_parse_next_shader_property(&sel
->info
, &shader
->key
);
1373 tgsi_binary
= si_get_tgsi_binary(sel
);
1375 /* Try to load the shader from the shader cache. */
1376 pipe_mutex_lock(sscreen
->shader_cache_mutex
);
1379 si_shader_cache_load_shader(sscreen
, tgsi_binary
, shader
)) {
1381 pipe_mutex_unlock(sscreen
->shader_cache_mutex
);
1383 pipe_mutex_unlock(sscreen
->shader_cache_mutex
);
1385 /* Compile the shader if it hasn't been loaded from the cache. */
1386 if (si_compile_tgsi_shader(sscreen
, tm
, shader
, false,
1390 fprintf(stderr
, "radeonsi: can't compile a main shader part\n");
1395 pipe_mutex_lock(sscreen
->shader_cache_mutex
);
1396 if (!si_shader_cache_insert_shader(sscreen
, tgsi_binary
, shader
))
1398 pipe_mutex_unlock(sscreen
->shader_cache_mutex
);
1402 sel
->main_shader_part
= shader
;
1404 /* Unset "outputs_written" flags for outputs converted to
1405 * DEFAULT_VAL, so that later inter-shader optimizations don't
1406 * try to eliminate outputs that don't exist in the final
1409 * This is only done if non-monolithic shaders are enabled.
1411 if ((sel
->type
== PIPE_SHADER_VERTEX
||
1412 sel
->type
== PIPE_SHADER_TESS_EVAL
) &&
1413 !shader
->key
.as_ls
&&
1414 !shader
->key
.as_es
) {
1417 for (i
= 0; i
< sel
->info
.num_outputs
; i
++) {
1418 unsigned offset
= shader
->info
.vs_output_param_offset
[i
];
1420 if (offset
<= EXP_PARAM_OFFSET_31
)
1423 unsigned name
= sel
->info
.output_semantic_name
[i
];
1424 unsigned index
= sel
->info
.output_semantic_index
[i
];
1428 case TGSI_SEMANTIC_GENERIC
:
1429 /* don't process indices the function can't handle */
1433 case TGSI_SEMANTIC_CLIPDIST
:
1434 id
= si_shader_io_get_unique_index(name
, index
);
1435 sel
->outputs_written
&= ~(1ull << id
);
1437 case TGSI_SEMANTIC_POSITION
: /* ignore these */
1438 case TGSI_SEMANTIC_PSIZE
:
1439 case TGSI_SEMANTIC_CLIPVERTEX
:
1440 case TGSI_SEMANTIC_EDGEFLAG
:
1443 id
= si_shader_io_get_unique_index2(name
, index
);
1444 sel
->outputs_written2
&= ~(1u << id
);
1450 /* Pre-compilation. */
1451 if (sscreen
->b
.debug_flags
& DBG_PRECOMPILE
) {
1452 struct si_shader_ctx_state state
= {sel
};
1453 struct si_shader_key key
;
1455 memset(&key
, 0, sizeof(key
));
1456 si_parse_next_shader_property(&sel
->info
, &key
);
1458 /* Set reasonable defaults, so that the shader key doesn't
1459 * cause any code to be eliminated.
1461 switch (sel
->type
) {
1462 case PIPE_SHADER_TESS_CTRL
:
1463 key
.part
.tcs
.epilog
.prim_mode
= PIPE_PRIM_TRIANGLES
;
1465 case PIPE_SHADER_FRAGMENT
:
1466 key
.part
.ps
.prolog
.bc_optimize_for_persp
=
1467 sel
->info
.uses_persp_center
&&
1468 sel
->info
.uses_persp_centroid
;
1469 key
.part
.ps
.prolog
.bc_optimize_for_linear
=
1470 sel
->info
.uses_linear_center
&&
1471 sel
->info
.uses_linear_centroid
;
1472 key
.part
.ps
.epilog
.alpha_func
= PIPE_FUNC_ALWAYS
;
1473 for (i
= 0; i
< 8; i
++)
1474 if (sel
->info
.colors_written
& (1 << i
))
1475 key
.part
.ps
.epilog
.spi_shader_col_format
|=
1476 V_028710_SPI_SHADER_FP16_ABGR
<< (i
* 4);
1480 if (si_shader_select_with_key(sscreen
, &state
,
1481 &sel
->compiler_ctx_state
, &key
,
1483 fprintf(stderr
, "radeonsi: can't create a monolithic shader\n");
1486 /* The GS copy shader is always pre-compiled. */
1487 if (sel
->type
== PIPE_SHADER_GEOMETRY
) {
1488 sel
->gs_copy_shader
= si_generate_gs_copy_shader(sscreen
, tm
, sel
, debug
);
1489 if (!sel
->gs_copy_shader
) {
1490 fprintf(stderr
, "radeonsi: can't create GS copy shader\n");
1494 si_shader_vs(sscreen
, sel
->gs_copy_shader
, sel
);
1498 static void *si_create_shader_selector(struct pipe_context
*ctx
,
1499 const struct pipe_shader_state
*state
)
1501 struct si_screen
*sscreen
= (struct si_screen
*)ctx
->screen
;
1502 struct si_context
*sctx
= (struct si_context
*)ctx
;
1503 struct si_shader_selector
*sel
= CALLOC_STRUCT(si_shader_selector
);
1509 sel
->screen
= sscreen
;
1510 sel
->compiler_ctx_state
.tm
= sctx
->tm
;
1511 sel
->compiler_ctx_state
.debug
= sctx
->b
.debug
;
1512 sel
->compiler_ctx_state
.is_debug_context
= sctx
->is_debug
;
1513 sel
->tokens
= tgsi_dup_tokens(state
->tokens
);
1519 sel
->so
= state
->stream_output
;
1520 tgsi_scan_shader(state
->tokens
, &sel
->info
);
1521 sel
->type
= sel
->info
.processor
;
1522 p_atomic_inc(&sscreen
->b
.num_shaders_created
);
1524 /* Set which opcode uses which (i,j) pair. */
1525 if (sel
->info
.uses_persp_opcode_interp_centroid
)
1526 sel
->info
.uses_persp_centroid
= true;
1528 if (sel
->info
.uses_linear_opcode_interp_centroid
)
1529 sel
->info
.uses_linear_centroid
= true;
1531 if (sel
->info
.uses_persp_opcode_interp_offset
||
1532 sel
->info
.uses_persp_opcode_interp_sample
)
1533 sel
->info
.uses_persp_center
= true;
1535 if (sel
->info
.uses_linear_opcode_interp_offset
||
1536 sel
->info
.uses_linear_opcode_interp_sample
)
1537 sel
->info
.uses_linear_center
= true;
1539 switch (sel
->type
) {
1540 case PIPE_SHADER_GEOMETRY
:
1541 sel
->gs_output_prim
=
1542 sel
->info
.properties
[TGSI_PROPERTY_GS_OUTPUT_PRIM
];
1543 sel
->gs_max_out_vertices
=
1544 sel
->info
.properties
[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES
];
1545 sel
->gs_num_invocations
=
1546 sel
->info
.properties
[TGSI_PROPERTY_GS_INVOCATIONS
];
1547 sel
->gsvs_vertex_size
= sel
->info
.num_outputs
* 16;
1548 sel
->max_gsvs_emit_size
= sel
->gsvs_vertex_size
*
1549 sel
->gs_max_out_vertices
;
1551 sel
->max_gs_stream
= 0;
1552 for (i
= 0; i
< sel
->so
.num_outputs
; i
++)
1553 sel
->max_gs_stream
= MAX2(sel
->max_gs_stream
,
1554 sel
->so
.output
[i
].stream
);
1556 sel
->gs_input_verts_per_prim
=
1557 u_vertices_per_prim(sel
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
]);
1560 case PIPE_SHADER_TESS_CTRL
:
1561 /* Always reserve space for these. */
1562 sel
->patch_outputs_written
|=
1563 (1llu << si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSINNER
, 0)) |
1564 (1llu << si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSOUTER
, 0));
1566 case PIPE_SHADER_VERTEX
:
1567 case PIPE_SHADER_TESS_EVAL
:
1568 for (i
= 0; i
< sel
->info
.num_outputs
; i
++) {
1569 unsigned name
= sel
->info
.output_semantic_name
[i
];
1570 unsigned index
= sel
->info
.output_semantic_index
[i
];
1573 case TGSI_SEMANTIC_TESSINNER
:
1574 case TGSI_SEMANTIC_TESSOUTER
:
1575 case TGSI_SEMANTIC_PATCH
:
1576 sel
->patch_outputs_written
|=
1577 1llu << si_shader_io_get_unique_index(name
, index
);
1580 case TGSI_SEMANTIC_GENERIC
:
1581 /* don't process indices the function can't handle */
1585 case TGSI_SEMANTIC_POSITION
:
1586 case TGSI_SEMANTIC_PSIZE
:
1587 case TGSI_SEMANTIC_CLIPDIST
:
1588 sel
->outputs_written
|=
1589 1llu << si_shader_io_get_unique_index(name
, index
);
1591 case TGSI_SEMANTIC_CLIPVERTEX
: /* ignore these */
1592 case TGSI_SEMANTIC_EDGEFLAG
:
1595 sel
->outputs_written2
|=
1596 1u << si_shader_io_get_unique_index2(name
, index
);
1599 sel
->esgs_itemsize
= util_last_bit64(sel
->outputs_written
) * 16;
1602 case PIPE_SHADER_FRAGMENT
:
1603 for (i
= 0; i
< sel
->info
.num_inputs
; i
++) {
1604 unsigned name
= sel
->info
.input_semantic_name
[i
];
1605 unsigned index
= sel
->info
.input_semantic_index
[i
];
1608 case TGSI_SEMANTIC_CLIPDIST
:
1609 case TGSI_SEMANTIC_GENERIC
:
1611 1llu << si_shader_io_get_unique_index(name
, index
);
1613 case TGSI_SEMANTIC_PCOORD
: /* ignore this */
1616 sel
->inputs_read2
|=
1617 1u << si_shader_io_get_unique_index2(name
, index
);
1621 for (i
= 0; i
< 8; i
++)
1622 if (sel
->info
.colors_written
& (1 << i
))
1623 sel
->colors_written_4bit
|= 0xf << (4 * i
);
1625 for (i
= 0; i
< sel
->info
.num_inputs
; i
++) {
1626 if (sel
->info
.input_semantic_name
[i
] == TGSI_SEMANTIC_COLOR
) {
1627 int index
= sel
->info
.input_semantic_index
[i
];
1628 sel
->color_attr_index
[index
] = i
;
1634 /* DB_SHADER_CONTROL */
1635 sel
->db_shader_control
=
1636 S_02880C_Z_EXPORT_ENABLE(sel
->info
.writes_z
) |
1637 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(sel
->info
.writes_stencil
) |
1638 S_02880C_MASK_EXPORT_ENABLE(sel
->info
.writes_samplemask
) |
1639 S_02880C_KILL_ENABLE(sel
->info
.uses_kill
);
1641 switch (sel
->info
.properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
]) {
1642 case TGSI_FS_DEPTH_LAYOUT_GREATER
:
1643 sel
->db_shader_control
|=
1644 S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_GREATER_THAN_Z
);
1646 case TGSI_FS_DEPTH_LAYOUT_LESS
:
1647 sel
->db_shader_control
|=
1648 S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_LESS_THAN_Z
);
1652 /* Z_ORDER, EXEC_ON_HIER_FAIL and EXEC_ON_NOOP should be set as following:
1654 * | early Z/S | writes_mem | allow_ReZ? | Z_ORDER | EXEC_ON_HIER_FAIL | EXEC_ON_NOOP
1655 * --|-----------|------------|------------|--------------------|-------------------|-------------
1656 * 1a| false | false | true | EarlyZ_Then_ReZ | 0 | 0
1657 * 1b| false | false | false | EarlyZ_Then_LateZ | 0 | 0
1658 * 2 | false | true | n/a | LateZ | 1 | 0
1659 * 3 | true | false | n/a | EarlyZ_Then_LateZ | 0 | 0
1660 * 4 | true | true | n/a | EarlyZ_Then_LateZ | 0 | 1
1662 * In cases 3 and 4, HW will force Z_ORDER to EarlyZ regardless of what's set in the register.
1663 * In case 2, NOOP_CULL is a don't care field. In case 2, 3 and 4, ReZ doesn't make sense.
1665 * Don't use ReZ without profiling !!!
1667 * ReZ decreases performance by 15% in DiRT: Showdown on Ultra settings, which has pretty complex
1670 if (sel
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
]) {
1672 sel
->db_shader_control
|= S_02880C_DEPTH_BEFORE_SHADER(1) |
1673 S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z
) |
1674 S_02880C_EXEC_ON_NOOP(sel
->info
.writes_memory
);
1675 } else if (sel
->info
.writes_memory
) {
1677 sel
->db_shader_control
|= S_02880C_Z_ORDER(V_02880C_LATE_Z
) |
1678 S_02880C_EXEC_ON_HIER_FAIL(1);
1681 sel
->db_shader_control
|= S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z
);
1684 pipe_mutex_init(sel
->mutex
);
1685 util_queue_fence_init(&sel
->ready
);
1687 if ((sctx
->b
.debug
.debug_message
&& !sctx
->b
.debug
.async
) ||
1689 r600_can_dump_shader(&sscreen
->b
, sel
->info
.processor
) ||
1690 !util_queue_is_initialized(&sscreen
->shader_compiler_queue
))
1691 si_init_shader_selector_async(sel
, -1);
1693 util_queue_add_job(&sscreen
->shader_compiler_queue
, sel
,
1694 &sel
->ready
, si_init_shader_selector_async
,
1700 static void si_bind_vs_shader(struct pipe_context
*ctx
, void *state
)
1702 struct si_context
*sctx
= (struct si_context
*)ctx
;
1703 struct si_shader_selector
*sel
= state
;
1705 if (sctx
->vs_shader
.cso
== sel
)
1708 sctx
->vs_shader
.cso
= sel
;
1709 sctx
->vs_shader
.current
= sel
? sel
->first_variant
: NULL
;
1710 sctx
->do_update_shaders
= true;
1711 si_mark_atom_dirty(sctx
, &sctx
->clip_regs
);
1712 r600_update_vs_writes_viewport_index(&sctx
->b
, si_get_vs_info(sctx
));
1715 static void si_bind_gs_shader(struct pipe_context
*ctx
, void *state
)
1717 struct si_context
*sctx
= (struct si_context
*)ctx
;
1718 struct si_shader_selector
*sel
= state
;
1719 bool enable_changed
= !!sctx
->gs_shader
.cso
!= !!sel
;
1721 if (sctx
->gs_shader
.cso
== sel
)
1724 sctx
->gs_shader
.cso
= sel
;
1725 sctx
->gs_shader
.current
= sel
? sel
->first_variant
: NULL
;
1726 sctx
->ia_multi_vgt_param_key
.u
.uses_gs
= sel
!= NULL
;
1727 sctx
->do_update_shaders
= true;
1728 si_mark_atom_dirty(sctx
, &sctx
->clip_regs
);
1729 sctx
->last_rast_prim
= -1; /* reset this so that it gets updated */
1732 si_shader_change_notify(sctx
);
1733 r600_update_vs_writes_viewport_index(&sctx
->b
, si_get_vs_info(sctx
));
1736 static void si_update_tcs_tes_uses_prim_id(struct si_context
*sctx
)
1738 sctx
->ia_multi_vgt_param_key
.u
.tcs_tes_uses_prim_id
=
1739 (sctx
->tes_shader
.cso
&&
1740 sctx
->tes_shader
.cso
->info
.uses_primid
) ||
1741 (sctx
->tcs_shader
.cso
&&
1742 sctx
->tcs_shader
.cso
->info
.uses_primid
);
1745 static void si_bind_tcs_shader(struct pipe_context
*ctx
, void *state
)
1747 struct si_context
*sctx
= (struct si_context
*)ctx
;
1748 struct si_shader_selector
*sel
= state
;
1749 bool enable_changed
= !!sctx
->tcs_shader
.cso
!= !!sel
;
1751 if (sctx
->tcs_shader
.cso
== sel
)
1754 sctx
->tcs_shader
.cso
= sel
;
1755 sctx
->tcs_shader
.current
= sel
? sel
->first_variant
: NULL
;
1756 si_update_tcs_tes_uses_prim_id(sctx
);
1757 sctx
->do_update_shaders
= true;
1760 sctx
->last_tcs
= NULL
; /* invalidate derived tess state */
1763 static void si_bind_tes_shader(struct pipe_context
*ctx
, void *state
)
1765 struct si_context
*sctx
= (struct si_context
*)ctx
;
1766 struct si_shader_selector
*sel
= state
;
1767 bool enable_changed
= !!sctx
->tes_shader
.cso
!= !!sel
;
1769 if (sctx
->tes_shader
.cso
== sel
)
1772 sctx
->tes_shader
.cso
= sel
;
1773 sctx
->tes_shader
.current
= sel
? sel
->first_variant
: NULL
;
1774 sctx
->ia_multi_vgt_param_key
.u
.uses_tess
= sel
!= NULL
;
1775 si_update_tcs_tes_uses_prim_id(sctx
);
1776 sctx
->do_update_shaders
= true;
1777 si_mark_atom_dirty(sctx
, &sctx
->clip_regs
);
1778 sctx
->last_rast_prim
= -1; /* reset this so that it gets updated */
1780 if (enable_changed
) {
1781 si_shader_change_notify(sctx
);
1782 sctx
->last_tes_sh_base
= -1; /* invalidate derived tess state */
1784 r600_update_vs_writes_viewport_index(&sctx
->b
, si_get_vs_info(sctx
));
1787 static void si_bind_ps_shader(struct pipe_context
*ctx
, void *state
)
1789 struct si_context
*sctx
= (struct si_context
*)ctx
;
1790 struct si_shader_selector
*sel
= state
;
1792 /* skip if supplied shader is one already in use */
1793 if (sctx
->ps_shader
.cso
== sel
)
1796 sctx
->ps_shader
.cso
= sel
;
1797 sctx
->ps_shader
.current
= sel
? sel
->first_variant
: NULL
;
1798 sctx
->do_update_shaders
= true;
1799 si_mark_atom_dirty(sctx
, &sctx
->cb_render_state
);
1802 static void si_delete_shader(struct si_context
*sctx
, struct si_shader
*shader
)
1804 if (shader
->is_optimized
) {
1805 util_queue_job_wait(&shader
->optimized_ready
);
1806 util_queue_fence_destroy(&shader
->optimized_ready
);
1810 switch (shader
->selector
->type
) {
1811 case PIPE_SHADER_VERTEX
:
1812 if (shader
->key
.as_ls
)
1813 si_pm4_delete_state(sctx
, ls
, shader
->pm4
);
1814 else if (shader
->key
.as_es
)
1815 si_pm4_delete_state(sctx
, es
, shader
->pm4
);
1817 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
1819 case PIPE_SHADER_TESS_CTRL
:
1820 si_pm4_delete_state(sctx
, hs
, shader
->pm4
);
1822 case PIPE_SHADER_TESS_EVAL
:
1823 if (shader
->key
.as_es
)
1824 si_pm4_delete_state(sctx
, es
, shader
->pm4
);
1826 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
1828 case PIPE_SHADER_GEOMETRY
:
1829 if (shader
->is_gs_copy_shader
)
1830 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
1832 si_pm4_delete_state(sctx
, gs
, shader
->pm4
);
1834 case PIPE_SHADER_FRAGMENT
:
1835 si_pm4_delete_state(sctx
, ps
, shader
->pm4
);
1840 si_shader_destroy(shader
);
1844 static void si_delete_shader_selector(struct pipe_context
*ctx
, void *state
)
1846 struct si_context
*sctx
= (struct si_context
*)ctx
;
1847 struct si_shader_selector
*sel
= (struct si_shader_selector
*)state
;
1848 struct si_shader
*p
= sel
->first_variant
, *c
;
1849 struct si_shader_ctx_state
*current_shader
[SI_NUM_SHADERS
] = {
1850 [PIPE_SHADER_VERTEX
] = &sctx
->vs_shader
,
1851 [PIPE_SHADER_TESS_CTRL
] = &sctx
->tcs_shader
,
1852 [PIPE_SHADER_TESS_EVAL
] = &sctx
->tes_shader
,
1853 [PIPE_SHADER_GEOMETRY
] = &sctx
->gs_shader
,
1854 [PIPE_SHADER_FRAGMENT
] = &sctx
->ps_shader
,
1857 util_queue_job_wait(&sel
->ready
);
1859 if (current_shader
[sel
->type
]->cso
== sel
) {
1860 current_shader
[sel
->type
]->cso
= NULL
;
1861 current_shader
[sel
->type
]->current
= NULL
;
1865 c
= p
->next_variant
;
1866 si_delete_shader(sctx
, p
);
1870 if (sel
->main_shader_part
)
1871 si_delete_shader(sctx
, sel
->main_shader_part
);
1872 if (sel
->gs_copy_shader
)
1873 si_delete_shader(sctx
, sel
->gs_copy_shader
);
1875 util_queue_fence_destroy(&sel
->ready
);
1876 pipe_mutex_destroy(sel
->mutex
);
1881 static unsigned si_get_ps_input_cntl(struct si_context
*sctx
,
1882 struct si_shader
*vs
, unsigned name
,
1883 unsigned index
, unsigned interpolate
)
1885 struct tgsi_shader_info
*vsinfo
= &vs
->selector
->info
;
1886 unsigned j
, offset
, ps_input_cntl
= 0;
1888 if (interpolate
== TGSI_INTERPOLATE_CONSTANT
||
1889 (interpolate
== TGSI_INTERPOLATE_COLOR
&& sctx
->flatshade
))
1890 ps_input_cntl
|= S_028644_FLAT_SHADE(1);
1892 if (name
== TGSI_SEMANTIC_PCOORD
||
1893 (name
== TGSI_SEMANTIC_TEXCOORD
&&
1894 sctx
->sprite_coord_enable
& (1 << index
))) {
1895 ps_input_cntl
|= S_028644_PT_SPRITE_TEX(1);
1898 for (j
= 0; j
< vsinfo
->num_outputs
; j
++) {
1899 if (name
== vsinfo
->output_semantic_name
[j
] &&
1900 index
== vsinfo
->output_semantic_index
[j
]) {
1901 offset
= vs
->info
.vs_output_param_offset
[j
];
1903 if (offset
<= EXP_PARAM_OFFSET_31
) {
1904 /* The input is loaded from parameter memory. */
1905 ps_input_cntl
|= S_028644_OFFSET(offset
);
1906 } else if (!G_028644_PT_SPRITE_TEX(ps_input_cntl
)) {
1907 if (offset
== EXP_PARAM_UNDEFINED
) {
1908 /* This can happen with depth-only rendering. */
1911 /* The input is a DEFAULT_VAL constant. */
1912 assert(offset
>= EXP_PARAM_DEFAULT_VAL_0000
&&
1913 offset
<= EXP_PARAM_DEFAULT_VAL_1111
);
1914 offset
-= EXP_PARAM_DEFAULT_VAL_0000
;
1917 ps_input_cntl
= S_028644_OFFSET(0x20) |
1918 S_028644_DEFAULT_VAL(offset
);
1924 if (name
== TGSI_SEMANTIC_PRIMID
)
1925 /* PrimID is written after the last output. */
1926 ps_input_cntl
|= S_028644_OFFSET(vs
->info
.vs_output_param_offset
[vsinfo
->num_outputs
]);
1927 else if (j
== vsinfo
->num_outputs
&& !G_028644_PT_SPRITE_TEX(ps_input_cntl
)) {
1928 /* No corresponding output found, load defaults into input.
1929 * Don't set any other bits.
1930 * (FLAT_SHADE=1 completely changes behavior) */
1931 ps_input_cntl
= S_028644_OFFSET(0x20);
1932 /* D3D 9 behaviour. GL is undefined */
1933 if (name
== TGSI_SEMANTIC_COLOR
&& index
== 0)
1934 ps_input_cntl
|= S_028644_DEFAULT_VAL(3);
1936 return ps_input_cntl
;
1939 static void si_emit_spi_map(struct si_context
*sctx
, struct r600_atom
*atom
)
1941 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
1942 struct si_shader
*ps
= sctx
->ps_shader
.current
;
1943 struct si_shader
*vs
= si_get_vs_state(sctx
);
1944 struct tgsi_shader_info
*psinfo
= ps
? &ps
->selector
->info
: NULL
;
1945 unsigned i
, num_interp
, num_written
= 0, bcol_interp
[2];
1947 if (!ps
|| !ps
->selector
->info
.num_inputs
)
1950 num_interp
= si_get_ps_num_interp(ps
);
1951 assert(num_interp
> 0);
1952 radeon_set_context_reg_seq(cs
, R_028644_SPI_PS_INPUT_CNTL_0
, num_interp
);
1954 for (i
= 0; i
< psinfo
->num_inputs
; i
++) {
1955 unsigned name
= psinfo
->input_semantic_name
[i
];
1956 unsigned index
= psinfo
->input_semantic_index
[i
];
1957 unsigned interpolate
= psinfo
->input_interpolate
[i
];
1959 radeon_emit(cs
, si_get_ps_input_cntl(sctx
, vs
, name
, index
,
1963 if (name
== TGSI_SEMANTIC_COLOR
) {
1964 assert(index
< ARRAY_SIZE(bcol_interp
));
1965 bcol_interp
[index
] = interpolate
;
1969 if (ps
->key
.part
.ps
.prolog
.color_two_side
) {
1970 unsigned bcol
= TGSI_SEMANTIC_BCOLOR
;
1972 for (i
= 0; i
< 2; i
++) {
1973 if (!(psinfo
->colors_read
& (0xf << (i
* 4))))
1976 radeon_emit(cs
, si_get_ps_input_cntl(sctx
, vs
, bcol
,
1977 i
, bcol_interp
[i
]));
1981 assert(num_interp
== num_written
);
1985 * Writing CONFIG or UCONFIG VGT registers requires VGT_FLUSH before that.
1987 static void si_init_config_add_vgt_flush(struct si_context
*sctx
)
1989 if (sctx
->init_config_has_vgt_flush
)
1992 /* Done by Vulkan before VGT_FLUSH. */
1993 si_pm4_cmd_begin(sctx
->init_config
, PKT3_EVENT_WRITE
);
1994 si_pm4_cmd_add(sctx
->init_config
,
1995 EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
1996 si_pm4_cmd_end(sctx
->init_config
, false);
1998 /* VGT_FLUSH is required even if VGT is idle. It resets VGT pointers. */
1999 si_pm4_cmd_begin(sctx
->init_config
, PKT3_EVENT_WRITE
);
2000 si_pm4_cmd_add(sctx
->init_config
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
2001 si_pm4_cmd_end(sctx
->init_config
, false);
2002 sctx
->init_config_has_vgt_flush
= true;
2005 /* Initialize state related to ESGS / GSVS ring buffers */
2006 static bool si_update_gs_ring_buffers(struct si_context
*sctx
)
2008 struct si_shader_selector
*es
=
2009 sctx
->tes_shader
.cso
? sctx
->tes_shader
.cso
: sctx
->vs_shader
.cso
;
2010 struct si_shader_selector
*gs
= sctx
->gs_shader
.cso
;
2011 struct si_pm4_state
*pm4
;
2013 /* Chip constants. */
2014 unsigned num_se
= sctx
->screen
->b
.info
.max_se
;
2015 unsigned wave_size
= 64;
2016 unsigned max_gs_waves
= 32 * num_se
; /* max 32 per SE on GCN */
2017 unsigned gs_vertex_reuse
= 16 * num_se
; /* GS_VERTEX_REUSE register (per SE) */
2018 unsigned alignment
= 256 * num_se
;
2019 /* The maximum size is 63.999 MB per SE. */
2020 unsigned max_size
= ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se
;
2022 /* Calculate the minimum size. */
2023 unsigned min_esgs_ring_size
= align(es
->esgs_itemsize
* gs_vertex_reuse
*
2024 wave_size
, alignment
);
2026 /* These are recommended sizes, not minimum sizes. */
2027 unsigned esgs_ring_size
= max_gs_waves
* 2 * wave_size
*
2028 es
->esgs_itemsize
* gs
->gs_input_verts_per_prim
;
2029 unsigned gsvs_ring_size
= max_gs_waves
* 2 * wave_size
*
2030 gs
->max_gsvs_emit_size
;
2032 min_esgs_ring_size
= align(min_esgs_ring_size
, alignment
);
2033 esgs_ring_size
= align(esgs_ring_size
, alignment
);
2034 gsvs_ring_size
= align(gsvs_ring_size
, alignment
);
2036 esgs_ring_size
= CLAMP(esgs_ring_size
, min_esgs_ring_size
, max_size
);
2037 gsvs_ring_size
= MIN2(gsvs_ring_size
, max_size
);
2039 /* Some rings don't have to be allocated if shaders don't use them.
2040 * (e.g. no varyings between ES and GS or GS and VS)
2042 bool update_esgs
= esgs_ring_size
&&
2043 (!sctx
->esgs_ring
||
2044 sctx
->esgs_ring
->width0
< esgs_ring_size
);
2045 bool update_gsvs
= gsvs_ring_size
&&
2046 (!sctx
->gsvs_ring
||
2047 sctx
->gsvs_ring
->width0
< gsvs_ring_size
);
2049 if (!update_esgs
&& !update_gsvs
)
2053 pipe_resource_reference(&sctx
->esgs_ring
, NULL
);
2054 sctx
->esgs_ring
= pipe_buffer_create(sctx
->b
.b
.screen
, 0,
2057 if (!sctx
->esgs_ring
)
2062 pipe_resource_reference(&sctx
->gsvs_ring
, NULL
);
2063 sctx
->gsvs_ring
= pipe_buffer_create(sctx
->b
.b
.screen
, 0,
2066 if (!sctx
->gsvs_ring
)
2070 /* Create the "init_config_gs_rings" state. */
2071 pm4
= CALLOC_STRUCT(si_pm4_state
);
2075 if (sctx
->b
.chip_class
>= CIK
) {
2076 if (sctx
->esgs_ring
)
2077 si_pm4_set_reg(pm4
, R_030900_VGT_ESGS_RING_SIZE
,
2078 sctx
->esgs_ring
->width0
/ 256);
2079 if (sctx
->gsvs_ring
)
2080 si_pm4_set_reg(pm4
, R_030904_VGT_GSVS_RING_SIZE
,
2081 sctx
->gsvs_ring
->width0
/ 256);
2083 if (sctx
->esgs_ring
)
2084 si_pm4_set_reg(pm4
, R_0088C8_VGT_ESGS_RING_SIZE
,
2085 sctx
->esgs_ring
->width0
/ 256);
2086 if (sctx
->gsvs_ring
)
2087 si_pm4_set_reg(pm4
, R_0088CC_VGT_GSVS_RING_SIZE
,
2088 sctx
->gsvs_ring
->width0
/ 256);
2091 /* Set the state. */
2092 if (sctx
->init_config_gs_rings
)
2093 si_pm4_free_state(sctx
, sctx
->init_config_gs_rings
, ~0);
2094 sctx
->init_config_gs_rings
= pm4
;
2096 if (!sctx
->init_config_has_vgt_flush
) {
2097 si_init_config_add_vgt_flush(sctx
);
2098 si_pm4_upload_indirect_buffer(sctx
, sctx
->init_config
);
2101 /* Flush the context to re-emit both init_config states. */
2102 sctx
->b
.initial_gfx_cs_size
= 0; /* force flush */
2103 si_context_gfx_flush(sctx
, RADEON_FLUSH_ASYNC
, NULL
);
2105 /* Set ring bindings. */
2106 if (sctx
->esgs_ring
) {
2107 si_set_ring_buffer(&sctx
->b
.b
, SI_ES_RING_ESGS
,
2108 sctx
->esgs_ring
, 0, sctx
->esgs_ring
->width0
,
2109 true, true, 4, 64, 0);
2110 si_set_ring_buffer(&sctx
->b
.b
, SI_GS_RING_ESGS
,
2111 sctx
->esgs_ring
, 0, sctx
->esgs_ring
->width0
,
2112 false, false, 0, 0, 0);
2114 if (sctx
->gsvs_ring
) {
2115 si_set_ring_buffer(&sctx
->b
.b
, SI_RING_GSVS
,
2116 sctx
->gsvs_ring
, 0, sctx
->gsvs_ring
->width0
,
2117 false, false, 0, 0, 0);
2124 * @returns 1 if \p sel has been updated to use a new scratch buffer
2126 * < 0 if there was a failure
2128 static int si_update_scratch_buffer(struct si_context
*sctx
,
2129 struct si_shader
*shader
)
2131 uint64_t scratch_va
= sctx
->scratch_buffer
->gpu_address
;
2137 /* This shader doesn't need a scratch buffer */
2138 if (shader
->config
.scratch_bytes_per_wave
== 0)
2141 /* This shader is already configured to use the current
2142 * scratch buffer. */
2143 if (shader
->scratch_bo
== sctx
->scratch_buffer
)
2146 assert(sctx
->scratch_buffer
);
2148 si_shader_apply_scratch_relocs(sctx
, shader
, &shader
->config
, scratch_va
);
2150 /* Replace the shader bo with a new bo that has the relocs applied. */
2151 r
= si_shader_binary_upload(sctx
->screen
, shader
);
2155 /* Update the shader state to use the new shader bo. */
2156 si_shader_init_pm4_state(sctx
->screen
, shader
);
2158 r600_resource_reference(&shader
->scratch_bo
, sctx
->scratch_buffer
);
2163 static unsigned si_get_current_scratch_buffer_size(struct si_context
*sctx
)
2165 return sctx
->scratch_buffer
? sctx
->scratch_buffer
->b
.b
.width0
: 0;
2168 static unsigned si_get_scratch_buffer_bytes_per_wave(struct si_shader
*shader
)
2170 return shader
? shader
->config
.scratch_bytes_per_wave
: 0;
2173 static unsigned si_get_max_scratch_bytes_per_wave(struct si_context
*sctx
)
2177 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->ps_shader
.current
));
2178 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->gs_shader
.current
));
2179 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->vs_shader
.current
));
2180 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->tcs_shader
.current
));
2181 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->tes_shader
.current
));
2185 static bool si_update_spi_tmpring_size(struct si_context
*sctx
)
2187 unsigned current_scratch_buffer_size
=
2188 si_get_current_scratch_buffer_size(sctx
);
2189 unsigned scratch_bytes_per_wave
=
2190 si_get_max_scratch_bytes_per_wave(sctx
);
2191 unsigned scratch_needed_size
= scratch_bytes_per_wave
*
2192 sctx
->scratch_waves
;
2193 unsigned spi_tmpring_size
;
2196 if (scratch_needed_size
> 0) {
2197 if (scratch_needed_size
> current_scratch_buffer_size
) {
2198 /* Create a bigger scratch buffer */
2199 r600_resource_reference(&sctx
->scratch_buffer
, NULL
);
2201 sctx
->scratch_buffer
= (struct r600_resource
*)
2202 pipe_buffer_create(&sctx
->screen
->b
.b
, 0,
2203 PIPE_USAGE_DEFAULT
, scratch_needed_size
);
2204 if (!sctx
->scratch_buffer
)
2207 si_mark_atom_dirty(sctx
, &sctx
->scratch_state
);
2208 r600_context_add_resource_size(&sctx
->b
.b
,
2209 &sctx
->scratch_buffer
->b
.b
);
2212 /* Update the shaders, so they are using the latest scratch. The
2213 * scratch buffer may have been changed since these shaders were
2214 * last used, so we still need to try to update them, even if
2215 * they require scratch buffers smaller than the current size.
2217 r
= si_update_scratch_buffer(sctx
, sctx
->ps_shader
.current
);
2221 si_pm4_bind_state(sctx
, ps
, sctx
->ps_shader
.current
->pm4
);
2223 r
= si_update_scratch_buffer(sctx
, sctx
->gs_shader
.current
);
2227 si_pm4_bind_state(sctx
, gs
, sctx
->gs_shader
.current
->pm4
);
2229 r
= si_update_scratch_buffer(sctx
, sctx
->tcs_shader
.current
);
2233 si_pm4_bind_state(sctx
, hs
, sctx
->tcs_shader
.current
->pm4
);
2235 /* VS can be bound as LS, ES, or VS. */
2236 r
= si_update_scratch_buffer(sctx
, sctx
->vs_shader
.current
);
2240 if (sctx
->tes_shader
.current
)
2241 si_pm4_bind_state(sctx
, ls
, sctx
->vs_shader
.current
->pm4
);
2242 else if (sctx
->gs_shader
.current
)
2243 si_pm4_bind_state(sctx
, es
, sctx
->vs_shader
.current
->pm4
);
2245 si_pm4_bind_state(sctx
, vs
, sctx
->vs_shader
.current
->pm4
);
2248 /* TES can be bound as ES or VS. */
2249 r
= si_update_scratch_buffer(sctx
, sctx
->tes_shader
.current
);
2253 if (sctx
->gs_shader
.current
)
2254 si_pm4_bind_state(sctx
, es
, sctx
->tes_shader
.current
->pm4
);
2256 si_pm4_bind_state(sctx
, vs
, sctx
->tes_shader
.current
->pm4
);
2260 /* The LLVM shader backend should be reporting aligned scratch_sizes. */
2261 assert((scratch_needed_size
& ~0x3FF) == scratch_needed_size
&&
2262 "scratch size should already be aligned correctly.");
2264 spi_tmpring_size
= S_0286E8_WAVES(sctx
->scratch_waves
) |
2265 S_0286E8_WAVESIZE(scratch_bytes_per_wave
>> 10);
2266 if (spi_tmpring_size
!= sctx
->spi_tmpring_size
) {
2267 sctx
->spi_tmpring_size
= spi_tmpring_size
;
2268 si_mark_atom_dirty(sctx
, &sctx
->scratch_state
);
2273 static void si_init_tess_factor_ring(struct si_context
*sctx
)
2275 bool double_offchip_buffers
= sctx
->b
.chip_class
>= CIK
;
2276 unsigned max_offchip_buffers_per_se
= double_offchip_buffers
? 128 : 64;
2277 unsigned max_offchip_buffers
= max_offchip_buffers_per_se
*
2278 sctx
->screen
->b
.info
.max_se
;
2279 unsigned offchip_granularity
;
2281 switch (sctx
->screen
->tess_offchip_block_dw_size
) {
2286 offchip_granularity
= V_03093C_X_8K_DWORDS
;
2289 offchip_granularity
= V_03093C_X_4K_DWORDS
;
2293 switch (sctx
->b
.chip_class
) {
2295 max_offchip_buffers
= MIN2(max_offchip_buffers
, 126);
2298 max_offchip_buffers
= MIN2(max_offchip_buffers
, 508);
2302 max_offchip_buffers
= MIN2(max_offchip_buffers
, 512);
2306 assert(!sctx
->tf_ring
);
2307 sctx
->tf_ring
= pipe_buffer_create(sctx
->b
.b
.screen
, 0,
2309 32768 * sctx
->screen
->b
.info
.max_se
);
2313 assert(((sctx
->tf_ring
->width0
/ 4) & C_030938_SIZE
) == 0);
2315 sctx
->tess_offchip_ring
= pipe_buffer_create(sctx
->b
.b
.screen
, 0,
2317 max_offchip_buffers
*
2318 sctx
->screen
->tess_offchip_block_dw_size
* 4);
2319 if (!sctx
->tess_offchip_ring
)
2322 si_init_config_add_vgt_flush(sctx
);
2324 /* Append these registers to the init config state. */
2325 if (sctx
->b
.chip_class
>= CIK
) {
2326 if (sctx
->b
.chip_class
>= VI
)
2327 --max_offchip_buffers
;
2329 si_pm4_set_reg(sctx
->init_config
, R_030938_VGT_TF_RING_SIZE
,
2330 S_030938_SIZE(sctx
->tf_ring
->width0
/ 4));
2331 si_pm4_set_reg(sctx
->init_config
, R_030940_VGT_TF_MEMORY_BASE
,
2332 r600_resource(sctx
->tf_ring
)->gpu_address
>> 8);
2333 si_pm4_set_reg(sctx
->init_config
, R_03093C_VGT_HS_OFFCHIP_PARAM
,
2334 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers
) |
2335 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity
));
2337 assert(offchip_granularity
== V_03093C_X_8K_DWORDS
);
2338 si_pm4_set_reg(sctx
->init_config
, R_008988_VGT_TF_RING_SIZE
,
2339 S_008988_SIZE(sctx
->tf_ring
->width0
/ 4));
2340 si_pm4_set_reg(sctx
->init_config
, R_0089B8_VGT_TF_MEMORY_BASE
,
2341 r600_resource(sctx
->tf_ring
)->gpu_address
>> 8);
2342 si_pm4_set_reg(sctx
->init_config
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
2343 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers
));
2346 /* Flush the context to re-emit the init_config state.
2347 * This is done only once in a lifetime of a context.
2349 si_pm4_upload_indirect_buffer(sctx
, sctx
->init_config
);
2350 sctx
->b
.initial_gfx_cs_size
= 0; /* force flush */
2351 si_context_gfx_flush(sctx
, RADEON_FLUSH_ASYNC
, NULL
);
2353 si_set_ring_buffer(&sctx
->b
.b
, SI_HS_RING_TESS_FACTOR
, sctx
->tf_ring
,
2354 0, sctx
->tf_ring
->width0
, false, false, 0, 0, 0);
2356 si_set_ring_buffer(&sctx
->b
.b
, SI_HS_RING_TESS_OFFCHIP
,
2357 sctx
->tess_offchip_ring
, 0,
2358 sctx
->tess_offchip_ring
->width0
, false, false, 0, 0, 0);
2362 * This is used when TCS is NULL in the VS->TCS->TES chain. In this case,
2363 * VS passes its outputs to TES directly, so the fixed-function shader only
2364 * has to write TESSOUTER and TESSINNER.
2366 static void si_generate_fixed_func_tcs(struct si_context
*sctx
)
2368 struct ureg_src outer
, inner
;
2369 struct ureg_dst tessouter
, tessinner
;
2370 struct ureg_program
*ureg
= ureg_create(PIPE_SHADER_TESS_CTRL
);
2373 return; /* if we get here, we're screwed */
2375 assert(!sctx
->fixed_func_tcs_shader
.cso
);
2377 outer
= ureg_DECL_system_value(ureg
,
2378 TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI
, 0);
2379 inner
= ureg_DECL_system_value(ureg
,
2380 TGSI_SEMANTIC_DEFAULT_TESSINNER_SI
, 0);
2382 tessouter
= ureg_DECL_output(ureg
, TGSI_SEMANTIC_TESSOUTER
, 0);
2383 tessinner
= ureg_DECL_output(ureg
, TGSI_SEMANTIC_TESSINNER
, 0);
2385 ureg_MOV(ureg
, tessouter
, outer
);
2386 ureg_MOV(ureg
, tessinner
, inner
);
2389 sctx
->fixed_func_tcs_shader
.cso
=
2390 ureg_create_shader_and_destroy(ureg
, &sctx
->b
.b
);
2393 static void si_update_vgt_shader_config(struct si_context
*sctx
)
2395 /* Calculate the index of the config.
2396 * 0 = VS, 1 = VS+GS, 2 = VS+Tess, 3 = VS+Tess+GS */
2397 unsigned index
= 2*!!sctx
->tes_shader
.cso
+ !!sctx
->gs_shader
.cso
;
2398 struct si_pm4_state
**pm4
= &sctx
->vgt_shader_config
[index
];
2401 uint32_t stages
= 0;
2403 *pm4
= CALLOC_STRUCT(si_pm4_state
);
2405 if (sctx
->tes_shader
.cso
) {
2406 stages
|= S_028B54_LS_EN(V_028B54_LS_STAGE_ON
) |
2407 S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
2409 if (sctx
->gs_shader
.cso
)
2410 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_DS
) |
2412 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
2414 stages
|= S_028B54_VS_EN(V_028B54_VS_STAGE_DS
);
2415 } else if (sctx
->gs_shader
.cso
) {
2416 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL
) |
2418 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
2421 si_pm4_set_reg(*pm4
, R_028B54_VGT_SHADER_STAGES_EN
, stages
);
2423 si_pm4_bind_state(sctx
, vgt_shader_config
, *pm4
);
2426 static void si_update_so(struct si_context
*sctx
, struct si_shader_selector
*shader
)
2428 struct pipe_stream_output_info
*so
= &shader
->so
;
2429 uint32_t enabled_stream_buffers_mask
= 0;
2432 for (i
= 0; i
< so
->num_outputs
; i
++)
2433 enabled_stream_buffers_mask
|= (1 << so
->output
[i
].output_buffer
) << (so
->output
[i
].stream
* 4);
2434 sctx
->b
.streamout
.enabled_stream_buffers_mask
= enabled_stream_buffers_mask
;
2435 sctx
->b
.streamout
.stride_in_dw
= shader
->so
.stride
;
2438 bool si_update_shaders(struct si_context
*sctx
)
2440 struct pipe_context
*ctx
= (struct pipe_context
*)sctx
;
2441 struct si_compiler_ctx_state compiler_state
;
2442 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
2443 struct si_shader
*old_vs
= si_get_vs_state(sctx
);
2444 bool old_clip_disable
= old_vs
? old_vs
->key
.opt
.hw_vs
.clip_disable
: false;
2447 compiler_state
.tm
= sctx
->tm
;
2448 compiler_state
.debug
= sctx
->b
.debug
;
2449 compiler_state
.is_debug_context
= sctx
->is_debug
;
2451 /* Update stages before GS. */
2452 if (sctx
->tes_shader
.cso
) {
2453 if (!sctx
->tf_ring
) {
2454 si_init_tess_factor_ring(sctx
);
2460 r
= si_shader_select(ctx
, &sctx
->vs_shader
, &compiler_state
);
2463 si_pm4_bind_state(sctx
, ls
, sctx
->vs_shader
.current
->pm4
);
2465 if (sctx
->tcs_shader
.cso
) {
2466 r
= si_shader_select(ctx
, &sctx
->tcs_shader
,
2470 si_pm4_bind_state(sctx
, hs
, sctx
->tcs_shader
.current
->pm4
);
2472 if (!sctx
->fixed_func_tcs_shader
.cso
) {
2473 si_generate_fixed_func_tcs(sctx
);
2474 if (!sctx
->fixed_func_tcs_shader
.cso
)
2478 r
= si_shader_select(ctx
, &sctx
->fixed_func_tcs_shader
,
2482 si_pm4_bind_state(sctx
, hs
,
2483 sctx
->fixed_func_tcs_shader
.current
->pm4
);
2486 r
= si_shader_select(ctx
, &sctx
->tes_shader
, &compiler_state
);
2490 if (sctx
->gs_shader
.cso
) {
2492 si_pm4_bind_state(sctx
, es
, sctx
->tes_shader
.current
->pm4
);
2495 si_pm4_bind_state(sctx
, vs
, sctx
->tes_shader
.current
->pm4
);
2496 si_update_so(sctx
, sctx
->tes_shader
.cso
);
2498 } else if (sctx
->gs_shader
.cso
) {
2500 r
= si_shader_select(ctx
, &sctx
->vs_shader
, &compiler_state
);
2503 si_pm4_bind_state(sctx
, es
, sctx
->vs_shader
.current
->pm4
);
2505 si_pm4_bind_state(sctx
, ls
, NULL
);
2506 si_pm4_bind_state(sctx
, hs
, NULL
);
2509 r
= si_shader_select(ctx
, &sctx
->vs_shader
, &compiler_state
);
2512 si_pm4_bind_state(sctx
, vs
, sctx
->vs_shader
.current
->pm4
);
2513 si_update_so(sctx
, sctx
->vs_shader
.cso
);
2515 si_pm4_bind_state(sctx
, ls
, NULL
);
2516 si_pm4_bind_state(sctx
, hs
, NULL
);
2520 if (sctx
->gs_shader
.cso
) {
2521 r
= si_shader_select(ctx
, &sctx
->gs_shader
, &compiler_state
);
2524 si_pm4_bind_state(sctx
, gs
, sctx
->gs_shader
.current
->pm4
);
2525 si_pm4_bind_state(sctx
, vs
, sctx
->gs_shader
.cso
->gs_copy_shader
->pm4
);
2526 si_update_so(sctx
, sctx
->gs_shader
.cso
);
2528 if (!si_update_gs_ring_buffers(sctx
))
2531 si_pm4_bind_state(sctx
, gs
, NULL
);
2532 si_pm4_bind_state(sctx
, es
, NULL
);
2535 si_update_vgt_shader_config(sctx
);
2537 if (old_clip_disable
!= si_get_vs_state(sctx
)->key
.opt
.hw_vs
.clip_disable
)
2538 si_mark_atom_dirty(sctx
, &sctx
->clip_regs
);
2540 if (sctx
->ps_shader
.cso
) {
2541 unsigned db_shader_control
;
2543 r
= si_shader_select(ctx
, &sctx
->ps_shader
, &compiler_state
);
2546 si_pm4_bind_state(sctx
, ps
, sctx
->ps_shader
.current
->pm4
);
2549 sctx
->ps_shader
.cso
->db_shader_control
|
2550 S_02880C_KILL_ENABLE(si_get_alpha_test_func(sctx
) != PIPE_FUNC_ALWAYS
);
2552 if (si_pm4_state_changed(sctx
, ps
) || si_pm4_state_changed(sctx
, vs
) ||
2553 sctx
->sprite_coord_enable
!= rs
->sprite_coord_enable
||
2554 sctx
->flatshade
!= rs
->flatshade
) {
2555 sctx
->sprite_coord_enable
= rs
->sprite_coord_enable
;
2556 sctx
->flatshade
= rs
->flatshade
;
2557 si_mark_atom_dirty(sctx
, &sctx
->spi_map
);
2560 if (sctx
->b
.family
== CHIP_STONEY
&& si_pm4_state_changed(sctx
, ps
))
2561 si_mark_atom_dirty(sctx
, &sctx
->cb_render_state
);
2563 if (sctx
->ps_db_shader_control
!= db_shader_control
) {
2564 sctx
->ps_db_shader_control
= db_shader_control
;
2565 si_mark_atom_dirty(sctx
, &sctx
->db_render_state
);
2568 if (sctx
->smoothing_enabled
!= sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.poly_line_smoothing
) {
2569 sctx
->smoothing_enabled
= sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.poly_line_smoothing
;
2570 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
2572 if (sctx
->b
.chip_class
== SI
)
2573 si_mark_atom_dirty(sctx
, &sctx
->db_render_state
);
2575 if (sctx
->framebuffer
.nr_samples
<= 1)
2576 si_mark_atom_dirty(sctx
, &sctx
->msaa_sample_locs
.atom
);
2580 if (si_pm4_state_changed(sctx
, ls
) ||
2581 si_pm4_state_changed(sctx
, hs
) ||
2582 si_pm4_state_changed(sctx
, es
) ||
2583 si_pm4_state_changed(sctx
, gs
) ||
2584 si_pm4_state_changed(sctx
, vs
) ||
2585 si_pm4_state_changed(sctx
, ps
)) {
2586 if (!si_update_spi_tmpring_size(sctx
))
2590 if (sctx
->b
.chip_class
>= CIK
)
2591 si_mark_atom_dirty(sctx
, &sctx
->prefetch_L2
);
2593 sctx
->do_update_shaders
= false;
2597 static void si_emit_scratch_state(struct si_context
*sctx
,
2598 struct r600_atom
*atom
)
2600 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
2602 radeon_set_context_reg(cs
, R_0286E8_SPI_TMPRING_SIZE
,
2603 sctx
->spi_tmpring_size
);
2605 if (sctx
->scratch_buffer
) {
2606 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
2607 sctx
->scratch_buffer
, RADEON_USAGE_READWRITE
,
2608 RADEON_PRIO_SCRATCH_BUFFER
);
2612 void si_init_shader_functions(struct si_context
*sctx
)
2614 si_init_atom(sctx
, &sctx
->spi_map
, &sctx
->atoms
.s
.spi_map
, si_emit_spi_map
);
2615 si_init_atom(sctx
, &sctx
->scratch_state
, &sctx
->atoms
.s
.scratch_state
,
2616 si_emit_scratch_state
);
2618 sctx
->b
.b
.create_vs_state
= si_create_shader_selector
;
2619 sctx
->b
.b
.create_tcs_state
= si_create_shader_selector
;
2620 sctx
->b
.b
.create_tes_state
= si_create_shader_selector
;
2621 sctx
->b
.b
.create_gs_state
= si_create_shader_selector
;
2622 sctx
->b
.b
.create_fs_state
= si_create_shader_selector
;
2624 sctx
->b
.b
.bind_vs_state
= si_bind_vs_shader
;
2625 sctx
->b
.b
.bind_tcs_state
= si_bind_tcs_shader
;
2626 sctx
->b
.b
.bind_tes_state
= si_bind_tes_shader
;
2627 sctx
->b
.b
.bind_gs_state
= si_bind_gs_shader
;
2628 sctx
->b
.b
.bind_fs_state
= si_bind_ps_shader
;
2630 sctx
->b
.b
.delete_vs_state
= si_delete_shader_selector
;
2631 sctx
->b
.b
.delete_tcs_state
= si_delete_shader_selector
;
2632 sctx
->b
.b
.delete_tes_state
= si_delete_shader_selector
;
2633 sctx
->b
.b
.delete_gs_state
= si_delete_shader_selector
;
2634 sctx
->b
.b
.delete_fs_state
= si_delete_shader_selector
;