2 * Copyright 2012 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 #include "si_build_pm4.h"
28 #include "compiler/nir/nir_serialize.h"
29 #include "tgsi/tgsi_parse.h"
30 #include "util/hash_table.h"
31 #include "util/crc32.h"
32 #include "util/u_async_debug.h"
33 #include "util/u_memory.h"
34 #include "util/u_prim.h"
36 #include "util/disk_cache.h"
37 #include "util/mesa-sha1.h"
38 #include "ac_exp_param.h"
39 #include "ac_shader_util.h"
44 * Return the IR binary in a buffer. For TGSI the first 4 bytes contain its
47 void *si_get_ir_binary(struct si_shader_selector
*sel
)
54 ir_binary
= sel
->tokens
;
55 ir_size
= tgsi_num_tokens(sel
->tokens
) *
56 sizeof(struct tgsi_token
);
61 nir_serialize(&blob
, sel
->nir
);
62 ir_binary
= blob
.data
;
66 unsigned size
= 4 + ir_size
+ sizeof(sel
->so
);
67 char *result
= (char*)MALLOC(size
);
71 *((uint32_t*)result
) = size
;
72 memcpy(result
+ 4, ir_binary
, ir_size
);
73 memcpy(result
+ 4 + ir_size
, &sel
->so
, sizeof(sel
->so
));
81 /** Copy "data" to "ptr" and return the next dword following copied data. */
82 static uint32_t *write_data(uint32_t *ptr
, const void *data
, unsigned size
)
84 /* data may be NULL if size == 0 */
86 memcpy(ptr
, data
, size
);
87 ptr
+= DIV_ROUND_UP(size
, 4);
91 /** Read data from "ptr". Return the next dword following the data. */
92 static uint32_t *read_data(uint32_t *ptr
, void *data
, unsigned size
)
94 memcpy(data
, ptr
, size
);
95 ptr
+= DIV_ROUND_UP(size
, 4);
100 * Write the size as uint followed by the data. Return the next dword
101 * following the copied data.
103 static uint32_t *write_chunk(uint32_t *ptr
, const void *data
, unsigned size
)
106 return write_data(ptr
, data
, size
);
110 * Read the size as uint followed by the data. Return both via parameters.
111 * Return the next dword following the data.
113 static uint32_t *read_chunk(uint32_t *ptr
, void **data
, unsigned *size
)
116 assert(*data
== NULL
);
119 *data
= malloc(*size
);
120 return read_data(ptr
, *data
, *size
);
124 * Return the shader binary in a buffer. The first 4 bytes contain its size
127 static void *si_get_shader_binary(struct si_shader
*shader
)
129 /* There is always a size of data followed by the data itself. */
130 unsigned llvm_ir_size
= shader
->binary
.llvm_ir_string
?
131 strlen(shader
->binary
.llvm_ir_string
) + 1 : 0;
133 /* Refuse to allocate overly large buffers and guard against integer
135 if (shader
->binary
.elf_size
> UINT_MAX
/ 4 ||
136 llvm_ir_size
> UINT_MAX
/ 4)
141 4 + /* CRC32 of the data below */
142 align(sizeof(shader
->config
), 4) +
143 align(sizeof(shader
->info
), 4) +
144 4 + align(shader
->binary
.elf_size
, 4) +
145 4 + align(llvm_ir_size
, 4);
146 void *buffer
= CALLOC(1, size
);
147 uint32_t *ptr
= (uint32_t*)buffer
;
153 ptr
++; /* CRC32 is calculated at the end. */
155 ptr
= write_data(ptr
, &shader
->config
, sizeof(shader
->config
));
156 ptr
= write_data(ptr
, &shader
->info
, sizeof(shader
->info
));
157 ptr
= write_chunk(ptr
, shader
->binary
.elf_buffer
, shader
->binary
.elf_size
);
158 ptr
= write_chunk(ptr
, shader
->binary
.llvm_ir_string
, llvm_ir_size
);
159 assert((char *)ptr
- (char *)buffer
== size
);
162 ptr
= (uint32_t*)buffer
;
164 *ptr
= util_hash_crc32(ptr
+ 1, size
- 8);
169 static bool si_load_shader_binary(struct si_shader
*shader
, void *binary
)
171 uint32_t *ptr
= (uint32_t*)binary
;
172 uint32_t size
= *ptr
++;
173 uint32_t crc32
= *ptr
++;
177 if (util_hash_crc32(ptr
, size
- 8) != crc32
) {
178 fprintf(stderr
, "radeonsi: binary shader has invalid CRC32\n");
182 ptr
= read_data(ptr
, &shader
->config
, sizeof(shader
->config
));
183 ptr
= read_data(ptr
, &shader
->info
, sizeof(shader
->info
));
184 ptr
= read_chunk(ptr
, (void**)&shader
->binary
.elf_buffer
,
186 shader
->binary
.elf_size
= elf_size
;
187 ptr
= read_chunk(ptr
, (void**)&shader
->binary
.llvm_ir_string
, &chunk_size
);
193 * Insert a shader into the cache. It's assumed the shader is not in the cache.
194 * Use si_shader_cache_load_shader before calling this.
196 * Returns false on failure, in which case the ir_binary should be freed.
198 bool si_shader_cache_insert_shader(struct si_screen
*sscreen
, void *ir_binary
,
199 struct si_shader
*shader
,
200 bool insert_into_disk_cache
)
203 struct hash_entry
*entry
;
204 uint8_t key
[CACHE_KEY_SIZE
];
206 entry
= _mesa_hash_table_search(sscreen
->shader_cache
, ir_binary
);
208 return false; /* already added */
210 hw_binary
= si_get_shader_binary(shader
);
214 if (_mesa_hash_table_insert(sscreen
->shader_cache
, ir_binary
,
215 hw_binary
) == NULL
) {
220 if (sscreen
->disk_shader_cache
&& insert_into_disk_cache
) {
221 disk_cache_compute_key(sscreen
->disk_shader_cache
, ir_binary
,
222 *((uint32_t *)ir_binary
), key
);
223 disk_cache_put(sscreen
->disk_shader_cache
, key
, hw_binary
,
224 *((uint32_t *) hw_binary
), NULL
);
230 bool si_shader_cache_load_shader(struct si_screen
*sscreen
, void *ir_binary
,
231 struct si_shader
*shader
)
233 struct hash_entry
*entry
=
234 _mesa_hash_table_search(sscreen
->shader_cache
, ir_binary
);
236 if (sscreen
->disk_shader_cache
) {
237 unsigned char sha1
[CACHE_KEY_SIZE
];
238 size_t tg_size
= *((uint32_t *) ir_binary
);
240 disk_cache_compute_key(sscreen
->disk_shader_cache
,
241 ir_binary
, tg_size
, sha1
);
245 disk_cache_get(sscreen
->disk_shader_cache
,
250 if (binary_size
< sizeof(uint32_t) ||
251 *((uint32_t*)buffer
) != binary_size
) {
252 /* Something has gone wrong discard the item
253 * from the cache and rebuild/link from
256 assert(!"Invalid radeonsi shader disk cache "
259 disk_cache_remove(sscreen
->disk_shader_cache
,
266 if (!si_load_shader_binary(shader
, buffer
)) {
272 if (!si_shader_cache_insert_shader(sscreen
, ir_binary
,
279 if (si_load_shader_binary(shader
, entry
->data
))
284 p_atomic_inc(&sscreen
->num_shader_cache_hits
);
288 static uint32_t si_shader_cache_key_hash(const void *key
)
290 /* The first dword is the key size. */
291 return util_hash_crc32(key
, *(uint32_t*)key
);
294 static bool si_shader_cache_key_equals(const void *a
, const void *b
)
296 uint32_t *keya
= (uint32_t*)a
;
297 uint32_t *keyb
= (uint32_t*)b
;
299 /* The first dword is the key size. */
303 return memcmp(keya
, keyb
, *keya
) == 0;
306 static void si_destroy_shader_cache_entry(struct hash_entry
*entry
)
308 FREE((void*)entry
->key
);
312 bool si_init_shader_cache(struct si_screen
*sscreen
)
314 (void) mtx_init(&sscreen
->shader_cache_mutex
, mtx_plain
);
315 sscreen
->shader_cache
=
316 _mesa_hash_table_create(NULL
,
317 si_shader_cache_key_hash
,
318 si_shader_cache_key_equals
);
320 return sscreen
->shader_cache
!= NULL
;
323 void si_destroy_shader_cache(struct si_screen
*sscreen
)
325 if (sscreen
->shader_cache
)
326 _mesa_hash_table_destroy(sscreen
->shader_cache
,
327 si_destroy_shader_cache_entry
);
328 mtx_destroy(&sscreen
->shader_cache_mutex
);
333 static void si_set_tesseval_regs(struct si_screen
*sscreen
,
334 const struct si_shader_selector
*tes
,
335 struct si_pm4_state
*pm4
)
337 const struct tgsi_shader_info
*info
= &tes
->info
;
338 unsigned tes_prim_mode
= info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
339 unsigned tes_spacing
= info
->properties
[TGSI_PROPERTY_TES_SPACING
];
340 bool tes_vertex_order_cw
= info
->properties
[TGSI_PROPERTY_TES_VERTEX_ORDER_CW
];
341 bool tes_point_mode
= info
->properties
[TGSI_PROPERTY_TES_POINT_MODE
];
342 unsigned type
, partitioning
, topology
, distribution_mode
;
344 switch (tes_prim_mode
) {
345 case PIPE_PRIM_LINES
:
346 type
= V_028B6C_TESS_ISOLINE
;
348 case PIPE_PRIM_TRIANGLES
:
349 type
= V_028B6C_TESS_TRIANGLE
;
351 case PIPE_PRIM_QUADS
:
352 type
= V_028B6C_TESS_QUAD
;
359 switch (tes_spacing
) {
360 case PIPE_TESS_SPACING_FRACTIONAL_ODD
:
361 partitioning
= V_028B6C_PART_FRAC_ODD
;
363 case PIPE_TESS_SPACING_FRACTIONAL_EVEN
:
364 partitioning
= V_028B6C_PART_FRAC_EVEN
;
366 case PIPE_TESS_SPACING_EQUAL
:
367 partitioning
= V_028B6C_PART_INTEGER
;
375 topology
= V_028B6C_OUTPUT_POINT
;
376 else if (tes_prim_mode
== PIPE_PRIM_LINES
)
377 topology
= V_028B6C_OUTPUT_LINE
;
378 else if (tes_vertex_order_cw
)
379 /* for some reason, this must be the other way around */
380 topology
= V_028B6C_OUTPUT_TRIANGLE_CCW
;
382 topology
= V_028B6C_OUTPUT_TRIANGLE_CW
;
384 if (sscreen
->has_distributed_tess
) {
385 if (sscreen
->info
.family
== CHIP_FIJI
||
386 sscreen
->info
.family
>= CHIP_POLARIS10
)
387 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS
;
389 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_DONUTS
;
391 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_NO_DIST
;
394 pm4
->shader
->vgt_tf_param
= S_028B6C_TYPE(type
) |
395 S_028B6C_PARTITIONING(partitioning
) |
396 S_028B6C_TOPOLOGY(topology
) |
397 S_028B6C_DISTRIBUTION_MODE(distribution_mode
);
400 /* Polaris needs different VTX_REUSE_DEPTH settings depending on
401 * whether the "fractional odd" tessellation spacing is used.
403 * Possible VGT configurations and which state should set the register:
405 * Reg set in | VGT shader configuration | Value
406 * ------------------------------------------------------
408 * VS as ES | ES -> GS -> VS | 30
409 * TES as VS | LS -> HS -> VS | 14 or 30
410 * TES as ES | LS -> HS -> ES -> GS -> VS | 14 or 30
412 * If "shader" is NULL, it's assumed it's not LS or GS copy shader.
414 static void polaris_set_vgt_vertex_reuse(struct si_screen
*sscreen
,
415 struct si_shader_selector
*sel
,
416 struct si_shader
*shader
,
417 struct si_pm4_state
*pm4
)
419 unsigned type
= sel
->type
;
421 if (sscreen
->info
.family
< CHIP_POLARIS10
)
424 /* VS as VS, or VS as ES: */
425 if ((type
== PIPE_SHADER_VERTEX
&&
427 (!shader
->key
.as_ls
&& !shader
->is_gs_copy_shader
))) ||
428 /* TES as VS, or TES as ES: */
429 type
== PIPE_SHADER_TESS_EVAL
) {
430 unsigned vtx_reuse_depth
= 30;
432 if (type
== PIPE_SHADER_TESS_EVAL
&&
433 sel
->info
.properties
[TGSI_PROPERTY_TES_SPACING
] ==
434 PIPE_TESS_SPACING_FRACTIONAL_ODD
)
435 vtx_reuse_depth
= 14;
438 pm4
->shader
->vgt_vertex_reuse_block_cntl
= vtx_reuse_depth
;
442 static struct si_pm4_state
*si_get_shader_pm4_state(struct si_shader
*shader
)
445 si_pm4_clear_state(shader
->pm4
);
447 shader
->pm4
= CALLOC_STRUCT(si_pm4_state
);
450 shader
->pm4
->shader
= shader
;
453 fprintf(stderr
, "radeonsi: Failed to create pm4 state.\n");
458 static unsigned si_get_num_vs_user_sgprs(unsigned num_always_on_user_sgprs
)
460 /* Add the pointer to VBO descriptors. */
461 return num_always_on_user_sgprs
+ 1;
464 static void si_shader_ls(struct si_screen
*sscreen
, struct si_shader
*shader
)
466 struct si_pm4_state
*pm4
;
467 unsigned vgpr_comp_cnt
;
470 assert(sscreen
->info
.chip_class
<= GFX8
);
472 pm4
= si_get_shader_pm4_state(shader
);
476 va
= shader
->bo
->gpu_address
;
477 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
479 /* We need at least 2 components for LS.
480 * VGPR0-3: (VertexID, RelAutoindex, InstanceID / StepRate0, InstanceID).
481 * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
483 vgpr_comp_cnt
= shader
->info
.uses_instanceid
? 2 : 1;
485 si_pm4_set_reg(pm4
, R_00B520_SPI_SHADER_PGM_LO_LS
, va
>> 8);
486 si_pm4_set_reg(pm4
, R_00B524_SPI_SHADER_PGM_HI_LS
, S_00B524_MEM_BASE(va
>> 40));
488 shader
->config
.rsrc1
= S_00B528_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
489 S_00B528_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
490 S_00B528_VGPR_COMP_CNT(vgpr_comp_cnt
) |
491 S_00B528_DX10_CLAMP(1) |
492 S_00B528_FLOAT_MODE(shader
->config
.float_mode
);
493 shader
->config
.rsrc2
= S_00B52C_USER_SGPR(si_get_num_vs_user_sgprs(SI_VS_NUM_USER_SGPR
)) |
494 S_00B52C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
497 static void si_shader_hs(struct si_screen
*sscreen
, struct si_shader
*shader
)
499 struct si_pm4_state
*pm4
;
501 unsigned ls_vgpr_comp_cnt
= 0;
503 pm4
= si_get_shader_pm4_state(shader
);
507 va
= shader
->bo
->gpu_address
;
508 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
510 if (sscreen
->info
.chip_class
>= GFX9
) {
511 if (sscreen
->info
.chip_class
>= GFX10
) {
512 si_pm4_set_reg(pm4
, R_00B520_SPI_SHADER_PGM_LO_LS
, va
>> 8);
513 si_pm4_set_reg(pm4
, R_00B524_SPI_SHADER_PGM_HI_LS
, S_00B524_MEM_BASE(va
>> 40));
515 si_pm4_set_reg(pm4
, R_00B410_SPI_SHADER_PGM_LO_LS
, va
>> 8);
516 si_pm4_set_reg(pm4
, R_00B414_SPI_SHADER_PGM_HI_LS
, S_00B414_MEM_BASE(va
>> 40));
519 /* We need at least 2 components for LS.
520 * GFX9 VGPR0-3: (VertexID, RelAutoindex, InstanceID / StepRate0, InstanceID).
521 * GFX10 VGPR0-3: (VertexID, RelAutoindex, UserVGPR1, InstanceID).
522 * On gfx9, StepRate0 is set to 1 so that VGPR3 doesn't have to
525 ls_vgpr_comp_cnt
= 1;
526 if (shader
->info
.uses_instanceid
) {
527 if (sscreen
->info
.chip_class
>= GFX10
)
528 ls_vgpr_comp_cnt
= 3;
530 ls_vgpr_comp_cnt
= 2;
533 unsigned num_user_sgprs
=
534 si_get_num_vs_user_sgprs(GFX9_TCS_NUM_USER_SGPR
);
536 shader
->config
.rsrc2
=
537 S_00B42C_USER_SGPR(num_user_sgprs
) |
538 S_00B42C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
540 if (sscreen
->info
.chip_class
>= GFX10
)
541 shader
->config
.rsrc2
|= S_00B42C_USER_SGPR_MSB_GFX10(num_user_sgprs
>> 5);
543 shader
->config
.rsrc2
|= S_00B42C_USER_SGPR_MSB_GFX9(num_user_sgprs
>> 5);
545 si_pm4_set_reg(pm4
, R_00B420_SPI_SHADER_PGM_LO_HS
, va
>> 8);
546 si_pm4_set_reg(pm4
, R_00B424_SPI_SHADER_PGM_HI_HS
, S_00B424_MEM_BASE(va
>> 40));
548 shader
->config
.rsrc2
=
549 S_00B42C_USER_SGPR(GFX6_TCS_NUM_USER_SGPR
) |
550 S_00B42C_OC_LDS_EN(1) |
551 S_00B42C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
554 si_pm4_set_reg(pm4
, R_00B428_SPI_SHADER_PGM_RSRC1_HS
,
555 S_00B428_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
556 (sscreen
->info
.chip_class
<= GFX9
?
557 S_00B428_SGPRS((shader
->config
.num_sgprs
- 1) / 8) : 0) |
558 S_00B428_DX10_CLAMP(1) |
559 S_00B428_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
560 S_00B428_FLOAT_MODE(shader
->config
.float_mode
) |
561 S_00B428_LS_VGPR_COMP_CNT(ls_vgpr_comp_cnt
));
563 if (sscreen
->info
.chip_class
<= GFX8
) {
564 si_pm4_set_reg(pm4
, R_00B42C_SPI_SHADER_PGM_RSRC2_HS
,
565 shader
->config
.rsrc2
);
569 static void si_emit_shader_es(struct si_context
*sctx
)
571 struct si_shader
*shader
= sctx
->queued
.named
.es
->shader
;
572 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
577 radeon_opt_set_context_reg(sctx
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
578 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE
,
579 shader
->selector
->esgs_itemsize
/ 4);
581 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
582 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
583 SI_TRACKED_VGT_TF_PARAM
,
584 shader
->vgt_tf_param
);
586 if (shader
->vgt_vertex_reuse_block_cntl
)
587 radeon_opt_set_context_reg(sctx
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
588 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL
,
589 shader
->vgt_vertex_reuse_block_cntl
);
591 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
592 sctx
->context_roll
= true;
595 static void si_shader_es(struct si_screen
*sscreen
, struct si_shader
*shader
)
597 struct si_pm4_state
*pm4
;
598 unsigned num_user_sgprs
;
599 unsigned vgpr_comp_cnt
;
603 assert(sscreen
->info
.chip_class
<= GFX8
);
605 pm4
= si_get_shader_pm4_state(shader
);
609 pm4
->atom
.emit
= si_emit_shader_es
;
610 va
= shader
->bo
->gpu_address
;
611 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
613 if (shader
->selector
->type
== PIPE_SHADER_VERTEX
) {
614 /* VGPR0-3: (VertexID, InstanceID / StepRate0, ...) */
615 vgpr_comp_cnt
= shader
->info
.uses_instanceid
? 1 : 0;
616 num_user_sgprs
= si_get_num_vs_user_sgprs(SI_VS_NUM_USER_SGPR
);
617 } else if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
618 vgpr_comp_cnt
= shader
->selector
->info
.uses_primid
? 3 : 2;
619 num_user_sgprs
= SI_TES_NUM_USER_SGPR
;
621 unreachable("invalid shader selector type");
623 oc_lds_en
= shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
? 1 : 0;
625 si_pm4_set_reg(pm4
, R_00B320_SPI_SHADER_PGM_LO_ES
, va
>> 8);
626 si_pm4_set_reg(pm4
, R_00B324_SPI_SHADER_PGM_HI_ES
, S_00B324_MEM_BASE(va
>> 40));
627 si_pm4_set_reg(pm4
, R_00B328_SPI_SHADER_PGM_RSRC1_ES
,
628 S_00B328_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
629 S_00B328_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
630 S_00B328_VGPR_COMP_CNT(vgpr_comp_cnt
) |
631 S_00B328_DX10_CLAMP(1) |
632 S_00B328_FLOAT_MODE(shader
->config
.float_mode
));
633 si_pm4_set_reg(pm4
, R_00B32C_SPI_SHADER_PGM_RSRC2_ES
,
634 S_00B32C_USER_SGPR(num_user_sgprs
) |
635 S_00B32C_OC_LDS_EN(oc_lds_en
) |
636 S_00B32C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
638 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
639 si_set_tesseval_regs(sscreen
, shader
->selector
, pm4
);
641 polaris_set_vgt_vertex_reuse(sscreen
, shader
->selector
, shader
, pm4
);
644 void gfx9_get_gs_info(struct si_shader_selector
*es
,
645 struct si_shader_selector
*gs
,
646 struct gfx9_gs_info
*out
)
648 unsigned gs_num_invocations
= MAX2(gs
->gs_num_invocations
, 1);
649 unsigned input_prim
= gs
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
];
650 bool uses_adjacency
= input_prim
>= PIPE_PRIM_LINES_ADJACENCY
&&
651 input_prim
<= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
;
653 /* All these are in dwords: */
654 /* We can't allow using the whole LDS, because GS waves compete with
655 * other shader stages for LDS space. */
656 const unsigned max_lds_size
= 8 * 1024;
657 const unsigned esgs_itemsize
= es
->esgs_itemsize
/ 4;
658 unsigned esgs_lds_size
;
660 /* All these are per subgroup: */
661 const unsigned max_out_prims
= 32 * 1024;
662 const unsigned max_es_verts
= 255;
663 const unsigned ideal_gs_prims
= 64;
664 unsigned max_gs_prims
, gs_prims
;
665 unsigned min_es_verts
, es_verts
, worst_case_es_verts
;
667 if (uses_adjacency
|| gs_num_invocations
> 1)
668 max_gs_prims
= 127 / gs_num_invocations
;
672 /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
673 * Make sure we don't go over the maximum value.
675 if (gs
->gs_max_out_vertices
> 0) {
676 max_gs_prims
= MIN2(max_gs_prims
,
678 (gs
->gs_max_out_vertices
* gs_num_invocations
));
680 assert(max_gs_prims
> 0);
682 /* If the primitive has adjacency, halve the number of vertices
683 * that will be reused in multiple primitives.
685 min_es_verts
= gs
->gs_input_verts_per_prim
/ (uses_adjacency
? 2 : 1);
687 gs_prims
= MIN2(ideal_gs_prims
, max_gs_prims
);
688 worst_case_es_verts
= MIN2(min_es_verts
* gs_prims
, max_es_verts
);
690 /* Compute ESGS LDS size based on the worst case number of ES vertices
691 * needed to create the target number of GS prims per subgroup.
693 esgs_lds_size
= esgs_itemsize
* worst_case_es_verts
;
695 /* If total LDS usage is too big, refactor partitions based on ratio
696 * of ESGS item sizes.
698 if (esgs_lds_size
> max_lds_size
) {
699 /* Our target GS Prims Per Subgroup was too large. Calculate
700 * the maximum number of GS Prims Per Subgroup that will fit
701 * into LDS, capped by the maximum that the hardware can support.
703 gs_prims
= MIN2((max_lds_size
/ (esgs_itemsize
* min_es_verts
)),
705 assert(gs_prims
> 0);
706 worst_case_es_verts
= MIN2(min_es_verts
* gs_prims
,
709 esgs_lds_size
= esgs_itemsize
* worst_case_es_verts
;
710 assert(esgs_lds_size
<= max_lds_size
);
713 /* Now calculate remaining ESGS information. */
715 es_verts
= MIN2(esgs_lds_size
/ esgs_itemsize
, max_es_verts
);
717 es_verts
= max_es_verts
;
719 /* Vertices for adjacency primitives are not always reused, so restore
720 * it for ES_VERTS_PER_SUBGRP.
722 min_es_verts
= gs
->gs_input_verts_per_prim
;
724 /* For normal primitives, the VGT only checks if they are past the ES
725 * verts per subgroup after allocating a full GS primitive and if they
726 * are, kick off a new subgroup. But if those additional ES verts are
727 * unique (e.g. not reused) we need to make sure there is enough LDS
728 * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
730 es_verts
-= min_es_verts
- 1;
732 out
->es_verts_per_subgroup
= es_verts
;
733 out
->gs_prims_per_subgroup
= gs_prims
;
734 out
->gs_inst_prims_in_subgroup
= gs_prims
* gs_num_invocations
;
735 out
->max_prims_per_subgroup
= out
->gs_inst_prims_in_subgroup
*
736 gs
->gs_max_out_vertices
;
737 out
->esgs_ring_size
= 4 * esgs_lds_size
;
739 assert(out
->max_prims_per_subgroup
<= max_out_prims
);
742 static void si_emit_shader_gs(struct si_context
*sctx
)
744 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
745 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
750 /* R_028A60_VGT_GSVS_RING_OFFSET_1, R_028A64_VGT_GSVS_RING_OFFSET_2
751 * R_028A68_VGT_GSVS_RING_OFFSET_3 */
752 radeon_opt_set_context_reg3(sctx
, R_028A60_VGT_GSVS_RING_OFFSET_1
,
753 SI_TRACKED_VGT_GSVS_RING_OFFSET_1
,
754 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_1
,
755 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_2
,
756 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_3
);
758 /* R_028AB0_VGT_GSVS_RING_ITEMSIZE */
759 radeon_opt_set_context_reg(sctx
, R_028AB0_VGT_GSVS_RING_ITEMSIZE
,
760 SI_TRACKED_VGT_GSVS_RING_ITEMSIZE
,
761 shader
->ctx_reg
.gs
.vgt_gsvs_ring_itemsize
);
763 /* R_028B38_VGT_GS_MAX_VERT_OUT */
764 radeon_opt_set_context_reg(sctx
, R_028B38_VGT_GS_MAX_VERT_OUT
,
765 SI_TRACKED_VGT_GS_MAX_VERT_OUT
,
766 shader
->ctx_reg
.gs
.vgt_gs_max_vert_out
);
768 /* R_028B5C_VGT_GS_VERT_ITEMSIZE, R_028B60_VGT_GS_VERT_ITEMSIZE_1
769 * R_028B64_VGT_GS_VERT_ITEMSIZE_2, R_028B68_VGT_GS_VERT_ITEMSIZE_3 */
770 radeon_opt_set_context_reg4(sctx
, R_028B5C_VGT_GS_VERT_ITEMSIZE
,
771 SI_TRACKED_VGT_GS_VERT_ITEMSIZE
,
772 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize
,
773 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_1
,
774 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_2
,
775 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_3
);
777 /* R_028B90_VGT_GS_INSTANCE_CNT */
778 radeon_opt_set_context_reg(sctx
, R_028B90_VGT_GS_INSTANCE_CNT
,
779 SI_TRACKED_VGT_GS_INSTANCE_CNT
,
780 shader
->ctx_reg
.gs
.vgt_gs_instance_cnt
);
782 if (sctx
->chip_class
>= GFX9
) {
783 /* R_028A44_VGT_GS_ONCHIP_CNTL */
784 radeon_opt_set_context_reg(sctx
, R_028A44_VGT_GS_ONCHIP_CNTL
,
785 SI_TRACKED_VGT_GS_ONCHIP_CNTL
,
786 shader
->ctx_reg
.gs
.vgt_gs_onchip_cntl
);
787 /* R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP */
788 radeon_opt_set_context_reg(sctx
, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP
,
789 SI_TRACKED_VGT_GS_MAX_PRIMS_PER_SUBGROUP
,
790 shader
->ctx_reg
.gs
.vgt_gs_max_prims_per_subgroup
);
791 /* R_028AAC_VGT_ESGS_RING_ITEMSIZE */
792 radeon_opt_set_context_reg(sctx
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
793 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE
,
794 shader
->ctx_reg
.gs
.vgt_esgs_ring_itemsize
);
796 if (shader
->key
.part
.gs
.es
->type
== PIPE_SHADER_TESS_EVAL
)
797 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
798 SI_TRACKED_VGT_TF_PARAM
,
799 shader
->vgt_tf_param
);
800 if (shader
->vgt_vertex_reuse_block_cntl
)
801 radeon_opt_set_context_reg(sctx
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
802 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL
,
803 shader
->vgt_vertex_reuse_block_cntl
);
806 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
807 sctx
->context_roll
= true;
810 static void si_shader_gs(struct si_screen
*sscreen
, struct si_shader
*shader
)
812 struct si_shader_selector
*sel
= shader
->selector
;
813 const ubyte
*num_components
= sel
->info
.num_stream_output_components
;
814 unsigned gs_num_invocations
= sel
->gs_num_invocations
;
815 struct si_pm4_state
*pm4
;
817 unsigned max_stream
= sel
->max_gs_stream
;
820 pm4
= si_get_shader_pm4_state(shader
);
824 pm4
->atom
.emit
= si_emit_shader_gs
;
826 offset
= num_components
[0] * sel
->gs_max_out_vertices
;
827 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_1
= offset
;
830 offset
+= num_components
[1] * sel
->gs_max_out_vertices
;
831 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_2
= offset
;
834 offset
+= num_components
[2] * sel
->gs_max_out_vertices
;
835 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_3
= offset
;
838 offset
+= num_components
[3] * sel
->gs_max_out_vertices
;
839 shader
->ctx_reg
.gs
.vgt_gsvs_ring_itemsize
= offset
;
841 /* The GSVS_RING_ITEMSIZE register takes 15 bits */
842 assert(offset
< (1 << 15));
844 shader
->ctx_reg
.gs
.vgt_gs_max_vert_out
= sel
->gs_max_out_vertices
;
846 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize
= num_components
[0];
847 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_1
= (max_stream
>= 1) ? num_components
[1] : 0;
848 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_2
= (max_stream
>= 2) ? num_components
[2] : 0;
849 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_3
= (max_stream
>= 3) ? num_components
[3] : 0;
851 shader
->ctx_reg
.gs
.vgt_gs_instance_cnt
= S_028B90_CNT(MIN2(gs_num_invocations
, 127)) |
852 S_028B90_ENABLE(gs_num_invocations
> 0);
854 va
= shader
->bo
->gpu_address
;
855 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
857 if (sscreen
->info
.chip_class
>= GFX9
) {
858 unsigned input_prim
= sel
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
];
859 unsigned es_type
= shader
->key
.part
.gs
.es
->type
;
860 unsigned es_vgpr_comp_cnt
, gs_vgpr_comp_cnt
;
862 if (es_type
== PIPE_SHADER_VERTEX
)
863 /* VGPR0-3: (VertexID, InstanceID / StepRate0, ...) */
864 es_vgpr_comp_cnt
= shader
->info
.uses_instanceid
? 1 : 0;
865 else if (es_type
== PIPE_SHADER_TESS_EVAL
)
866 es_vgpr_comp_cnt
= shader
->key
.part
.gs
.es
->info
.uses_primid
? 3 : 2;
868 unreachable("invalid shader selector type");
870 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
871 * VGPR[0:4] are always loaded.
873 if (sel
->info
.uses_invocationid
)
874 gs_vgpr_comp_cnt
= 3; /* VGPR3 contains InvocationID. */
875 else if (sel
->info
.uses_primid
)
876 gs_vgpr_comp_cnt
= 2; /* VGPR2 contains PrimitiveID. */
877 else if (input_prim
>= PIPE_PRIM_TRIANGLES
)
878 gs_vgpr_comp_cnt
= 1; /* VGPR1 contains offsets 2, 3 */
880 gs_vgpr_comp_cnt
= 0; /* VGPR0 contains offsets 0, 1 */
882 unsigned num_user_sgprs
;
883 if (es_type
== PIPE_SHADER_VERTEX
)
884 num_user_sgprs
= si_get_num_vs_user_sgprs(GFX9_VSGS_NUM_USER_SGPR
);
886 num_user_sgprs
= GFX9_TESGS_NUM_USER_SGPR
;
888 if (sscreen
->info
.chip_class
>= GFX10
) {
889 si_pm4_set_reg(pm4
, R_00B320_SPI_SHADER_PGM_LO_ES
, va
>> 8);
890 si_pm4_set_reg(pm4
, R_00B324_SPI_SHADER_PGM_HI_ES
, S_00B324_MEM_BASE(va
>> 40));
892 si_pm4_set_reg(pm4
, R_00B210_SPI_SHADER_PGM_LO_ES
, va
>> 8);
893 si_pm4_set_reg(pm4
, R_00B214_SPI_SHADER_PGM_HI_ES
, S_00B214_MEM_BASE(va
>> 40));
897 S_00B228_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
898 S_00B228_DX10_CLAMP(1) |
899 S_00B228_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
900 S_00B228_FLOAT_MODE(shader
->config
.float_mode
) |
901 S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt
);
903 S_00B22C_USER_SGPR(num_user_sgprs
) |
904 S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt
) |
905 S_00B22C_OC_LDS_EN(es_type
== PIPE_SHADER_TESS_EVAL
) |
906 S_00B22C_LDS_SIZE(shader
->config
.lds_size
) |
907 S_00B22C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
909 if (sscreen
->info
.chip_class
>= GFX10
) {
910 rsrc2
|= S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs
>> 5);
912 rsrc1
|= S_00B228_SGPRS((shader
->config
.num_sgprs
- 1) / 8);
913 rsrc2
|= S_00B22C_USER_SGPR_MSB_GFX9(num_user_sgprs
>> 5);
916 si_pm4_set_reg(pm4
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
, rsrc1
);
917 si_pm4_set_reg(pm4
, R_00B22C_SPI_SHADER_PGM_RSRC2_GS
, rsrc2
);
919 shader
->ctx_reg
.gs
.vgt_gs_onchip_cntl
=
920 S_028A44_ES_VERTS_PER_SUBGRP(shader
->gs_info
.es_verts_per_subgroup
) |
921 S_028A44_GS_PRIMS_PER_SUBGRP(shader
->gs_info
.gs_prims_per_subgroup
) |
922 S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader
->gs_info
.gs_inst_prims_in_subgroup
);
923 shader
->ctx_reg
.gs
.vgt_gs_max_prims_per_subgroup
=
924 S_028A94_MAX_PRIMS_PER_SUBGROUP(shader
->gs_info
.max_prims_per_subgroup
);
925 shader
->ctx_reg
.gs
.vgt_esgs_ring_itemsize
=
926 shader
->key
.part
.gs
.es
->esgs_itemsize
/ 4;
928 if (es_type
== PIPE_SHADER_TESS_EVAL
)
929 si_set_tesseval_regs(sscreen
, shader
->key
.part
.gs
.es
, pm4
);
931 polaris_set_vgt_vertex_reuse(sscreen
, shader
->key
.part
.gs
.es
,
934 si_pm4_set_reg(pm4
, R_00B220_SPI_SHADER_PGM_LO_GS
, va
>> 8);
935 si_pm4_set_reg(pm4
, R_00B224_SPI_SHADER_PGM_HI_GS
, S_00B224_MEM_BASE(va
>> 40));
937 si_pm4_set_reg(pm4
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
,
938 S_00B228_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
939 S_00B228_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
940 S_00B228_DX10_CLAMP(1) |
941 S_00B228_FLOAT_MODE(shader
->config
.float_mode
));
942 si_pm4_set_reg(pm4
, R_00B22C_SPI_SHADER_PGM_RSRC2_GS
,
943 S_00B22C_USER_SGPR(GFX6_GS_NUM_USER_SGPR
) |
944 S_00B22C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
948 /* Common tail code for NGG primitive shaders. */
949 static void gfx10_emit_shader_ngg_tail(struct si_context
*sctx
,
950 struct si_shader
*shader
,
951 unsigned initial_cdw
)
953 radeon_opt_set_context_reg(sctx
, R_0287FC_GE_MAX_OUTPUT_PER_SUBGROUP
,
954 SI_TRACKED_GE_MAX_OUTPUT_PER_SUBGROUP
,
955 shader
->ctx_reg
.ngg
.ge_max_output_per_subgroup
);
956 radeon_opt_set_context_reg(sctx
, R_028B4C_GE_NGG_SUBGRP_CNTL
,
957 SI_TRACKED_GE_NGG_SUBGRP_CNTL
,
958 shader
->ctx_reg
.ngg
.ge_ngg_subgrp_cntl
);
959 radeon_opt_set_context_reg(sctx
, R_028A84_VGT_PRIMITIVEID_EN
,
960 SI_TRACKED_VGT_PRIMITIVEID_EN
,
961 shader
->ctx_reg
.ngg
.vgt_primitiveid_en
);
962 radeon_opt_set_context_reg(sctx
, R_028A44_VGT_GS_ONCHIP_CNTL
,
963 SI_TRACKED_VGT_GS_ONCHIP_CNTL
,
964 shader
->ctx_reg
.ngg
.vgt_gs_onchip_cntl
);
965 radeon_opt_set_context_reg(sctx
, R_028B90_VGT_GS_INSTANCE_CNT
,
966 SI_TRACKED_VGT_GS_INSTANCE_CNT
,
967 shader
->ctx_reg
.ngg
.vgt_gs_instance_cnt
);
968 radeon_opt_set_context_reg(sctx
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
969 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE
,
970 shader
->ctx_reg
.ngg
.vgt_esgs_ring_itemsize
);
971 radeon_opt_set_context_reg(sctx
, R_028AB4_VGT_REUSE_OFF
,
972 SI_TRACKED_VGT_REUSE_OFF
,
973 shader
->ctx_reg
.ngg
.vgt_reuse_off
);
974 radeon_opt_set_context_reg(sctx
, R_0286C4_SPI_VS_OUT_CONFIG
,
975 SI_TRACKED_SPI_VS_OUT_CONFIG
,
976 shader
->ctx_reg
.ngg
.spi_vs_out_config
);
977 radeon_opt_set_context_reg2(sctx
, R_028708_SPI_SHADER_IDX_FORMAT
,
978 SI_TRACKED_SPI_SHADER_IDX_FORMAT
,
979 shader
->ctx_reg
.ngg
.spi_shader_idx_format
,
980 shader
->ctx_reg
.ngg
.spi_shader_pos_format
);
981 radeon_opt_set_context_reg(sctx
, R_028818_PA_CL_VTE_CNTL
,
982 SI_TRACKED_PA_CL_VTE_CNTL
,
983 shader
->ctx_reg
.ngg
.pa_cl_vte_cntl
);
984 radeon_opt_set_context_reg(sctx
, R_028838_PA_CL_NGG_CNTL
,
985 SI_TRACKED_PA_CL_NGG_CNTL
,
986 shader
->ctx_reg
.ngg
.pa_cl_ngg_cntl
);
988 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
989 sctx
->context_roll
= true;
991 if (shader
->ge_cntl
!= sctx
->last_multi_vgt_param
) {
992 radeon_set_uconfig_reg(sctx
->gfx_cs
, R_03096C_GE_CNTL
, shader
->ge_cntl
);
993 sctx
->last_multi_vgt_param
= shader
->ge_cntl
;
997 static void gfx10_emit_shader_ngg_notess_nogs(struct si_context
*sctx
)
999 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1000 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1005 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1008 static void gfx10_emit_shader_ngg_tess_nogs(struct si_context
*sctx
)
1010 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1011 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1016 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
1017 SI_TRACKED_VGT_TF_PARAM
,
1018 shader
->vgt_tf_param
);
1020 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1023 static void gfx10_emit_shader_ngg_notess_gs(struct si_context
*sctx
)
1025 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1026 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1031 radeon_opt_set_context_reg(sctx
, R_028B38_VGT_GS_MAX_VERT_OUT
,
1032 SI_TRACKED_VGT_GS_MAX_VERT_OUT
,
1033 shader
->ctx_reg
.ngg
.vgt_gs_max_vert_out
);
1035 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1038 static void gfx10_emit_shader_ngg_tess_gs(struct si_context
*sctx
)
1040 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1041 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1046 radeon_opt_set_context_reg(sctx
, R_028B38_VGT_GS_MAX_VERT_OUT
,
1047 SI_TRACKED_VGT_GS_MAX_VERT_OUT
,
1048 shader
->ctx_reg
.ngg
.vgt_gs_max_vert_out
);
1049 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
1050 SI_TRACKED_VGT_TF_PARAM
,
1051 shader
->vgt_tf_param
);
1053 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1057 * Prepare the PM4 image for \p shader, which will run as a merged ESGS shader
1060 static void gfx10_shader_ngg(struct si_screen
*sscreen
, struct si_shader
*shader
)
1062 const struct si_shader_selector
*gs_sel
= shader
->selector
;
1063 const struct tgsi_shader_info
*gs_info
= &gs_sel
->info
;
1064 enum pipe_shader_type gs_type
= shader
->selector
->type
;
1065 const struct si_shader_selector
*es_sel
=
1066 shader
->previous_stage_sel
? shader
->previous_stage_sel
: shader
->selector
;
1067 const struct tgsi_shader_info
*es_info
= &es_sel
->info
;
1068 enum pipe_shader_type es_type
= es_sel
->type
;
1069 unsigned num_user_sgprs
;
1070 unsigned nparams
, es_vgpr_comp_cnt
, gs_vgpr_comp_cnt
;
1072 unsigned window_space
=
1073 gs_info
->properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
];
1074 bool es_enable_prim_id
= shader
->key
.mono
.u
.vs_export_prim_id
|| es_info
->uses_primid
;
1075 unsigned gs_num_invocations
= MAX2(gs_sel
->gs_num_invocations
, 1);
1076 unsigned input_prim
=
1077 gs_type
== PIPE_SHADER_GEOMETRY
?
1078 gs_info
->properties
[TGSI_PROPERTY_GS_INPUT_PRIM
] :
1079 PIPE_PRIM_TRIANGLES
; /* TODO: Optimize when primtype is known */
1080 bool break_wave_at_eoi
= false;
1081 struct si_pm4_state
*pm4
= si_get_shader_pm4_state(shader
);
1085 if (es_type
== PIPE_SHADER_TESS_EVAL
) {
1086 pm4
->atom
.emit
= gs_type
== PIPE_SHADER_GEOMETRY
? gfx10_emit_shader_ngg_tess_gs
1087 : gfx10_emit_shader_ngg_tess_nogs
;
1089 pm4
->atom
.emit
= gs_type
== PIPE_SHADER_GEOMETRY
? gfx10_emit_shader_ngg_notess_gs
1090 : gfx10_emit_shader_ngg_notess_nogs
;
1093 va
= shader
->bo
->gpu_address
;
1094 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
1096 if (es_type
== PIPE_SHADER_VERTEX
) {
1097 /* VGPR5-8: (VertexID, UserVGPR0, UserVGPR1, UserVGPR2 / InstanceID) */
1098 es_vgpr_comp_cnt
= shader
->info
.uses_instanceid
? 3 : 0;
1100 if (es_info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
]) {
1101 num_user_sgprs
= SI_SGPR_VS_BLIT_DATA
+
1102 es_info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
];
1104 num_user_sgprs
= si_get_num_vs_user_sgprs(GFX9_VSGS_NUM_USER_SGPR
);
1107 assert(es_type
== PIPE_SHADER_TESS_EVAL
);
1108 es_vgpr_comp_cnt
= es_enable_prim_id
? 3 : 2;
1109 num_user_sgprs
= GFX9_TESGS_NUM_USER_SGPR
;
1111 if (es_enable_prim_id
|| gs_info
->uses_primid
)
1112 break_wave_at_eoi
= true;
1115 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
1116 * VGPR[0:4] are always loaded.
1118 * Vertex shaders always need to load VGPR3, because they need to
1119 * pass edge flags for decomposed primitives (such as quads) to the PA
1120 * for the GL_LINE polygon mode to skip rendering lines on inner edges.
1122 if (gs_info
->uses_invocationid
|| gs_type
== PIPE_SHADER_VERTEX
)
1123 gs_vgpr_comp_cnt
= 3; /* VGPR3 contains InvocationID, edge flags. */
1124 else if (gs_info
->uses_primid
)
1125 gs_vgpr_comp_cnt
= 2; /* VGPR2 contains PrimitiveID. */
1126 else if (input_prim
>= PIPE_PRIM_TRIANGLES
)
1127 gs_vgpr_comp_cnt
= 1; /* VGPR1 contains offsets 2, 3 */
1129 gs_vgpr_comp_cnt
= 0; /* VGPR0 contains offsets 0, 1 */
1131 si_pm4_set_reg(pm4
, R_00B320_SPI_SHADER_PGM_LO_ES
, va
>> 8);
1132 si_pm4_set_reg(pm4
, R_00B324_SPI_SHADER_PGM_HI_ES
, va
>> 40);
1133 si_pm4_set_reg(pm4
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
,
1134 S_00B228_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
1135 S_00B228_FLOAT_MODE(shader
->config
.float_mode
) |
1136 S_00B228_DX10_CLAMP(1) |
1137 S_00B228_MEM_ORDERED(1) |
1138 S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt
));
1139 si_pm4_set_reg(pm4
, R_00B22C_SPI_SHADER_PGM_RSRC2_GS
,
1140 S_00B22C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0) |
1141 S_00B22C_USER_SGPR(num_user_sgprs
) |
1142 S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt
) |
1143 S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs
>> 5) |
1144 S_00B22C_OC_LDS_EN(es_type
== PIPE_SHADER_TESS_EVAL
) |
1145 S_00B22C_LDS_SIZE(shader
->config
.lds_size
));
1147 /* TODO: Use NO_PC_EXPORT when applicable. */
1148 nparams
= MAX2(shader
->info
.nr_param_exports
, 1);
1149 shader
->ctx_reg
.ngg
.spi_vs_out_config
=
1150 S_0286C4_VS_EXPORT_COUNT(nparams
- 1);
1152 shader
->ctx_reg
.ngg
.spi_shader_idx_format
=
1153 S_028708_IDX0_EXPORT_FORMAT(V_028708_SPI_SHADER_1COMP
);
1154 shader
->ctx_reg
.ngg
.spi_shader_pos_format
=
1155 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP
) |
1156 S_02870C_POS1_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 1 ?
1157 V_02870C_SPI_SHADER_4COMP
:
1158 V_02870C_SPI_SHADER_NONE
) |
1159 S_02870C_POS2_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 2 ?
1160 V_02870C_SPI_SHADER_4COMP
:
1161 V_02870C_SPI_SHADER_NONE
) |
1162 S_02870C_POS3_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 3 ?
1163 V_02870C_SPI_SHADER_4COMP
:
1164 V_02870C_SPI_SHADER_NONE
);
1166 shader
->ctx_reg
.ngg
.vgt_primitiveid_en
=
1167 S_028A84_PRIMITIVEID_EN(es_enable_prim_id
) |
1168 S_028A84_NGG_DISABLE_PROVOK_REUSE(es_enable_prim_id
);
1170 if (gs_type
== PIPE_SHADER_GEOMETRY
) {
1171 shader
->ctx_reg
.ngg
.vgt_esgs_ring_itemsize
= es_sel
->esgs_itemsize
/ 4;
1172 shader
->ctx_reg
.ngg
.vgt_gs_max_vert_out
= gs_sel
->gs_max_out_vertices
;
1174 shader
->ctx_reg
.ngg
.vgt_esgs_ring_itemsize
= 1;
1177 if (es_type
== PIPE_SHADER_TESS_EVAL
)
1178 si_set_tesseval_regs(sscreen
, es_sel
, pm4
);
1180 shader
->ctx_reg
.ngg
.vgt_gs_onchip_cntl
=
1181 S_028A44_ES_VERTS_PER_SUBGRP(shader
->ngg
.hw_max_esverts
) |
1182 S_028A44_GS_PRIMS_PER_SUBGRP(shader
->ngg
.max_gsprims
) |
1183 S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader
->ngg
.max_gsprims
* gs_num_invocations
);
1184 shader
->ctx_reg
.ngg
.ge_max_output_per_subgroup
=
1185 S_0287FC_MAX_VERTS_PER_SUBGROUP(shader
->ngg
.max_out_verts
);
1186 shader
->ctx_reg
.ngg
.ge_ngg_subgrp_cntl
=
1187 S_028B4C_PRIM_AMP_FACTOR(shader
->ngg
.prim_amp_factor
) |
1188 S_028B4C_THDS_PER_SUBGRP(0); /* for fast launch */
1189 shader
->ctx_reg
.ngg
.vgt_gs_instance_cnt
=
1190 S_028B90_CNT(gs_num_invocations
) |
1191 S_028B90_ENABLE(gs_num_invocations
> 1) |
1192 S_028B90_EN_MAX_VERT_OUT_PER_GS_INSTANCE(
1193 shader
->ngg
.max_vert_out_per_gs_instance
);
1195 /* User edge flags are set by the pos exports. If user edge flags are
1196 * not used, we must use hw-generated edge flags and pass them via
1197 * the prim export to prevent drawing lines on internal edges of
1198 * decomposed primitives (such as quads) with polygon mode = lines.
1200 * TODO: We should combine hw-generated edge flags with user edge
1201 * flags in the shader.
1203 shader
->ctx_reg
.ngg
.pa_cl_ngg_cntl
=
1204 S_028838_INDEX_BUF_EDGE_FLAG_ENA(gs_type
== PIPE_SHADER_VERTEX
&&
1205 !gs_info
->writes_edgeflag
);
1208 S_03096C_PRIM_GRP_SIZE(shader
->ngg
.max_gsprims
) |
1209 S_03096C_VERT_GRP_SIZE(shader
->ngg
.hw_max_esverts
) |
1210 S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi
);
1213 shader
->ctx_reg
.ngg
.pa_cl_vte_cntl
=
1214 S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
1216 shader
->ctx_reg
.ngg
.pa_cl_vte_cntl
=
1217 S_028818_VTX_W0_FMT(1) |
1218 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
1219 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
1220 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
1223 shader
->ctx_reg
.ngg
.vgt_reuse_off
=
1224 S_028AB4_REUSE_OFF(sscreen
->info
.family
== CHIP_NAVI10
&&
1225 sscreen
->info
.chip_external_rev
== 0x1 &&
1226 es_type
== PIPE_SHADER_TESS_EVAL
);
1229 static void si_emit_shader_vs(struct si_context
*sctx
)
1231 struct si_shader
*shader
= sctx
->queued
.named
.vs
->shader
;
1232 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1237 radeon_opt_set_context_reg(sctx
, R_028A40_VGT_GS_MODE
,
1238 SI_TRACKED_VGT_GS_MODE
,
1239 shader
->ctx_reg
.vs
.vgt_gs_mode
);
1240 radeon_opt_set_context_reg(sctx
, R_028A84_VGT_PRIMITIVEID_EN
,
1241 SI_TRACKED_VGT_PRIMITIVEID_EN
,
1242 shader
->ctx_reg
.vs
.vgt_primitiveid_en
);
1244 if (sctx
->chip_class
<= GFX8
) {
1245 radeon_opt_set_context_reg(sctx
, R_028AB4_VGT_REUSE_OFF
,
1246 SI_TRACKED_VGT_REUSE_OFF
,
1247 shader
->ctx_reg
.vs
.vgt_reuse_off
);
1250 radeon_opt_set_context_reg(sctx
, R_0286C4_SPI_VS_OUT_CONFIG
,
1251 SI_TRACKED_SPI_VS_OUT_CONFIG
,
1252 shader
->ctx_reg
.vs
.spi_vs_out_config
);
1254 radeon_opt_set_context_reg(sctx
, R_02870C_SPI_SHADER_POS_FORMAT
,
1255 SI_TRACKED_SPI_SHADER_POS_FORMAT
,
1256 shader
->ctx_reg
.vs
.spi_shader_pos_format
);
1258 radeon_opt_set_context_reg(sctx
, R_028818_PA_CL_VTE_CNTL
,
1259 SI_TRACKED_PA_CL_VTE_CNTL
,
1260 shader
->ctx_reg
.vs
.pa_cl_vte_cntl
);
1262 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
1263 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
1264 SI_TRACKED_VGT_TF_PARAM
,
1265 shader
->vgt_tf_param
);
1267 if (shader
->vgt_vertex_reuse_block_cntl
)
1268 radeon_opt_set_context_reg(sctx
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
1269 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL
,
1270 shader
->vgt_vertex_reuse_block_cntl
);
1272 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
1273 sctx
->context_roll
= true;
1277 * Compute the state for \p shader, which will run as a vertex shader on the
1280 * If \p gs is non-NULL, it points to the geometry shader for which this shader
1281 * is the copy shader.
1283 static void si_shader_vs(struct si_screen
*sscreen
, struct si_shader
*shader
,
1284 struct si_shader_selector
*gs
)
1286 const struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1287 struct si_pm4_state
*pm4
;
1288 unsigned num_user_sgprs
, vgpr_comp_cnt
;
1290 unsigned nparams
, oc_lds_en
;
1291 unsigned window_space
=
1292 info
->properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
];
1293 bool enable_prim_id
= shader
->key
.mono
.u
.vs_export_prim_id
|| info
->uses_primid
;
1295 pm4
= si_get_shader_pm4_state(shader
);
1299 pm4
->atom
.emit
= si_emit_shader_vs
;
1301 /* We always write VGT_GS_MODE in the VS state, because every switch
1302 * between different shader pipelines involving a different GS or no
1303 * GS at all involves a switch of the VS (different GS use different
1304 * copy shaders). On the other hand, when the API switches from a GS to
1305 * no GS and then back to the same GS used originally, the GS state is
1309 unsigned mode
= V_028A40_GS_OFF
;
1311 /* PrimID needs GS scenario A. */
1313 mode
= V_028A40_GS_SCENARIO_A
;
1315 shader
->ctx_reg
.vs
.vgt_gs_mode
= S_028A40_MODE(mode
);
1316 shader
->ctx_reg
.vs
.vgt_primitiveid_en
= enable_prim_id
;
1318 shader
->ctx_reg
.vs
.vgt_gs_mode
= ac_vgt_gs_mode(gs
->gs_max_out_vertices
,
1319 sscreen
->info
.chip_class
);
1320 shader
->ctx_reg
.vs
.vgt_primitiveid_en
= 0;
1323 if (sscreen
->info
.chip_class
<= GFX8
) {
1324 /* Reuse needs to be set off if we write oViewport. */
1325 shader
->ctx_reg
.vs
.vgt_reuse_off
=
1326 S_028AB4_REUSE_OFF(info
->writes_viewport_index
);
1329 va
= shader
->bo
->gpu_address
;
1330 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
1333 vgpr_comp_cnt
= 0; /* only VertexID is needed for GS-COPY. */
1334 num_user_sgprs
= SI_GSCOPY_NUM_USER_SGPR
;
1335 } else if (shader
->selector
->type
== PIPE_SHADER_VERTEX
) {
1336 /* VGPR0-3: (VertexID, InstanceID / StepRate0, PrimID, InstanceID)
1337 * If PrimID is disabled. InstanceID / StepRate1 is loaded instead.
1338 * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
1340 vgpr_comp_cnt
= enable_prim_id
? 2 : (shader
->info
.uses_instanceid
? 1 : 0);
1342 if (info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
]) {
1343 num_user_sgprs
= SI_SGPR_VS_BLIT_DATA
+
1344 info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
];
1346 num_user_sgprs
= si_get_num_vs_user_sgprs(SI_VS_NUM_USER_SGPR
);
1348 } else if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
1349 vgpr_comp_cnt
= enable_prim_id
? 3 : 2;
1350 num_user_sgprs
= SI_TES_NUM_USER_SGPR
;
1352 unreachable("invalid shader selector type");
1354 /* VS is required to export at least one param. */
1355 nparams
= MAX2(shader
->info
.nr_param_exports
, 1);
1356 shader
->ctx_reg
.vs
.spi_vs_out_config
= S_0286C4_VS_EXPORT_COUNT(nparams
- 1);
1358 shader
->ctx_reg
.vs
.spi_shader_pos_format
=
1359 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP
) |
1360 S_02870C_POS1_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 1 ?
1361 V_02870C_SPI_SHADER_4COMP
:
1362 V_02870C_SPI_SHADER_NONE
) |
1363 S_02870C_POS2_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 2 ?
1364 V_02870C_SPI_SHADER_4COMP
:
1365 V_02870C_SPI_SHADER_NONE
) |
1366 S_02870C_POS3_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 3 ?
1367 V_02870C_SPI_SHADER_4COMP
:
1368 V_02870C_SPI_SHADER_NONE
);
1370 oc_lds_en
= shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
? 1 : 0;
1372 si_pm4_set_reg(pm4
, R_00B120_SPI_SHADER_PGM_LO_VS
, va
>> 8);
1373 si_pm4_set_reg(pm4
, R_00B124_SPI_SHADER_PGM_HI_VS
, S_00B124_MEM_BASE(va
>> 40));
1374 si_pm4_set_reg(pm4
, R_00B128_SPI_SHADER_PGM_RSRC1_VS
,
1375 S_00B128_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
1376 S_00B128_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
1377 S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt
) |
1378 S_00B128_DX10_CLAMP(1) |
1379 S_00B128_FLOAT_MODE(shader
->config
.float_mode
));
1380 si_pm4_set_reg(pm4
, R_00B12C_SPI_SHADER_PGM_RSRC2_VS
,
1381 S_00B12C_USER_SGPR(num_user_sgprs
) |
1382 S_00B12C_OC_LDS_EN(oc_lds_en
) |
1383 S_00B12C_SO_BASE0_EN(!!shader
->selector
->so
.stride
[0]) |
1384 S_00B12C_SO_BASE1_EN(!!shader
->selector
->so
.stride
[1]) |
1385 S_00B12C_SO_BASE2_EN(!!shader
->selector
->so
.stride
[2]) |
1386 S_00B12C_SO_BASE3_EN(!!shader
->selector
->so
.stride
[3]) |
1387 S_00B12C_SO_EN(!!shader
->selector
->so
.num_outputs
) |
1388 S_00B12C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
1391 shader
->ctx_reg
.vs
.pa_cl_vte_cntl
=
1392 S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
1394 shader
->ctx_reg
.vs
.pa_cl_vte_cntl
=
1395 S_028818_VTX_W0_FMT(1) |
1396 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
1397 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
1398 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
1400 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
1401 si_set_tesseval_regs(sscreen
, shader
->selector
, pm4
);
1403 polaris_set_vgt_vertex_reuse(sscreen
, shader
->selector
, shader
, pm4
);
1406 static unsigned si_get_ps_num_interp(struct si_shader
*ps
)
1408 struct tgsi_shader_info
*info
= &ps
->selector
->info
;
1409 unsigned num_colors
= !!(info
->colors_read
& 0x0f) +
1410 !!(info
->colors_read
& 0xf0);
1411 unsigned num_interp
= ps
->selector
->info
.num_inputs
+
1412 (ps
->key
.part
.ps
.prolog
.color_two_side
? num_colors
: 0);
1414 assert(num_interp
<= 32);
1415 return MIN2(num_interp
, 32);
1418 static unsigned si_get_spi_shader_col_format(struct si_shader
*shader
)
1420 unsigned value
= shader
->key
.part
.ps
.epilog
.spi_shader_col_format
;
1421 unsigned i
, num_targets
= (util_last_bit(value
) + 3) / 4;
1423 /* If the i-th target format is set, all previous target formats must
1424 * be non-zero to avoid hangs.
1426 for (i
= 0; i
< num_targets
; i
++)
1427 if (!(value
& (0xf << (i
* 4))))
1428 value
|= V_028714_SPI_SHADER_32_R
<< (i
* 4);
1433 static void si_emit_shader_ps(struct si_context
*sctx
)
1435 struct si_shader
*shader
= sctx
->queued
.named
.ps
->shader
;
1436 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1441 /* R_0286CC_SPI_PS_INPUT_ENA, R_0286D0_SPI_PS_INPUT_ADDR*/
1442 radeon_opt_set_context_reg2(sctx
, R_0286CC_SPI_PS_INPUT_ENA
,
1443 SI_TRACKED_SPI_PS_INPUT_ENA
,
1444 shader
->ctx_reg
.ps
.spi_ps_input_ena
,
1445 shader
->ctx_reg
.ps
.spi_ps_input_addr
);
1447 radeon_opt_set_context_reg(sctx
, R_0286E0_SPI_BARYC_CNTL
,
1448 SI_TRACKED_SPI_BARYC_CNTL
,
1449 shader
->ctx_reg
.ps
.spi_baryc_cntl
);
1450 radeon_opt_set_context_reg(sctx
, R_0286D8_SPI_PS_IN_CONTROL
,
1451 SI_TRACKED_SPI_PS_IN_CONTROL
,
1452 shader
->ctx_reg
.ps
.spi_ps_in_control
);
1454 /* R_028710_SPI_SHADER_Z_FORMAT, R_028714_SPI_SHADER_COL_FORMAT */
1455 radeon_opt_set_context_reg2(sctx
, R_028710_SPI_SHADER_Z_FORMAT
,
1456 SI_TRACKED_SPI_SHADER_Z_FORMAT
,
1457 shader
->ctx_reg
.ps
.spi_shader_z_format
,
1458 shader
->ctx_reg
.ps
.spi_shader_col_format
);
1460 radeon_opt_set_context_reg(sctx
, R_02823C_CB_SHADER_MASK
,
1461 SI_TRACKED_CB_SHADER_MASK
,
1462 shader
->ctx_reg
.ps
.cb_shader_mask
);
1464 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
1465 sctx
->context_roll
= true;
1468 static void si_shader_ps(struct si_screen
*sscreen
, struct si_shader
*shader
)
1470 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1471 struct si_pm4_state
*pm4
;
1472 unsigned spi_ps_in_control
, spi_shader_col_format
, cb_shader_mask
;
1473 unsigned spi_baryc_cntl
= S_0286E0_FRONT_FACE_ALL_BITS(1);
1475 unsigned input_ena
= shader
->config
.spi_ps_input_ena
;
1477 /* we need to enable at least one of them, otherwise we hang the GPU */
1478 assert(G_0286CC_PERSP_SAMPLE_ENA(input_ena
) ||
1479 G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
1480 G_0286CC_PERSP_CENTROID_ENA(input_ena
) ||
1481 G_0286CC_PERSP_PULL_MODEL_ENA(input_ena
) ||
1482 G_0286CC_LINEAR_SAMPLE_ENA(input_ena
) ||
1483 G_0286CC_LINEAR_CENTER_ENA(input_ena
) ||
1484 G_0286CC_LINEAR_CENTROID_ENA(input_ena
) ||
1485 G_0286CC_LINE_STIPPLE_TEX_ENA(input_ena
));
1486 /* POS_W_FLOAT_ENA requires one of the perspective weights. */
1487 assert(!G_0286CC_POS_W_FLOAT_ENA(input_ena
) ||
1488 G_0286CC_PERSP_SAMPLE_ENA(input_ena
) ||
1489 G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
1490 G_0286CC_PERSP_CENTROID_ENA(input_ena
) ||
1491 G_0286CC_PERSP_PULL_MODEL_ENA(input_ena
));
1493 /* Validate interpolation optimization flags (read as implications). */
1494 assert(!shader
->key
.part
.ps
.prolog
.bc_optimize_for_persp
||
1495 (G_0286CC_PERSP_CENTER_ENA(input_ena
) &&
1496 G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
1497 assert(!shader
->key
.part
.ps
.prolog
.bc_optimize_for_linear
||
1498 (G_0286CC_LINEAR_CENTER_ENA(input_ena
) &&
1499 G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
1500 assert(!shader
->key
.part
.ps
.prolog
.force_persp_center_interp
||
1501 (!G_0286CC_PERSP_SAMPLE_ENA(input_ena
) &&
1502 !G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
1503 assert(!shader
->key
.part
.ps
.prolog
.force_linear_center_interp
||
1504 (!G_0286CC_LINEAR_SAMPLE_ENA(input_ena
) &&
1505 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
1506 assert(!shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
||
1507 (!G_0286CC_PERSP_CENTER_ENA(input_ena
) &&
1508 !G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
1509 assert(!shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
||
1510 (!G_0286CC_LINEAR_CENTER_ENA(input_ena
) &&
1511 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
1513 /* Validate cases when the optimizations are off (read as implications). */
1514 assert(shader
->key
.part
.ps
.prolog
.bc_optimize_for_persp
||
1515 !G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
1516 !G_0286CC_PERSP_CENTROID_ENA(input_ena
));
1517 assert(shader
->key
.part
.ps
.prolog
.bc_optimize_for_linear
||
1518 !G_0286CC_LINEAR_CENTER_ENA(input_ena
) ||
1519 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
));
1521 pm4
= si_get_shader_pm4_state(shader
);
1525 pm4
->atom
.emit
= si_emit_shader_ps
;
1527 /* SPI_BARYC_CNTL.POS_FLOAT_LOCATION
1529 * 0 -> Position = pixel center
1530 * 1 -> Position = pixel centroid
1531 * 2 -> Position = at sample position
1533 * From GLSL 4.5 specification, section 7.1:
1534 * "The variable gl_FragCoord is available as an input variable from
1535 * within fragment shaders and it holds the window relative coordinates
1536 * (x, y, z, 1/w) values for the fragment. If multi-sampling, this
1537 * value can be for any location within the pixel, or one of the
1538 * fragment samples. The use of centroid does not further restrict
1539 * this value to be inside the current primitive."
1541 * Meaning that centroid has no effect and we can return anything within
1542 * the pixel. Thus, return the value at sample position, because that's
1543 * the most accurate one shaders can get.
1545 spi_baryc_cntl
|= S_0286E0_POS_FLOAT_LOCATION(2);
1547 if (info
->properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
] ==
1548 TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)
1549 spi_baryc_cntl
|= S_0286E0_POS_FLOAT_ULC(1);
1551 spi_shader_col_format
= si_get_spi_shader_col_format(shader
);
1552 cb_shader_mask
= ac_get_cb_shader_mask(spi_shader_col_format
);
1554 /* Ensure that some export memory is always allocated, for two reasons:
1556 * 1) Correctness: The hardware ignores the EXEC mask if no export
1557 * memory is allocated, so KILL and alpha test do not work correctly
1559 * 2) Performance: Every shader needs at least a NULL export, even when
1560 * it writes no color/depth output. The NULL export instruction
1561 * stalls without this setting.
1563 * Don't add this to CB_SHADER_MASK.
1565 if (!spi_shader_col_format
&&
1566 !info
->writes_z
&& !info
->writes_stencil
&& !info
->writes_samplemask
)
1567 spi_shader_col_format
= V_028714_SPI_SHADER_32_R
;
1569 shader
->ctx_reg
.ps
.spi_ps_input_ena
= input_ena
;
1570 shader
->ctx_reg
.ps
.spi_ps_input_addr
= shader
->config
.spi_ps_input_addr
;
1572 /* Set interpolation controls. */
1573 spi_ps_in_control
= S_0286D8_NUM_INTERP(si_get_ps_num_interp(shader
));
1575 shader
->ctx_reg
.ps
.spi_baryc_cntl
= spi_baryc_cntl
;
1576 shader
->ctx_reg
.ps
.spi_ps_in_control
= spi_ps_in_control
;
1577 shader
->ctx_reg
.ps
.spi_shader_z_format
=
1578 ac_get_spi_shader_z_format(info
->writes_z
,
1579 info
->writes_stencil
,
1580 info
->writes_samplemask
);
1581 shader
->ctx_reg
.ps
.spi_shader_col_format
= spi_shader_col_format
;
1582 shader
->ctx_reg
.ps
.cb_shader_mask
= cb_shader_mask
;
1584 va
= shader
->bo
->gpu_address
;
1585 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
1586 si_pm4_set_reg(pm4
, R_00B020_SPI_SHADER_PGM_LO_PS
, va
>> 8);
1587 si_pm4_set_reg(pm4
, R_00B024_SPI_SHADER_PGM_HI_PS
, S_00B024_MEM_BASE(va
>> 40));
1590 S_00B028_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
1591 S_00B028_DX10_CLAMP(1) |
1592 S_00B028_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
1593 S_00B028_FLOAT_MODE(shader
->config
.float_mode
);
1595 if (sscreen
->info
.chip_class
< GFX10
) {
1596 rsrc1
|= S_00B028_SGPRS((shader
->config
.num_sgprs
- 1) / 8);
1599 si_pm4_set_reg(pm4
, R_00B028_SPI_SHADER_PGM_RSRC1_PS
, rsrc1
);
1600 si_pm4_set_reg(pm4
, R_00B02C_SPI_SHADER_PGM_RSRC2_PS
,
1601 S_00B02C_EXTRA_LDS_SIZE(shader
->config
.lds_size
) |
1602 S_00B02C_USER_SGPR(SI_PS_NUM_USER_SGPR
) |
1603 S_00B32C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
1606 static void si_shader_init_pm4_state(struct si_screen
*sscreen
,
1607 struct si_shader
*shader
)
1609 switch (shader
->selector
->type
) {
1610 case PIPE_SHADER_VERTEX
:
1611 if (shader
->key
.as_ls
)
1612 si_shader_ls(sscreen
, shader
);
1613 else if (shader
->key
.as_es
)
1614 si_shader_es(sscreen
, shader
);
1615 else if (shader
->key
.as_ngg
)
1616 gfx10_shader_ngg(sscreen
, shader
);
1618 si_shader_vs(sscreen
, shader
, NULL
);
1620 case PIPE_SHADER_TESS_CTRL
:
1621 si_shader_hs(sscreen
, shader
);
1623 case PIPE_SHADER_TESS_EVAL
:
1624 if (shader
->key
.as_es
)
1625 si_shader_es(sscreen
, shader
);
1626 else if (shader
->key
.as_ngg
)
1627 gfx10_shader_ngg(sscreen
, shader
);
1629 si_shader_vs(sscreen
, shader
, NULL
);
1631 case PIPE_SHADER_GEOMETRY
:
1632 if (shader
->key
.as_ngg
)
1633 gfx10_shader_ngg(sscreen
, shader
);
1635 si_shader_gs(sscreen
, shader
);
1637 case PIPE_SHADER_FRAGMENT
:
1638 si_shader_ps(sscreen
, shader
);
1645 static unsigned si_get_alpha_test_func(struct si_context
*sctx
)
1647 /* Alpha-test should be disabled if colorbuffer 0 is integer. */
1648 if (sctx
->queued
.named
.dsa
)
1649 return sctx
->queued
.named
.dsa
->alpha_func
;
1651 return PIPE_FUNC_ALWAYS
;
1654 void si_shader_selector_key_vs(struct si_context
*sctx
,
1655 struct si_shader_selector
*vs
,
1656 struct si_shader_key
*key
,
1657 struct si_vs_prolog_bits
*prolog_key
)
1659 if (!sctx
->vertex_elements
||
1660 vs
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
])
1663 struct si_vertex_elements
*elts
= sctx
->vertex_elements
;
1665 prolog_key
->instance_divisor_is_one
= elts
->instance_divisor_is_one
;
1666 prolog_key
->instance_divisor_is_fetched
= elts
->instance_divisor_is_fetched
;
1667 prolog_key
->unpack_instance_id_from_vertex_id
=
1668 sctx
->prim_discard_cs_instancing
;
1670 /* Prefer a monolithic shader to allow scheduling divisions around
1672 if (prolog_key
->instance_divisor_is_fetched
)
1673 key
->opt
.prefer_mono
= 1;
1675 unsigned count
= MIN2(vs
->info
.num_inputs
, elts
->count
);
1676 unsigned count_mask
= (1 << count
) - 1;
1677 unsigned fix
= elts
->fix_fetch_always
& count_mask
;
1678 unsigned opencode
= elts
->fix_fetch_opencode
& count_mask
;
1680 if (sctx
->vertex_buffer_unaligned
& elts
->vb_alignment_check_mask
) {
1681 uint32_t mask
= elts
->fix_fetch_unaligned
& count_mask
;
1683 unsigned i
= u_bit_scan(&mask
);
1684 unsigned log_hw_load_size
= 1 + ((elts
->hw_load_is_dword
>> i
) & 1);
1685 unsigned vbidx
= elts
->vertex_buffer_index
[i
];
1686 struct pipe_vertex_buffer
*vb
= &sctx
->vertex_buffer
[vbidx
];
1687 unsigned align_mask
= (1 << log_hw_load_size
) - 1;
1688 if (vb
->buffer_offset
& align_mask
||
1689 vb
->stride
& align_mask
) {
1697 unsigned i
= u_bit_scan(&fix
);
1698 key
->mono
.vs_fix_fetch
[i
].bits
= elts
->fix_fetch
[i
];
1700 key
->mono
.vs_fetch_opencode
= opencode
;
1703 static void si_shader_selector_key_hw_vs(struct si_context
*sctx
,
1704 struct si_shader_selector
*vs
,
1705 struct si_shader_key
*key
)
1707 struct si_shader_selector
*ps
= sctx
->ps_shader
.cso
;
1709 key
->opt
.clip_disable
=
1710 sctx
->queued
.named
.rasterizer
->clip_plane_enable
== 0 &&
1711 (vs
->info
.clipdist_writemask
||
1712 vs
->info
.writes_clipvertex
) &&
1713 !vs
->info
.culldist_writemask
;
1715 /* Find out if PS is disabled. */
1716 bool ps_disabled
= true;
1718 const struct si_state_blend
*blend
= sctx
->queued
.named
.blend
;
1719 bool alpha_to_coverage
= blend
&& blend
->alpha_to_coverage
;
1720 bool ps_modifies_zs
= ps
->info
.uses_kill
||
1721 ps
->info
.writes_z
||
1722 ps
->info
.writes_stencil
||
1723 ps
->info
.writes_samplemask
||
1724 alpha_to_coverage
||
1725 si_get_alpha_test_func(sctx
) != PIPE_FUNC_ALWAYS
;
1726 unsigned ps_colormask
= si_get_total_colormask(sctx
);
1728 ps_disabled
= sctx
->queued
.named
.rasterizer
->rasterizer_discard
||
1731 !ps
->info
.writes_memory
);
1734 /* Find out which VS outputs aren't used by the PS. */
1735 uint64_t outputs_written
= vs
->outputs_written_before_ps
;
1736 uint64_t inputs_read
= 0;
1738 /* Ignore outputs that are not passed from VS to PS. */
1739 outputs_written
&= ~((1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_POSITION
, 0, true)) |
1740 (1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_PSIZE
, 0, true)) |
1741 (1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_CLIPVERTEX
, 0, true)));
1744 inputs_read
= ps
->inputs_read
;
1747 uint64_t linked
= outputs_written
& inputs_read
;
1749 key
->opt
.kill_outputs
= ~linked
& outputs_written
;
1752 /* Compute the key for the hw shader variant */
1753 static inline void si_shader_selector_key(struct pipe_context
*ctx
,
1754 struct si_shader_selector
*sel
,
1755 union si_vgt_stages_key stages_key
,
1756 struct si_shader_key
*key
)
1758 struct si_context
*sctx
= (struct si_context
*)ctx
;
1760 memset(key
, 0, sizeof(*key
));
1762 switch (sel
->type
) {
1763 case PIPE_SHADER_VERTEX
:
1764 si_shader_selector_key_vs(sctx
, sel
, key
, &key
->part
.vs
.prolog
);
1766 if (sctx
->tes_shader
.cso
)
1768 else if (sctx
->gs_shader
.cso
)
1771 key
->as_ngg
= stages_key
.u
.ngg
;
1772 si_shader_selector_key_hw_vs(sctx
, sel
, key
);
1774 if (sctx
->ps_shader
.cso
&& sctx
->ps_shader
.cso
->info
.uses_primid
)
1775 key
->mono
.u
.vs_export_prim_id
= 1;
1778 case PIPE_SHADER_TESS_CTRL
:
1779 if (sctx
->chip_class
>= GFX9
) {
1780 si_shader_selector_key_vs(sctx
, sctx
->vs_shader
.cso
,
1781 key
, &key
->part
.tcs
.ls_prolog
);
1782 key
->part
.tcs
.ls
= sctx
->vs_shader
.cso
;
1784 /* When the LS VGPR fix is needed, monolithic shaders
1786 * - avoid initializing EXEC in both the LS prolog
1787 * and the LS main part when !vs_needs_prolog
1788 * - remove the fixup for unused input VGPRs
1790 key
->part
.tcs
.ls_prolog
.ls_vgpr_fix
= sctx
->ls_vgpr_fix
;
1792 /* The LS output / HS input layout can be communicated
1793 * directly instead of via user SGPRs for merged LS-HS.
1794 * The LS VGPR fix prefers this too.
1796 key
->opt
.prefer_mono
= 1;
1799 key
->part
.tcs
.epilog
.prim_mode
=
1800 sctx
->tes_shader
.cso
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
1801 key
->part
.tcs
.epilog
.invoc0_tess_factors_are_def
=
1802 sel
->tcs_info
.tessfactors_are_def_in_all_invocs
;
1803 key
->part
.tcs
.epilog
.tes_reads_tess_factors
=
1804 sctx
->tes_shader
.cso
->info
.reads_tess_factors
;
1806 if (sel
== sctx
->fixed_func_tcs_shader
.cso
)
1807 key
->mono
.u
.ff_tcs_inputs_to_copy
= sctx
->vs_shader
.cso
->outputs_written
;
1809 case PIPE_SHADER_TESS_EVAL
:
1810 if (sctx
->gs_shader
.cso
)
1813 key
->as_ngg
= stages_key
.u
.ngg
;
1814 si_shader_selector_key_hw_vs(sctx
, sel
, key
);
1816 if (sctx
->ps_shader
.cso
&& sctx
->ps_shader
.cso
->info
.uses_primid
)
1817 key
->mono
.u
.vs_export_prim_id
= 1;
1820 case PIPE_SHADER_GEOMETRY
:
1821 if (sctx
->chip_class
>= GFX9
) {
1822 if (sctx
->tes_shader
.cso
) {
1823 key
->part
.gs
.es
= sctx
->tes_shader
.cso
;
1825 si_shader_selector_key_vs(sctx
, sctx
->vs_shader
.cso
,
1826 key
, &key
->part
.gs
.vs_prolog
);
1827 key
->part
.gs
.es
= sctx
->vs_shader
.cso
;
1828 key
->part
.gs
.prolog
.gfx9_prev_is_vs
= 1;
1831 key
->as_ngg
= stages_key
.u
.ngg
;
1833 /* Merged ES-GS can have unbalanced wave usage.
1835 * ES threads are per-vertex, while GS threads are
1836 * per-primitive. So without any amplification, there
1837 * are fewer GS threads than ES threads, which can result
1838 * in empty (no-op) GS waves. With too much amplification,
1839 * there are more GS threads than ES threads, which
1840 * can result in empty (no-op) ES waves.
1842 * Non-monolithic shaders are implemented by setting EXEC
1843 * at the beginning of shader parts, and don't jump to
1844 * the end if EXEC is 0.
1846 * Monolithic shaders use conditional blocks, so they can
1847 * jump and skip empty waves of ES or GS. So set this to
1848 * always use optimized variants, which are monolithic.
1850 key
->opt
.prefer_mono
= 1;
1852 key
->part
.gs
.prolog
.tri_strip_adj_fix
= sctx
->gs_tri_strip_adj_fix
;
1854 case PIPE_SHADER_FRAGMENT
: {
1855 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
1856 struct si_state_blend
*blend
= sctx
->queued
.named
.blend
;
1858 if (sel
->info
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
] &&
1859 sel
->info
.colors_written
== 0x1)
1860 key
->part
.ps
.epilog
.last_cbuf
= MAX2(sctx
->framebuffer
.state
.nr_cbufs
, 1) - 1;
1863 /* Select the shader color format based on whether
1864 * blending or alpha are needed.
1866 key
->part
.ps
.epilog
.spi_shader_col_format
=
1867 (blend
->blend_enable_4bit
& blend
->need_src_alpha_4bit
&
1868 sctx
->framebuffer
.spi_shader_col_format_blend_alpha
) |
1869 (blend
->blend_enable_4bit
& ~blend
->need_src_alpha_4bit
&
1870 sctx
->framebuffer
.spi_shader_col_format_blend
) |
1871 (~blend
->blend_enable_4bit
& blend
->need_src_alpha_4bit
&
1872 sctx
->framebuffer
.spi_shader_col_format_alpha
) |
1873 (~blend
->blend_enable_4bit
& ~blend
->need_src_alpha_4bit
&
1874 sctx
->framebuffer
.spi_shader_col_format
);
1875 key
->part
.ps
.epilog
.spi_shader_col_format
&= blend
->cb_target_enabled_4bit
;
1877 /* The output for dual source blending should have
1878 * the same format as the first output.
1880 if (blend
->dual_src_blend
)
1881 key
->part
.ps
.epilog
.spi_shader_col_format
|=
1882 (key
->part
.ps
.epilog
.spi_shader_col_format
& 0xf) << 4;
1884 key
->part
.ps
.epilog
.spi_shader_col_format
= sctx
->framebuffer
.spi_shader_col_format
;
1886 /* If alpha-to-coverage is enabled, we have to export alpha
1887 * even if there is no color buffer.
1889 if (!(key
->part
.ps
.epilog
.spi_shader_col_format
& 0xf) &&
1890 blend
&& blend
->alpha_to_coverage
)
1891 key
->part
.ps
.epilog
.spi_shader_col_format
|= V_028710_SPI_SHADER_32_AR
;
1893 /* On GFX6 and GFX7 except Hawaii, the CB doesn't clamp outputs
1894 * to the range supported by the type if a channel has less
1895 * than 16 bits and the export format is 16_ABGR.
1897 if (sctx
->chip_class
<= GFX7
&& sctx
->family
!= CHIP_HAWAII
) {
1898 key
->part
.ps
.epilog
.color_is_int8
= sctx
->framebuffer
.color_is_int8
;
1899 key
->part
.ps
.epilog
.color_is_int10
= sctx
->framebuffer
.color_is_int10
;
1902 /* Disable unwritten outputs (if WRITE_ALL_CBUFS isn't enabled). */
1903 if (!key
->part
.ps
.epilog
.last_cbuf
) {
1904 key
->part
.ps
.epilog
.spi_shader_col_format
&= sel
->colors_written_4bit
;
1905 key
->part
.ps
.epilog
.color_is_int8
&= sel
->info
.colors_written
;
1906 key
->part
.ps
.epilog
.color_is_int10
&= sel
->info
.colors_written
;
1909 bool is_poly
= !util_prim_is_points_or_lines(sctx
->current_rast_prim
);
1910 bool is_line
= util_prim_is_lines(sctx
->current_rast_prim
);
1912 key
->part
.ps
.prolog
.color_two_side
= rs
->two_side
&& sel
->info
.colors_read
;
1913 key
->part
.ps
.prolog
.flatshade_colors
= rs
->flatshade
&& sel
->info
.colors_read
;
1915 if (sctx
->queued
.named
.blend
) {
1916 key
->part
.ps
.epilog
.alpha_to_one
= sctx
->queued
.named
.blend
->alpha_to_one
&&
1917 rs
->multisample_enable
;
1920 key
->part
.ps
.prolog
.poly_stipple
= rs
->poly_stipple_enable
&& is_poly
;
1921 key
->part
.ps
.epilog
.poly_line_smoothing
= ((is_poly
&& rs
->poly_smooth
) ||
1922 (is_line
&& rs
->line_smooth
)) &&
1923 sctx
->framebuffer
.nr_samples
<= 1;
1924 key
->part
.ps
.epilog
.clamp_color
= rs
->clamp_fragment_color
;
1926 if (sctx
->ps_iter_samples
> 1 &&
1927 sel
->info
.reads_samplemask
) {
1928 key
->part
.ps
.prolog
.samplemask_log_ps_iter
=
1929 util_logbase2(sctx
->ps_iter_samples
);
1932 if (rs
->force_persample_interp
&&
1933 rs
->multisample_enable
&&
1934 sctx
->framebuffer
.nr_samples
> 1 &&
1935 sctx
->ps_iter_samples
> 1) {
1936 key
->part
.ps
.prolog
.force_persp_sample_interp
=
1937 sel
->info
.uses_persp_center
||
1938 sel
->info
.uses_persp_centroid
;
1940 key
->part
.ps
.prolog
.force_linear_sample_interp
=
1941 sel
->info
.uses_linear_center
||
1942 sel
->info
.uses_linear_centroid
;
1943 } else if (rs
->multisample_enable
&&
1944 sctx
->framebuffer
.nr_samples
> 1) {
1945 key
->part
.ps
.prolog
.bc_optimize_for_persp
=
1946 sel
->info
.uses_persp_center
&&
1947 sel
->info
.uses_persp_centroid
;
1948 key
->part
.ps
.prolog
.bc_optimize_for_linear
=
1949 sel
->info
.uses_linear_center
&&
1950 sel
->info
.uses_linear_centroid
;
1952 /* Make sure SPI doesn't compute more than 1 pair
1953 * of (i,j), which is the optimization here. */
1954 key
->part
.ps
.prolog
.force_persp_center_interp
=
1955 sel
->info
.uses_persp_center
+
1956 sel
->info
.uses_persp_centroid
+
1957 sel
->info
.uses_persp_sample
> 1;
1959 key
->part
.ps
.prolog
.force_linear_center_interp
=
1960 sel
->info
.uses_linear_center
+
1961 sel
->info
.uses_linear_centroid
+
1962 sel
->info
.uses_linear_sample
> 1;
1964 if (sel
->info
.opcode_count
[TGSI_OPCODE_INTERP_SAMPLE
])
1965 key
->mono
.u
.ps
.interpolate_at_sample_force_center
= 1;
1968 key
->part
.ps
.epilog
.alpha_func
= si_get_alpha_test_func(sctx
);
1970 /* ps_uses_fbfetch is true only if the color buffer is bound. */
1971 if (sctx
->ps_uses_fbfetch
&& !sctx
->blitter
->running
) {
1972 struct pipe_surface
*cb0
= sctx
->framebuffer
.state
.cbufs
[0];
1973 struct pipe_resource
*tex
= cb0
->texture
;
1975 /* 1D textures are allocated and used as 2D on GFX9. */
1976 key
->mono
.u
.ps
.fbfetch_msaa
= sctx
->framebuffer
.nr_samples
> 1;
1977 key
->mono
.u
.ps
.fbfetch_is_1D
= sctx
->chip_class
!= GFX9
&&
1978 (tex
->target
== PIPE_TEXTURE_1D
||
1979 tex
->target
== PIPE_TEXTURE_1D_ARRAY
);
1980 key
->mono
.u
.ps
.fbfetch_layered
= tex
->target
== PIPE_TEXTURE_1D_ARRAY
||
1981 tex
->target
== PIPE_TEXTURE_2D_ARRAY
||
1982 tex
->target
== PIPE_TEXTURE_CUBE
||
1983 tex
->target
== PIPE_TEXTURE_CUBE_ARRAY
||
1984 tex
->target
== PIPE_TEXTURE_3D
;
1992 if (unlikely(sctx
->screen
->debug_flags
& DBG(NO_OPT_VARIANT
)))
1993 memset(&key
->opt
, 0, sizeof(key
->opt
));
1996 static void si_build_shader_variant(struct si_shader
*shader
,
2000 struct si_shader_selector
*sel
= shader
->selector
;
2001 struct si_screen
*sscreen
= sel
->screen
;
2002 struct ac_llvm_compiler
*compiler
;
2003 struct pipe_debug_callback
*debug
= &shader
->compiler_ctx_state
.debug
;
2005 if (thread_index
>= 0) {
2007 assert(thread_index
< ARRAY_SIZE(sscreen
->compiler_lowp
));
2008 compiler
= &sscreen
->compiler_lowp
[thread_index
];
2010 assert(thread_index
< ARRAY_SIZE(sscreen
->compiler
));
2011 compiler
= &sscreen
->compiler
[thread_index
];
2016 assert(!low_priority
);
2017 compiler
= shader
->compiler_ctx_state
.compiler
;
2020 if (unlikely(!si_shader_create(sscreen
, compiler
, shader
, debug
))) {
2021 PRINT_ERR("Failed to build shader variant (type=%u)\n",
2023 shader
->compilation_failed
= true;
2027 if (shader
->compiler_ctx_state
.is_debug_context
) {
2028 FILE *f
= open_memstream(&shader
->shader_log
,
2029 &shader
->shader_log_size
);
2031 si_shader_dump(sscreen
, shader
, NULL
, sel
->type
, f
, false);
2036 si_shader_init_pm4_state(sscreen
, shader
);
2039 static void si_build_shader_variant_low_priority(void *job
, int thread_index
)
2041 struct si_shader
*shader
= (struct si_shader
*)job
;
2043 assert(thread_index
>= 0);
2045 si_build_shader_variant(shader
, thread_index
, true);
2048 static const struct si_shader_key zeroed
;
2050 static bool si_check_missing_main_part(struct si_screen
*sscreen
,
2051 struct si_shader_selector
*sel
,
2052 struct si_compiler_ctx_state
*compiler_state
,
2053 struct si_shader_key
*key
)
2055 struct si_shader
**mainp
= si_get_main_shader_part(sel
, key
);
2058 struct si_shader
*main_part
= CALLOC_STRUCT(si_shader
);
2063 /* We can leave the fence as permanently signaled because the
2064 * main part becomes visible globally only after it has been
2066 util_queue_fence_init(&main_part
->ready
);
2068 main_part
->selector
= sel
;
2069 main_part
->key
.as_es
= key
->as_es
;
2070 main_part
->key
.as_ls
= key
->as_ls
;
2071 main_part
->key
.as_ngg
= key
->as_ngg
;
2072 main_part
->is_monolithic
= false;
2074 if (si_compile_tgsi_shader(sscreen
, compiler_state
->compiler
,
2075 main_part
, &compiler_state
->debug
) != 0) {
2085 * Select a shader variant according to the shader key.
2087 * \param optimized_or_none If the key describes an optimized shader variant and
2088 * the compilation isn't finished, don't select any
2089 * shader and return an error.
2091 int si_shader_select_with_key(struct si_screen
*sscreen
,
2092 struct si_shader_ctx_state
*state
,
2093 struct si_compiler_ctx_state
*compiler_state
,
2094 struct si_shader_key
*key
,
2096 bool optimized_or_none
)
2098 struct si_shader_selector
*sel
= state
->cso
;
2099 struct si_shader_selector
*previous_stage_sel
= NULL
;
2100 struct si_shader
*current
= state
->current
;
2101 struct si_shader
*iter
, *shader
= NULL
;
2104 /* Check if we don't need to change anything.
2105 * This path is also used for most shaders that don't need multiple
2106 * variants, it will cost just a computation of the key and this
2108 if (likely(current
&&
2109 memcmp(¤t
->key
, key
, sizeof(*key
)) == 0)) {
2110 if (unlikely(!util_queue_fence_is_signalled(¤t
->ready
))) {
2111 if (current
->is_optimized
) {
2112 if (optimized_or_none
)
2115 memset(&key
->opt
, 0, sizeof(key
->opt
));
2116 goto current_not_ready
;
2119 util_queue_fence_wait(¤t
->ready
);
2122 return current
->compilation_failed
? -1 : 0;
2126 /* This must be done before the mutex is locked, because async GS
2127 * compilation calls this function too, and therefore must enter
2130 * Only wait if we are in a draw call. Don't wait if we are
2131 * in a compiler thread.
2133 if (thread_index
< 0)
2134 util_queue_fence_wait(&sel
->ready
);
2136 mtx_lock(&sel
->mutex
);
2138 /* Find the shader variant. */
2139 for (iter
= sel
->first_variant
; iter
; iter
= iter
->next_variant
) {
2140 /* Don't check the "current" shader. We checked it above. */
2141 if (current
!= iter
&&
2142 memcmp(&iter
->key
, key
, sizeof(*key
)) == 0) {
2143 mtx_unlock(&sel
->mutex
);
2145 if (unlikely(!util_queue_fence_is_signalled(&iter
->ready
))) {
2146 /* If it's an optimized shader and its compilation has
2147 * been started but isn't done, use the unoptimized
2148 * shader so as not to cause a stall due to compilation.
2150 if (iter
->is_optimized
) {
2151 if (optimized_or_none
)
2153 memset(&key
->opt
, 0, sizeof(key
->opt
));
2157 util_queue_fence_wait(&iter
->ready
);
2160 if (iter
->compilation_failed
) {
2161 return -1; /* skip the draw call */
2164 state
->current
= iter
;
2169 /* Build a new shader. */
2170 shader
= CALLOC_STRUCT(si_shader
);
2172 mtx_unlock(&sel
->mutex
);
2176 util_queue_fence_init(&shader
->ready
);
2178 shader
->selector
= sel
;
2180 shader
->compiler_ctx_state
= *compiler_state
;
2182 /* If this is a merged shader, get the first shader's selector. */
2183 if (sscreen
->info
.chip_class
>= GFX9
) {
2184 if (sel
->type
== PIPE_SHADER_TESS_CTRL
)
2185 previous_stage_sel
= key
->part
.tcs
.ls
;
2186 else if (sel
->type
== PIPE_SHADER_GEOMETRY
)
2187 previous_stage_sel
= key
->part
.gs
.es
;
2189 /* We need to wait for the previous shader. */
2190 if (previous_stage_sel
&& thread_index
< 0)
2191 util_queue_fence_wait(&previous_stage_sel
->ready
);
2194 bool is_pure_monolithic
=
2195 sscreen
->use_monolithic_shaders
||
2196 memcmp(&key
->mono
, &zeroed
.mono
, sizeof(key
->mono
)) != 0;
2198 /* Compile the main shader part if it doesn't exist. This can happen
2199 * if the initial guess was wrong.
2201 * The prim discard CS doesn't need the main shader part.
2203 if (!is_pure_monolithic
&&
2204 !key
->opt
.vs_as_prim_discard_cs
) {
2207 /* Make sure the main shader part is present. This is needed
2208 * for shaders that can be compiled as VS, LS, or ES, and only
2209 * one of them is compiled at creation.
2211 * It is also needed for GS, which can be compiled as non-NGG
2214 * For merged shaders, check that the starting shader's main
2217 if (previous_stage_sel
) {
2218 struct si_shader_key shader1_key
= zeroed
;
2220 if (sel
->type
== PIPE_SHADER_TESS_CTRL
)
2221 shader1_key
.as_ls
= 1;
2222 else if (sel
->type
== PIPE_SHADER_GEOMETRY
)
2223 shader1_key
.as_es
= 1;
2227 mtx_lock(&previous_stage_sel
->mutex
);
2228 ok
= si_check_missing_main_part(sscreen
,
2230 compiler_state
, &shader1_key
);
2231 mtx_unlock(&previous_stage_sel
->mutex
);
2235 ok
= si_check_missing_main_part(sscreen
, sel
,
2236 compiler_state
, key
);
2241 mtx_unlock(&sel
->mutex
);
2242 return -ENOMEM
; /* skip the draw call */
2246 /* Keep the reference to the 1st shader of merged shaders, so that
2247 * Gallium can't destroy it before we destroy the 2nd shader.
2249 * Set sctx = NULL, because it's unused if we're not releasing
2250 * the shader, and we don't have any sctx here.
2252 si_shader_selector_reference(NULL
, &shader
->previous_stage_sel
,
2253 previous_stage_sel
);
2255 /* Monolithic-only shaders don't make a distinction between optimized
2256 * and unoptimized. */
2257 shader
->is_monolithic
=
2258 is_pure_monolithic
||
2259 memcmp(&key
->opt
, &zeroed
.opt
, sizeof(key
->opt
)) != 0;
2261 /* The prim discard CS is always optimized. */
2262 shader
->is_optimized
=
2263 (!is_pure_monolithic
|| key
->opt
.vs_as_prim_discard_cs
) &&
2264 memcmp(&key
->opt
, &zeroed
.opt
, sizeof(key
->opt
)) != 0;
2266 /* If it's an optimized shader, compile it asynchronously. */
2267 if (shader
->is_optimized
&& thread_index
< 0) {
2268 /* Compile it asynchronously. */
2269 util_queue_add_job(&sscreen
->shader_compiler_queue_low_priority
,
2270 shader
, &shader
->ready
,
2271 si_build_shader_variant_low_priority
, NULL
);
2273 /* Add only after the ready fence was reset, to guard against a
2274 * race with si_bind_XX_shader. */
2275 if (!sel
->last_variant
) {
2276 sel
->first_variant
= shader
;
2277 sel
->last_variant
= shader
;
2279 sel
->last_variant
->next_variant
= shader
;
2280 sel
->last_variant
= shader
;
2283 /* Use the default (unoptimized) shader for now. */
2284 memset(&key
->opt
, 0, sizeof(key
->opt
));
2285 mtx_unlock(&sel
->mutex
);
2287 if (sscreen
->options
.sync_compile
)
2288 util_queue_fence_wait(&shader
->ready
);
2290 if (optimized_or_none
)
2295 /* Reset the fence before adding to the variant list. */
2296 util_queue_fence_reset(&shader
->ready
);
2298 if (!sel
->last_variant
) {
2299 sel
->first_variant
= shader
;
2300 sel
->last_variant
= shader
;
2302 sel
->last_variant
->next_variant
= shader
;
2303 sel
->last_variant
= shader
;
2306 mtx_unlock(&sel
->mutex
);
2308 assert(!shader
->is_optimized
);
2309 si_build_shader_variant(shader
, thread_index
, false);
2311 util_queue_fence_signal(&shader
->ready
);
2313 if (!shader
->compilation_failed
)
2314 state
->current
= shader
;
2316 return shader
->compilation_failed
? -1 : 0;
2319 static int si_shader_select(struct pipe_context
*ctx
,
2320 struct si_shader_ctx_state
*state
,
2321 union si_vgt_stages_key stages_key
,
2322 struct si_compiler_ctx_state
*compiler_state
)
2324 struct si_context
*sctx
= (struct si_context
*)ctx
;
2325 struct si_shader_key key
;
2327 si_shader_selector_key(ctx
, state
->cso
, stages_key
, &key
);
2328 return si_shader_select_with_key(sctx
->screen
, state
, compiler_state
,
2332 static void si_parse_next_shader_property(const struct tgsi_shader_info
*info
,
2334 struct si_shader_key
*key
)
2336 unsigned next_shader
= info
->properties
[TGSI_PROPERTY_NEXT_SHADER
];
2338 switch (info
->processor
) {
2339 case PIPE_SHADER_VERTEX
:
2340 switch (next_shader
) {
2341 case PIPE_SHADER_GEOMETRY
:
2344 case PIPE_SHADER_TESS_CTRL
:
2345 case PIPE_SHADER_TESS_EVAL
:
2349 /* If POSITION isn't written, it can only be a HW VS
2350 * if streamout is used. If streamout isn't used,
2351 * assume that it's a HW LS. (the next shader is TCS)
2352 * This heuristic is needed for separate shader objects.
2354 if (!info
->writes_position
&& !streamout
)
2359 case PIPE_SHADER_TESS_EVAL
:
2360 if (next_shader
== PIPE_SHADER_GEOMETRY
||
2361 !info
->writes_position
)
2368 * Compile the main shader part or the monolithic shader as part of
2369 * si_shader_selector initialization. Since it can be done asynchronously,
2370 * there is no way to report compile failures to applications.
2372 static void si_init_shader_selector_async(void *job
, int thread_index
)
2374 struct si_shader_selector
*sel
= (struct si_shader_selector
*)job
;
2375 struct si_screen
*sscreen
= sel
->screen
;
2376 struct ac_llvm_compiler
*compiler
;
2377 struct pipe_debug_callback
*debug
= &sel
->compiler_ctx_state
.debug
;
2379 assert(!debug
->debug_message
|| debug
->async
);
2380 assert(thread_index
>= 0);
2381 assert(thread_index
< ARRAY_SIZE(sscreen
->compiler
));
2382 compiler
= &sscreen
->compiler
[thread_index
];
2387 /* Compile the main shader part for use with a prolog and/or epilog.
2388 * If this fails, the driver will try to compile a monolithic shader
2391 if (!sscreen
->use_monolithic_shaders
) {
2392 struct si_shader
*shader
= CALLOC_STRUCT(si_shader
);
2393 void *ir_binary
= NULL
;
2396 fprintf(stderr
, "radeonsi: can't allocate a main shader part\n");
2400 /* We can leave the fence signaled because use of the default
2401 * main part is guarded by the selector's ready fence. */
2402 util_queue_fence_init(&shader
->ready
);
2404 shader
->selector
= sel
;
2405 shader
->is_monolithic
= false;
2406 si_parse_next_shader_property(&sel
->info
,
2407 sel
->so
.num_outputs
!= 0,
2409 if (sscreen
->info
.chip_class
>= GFX10
&&
2410 !sscreen
->options
.disable_ngg
&&
2411 (((sel
->type
== PIPE_SHADER_VERTEX
||
2412 sel
->type
== PIPE_SHADER_TESS_EVAL
) &&
2413 !shader
->key
.as_ls
&& !shader
->key
.as_es
) ||
2414 sel
->type
== PIPE_SHADER_GEOMETRY
))
2415 shader
->key
.as_ngg
= 1;
2417 if (sel
->tokens
|| sel
->nir
)
2418 ir_binary
= si_get_ir_binary(sel
);
2420 /* Try to load the shader from the shader cache. */
2421 mtx_lock(&sscreen
->shader_cache_mutex
);
2424 si_shader_cache_load_shader(sscreen
, ir_binary
, shader
)) {
2425 mtx_unlock(&sscreen
->shader_cache_mutex
);
2426 si_shader_dump_stats_for_shader_db(sscreen
, shader
, debug
);
2428 mtx_unlock(&sscreen
->shader_cache_mutex
);
2430 /* Compile the shader if it hasn't been loaded from the cache. */
2431 if (si_compile_tgsi_shader(sscreen
, compiler
, shader
,
2435 fprintf(stderr
, "radeonsi: can't compile a main shader part\n");
2440 mtx_lock(&sscreen
->shader_cache_mutex
);
2441 if (!si_shader_cache_insert_shader(sscreen
, ir_binary
, shader
, true))
2443 mtx_unlock(&sscreen
->shader_cache_mutex
);
2447 *si_get_main_shader_part(sel
, &shader
->key
) = shader
;
2449 /* Unset "outputs_written" flags for outputs converted to
2450 * DEFAULT_VAL, so that later inter-shader optimizations don't
2451 * try to eliminate outputs that don't exist in the final
2454 * This is only done if non-monolithic shaders are enabled.
2456 if ((sel
->type
== PIPE_SHADER_VERTEX
||
2457 sel
->type
== PIPE_SHADER_TESS_EVAL
) &&
2458 !shader
->key
.as_ls
&&
2459 !shader
->key
.as_es
) {
2462 for (i
= 0; i
< sel
->info
.num_outputs
; i
++) {
2463 unsigned offset
= shader
->info
.vs_output_param_offset
[i
];
2465 if (offset
<= AC_EXP_PARAM_OFFSET_31
)
2468 unsigned name
= sel
->info
.output_semantic_name
[i
];
2469 unsigned index
= sel
->info
.output_semantic_index
[i
];
2473 case TGSI_SEMANTIC_GENERIC
:
2474 /* don't process indices the function can't handle */
2475 if (index
>= SI_MAX_IO_GENERIC
)
2479 id
= si_shader_io_get_unique_index(name
, index
, true);
2480 sel
->outputs_written_before_ps
&= ~(1ull << id
);
2482 case TGSI_SEMANTIC_POSITION
: /* ignore these */
2483 case TGSI_SEMANTIC_PSIZE
:
2484 case TGSI_SEMANTIC_CLIPVERTEX
:
2485 case TGSI_SEMANTIC_EDGEFLAG
:
2492 /* The GS copy shader is always pre-compiled.
2494 * TODO-GFX10: We could compile the GS copy shader on demand, since it
2495 * is only used in the (rare) non-NGG case.
2497 if (sel
->type
== PIPE_SHADER_GEOMETRY
) {
2498 sel
->gs_copy_shader
= si_generate_gs_copy_shader(sscreen
, compiler
, sel
, debug
);
2499 if (!sel
->gs_copy_shader
) {
2500 fprintf(stderr
, "radeonsi: can't create GS copy shader\n");
2504 si_shader_vs(sscreen
, sel
->gs_copy_shader
, sel
);
2508 void si_schedule_initial_compile(struct si_context
*sctx
, unsigned processor
,
2509 struct util_queue_fence
*ready_fence
,
2510 struct si_compiler_ctx_state
*compiler_ctx_state
,
2511 void *job
, util_queue_execute_func execute
)
2513 util_queue_fence_init(ready_fence
);
2515 struct util_async_debug_callback async_debug
;
2517 (sctx
->debug
.debug_message
&& !sctx
->debug
.async
) ||
2519 si_can_dump_shader(sctx
->screen
, processor
);
2522 u_async_debug_init(&async_debug
);
2523 compiler_ctx_state
->debug
= async_debug
.base
;
2526 util_queue_add_job(&sctx
->screen
->shader_compiler_queue
, job
,
2527 ready_fence
, execute
, NULL
);
2530 util_queue_fence_wait(ready_fence
);
2531 u_async_debug_drain(&async_debug
, &sctx
->debug
);
2532 u_async_debug_cleanup(&async_debug
);
2535 if (sctx
->screen
->options
.sync_compile
)
2536 util_queue_fence_wait(ready_fence
);
2539 /* Return descriptor slot usage masks from the given shader info. */
2540 void si_get_active_slot_masks(const struct tgsi_shader_info
*info
,
2541 uint32_t *const_and_shader_buffers
,
2542 uint64_t *samplers_and_images
)
2544 unsigned start
, num_shaderbufs
, num_constbufs
, num_images
, num_samplers
;
2546 num_shaderbufs
= util_last_bit(info
->shader_buffers_declared
);
2547 num_constbufs
= util_last_bit(info
->const_buffers_declared
);
2548 /* two 8-byte images share one 16-byte slot */
2549 num_images
= align(util_last_bit(info
->images_declared
), 2);
2550 num_samplers
= util_last_bit(info
->samplers_declared
);
2552 /* The layout is: sb[last] ... sb[0], cb[0] ... cb[last] */
2553 start
= si_get_shaderbuf_slot(num_shaderbufs
- 1);
2554 *const_and_shader_buffers
=
2555 u_bit_consecutive(start
, num_shaderbufs
+ num_constbufs
);
2557 /* The layout is: image[last] ... image[0], sampler[0] ... sampler[last] */
2558 start
= si_get_image_slot(num_images
- 1) / 2;
2559 *samplers_and_images
=
2560 u_bit_consecutive64(start
, num_images
/ 2 + num_samplers
);
2563 static void *si_create_shader_selector(struct pipe_context
*ctx
,
2564 const struct pipe_shader_state
*state
)
2566 struct si_screen
*sscreen
= (struct si_screen
*)ctx
->screen
;
2567 struct si_context
*sctx
= (struct si_context
*)ctx
;
2568 struct si_shader_selector
*sel
= CALLOC_STRUCT(si_shader_selector
);
2574 pipe_reference_init(&sel
->reference
, 1);
2575 sel
->screen
= sscreen
;
2576 sel
->compiler_ctx_state
.debug
= sctx
->debug
;
2577 sel
->compiler_ctx_state
.is_debug_context
= sctx
->is_debug
;
2579 sel
->so
= state
->stream_output
;
2581 if (state
->type
== PIPE_SHADER_IR_TGSI
) {
2582 sel
->tokens
= tgsi_dup_tokens(state
->tokens
);
2588 tgsi_scan_shader(state
->tokens
, &sel
->info
);
2589 tgsi_scan_tess_ctrl(state
->tokens
, &sel
->info
, &sel
->tcs_info
);
2591 assert(state
->type
== PIPE_SHADER_IR_NIR
);
2593 sel
->nir
= state
->ir
.nir
;
2595 si_nir_opts(sel
->nir
);
2596 si_nir_scan_shader(sel
->nir
, &sel
->info
);
2597 si_nir_scan_tess_ctrl(sel
->nir
, &sel
->tcs_info
);
2600 sel
->type
= sel
->info
.processor
;
2601 p_atomic_inc(&sscreen
->num_shaders_created
);
2602 si_get_active_slot_masks(&sel
->info
,
2603 &sel
->active_const_and_shader_buffers
,
2604 &sel
->active_samplers_and_images
);
2606 /* Record which streamout buffers are enabled. */
2607 for (i
= 0; i
< sel
->so
.num_outputs
; i
++) {
2608 sel
->enabled_streamout_buffer_mask
|=
2609 (1 << sel
->so
.output
[i
].output_buffer
) <<
2610 (sel
->so
.output
[i
].stream
* 4);
2613 /* The prolog is a no-op if there are no inputs. */
2614 sel
->vs_needs_prolog
= sel
->type
== PIPE_SHADER_VERTEX
&&
2615 sel
->info
.num_inputs
&&
2616 !sel
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
];
2618 sel
->force_correct_derivs_after_kill
=
2619 sel
->type
== PIPE_SHADER_FRAGMENT
&&
2620 sel
->info
.uses_derivatives
&&
2621 sel
->info
.uses_kill
&&
2622 sctx
->screen
->debug_flags
& DBG(FS_CORRECT_DERIVS_AFTER_KILL
);
2624 sel
->prim_discard_cs_allowed
=
2625 sel
->type
== PIPE_SHADER_VERTEX
&&
2626 !sel
->info
.uses_bindless_images
&&
2627 !sel
->info
.uses_bindless_samplers
&&
2628 !sel
->info
.writes_memory
&&
2629 !sel
->info
.writes_viewport_index
&&
2630 !sel
->info
.properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
] &&
2631 !sel
->so
.num_outputs
;
2633 /* Set which opcode uses which (i,j) pair. */
2634 if (sel
->info
.uses_persp_opcode_interp_centroid
)
2635 sel
->info
.uses_persp_centroid
= true;
2637 if (sel
->info
.uses_linear_opcode_interp_centroid
)
2638 sel
->info
.uses_linear_centroid
= true;
2640 if (sel
->info
.uses_persp_opcode_interp_offset
||
2641 sel
->info
.uses_persp_opcode_interp_sample
)
2642 sel
->info
.uses_persp_center
= true;
2644 if (sel
->info
.uses_linear_opcode_interp_offset
||
2645 sel
->info
.uses_linear_opcode_interp_sample
)
2646 sel
->info
.uses_linear_center
= true;
2648 switch (sel
->type
) {
2649 case PIPE_SHADER_GEOMETRY
:
2650 sel
->gs_output_prim
=
2651 sel
->info
.properties
[TGSI_PROPERTY_GS_OUTPUT_PRIM
];
2652 sel
->gs_max_out_vertices
=
2653 sel
->info
.properties
[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES
];
2654 sel
->gs_num_invocations
=
2655 sel
->info
.properties
[TGSI_PROPERTY_GS_INVOCATIONS
];
2656 sel
->gsvs_vertex_size
= sel
->info
.num_outputs
* 16;
2657 sel
->max_gsvs_emit_size
= sel
->gsvs_vertex_size
*
2658 sel
->gs_max_out_vertices
;
2660 sel
->max_gs_stream
= 0;
2661 for (i
= 0; i
< sel
->so
.num_outputs
; i
++)
2662 sel
->max_gs_stream
= MAX2(sel
->max_gs_stream
,
2663 sel
->so
.output
[i
].stream
);
2665 sel
->gs_input_verts_per_prim
=
2666 u_vertices_per_prim(sel
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
]);
2669 case PIPE_SHADER_TESS_CTRL
:
2670 /* Always reserve space for these. */
2671 sel
->patch_outputs_written
|=
2672 (1ull << si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSINNER
, 0)) |
2673 (1ull << si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSOUTER
, 0));
2675 case PIPE_SHADER_VERTEX
:
2676 case PIPE_SHADER_TESS_EVAL
:
2677 for (i
= 0; i
< sel
->info
.num_outputs
; i
++) {
2678 unsigned name
= sel
->info
.output_semantic_name
[i
];
2679 unsigned index
= sel
->info
.output_semantic_index
[i
];
2682 case TGSI_SEMANTIC_TESSINNER
:
2683 case TGSI_SEMANTIC_TESSOUTER
:
2684 case TGSI_SEMANTIC_PATCH
:
2685 sel
->patch_outputs_written
|=
2686 1ull << si_shader_io_get_unique_index_patch(name
, index
);
2689 case TGSI_SEMANTIC_GENERIC
:
2690 /* don't process indices the function can't handle */
2691 if (index
>= SI_MAX_IO_GENERIC
)
2695 sel
->outputs_written
|=
2696 1ull << si_shader_io_get_unique_index(name
, index
, false);
2697 sel
->outputs_written_before_ps
|=
2698 1ull << si_shader_io_get_unique_index(name
, index
, true);
2700 case TGSI_SEMANTIC_EDGEFLAG
:
2704 sel
->esgs_itemsize
= util_last_bit64(sel
->outputs_written
) * 16;
2705 sel
->lshs_vertex_stride
= sel
->esgs_itemsize
;
2707 /* Add 1 dword to reduce LDS bank conflicts, so that each vertex
2708 * will start on a different bank. (except for the maximum 32*16).
2710 if (sel
->lshs_vertex_stride
< 32*16)
2711 sel
->lshs_vertex_stride
+= 4;
2713 /* For the ESGS ring in LDS, add 1 dword to reduce LDS bank
2714 * conflicts, i.e. each vertex will start at a different bank.
2716 if (sctx
->chip_class
>= GFX9
)
2717 sel
->esgs_itemsize
+= 4;
2719 assert(((sel
->esgs_itemsize
/ 4) & C_028AAC_ITEMSIZE
) == 0);
2722 case PIPE_SHADER_FRAGMENT
:
2723 for (i
= 0; i
< sel
->info
.num_inputs
; i
++) {
2724 unsigned name
= sel
->info
.input_semantic_name
[i
];
2725 unsigned index
= sel
->info
.input_semantic_index
[i
];
2728 case TGSI_SEMANTIC_GENERIC
:
2729 /* don't process indices the function can't handle */
2730 if (index
>= SI_MAX_IO_GENERIC
)
2735 1ull << si_shader_io_get_unique_index(name
, index
, true);
2737 case TGSI_SEMANTIC_PCOORD
: /* ignore this */
2742 for (i
= 0; i
< 8; i
++)
2743 if (sel
->info
.colors_written
& (1 << i
))
2744 sel
->colors_written_4bit
|= 0xf << (4 * i
);
2746 for (i
= 0; i
< sel
->info
.num_inputs
; i
++) {
2747 if (sel
->info
.input_semantic_name
[i
] == TGSI_SEMANTIC_COLOR
) {
2748 int index
= sel
->info
.input_semantic_index
[i
];
2749 sel
->color_attr_index
[index
] = i
;
2755 /* PA_CL_VS_OUT_CNTL */
2757 sel
->info
.writes_psize
|| sel
->info
.writes_edgeflag
||
2758 sel
->info
.writes_layer
|| sel
->info
.writes_viewport_index
;
2759 sel
->pa_cl_vs_out_cntl
=
2760 S_02881C_USE_VTX_POINT_SIZE(sel
->info
.writes_psize
) |
2761 S_02881C_USE_VTX_EDGE_FLAG(sel
->info
.writes_edgeflag
) |
2762 S_02881C_USE_VTX_RENDER_TARGET_INDX(sel
->info
.writes_layer
) |
2763 S_02881C_USE_VTX_VIEWPORT_INDX(sel
->info
.writes_viewport_index
) |
2764 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena
) |
2765 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena
);
2766 sel
->clipdist_mask
= sel
->info
.writes_clipvertex
?
2767 SIX_BITS
: sel
->info
.clipdist_writemask
;
2768 sel
->culldist_mask
= sel
->info
.culldist_writemask
<<
2769 sel
->info
.num_written_clipdistance
;
2771 /* DB_SHADER_CONTROL */
2772 sel
->db_shader_control
=
2773 S_02880C_Z_EXPORT_ENABLE(sel
->info
.writes_z
) |
2774 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(sel
->info
.writes_stencil
) |
2775 S_02880C_MASK_EXPORT_ENABLE(sel
->info
.writes_samplemask
) |
2776 S_02880C_KILL_ENABLE(sel
->info
.uses_kill
);
2778 switch (sel
->info
.properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
]) {
2779 case TGSI_FS_DEPTH_LAYOUT_GREATER
:
2780 sel
->db_shader_control
|=
2781 S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_GREATER_THAN_Z
);
2783 case TGSI_FS_DEPTH_LAYOUT_LESS
:
2784 sel
->db_shader_control
|=
2785 S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_LESS_THAN_Z
);
2789 /* Z_ORDER, EXEC_ON_HIER_FAIL and EXEC_ON_NOOP should be set as following:
2791 * | early Z/S | writes_mem | allow_ReZ? | Z_ORDER | EXEC_ON_HIER_FAIL | EXEC_ON_NOOP
2792 * --|-----------|------------|------------|--------------------|-------------------|-------------
2793 * 1a| false | false | true | EarlyZ_Then_ReZ | 0 | 0
2794 * 1b| false | false | false | EarlyZ_Then_LateZ | 0 | 0
2795 * 2 | false | true | n/a | LateZ | 1 | 0
2796 * 3 | true | false | n/a | EarlyZ_Then_LateZ | 0 | 0
2797 * 4 | true | true | n/a | EarlyZ_Then_LateZ | 0 | 1
2799 * In cases 3 and 4, HW will force Z_ORDER to EarlyZ regardless of what's set in the register.
2800 * In case 2, NOOP_CULL is a don't care field. In case 2, 3 and 4, ReZ doesn't make sense.
2802 * Don't use ReZ without profiling !!!
2804 * ReZ decreases performance by 15% in DiRT: Showdown on Ultra settings, which has pretty complex
2807 if (sel
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
]) {
2809 sel
->db_shader_control
|= S_02880C_DEPTH_BEFORE_SHADER(1) |
2810 S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z
) |
2811 S_02880C_EXEC_ON_NOOP(sel
->info
.writes_memory
);
2812 } else if (sel
->info
.writes_memory
) {
2814 sel
->db_shader_control
|= S_02880C_Z_ORDER(V_02880C_LATE_Z
) |
2815 S_02880C_EXEC_ON_HIER_FAIL(1);
2818 sel
->db_shader_control
|= S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z
);
2821 (void) mtx_init(&sel
->mutex
, mtx_plain
);
2823 si_schedule_initial_compile(sctx
, sel
->info
.processor
, &sel
->ready
,
2824 &sel
->compiler_ctx_state
, sel
,
2825 si_init_shader_selector_async
);
2829 static void si_update_streamout_state(struct si_context
*sctx
)
2831 struct si_shader_selector
*shader_with_so
= si_get_vs(sctx
)->cso
;
2833 if (!shader_with_so
)
2836 sctx
->streamout
.enabled_stream_buffers_mask
=
2837 shader_with_so
->enabled_streamout_buffer_mask
;
2838 sctx
->streamout
.stride_in_dw
= shader_with_so
->so
.stride
;
2841 static void si_update_clip_regs(struct si_context
*sctx
,
2842 struct si_shader_selector
*old_hw_vs
,
2843 struct si_shader
*old_hw_vs_variant
,
2844 struct si_shader_selector
*next_hw_vs
,
2845 struct si_shader
*next_hw_vs_variant
)
2849 old_hw_vs
->info
.properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
] !=
2850 next_hw_vs
->info
.properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
] ||
2851 old_hw_vs
->pa_cl_vs_out_cntl
!= next_hw_vs
->pa_cl_vs_out_cntl
||
2852 old_hw_vs
->clipdist_mask
!= next_hw_vs
->clipdist_mask
||
2853 old_hw_vs
->culldist_mask
!= next_hw_vs
->culldist_mask
||
2854 !old_hw_vs_variant
||
2855 !next_hw_vs_variant
||
2856 old_hw_vs_variant
->key
.opt
.clip_disable
!=
2857 next_hw_vs_variant
->key
.opt
.clip_disable
))
2858 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.clip_regs
);
2861 static void si_update_common_shader_state(struct si_context
*sctx
)
2863 sctx
->uses_bindless_samplers
=
2864 si_shader_uses_bindless_samplers(sctx
->vs_shader
.cso
) ||
2865 si_shader_uses_bindless_samplers(sctx
->gs_shader
.cso
) ||
2866 si_shader_uses_bindless_samplers(sctx
->ps_shader
.cso
) ||
2867 si_shader_uses_bindless_samplers(sctx
->tcs_shader
.cso
) ||
2868 si_shader_uses_bindless_samplers(sctx
->tes_shader
.cso
);
2869 sctx
->uses_bindless_images
=
2870 si_shader_uses_bindless_images(sctx
->vs_shader
.cso
) ||
2871 si_shader_uses_bindless_images(sctx
->gs_shader
.cso
) ||
2872 si_shader_uses_bindless_images(sctx
->ps_shader
.cso
) ||
2873 si_shader_uses_bindless_images(sctx
->tcs_shader
.cso
) ||
2874 si_shader_uses_bindless_images(sctx
->tes_shader
.cso
);
2875 sctx
->do_update_shaders
= true;
2878 static void si_bind_vs_shader(struct pipe_context
*ctx
, void *state
)
2880 struct si_context
*sctx
= (struct si_context
*)ctx
;
2881 struct si_shader_selector
*old_hw_vs
= si_get_vs(sctx
)->cso
;
2882 struct si_shader
*old_hw_vs_variant
= si_get_vs_state(sctx
);
2883 struct si_shader_selector
*sel
= state
;
2885 if (sctx
->vs_shader
.cso
== sel
)
2888 sctx
->vs_shader
.cso
= sel
;
2889 sctx
->vs_shader
.current
= sel
? sel
->first_variant
: NULL
;
2890 sctx
->num_vs_blit_sgprs
= sel
? sel
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS
] : 0;
2892 si_update_common_shader_state(sctx
);
2893 si_update_vs_viewport_state(sctx
);
2894 si_set_active_descriptors_for_shader(sctx
, sel
);
2895 si_update_streamout_state(sctx
);
2896 si_update_clip_regs(sctx
, old_hw_vs
, old_hw_vs_variant
,
2897 si_get_vs(sctx
)->cso
, si_get_vs_state(sctx
));
2900 static void si_update_tess_uses_prim_id(struct si_context
*sctx
)
2902 sctx
->ia_multi_vgt_param_key
.u
.tess_uses_prim_id
=
2903 (sctx
->tes_shader
.cso
&&
2904 sctx
->tes_shader
.cso
->info
.uses_primid
) ||
2905 (sctx
->tcs_shader
.cso
&&
2906 sctx
->tcs_shader
.cso
->info
.uses_primid
) ||
2907 (sctx
->gs_shader
.cso
&&
2908 sctx
->gs_shader
.cso
->info
.uses_primid
) ||
2909 (sctx
->ps_shader
.cso
&& !sctx
->gs_shader
.cso
&&
2910 sctx
->ps_shader
.cso
->info
.uses_primid
);
2913 static bool si_update_ngg(struct si_context
*sctx
)
2915 if (sctx
->chip_class
<= GFX9
||
2916 sctx
->screen
->options
.disable_ngg
)
2919 bool new_ngg
= true;
2921 /* EN_MAX_VERT_OUT_PER_GS_INSTANCE does not work with tesselation. */
2922 if (sctx
->gs_shader
.cso
&& sctx
->tes_shader
.cso
&&
2923 sctx
->gs_shader
.cso
->gs_num_invocations
* sctx
->gs_shader
.cso
->gs_max_out_vertices
> 256)
2926 if (new_ngg
!= sctx
->ngg
) {
2927 sctx
->ngg
= new_ngg
;
2928 sctx
->last_rast_prim
= -1; /* reset this so that it gets updated */
2934 static void si_bind_gs_shader(struct pipe_context
*ctx
, void *state
)
2936 struct si_context
*sctx
= (struct si_context
*)ctx
;
2937 struct si_shader_selector
*old_hw_vs
= si_get_vs(sctx
)->cso
;
2938 struct si_shader
*old_hw_vs_variant
= si_get_vs_state(sctx
);
2939 struct si_shader_selector
*sel
= state
;
2940 bool enable_changed
= !!sctx
->gs_shader
.cso
!= !!sel
;
2943 if (sctx
->gs_shader
.cso
== sel
)
2946 sctx
->gs_shader
.cso
= sel
;
2947 sctx
->gs_shader
.current
= sel
? sel
->first_variant
: NULL
;
2948 sctx
->ia_multi_vgt_param_key
.u
.uses_gs
= sel
!= NULL
;
2950 si_update_common_shader_state(sctx
);
2951 sctx
->last_rast_prim
= -1; /* reset this so that it gets updated */
2953 ngg_changed
= si_update_ngg(sctx
);
2954 if (ngg_changed
|| enable_changed
)
2955 si_shader_change_notify(sctx
);
2956 if (enable_changed
) {
2957 if (sctx
->ia_multi_vgt_param_key
.u
.uses_tess
)
2958 si_update_tess_uses_prim_id(sctx
);
2960 si_update_vs_viewport_state(sctx
);
2961 si_set_active_descriptors_for_shader(sctx
, sel
);
2962 si_update_streamout_state(sctx
);
2963 si_update_clip_regs(sctx
, old_hw_vs
, old_hw_vs_variant
,
2964 si_get_vs(sctx
)->cso
, si_get_vs_state(sctx
));
2967 static void si_bind_tcs_shader(struct pipe_context
*ctx
, void *state
)
2969 struct si_context
*sctx
= (struct si_context
*)ctx
;
2970 struct si_shader_selector
*sel
= state
;
2971 bool enable_changed
= !!sctx
->tcs_shader
.cso
!= !!sel
;
2973 if (sctx
->tcs_shader
.cso
== sel
)
2976 sctx
->tcs_shader
.cso
= sel
;
2977 sctx
->tcs_shader
.current
= sel
? sel
->first_variant
: NULL
;
2978 si_update_tess_uses_prim_id(sctx
);
2980 si_update_common_shader_state(sctx
);
2983 sctx
->last_tcs
= NULL
; /* invalidate derived tess state */
2985 si_set_active_descriptors_for_shader(sctx
, sel
);
2988 static void si_bind_tes_shader(struct pipe_context
*ctx
, void *state
)
2990 struct si_context
*sctx
= (struct si_context
*)ctx
;
2991 struct si_shader_selector
*old_hw_vs
= si_get_vs(sctx
)->cso
;
2992 struct si_shader
*old_hw_vs_variant
= si_get_vs_state(sctx
);
2993 struct si_shader_selector
*sel
= state
;
2994 bool enable_changed
= !!sctx
->tes_shader
.cso
!= !!sel
;
2996 if (sctx
->tes_shader
.cso
== sel
)
2999 sctx
->tes_shader
.cso
= sel
;
3000 sctx
->tes_shader
.current
= sel
? sel
->first_variant
: NULL
;
3001 sctx
->ia_multi_vgt_param_key
.u
.uses_tess
= sel
!= NULL
;
3002 si_update_tess_uses_prim_id(sctx
);
3004 si_update_common_shader_state(sctx
);
3005 sctx
->last_rast_prim
= -1; /* reset this so that it gets updated */
3007 if (enable_changed
) {
3008 si_update_ngg(sctx
);
3009 si_shader_change_notify(sctx
);
3010 sctx
->last_tes_sh_base
= -1; /* invalidate derived tess state */
3012 si_update_vs_viewport_state(sctx
);
3013 si_set_active_descriptors_for_shader(sctx
, sel
);
3014 si_update_streamout_state(sctx
);
3015 si_update_clip_regs(sctx
, old_hw_vs
, old_hw_vs_variant
,
3016 si_get_vs(sctx
)->cso
, si_get_vs_state(sctx
));
3019 static void si_bind_ps_shader(struct pipe_context
*ctx
, void *state
)
3021 struct si_context
*sctx
= (struct si_context
*)ctx
;
3022 struct si_shader_selector
*old_sel
= sctx
->ps_shader
.cso
;
3023 struct si_shader_selector
*sel
= state
;
3025 /* skip if supplied shader is one already in use */
3029 sctx
->ps_shader
.cso
= sel
;
3030 sctx
->ps_shader
.current
= sel
? sel
->first_variant
: NULL
;
3032 si_update_common_shader_state(sctx
);
3034 if (sctx
->ia_multi_vgt_param_key
.u
.uses_tess
)
3035 si_update_tess_uses_prim_id(sctx
);
3038 old_sel
->info
.colors_written
!= sel
->info
.colors_written
)
3039 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.cb_render_state
);
3041 if (sctx
->screen
->has_out_of_order_rast
&&
3043 old_sel
->info
.writes_memory
!= sel
->info
.writes_memory
||
3044 old_sel
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
] !=
3045 sel
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
]))
3046 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.msaa_config
);
3048 si_set_active_descriptors_for_shader(sctx
, sel
);
3049 si_update_ps_colorbuf0_slot(sctx
);
3052 static void si_delete_shader(struct si_context
*sctx
, struct si_shader
*shader
)
3054 if (shader
->is_optimized
) {
3055 util_queue_drop_job(&sctx
->screen
->shader_compiler_queue_low_priority
,
3059 util_queue_fence_destroy(&shader
->ready
);
3062 /* If destroyed shaders were not unbound, the next compiled
3063 * shader variant could get the same pointer address and so
3064 * binding it to the same shader stage would be considered
3065 * a no-op, causing random behavior.
3067 switch (shader
->selector
->type
) {
3068 case PIPE_SHADER_VERTEX
:
3069 if (shader
->key
.as_ls
) {
3070 assert(sctx
->chip_class
<= GFX8
);
3071 si_pm4_delete_state(sctx
, ls
, shader
->pm4
);
3072 } else if (shader
->key
.as_es
) {
3073 assert(sctx
->chip_class
<= GFX8
);
3074 si_pm4_delete_state(sctx
, es
, shader
->pm4
);
3075 } else if (shader
->key
.as_ngg
) {
3076 si_pm4_delete_state(sctx
, gs
, shader
->pm4
);
3078 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
3081 case PIPE_SHADER_TESS_CTRL
:
3082 si_pm4_delete_state(sctx
, hs
, shader
->pm4
);
3084 case PIPE_SHADER_TESS_EVAL
:
3085 if (shader
->key
.as_es
) {
3086 assert(sctx
->chip_class
<= GFX8
);
3087 si_pm4_delete_state(sctx
, es
, shader
->pm4
);
3088 } else if (shader
->key
.as_ngg
) {
3089 si_pm4_delete_state(sctx
, gs
, shader
->pm4
);
3091 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
3094 case PIPE_SHADER_GEOMETRY
:
3095 if (shader
->is_gs_copy_shader
)
3096 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
3098 si_pm4_delete_state(sctx
, gs
, shader
->pm4
);
3100 case PIPE_SHADER_FRAGMENT
:
3101 si_pm4_delete_state(sctx
, ps
, shader
->pm4
);
3106 si_shader_selector_reference(sctx
, &shader
->previous_stage_sel
, NULL
);
3107 si_shader_destroy(shader
);
3111 void si_destroy_shader_selector(struct si_context
*sctx
,
3112 struct si_shader_selector
*sel
)
3114 struct si_shader
*p
= sel
->first_variant
, *c
;
3115 struct si_shader_ctx_state
*current_shader
[SI_NUM_SHADERS
] = {
3116 [PIPE_SHADER_VERTEX
] = &sctx
->vs_shader
,
3117 [PIPE_SHADER_TESS_CTRL
] = &sctx
->tcs_shader
,
3118 [PIPE_SHADER_TESS_EVAL
] = &sctx
->tes_shader
,
3119 [PIPE_SHADER_GEOMETRY
] = &sctx
->gs_shader
,
3120 [PIPE_SHADER_FRAGMENT
] = &sctx
->ps_shader
,
3123 util_queue_drop_job(&sctx
->screen
->shader_compiler_queue
, &sel
->ready
);
3125 if (current_shader
[sel
->type
]->cso
== sel
) {
3126 current_shader
[sel
->type
]->cso
= NULL
;
3127 current_shader
[sel
->type
]->current
= NULL
;
3131 c
= p
->next_variant
;
3132 si_delete_shader(sctx
, p
);
3136 if (sel
->main_shader_part
)
3137 si_delete_shader(sctx
, sel
->main_shader_part
);
3138 if (sel
->main_shader_part_ls
)
3139 si_delete_shader(sctx
, sel
->main_shader_part_ls
);
3140 if (sel
->main_shader_part_es
)
3141 si_delete_shader(sctx
, sel
->main_shader_part_es
);
3142 if (sel
->main_shader_part_ngg
)
3143 si_delete_shader(sctx
, sel
->main_shader_part_ngg
);
3144 if (sel
->gs_copy_shader
)
3145 si_delete_shader(sctx
, sel
->gs_copy_shader
);
3147 util_queue_fence_destroy(&sel
->ready
);
3148 mtx_destroy(&sel
->mutex
);
3150 ralloc_free(sel
->nir
);
3154 static void si_delete_shader_selector(struct pipe_context
*ctx
, void *state
)
3156 struct si_context
*sctx
= (struct si_context
*)ctx
;
3157 struct si_shader_selector
*sel
= (struct si_shader_selector
*)state
;
3159 si_shader_selector_reference(sctx
, &sel
, NULL
);
3162 static unsigned si_get_ps_input_cntl(struct si_context
*sctx
,
3163 struct si_shader
*vs
, unsigned name
,
3164 unsigned index
, unsigned interpolate
)
3166 struct tgsi_shader_info
*vsinfo
= &vs
->selector
->info
;
3167 unsigned j
, offset
, ps_input_cntl
= 0;
3169 if (interpolate
== TGSI_INTERPOLATE_CONSTANT
||
3170 (interpolate
== TGSI_INTERPOLATE_COLOR
&& sctx
->flatshade
) ||
3171 name
== TGSI_SEMANTIC_PRIMID
)
3172 ps_input_cntl
|= S_028644_FLAT_SHADE(1);
3174 if (name
== TGSI_SEMANTIC_PCOORD
||
3175 (name
== TGSI_SEMANTIC_TEXCOORD
&&
3176 sctx
->sprite_coord_enable
& (1 << index
))) {
3177 ps_input_cntl
|= S_028644_PT_SPRITE_TEX(1);
3180 for (j
= 0; j
< vsinfo
->num_outputs
; j
++) {
3181 if (name
== vsinfo
->output_semantic_name
[j
] &&
3182 index
== vsinfo
->output_semantic_index
[j
]) {
3183 offset
= vs
->info
.vs_output_param_offset
[j
];
3185 if (offset
<= AC_EXP_PARAM_OFFSET_31
) {
3186 /* The input is loaded from parameter memory. */
3187 ps_input_cntl
|= S_028644_OFFSET(offset
);
3188 } else if (!G_028644_PT_SPRITE_TEX(ps_input_cntl
)) {
3189 if (offset
== AC_EXP_PARAM_UNDEFINED
) {
3190 /* This can happen with depth-only rendering. */
3193 /* The input is a DEFAULT_VAL constant. */
3194 assert(offset
>= AC_EXP_PARAM_DEFAULT_VAL_0000
&&
3195 offset
<= AC_EXP_PARAM_DEFAULT_VAL_1111
);
3196 offset
-= AC_EXP_PARAM_DEFAULT_VAL_0000
;
3199 ps_input_cntl
= S_028644_OFFSET(0x20) |
3200 S_028644_DEFAULT_VAL(offset
);
3206 if (j
== vsinfo
->num_outputs
&& name
== TGSI_SEMANTIC_PRIMID
)
3207 /* PrimID is written after the last output when HW VS is used. */
3208 ps_input_cntl
|= S_028644_OFFSET(vs
->info
.vs_output_param_offset
[vsinfo
->num_outputs
]);
3209 else if (j
== vsinfo
->num_outputs
&& !G_028644_PT_SPRITE_TEX(ps_input_cntl
)) {
3210 /* No corresponding output found, load defaults into input.
3211 * Don't set any other bits.
3212 * (FLAT_SHADE=1 completely changes behavior) */
3213 ps_input_cntl
= S_028644_OFFSET(0x20);
3214 /* D3D 9 behaviour. GL is undefined */
3215 if (name
== TGSI_SEMANTIC_COLOR
&& index
== 0)
3216 ps_input_cntl
|= S_028644_DEFAULT_VAL(3);
3218 return ps_input_cntl
;
3221 static void si_emit_spi_map(struct si_context
*sctx
)
3223 struct si_shader
*ps
= sctx
->ps_shader
.current
;
3224 struct si_shader
*vs
= si_get_vs_state(sctx
);
3225 struct tgsi_shader_info
*psinfo
= ps
? &ps
->selector
->info
: NULL
;
3226 unsigned i
, num_interp
, num_written
= 0, bcol_interp
[2];
3227 unsigned spi_ps_input_cntl
[32];
3229 if (!ps
|| !ps
->selector
->info
.num_inputs
)
3232 num_interp
= si_get_ps_num_interp(ps
);
3233 assert(num_interp
> 0);
3235 for (i
= 0; i
< psinfo
->num_inputs
; i
++) {
3236 unsigned name
= psinfo
->input_semantic_name
[i
];
3237 unsigned index
= psinfo
->input_semantic_index
[i
];
3238 unsigned interpolate
= psinfo
->input_interpolate
[i
];
3240 spi_ps_input_cntl
[num_written
++] = si_get_ps_input_cntl(sctx
, vs
, name
,
3241 index
, interpolate
);
3243 if (name
== TGSI_SEMANTIC_COLOR
) {
3244 assert(index
< ARRAY_SIZE(bcol_interp
));
3245 bcol_interp
[index
] = interpolate
;
3249 if (ps
->key
.part
.ps
.prolog
.color_two_side
) {
3250 unsigned bcol
= TGSI_SEMANTIC_BCOLOR
;
3252 for (i
= 0; i
< 2; i
++) {
3253 if (!(psinfo
->colors_read
& (0xf << (i
* 4))))
3256 spi_ps_input_cntl
[num_written
++] =
3257 si_get_ps_input_cntl(sctx
, vs
, bcol
, i
, bcol_interp
[i
]);
3261 assert(num_interp
== num_written
);
3263 /* R_028644_SPI_PS_INPUT_CNTL_0 */
3264 /* Dota 2: Only ~16% of SPI map updates set different values. */
3265 /* Talos: Only ~9% of SPI map updates set different values. */
3266 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
3267 radeon_opt_set_context_regn(sctx
, R_028644_SPI_PS_INPUT_CNTL_0
,
3269 sctx
->tracked_regs
.spi_ps_input_cntl
, num_interp
);
3271 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
3272 sctx
->context_roll
= true;
3276 * Writing CONFIG or UCONFIG VGT registers requires VGT_FLUSH before that.
3278 static void si_init_config_add_vgt_flush(struct si_context
*sctx
)
3280 if (sctx
->init_config_has_vgt_flush
)
3283 /* Done by Vulkan before VGT_FLUSH. */
3284 si_pm4_cmd_begin(sctx
->init_config
, PKT3_EVENT_WRITE
);
3285 si_pm4_cmd_add(sctx
->init_config
,
3286 EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
3287 si_pm4_cmd_end(sctx
->init_config
, false);
3289 /* VGT_FLUSH is required even if VGT is idle. It resets VGT pointers. */
3290 si_pm4_cmd_begin(sctx
->init_config
, PKT3_EVENT_WRITE
);
3291 si_pm4_cmd_add(sctx
->init_config
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
3292 si_pm4_cmd_end(sctx
->init_config
, false);
3293 sctx
->init_config_has_vgt_flush
= true;
3296 /* Initialize state related to ESGS / GSVS ring buffers */
3297 static bool si_update_gs_ring_buffers(struct si_context
*sctx
)
3299 struct si_shader_selector
*es
=
3300 sctx
->tes_shader
.cso
? sctx
->tes_shader
.cso
: sctx
->vs_shader
.cso
;
3301 struct si_shader_selector
*gs
= sctx
->gs_shader
.cso
;
3302 struct si_pm4_state
*pm4
;
3304 /* Chip constants. */
3305 unsigned num_se
= sctx
->screen
->info
.max_se
;
3306 unsigned wave_size
= 64;
3307 unsigned max_gs_waves
= 32 * num_se
; /* max 32 per SE on GCN */
3308 /* On GFX6-GFX7, the value comes from VGT_GS_VERTEX_REUSE = 16.
3309 * On GFX8+, the value comes from VGT_VERTEX_REUSE_BLOCK_CNTL = 30 (+2).
3311 unsigned gs_vertex_reuse
= (sctx
->chip_class
>= GFX8
? 32 : 16) * num_se
;
3312 unsigned alignment
= 256 * num_se
;
3313 /* The maximum size is 63.999 MB per SE. */
3314 unsigned max_size
= ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se
;
3316 /* Calculate the minimum size. */
3317 unsigned min_esgs_ring_size
= align(es
->esgs_itemsize
* gs_vertex_reuse
*
3318 wave_size
, alignment
);
3320 /* These are recommended sizes, not minimum sizes. */
3321 unsigned esgs_ring_size
= max_gs_waves
* 2 * wave_size
*
3322 es
->esgs_itemsize
* gs
->gs_input_verts_per_prim
;
3323 unsigned gsvs_ring_size
= max_gs_waves
* 2 * wave_size
*
3324 gs
->max_gsvs_emit_size
;
3326 min_esgs_ring_size
= align(min_esgs_ring_size
, alignment
);
3327 esgs_ring_size
= align(esgs_ring_size
, alignment
);
3328 gsvs_ring_size
= align(gsvs_ring_size
, alignment
);
3330 esgs_ring_size
= CLAMP(esgs_ring_size
, min_esgs_ring_size
, max_size
);
3331 gsvs_ring_size
= MIN2(gsvs_ring_size
, max_size
);
3333 /* Some rings don't have to be allocated if shaders don't use them.
3334 * (e.g. no varyings between ES and GS or GS and VS)
3336 * GFX9 doesn't have the ESGS ring.
3338 bool update_esgs
= sctx
->chip_class
<= GFX8
&&
3340 (!sctx
->esgs_ring
||
3341 sctx
->esgs_ring
->width0
< esgs_ring_size
);
3342 bool update_gsvs
= gsvs_ring_size
&&
3343 (!sctx
->gsvs_ring
||
3344 sctx
->gsvs_ring
->width0
< gsvs_ring_size
);
3346 if (!update_esgs
&& !update_gsvs
)
3350 pipe_resource_reference(&sctx
->esgs_ring
, NULL
);
3352 pipe_aligned_buffer_create(sctx
->b
.screen
,
3353 SI_RESOURCE_FLAG_UNMAPPABLE
,
3355 esgs_ring_size
, alignment
);
3356 if (!sctx
->esgs_ring
)
3361 pipe_resource_reference(&sctx
->gsvs_ring
, NULL
);
3363 pipe_aligned_buffer_create(sctx
->b
.screen
,
3364 SI_RESOURCE_FLAG_UNMAPPABLE
,
3366 gsvs_ring_size
, alignment
);
3367 if (!sctx
->gsvs_ring
)
3371 /* Create the "init_config_gs_rings" state. */
3372 pm4
= CALLOC_STRUCT(si_pm4_state
);
3376 if (sctx
->chip_class
>= GFX7
) {
3377 if (sctx
->esgs_ring
) {
3378 assert(sctx
->chip_class
<= GFX8
);
3379 si_pm4_set_reg(pm4
, R_030900_VGT_ESGS_RING_SIZE
,
3380 sctx
->esgs_ring
->width0
/ 256);
3382 if (sctx
->gsvs_ring
)
3383 si_pm4_set_reg(pm4
, R_030904_VGT_GSVS_RING_SIZE
,
3384 sctx
->gsvs_ring
->width0
/ 256);
3386 if (sctx
->esgs_ring
)
3387 si_pm4_set_reg(pm4
, R_0088C8_VGT_ESGS_RING_SIZE
,
3388 sctx
->esgs_ring
->width0
/ 256);
3389 if (sctx
->gsvs_ring
)
3390 si_pm4_set_reg(pm4
, R_0088CC_VGT_GSVS_RING_SIZE
,
3391 sctx
->gsvs_ring
->width0
/ 256);
3394 /* Set the state. */
3395 if (sctx
->init_config_gs_rings
)
3396 si_pm4_free_state(sctx
, sctx
->init_config_gs_rings
, ~0);
3397 sctx
->init_config_gs_rings
= pm4
;
3399 if (!sctx
->init_config_has_vgt_flush
) {
3400 si_init_config_add_vgt_flush(sctx
);
3401 si_pm4_upload_indirect_buffer(sctx
, sctx
->init_config
);
3404 /* Flush the context to re-emit both init_config states. */
3405 sctx
->initial_gfx_cs_size
= 0; /* force flush */
3406 si_flush_gfx_cs(sctx
, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW
, NULL
);
3408 /* Set ring bindings. */
3409 if (sctx
->esgs_ring
) {
3410 assert(sctx
->chip_class
<= GFX8
);
3411 si_set_ring_buffer(sctx
, SI_ES_RING_ESGS
,
3412 sctx
->esgs_ring
, 0, sctx
->esgs_ring
->width0
,
3413 true, true, 4, 64, 0);
3414 si_set_ring_buffer(sctx
, SI_GS_RING_ESGS
,
3415 sctx
->esgs_ring
, 0, sctx
->esgs_ring
->width0
,
3416 false, false, 0, 0, 0);
3418 if (sctx
->gsvs_ring
) {
3419 si_set_ring_buffer(sctx
, SI_RING_GSVS
,
3420 sctx
->gsvs_ring
, 0, sctx
->gsvs_ring
->width0
,
3421 false, false, 0, 0, 0);
3427 static void si_shader_lock(struct si_shader
*shader
)
3429 mtx_lock(&shader
->selector
->mutex
);
3430 if (shader
->previous_stage_sel
) {
3431 assert(shader
->previous_stage_sel
!= shader
->selector
);
3432 mtx_lock(&shader
->previous_stage_sel
->mutex
);
3436 static void si_shader_unlock(struct si_shader
*shader
)
3438 if (shader
->previous_stage_sel
)
3439 mtx_unlock(&shader
->previous_stage_sel
->mutex
);
3440 mtx_unlock(&shader
->selector
->mutex
);
3444 * @returns 1 if \p sel has been updated to use a new scratch buffer
3446 * < 0 if there was a failure
3448 static int si_update_scratch_buffer(struct si_context
*sctx
,
3449 struct si_shader
*shader
)
3451 uint64_t scratch_va
= sctx
->scratch_buffer
->gpu_address
;
3456 /* This shader doesn't need a scratch buffer */
3457 if (shader
->config
.scratch_bytes_per_wave
== 0)
3460 /* Prevent race conditions when updating:
3461 * - si_shader::scratch_bo
3462 * - si_shader::binary::code
3463 * - si_shader::previous_stage::binary::code.
3465 si_shader_lock(shader
);
3467 /* This shader is already configured to use the current
3468 * scratch buffer. */
3469 if (shader
->scratch_bo
== sctx
->scratch_buffer
) {
3470 si_shader_unlock(shader
);
3474 assert(sctx
->scratch_buffer
);
3476 /* Replace the shader bo with a new bo that has the relocs applied. */
3477 if (!si_shader_binary_upload(sctx
->screen
, shader
, scratch_va
)) {
3478 si_shader_unlock(shader
);
3482 /* Update the shader state to use the new shader bo. */
3483 si_shader_init_pm4_state(sctx
->screen
, shader
);
3485 si_resource_reference(&shader
->scratch_bo
, sctx
->scratch_buffer
);
3487 si_shader_unlock(shader
);
3491 static unsigned si_get_current_scratch_buffer_size(struct si_context
*sctx
)
3493 return sctx
->scratch_buffer
? sctx
->scratch_buffer
->b
.b
.width0
: 0;
3496 static unsigned si_get_scratch_buffer_bytes_per_wave(struct si_shader
*shader
)
3498 return shader
? shader
->config
.scratch_bytes_per_wave
: 0;
3501 static struct si_shader
*si_get_tcs_current(struct si_context
*sctx
)
3503 if (!sctx
->tes_shader
.cso
)
3504 return NULL
; /* tessellation disabled */
3506 return sctx
->tcs_shader
.cso
? sctx
->tcs_shader
.current
:
3507 sctx
->fixed_func_tcs_shader
.current
;
3510 static unsigned si_get_max_scratch_bytes_per_wave(struct si_context
*sctx
)
3514 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->ps_shader
.current
));
3515 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->gs_shader
.current
));
3516 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->vs_shader
.current
));
3517 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->tes_shader
.current
));
3519 if (sctx
->tes_shader
.cso
) {
3520 struct si_shader
*tcs
= si_get_tcs_current(sctx
);
3522 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(tcs
));
3527 static bool si_update_scratch_relocs(struct si_context
*sctx
)
3529 struct si_shader
*tcs
= si_get_tcs_current(sctx
);
3532 /* Update the shaders, so that they are using the latest scratch.
3533 * The scratch buffer may have been changed since these shaders were
3534 * last used, so we still need to try to update them, even if they
3535 * require scratch buffers smaller than the current size.
3537 r
= si_update_scratch_buffer(sctx
, sctx
->ps_shader
.current
);
3541 si_pm4_bind_state(sctx
, ps
, sctx
->ps_shader
.current
->pm4
);
3543 r
= si_update_scratch_buffer(sctx
, sctx
->gs_shader
.current
);
3547 si_pm4_bind_state(sctx
, gs
, sctx
->gs_shader
.current
->pm4
);
3549 r
= si_update_scratch_buffer(sctx
, tcs
);
3553 si_pm4_bind_state(sctx
, hs
, tcs
->pm4
);
3555 /* VS can be bound as LS, ES, or VS. */
3556 r
= si_update_scratch_buffer(sctx
, sctx
->vs_shader
.current
);
3560 if (sctx
->vs_shader
.current
->key
.as_ls
)
3561 si_pm4_bind_state(sctx
, ls
, sctx
->vs_shader
.current
->pm4
);
3562 else if (sctx
->vs_shader
.current
->key
.as_es
)
3563 si_pm4_bind_state(sctx
, es
, sctx
->vs_shader
.current
->pm4
);
3564 else if (sctx
->vs_shader
.current
->key
.as_ngg
)
3565 si_pm4_bind_state(sctx
, gs
, sctx
->vs_shader
.current
->pm4
);
3567 si_pm4_bind_state(sctx
, vs
, sctx
->vs_shader
.current
->pm4
);
3570 /* TES can be bound as ES or VS. */
3571 r
= si_update_scratch_buffer(sctx
, sctx
->tes_shader
.current
);
3575 if (sctx
->tes_shader
.current
->key
.as_es
)
3576 si_pm4_bind_state(sctx
, es
, sctx
->tes_shader
.current
->pm4
);
3577 else if (sctx
->tes_shader
.current
->key
.as_ngg
)
3578 si_pm4_bind_state(sctx
, gs
, sctx
->tes_shader
.current
->pm4
);
3580 si_pm4_bind_state(sctx
, vs
, sctx
->tes_shader
.current
->pm4
);
3586 static bool si_update_spi_tmpring_size(struct si_context
*sctx
)
3588 unsigned current_scratch_buffer_size
=
3589 si_get_current_scratch_buffer_size(sctx
);
3590 unsigned scratch_bytes_per_wave
=
3591 si_get_max_scratch_bytes_per_wave(sctx
);
3592 unsigned scratch_needed_size
= scratch_bytes_per_wave
*
3593 sctx
->scratch_waves
;
3594 unsigned spi_tmpring_size
;
3596 if (scratch_needed_size
> 0) {
3597 if (scratch_needed_size
> current_scratch_buffer_size
) {
3598 /* Create a bigger scratch buffer */
3599 si_resource_reference(&sctx
->scratch_buffer
, NULL
);
3601 sctx
->scratch_buffer
=
3602 si_aligned_buffer_create(&sctx
->screen
->b
,
3603 SI_RESOURCE_FLAG_UNMAPPABLE
,
3605 scratch_needed_size
, 256);
3606 if (!sctx
->scratch_buffer
)
3609 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.scratch_state
);
3610 si_context_add_resource_size(sctx
,
3611 &sctx
->scratch_buffer
->b
.b
);
3614 if (!si_update_scratch_relocs(sctx
))
3618 /* The LLVM shader backend should be reporting aligned scratch_sizes. */
3619 assert((scratch_needed_size
& ~0x3FF) == scratch_needed_size
&&
3620 "scratch size should already be aligned correctly.");
3622 spi_tmpring_size
= S_0286E8_WAVES(sctx
->scratch_waves
) |
3623 S_0286E8_WAVESIZE(scratch_bytes_per_wave
>> 10);
3624 if (spi_tmpring_size
!= sctx
->spi_tmpring_size
) {
3625 sctx
->spi_tmpring_size
= spi_tmpring_size
;
3626 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.scratch_state
);
3631 static void si_init_tess_factor_ring(struct si_context
*sctx
)
3633 assert(!sctx
->tess_rings
);
3635 /* The address must be aligned to 2^19, because the shader only
3636 * receives the high 13 bits.
3638 sctx
->tess_rings
= pipe_aligned_buffer_create(sctx
->b
.screen
,
3639 SI_RESOURCE_FLAG_32BIT
,
3641 sctx
->screen
->tess_offchip_ring_size
+
3642 sctx
->screen
->tess_factor_ring_size
,
3644 if (!sctx
->tess_rings
)
3647 si_init_config_add_vgt_flush(sctx
);
3649 si_pm4_add_bo(sctx
->init_config
, si_resource(sctx
->tess_rings
),
3650 RADEON_USAGE_READWRITE
, RADEON_PRIO_SHADER_RINGS
);
3652 uint64_t factor_va
= si_resource(sctx
->tess_rings
)->gpu_address
+
3653 sctx
->screen
->tess_offchip_ring_size
;
3655 /* Append these registers to the init config state. */
3656 if (sctx
->chip_class
>= GFX7
) {
3657 si_pm4_set_reg(sctx
->init_config
, R_030938_VGT_TF_RING_SIZE
,
3658 S_030938_SIZE(sctx
->screen
->tess_factor_ring_size
/ 4));
3659 si_pm4_set_reg(sctx
->init_config
, R_030940_VGT_TF_MEMORY_BASE
,
3661 if (sctx
->chip_class
>= GFX10
)
3662 si_pm4_set_reg(sctx
->init_config
, R_030984_VGT_TF_MEMORY_BASE_HI_UMD
,
3663 S_030984_BASE_HI(factor_va
>> 40));
3664 else if (sctx
->chip_class
== GFX9
)
3665 si_pm4_set_reg(sctx
->init_config
, R_030944_VGT_TF_MEMORY_BASE_HI
,
3666 S_030944_BASE_HI(factor_va
>> 40));
3667 si_pm4_set_reg(sctx
->init_config
, R_03093C_VGT_HS_OFFCHIP_PARAM
,
3668 sctx
->screen
->vgt_hs_offchip_param
);
3670 si_pm4_set_reg(sctx
->init_config
, R_008988_VGT_TF_RING_SIZE
,
3671 S_008988_SIZE(sctx
->screen
->tess_factor_ring_size
/ 4));
3672 si_pm4_set_reg(sctx
->init_config
, R_0089B8_VGT_TF_MEMORY_BASE
,
3674 si_pm4_set_reg(sctx
->init_config
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
3675 sctx
->screen
->vgt_hs_offchip_param
);
3678 /* Flush the context to re-emit the init_config state.
3679 * This is done only once in a lifetime of a context.
3681 si_pm4_upload_indirect_buffer(sctx
, sctx
->init_config
);
3682 sctx
->initial_gfx_cs_size
= 0; /* force flush */
3683 si_flush_gfx_cs(sctx
, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW
, NULL
);
3686 static struct si_pm4_state
*si_build_vgt_shader_config(struct si_screen
*screen
,
3687 union si_vgt_stages_key key
)
3689 struct si_pm4_state
*pm4
= CALLOC_STRUCT(si_pm4_state
);
3690 uint32_t stages
= 0;
3693 stages
|= S_028B54_LS_EN(V_028B54_LS_STAGE_ON
) |
3694 S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
3697 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_DS
) |
3700 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_DS
);
3702 stages
|= S_028B54_VS_EN(V_028B54_VS_STAGE_DS
);
3703 } else if (key
.u
.gs
) {
3704 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL
) |
3706 } else if (key
.u
.ngg
) {
3707 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL
);
3711 stages
|= S_028B54_PRIMGEN_EN(1);
3712 if (key
.u
.streamout
)
3713 stages
|= S_028B54_NGG_WAVE_ID_EN(1);
3714 } else if (key
.u
.gs
)
3715 stages
|= S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
3717 if (screen
->info
.chip_class
>= GFX9
)
3718 stages
|= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
3720 si_pm4_set_reg(pm4
, R_028B54_VGT_SHADER_STAGES_EN
, stages
);
3724 static void si_update_vgt_shader_config(struct si_context
*sctx
,
3725 union si_vgt_stages_key key
)
3727 struct si_pm4_state
**pm4
= &sctx
->vgt_shader_config
[key
.index
];
3729 if (unlikely(!*pm4
))
3730 *pm4
= si_build_vgt_shader_config(sctx
->screen
, key
);
3731 si_pm4_bind_state(sctx
, vgt_shader_config
, *pm4
);
3734 bool si_update_shaders(struct si_context
*sctx
)
3736 struct pipe_context
*ctx
= (struct pipe_context
*)sctx
;
3737 struct si_compiler_ctx_state compiler_state
;
3738 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
3739 struct si_shader
*old_vs
= si_get_vs_state(sctx
);
3740 bool old_clip_disable
= old_vs
? old_vs
->key
.opt
.clip_disable
: false;
3741 struct si_shader
*old_ps
= sctx
->ps_shader
.current
;
3742 union si_vgt_stages_key key
;
3743 unsigned old_spi_shader_col_format
=
3744 old_ps
? old_ps
->key
.part
.ps
.epilog
.spi_shader_col_format
: 0;
3747 compiler_state
.compiler
= &sctx
->compiler
;
3748 compiler_state
.debug
= sctx
->debug
;
3749 compiler_state
.is_debug_context
= sctx
->is_debug
;
3753 if (sctx
->tes_shader
.cso
)
3755 if (sctx
->gs_shader
.cso
)
3758 if (sctx
->chip_class
>= GFX10
) {
3759 key
.u
.ngg
= sctx
->ngg
;
3761 if (sctx
->gs_shader
.cso
)
3762 key
.u
.streamout
= !!sctx
->gs_shader
.cso
->so
.num_outputs
;
3763 else if (sctx
->tes_shader
.cso
)
3764 key
.u
.streamout
= !!sctx
->tes_shader
.cso
->so
.num_outputs
;
3766 key
.u
.streamout
= !!sctx
->vs_shader
.cso
->so
.num_outputs
;
3769 /* Update TCS and TES. */
3770 if (sctx
->tes_shader
.cso
) {
3771 if (!sctx
->tess_rings
) {
3772 si_init_tess_factor_ring(sctx
);
3773 if (!sctx
->tess_rings
)
3777 if (sctx
->tcs_shader
.cso
) {
3778 r
= si_shader_select(ctx
, &sctx
->tcs_shader
, key
,
3782 si_pm4_bind_state(sctx
, hs
, sctx
->tcs_shader
.current
->pm4
);
3784 if (!sctx
->fixed_func_tcs_shader
.cso
) {
3785 sctx
->fixed_func_tcs_shader
.cso
=
3786 si_create_fixed_func_tcs(sctx
);
3787 if (!sctx
->fixed_func_tcs_shader
.cso
)
3791 r
= si_shader_select(ctx
, &sctx
->fixed_func_tcs_shader
,
3792 key
, &compiler_state
);
3795 si_pm4_bind_state(sctx
, hs
,
3796 sctx
->fixed_func_tcs_shader
.current
->pm4
);
3799 if (!sctx
->gs_shader
.cso
|| sctx
->chip_class
<= GFX8
) {
3800 r
= si_shader_select(ctx
, &sctx
->tes_shader
, key
, &compiler_state
);
3804 if (sctx
->gs_shader
.cso
) {
3806 assert(sctx
->chip_class
<= GFX8
);
3807 si_pm4_bind_state(sctx
, es
, sctx
->tes_shader
.current
->pm4
);
3808 } else if (key
.u
.ngg
) {
3809 si_pm4_bind_state(sctx
, gs
, sctx
->tes_shader
.current
->pm4
);
3811 si_pm4_bind_state(sctx
, vs
, sctx
->tes_shader
.current
->pm4
);
3815 if (sctx
->chip_class
<= GFX8
)
3816 si_pm4_bind_state(sctx
, ls
, NULL
);
3817 si_pm4_bind_state(sctx
, hs
, NULL
);
3821 if (sctx
->gs_shader
.cso
) {
3822 r
= si_shader_select(ctx
, &sctx
->gs_shader
, key
, &compiler_state
);
3825 si_pm4_bind_state(sctx
, gs
, sctx
->gs_shader
.current
->pm4
);
3827 si_pm4_bind_state(sctx
, vs
, sctx
->gs_shader
.cso
->gs_copy_shader
->pm4
);
3829 if (!si_update_gs_ring_buffers(sctx
))
3832 si_pm4_bind_state(sctx
, vs
, NULL
);
3836 si_pm4_bind_state(sctx
, gs
, NULL
);
3837 if (sctx
->chip_class
<= GFX8
)
3838 si_pm4_bind_state(sctx
, es
, NULL
);
3843 if ((!key
.u
.tess
&& !key
.u
.gs
) || sctx
->chip_class
<= GFX8
) {
3844 r
= si_shader_select(ctx
, &sctx
->vs_shader
, key
, &compiler_state
);
3848 if (!key
.u
.tess
&& !key
.u
.gs
) {
3850 si_pm4_bind_state(sctx
, gs
, sctx
->vs_shader
.current
->pm4
);
3851 si_pm4_bind_state(sctx
, vs
, NULL
);
3853 si_pm4_bind_state(sctx
, vs
, sctx
->vs_shader
.current
->pm4
);
3855 } else if (sctx
->tes_shader
.cso
) {
3856 si_pm4_bind_state(sctx
, ls
, sctx
->vs_shader
.current
->pm4
);
3858 assert(sctx
->gs_shader
.cso
);
3859 si_pm4_bind_state(sctx
, es
, sctx
->vs_shader
.current
->pm4
);
3863 si_update_vgt_shader_config(sctx
, key
);
3865 if (old_clip_disable
!= si_get_vs_state(sctx
)->key
.opt
.clip_disable
)
3866 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.clip_regs
);
3868 if (sctx
->ps_shader
.cso
) {
3869 unsigned db_shader_control
;
3871 r
= si_shader_select(ctx
, &sctx
->ps_shader
, key
, &compiler_state
);
3874 si_pm4_bind_state(sctx
, ps
, sctx
->ps_shader
.current
->pm4
);
3877 sctx
->ps_shader
.cso
->db_shader_control
|
3878 S_02880C_KILL_ENABLE(si_get_alpha_test_func(sctx
) != PIPE_FUNC_ALWAYS
);
3880 if (si_pm4_state_changed(sctx
, ps
) ||
3881 si_pm4_state_changed(sctx
, vs
) ||
3882 (key
.u
.ngg
&& si_pm4_state_changed(sctx
, gs
)) ||
3883 sctx
->sprite_coord_enable
!= rs
->sprite_coord_enable
||
3884 sctx
->flatshade
!= rs
->flatshade
) {
3885 sctx
->sprite_coord_enable
= rs
->sprite_coord_enable
;
3886 sctx
->flatshade
= rs
->flatshade
;
3887 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.spi_map
);
3890 if (sctx
->screen
->rbplus_allowed
&&
3891 si_pm4_state_changed(sctx
, ps
) &&
3893 old_spi_shader_col_format
!=
3894 sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.spi_shader_col_format
))
3895 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.cb_render_state
);
3897 if (sctx
->ps_db_shader_control
!= db_shader_control
) {
3898 sctx
->ps_db_shader_control
= db_shader_control
;
3899 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.db_render_state
);
3900 if (sctx
->screen
->dpbb_allowed
)
3901 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.dpbb_state
);
3904 if (sctx
->smoothing_enabled
!= sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.poly_line_smoothing
) {
3905 sctx
->smoothing_enabled
= sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.poly_line_smoothing
;
3906 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.msaa_config
);
3908 if (sctx
->chip_class
== GFX6
)
3909 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.db_render_state
);
3911 if (sctx
->framebuffer
.nr_samples
<= 1)
3912 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.msaa_sample_locs
);
3916 if (si_pm4_state_enabled_and_changed(sctx
, ls
) ||
3917 si_pm4_state_enabled_and_changed(sctx
, hs
) ||
3918 si_pm4_state_enabled_and_changed(sctx
, es
) ||
3919 si_pm4_state_enabled_and_changed(sctx
, gs
) ||
3920 si_pm4_state_enabled_and_changed(sctx
, vs
) ||
3921 si_pm4_state_enabled_and_changed(sctx
, ps
)) {
3922 if (!si_update_spi_tmpring_size(sctx
))
3926 if (sctx
->chip_class
>= GFX7
) {
3927 if (si_pm4_state_enabled_and_changed(sctx
, ls
))
3928 sctx
->prefetch_L2_mask
|= SI_PREFETCH_LS
;
3929 else if (!sctx
->queued
.named
.ls
)
3930 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_LS
;
3932 if (si_pm4_state_enabled_and_changed(sctx
, hs
))
3933 sctx
->prefetch_L2_mask
|= SI_PREFETCH_HS
;
3934 else if (!sctx
->queued
.named
.hs
)
3935 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_HS
;
3937 if (si_pm4_state_enabled_and_changed(sctx
, es
))
3938 sctx
->prefetch_L2_mask
|= SI_PREFETCH_ES
;
3939 else if (!sctx
->queued
.named
.es
)
3940 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_ES
;
3942 if (si_pm4_state_enabled_and_changed(sctx
, gs
))
3943 sctx
->prefetch_L2_mask
|= SI_PREFETCH_GS
;
3944 else if (!sctx
->queued
.named
.gs
)
3945 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_GS
;
3947 if (si_pm4_state_enabled_and_changed(sctx
, vs
))
3948 sctx
->prefetch_L2_mask
|= SI_PREFETCH_VS
;
3949 else if (!sctx
->queued
.named
.vs
)
3950 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_VS
;
3952 if (si_pm4_state_enabled_and_changed(sctx
, ps
))
3953 sctx
->prefetch_L2_mask
|= SI_PREFETCH_PS
;
3954 else if (!sctx
->queued
.named
.ps
)
3955 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_PS
;
3958 sctx
->do_update_shaders
= false;
3962 static void si_emit_scratch_state(struct si_context
*sctx
)
3964 struct radeon_cmdbuf
*cs
= sctx
->gfx_cs
;
3966 radeon_set_context_reg(cs
, R_0286E8_SPI_TMPRING_SIZE
,
3967 sctx
->spi_tmpring_size
);
3969 if (sctx
->scratch_buffer
) {
3970 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
3971 sctx
->scratch_buffer
, RADEON_USAGE_READWRITE
,
3972 RADEON_PRIO_SCRATCH_BUFFER
);
3976 void si_init_shader_functions(struct si_context
*sctx
)
3978 sctx
->atoms
.s
.spi_map
.emit
= si_emit_spi_map
;
3979 sctx
->atoms
.s
.scratch_state
.emit
= si_emit_scratch_state
;
3981 sctx
->b
.create_vs_state
= si_create_shader_selector
;
3982 sctx
->b
.create_tcs_state
= si_create_shader_selector
;
3983 sctx
->b
.create_tes_state
= si_create_shader_selector
;
3984 sctx
->b
.create_gs_state
= si_create_shader_selector
;
3985 sctx
->b
.create_fs_state
= si_create_shader_selector
;
3987 sctx
->b
.bind_vs_state
= si_bind_vs_shader
;
3988 sctx
->b
.bind_tcs_state
= si_bind_tcs_shader
;
3989 sctx
->b
.bind_tes_state
= si_bind_tes_shader
;
3990 sctx
->b
.bind_gs_state
= si_bind_gs_shader
;
3991 sctx
->b
.bind_fs_state
= si_bind_ps_shader
;
3993 sctx
->b
.delete_vs_state
= si_delete_shader_selector
;
3994 sctx
->b
.delete_tcs_state
= si_delete_shader_selector
;
3995 sctx
->b
.delete_tes_state
= si_delete_shader_selector
;
3996 sctx
->b
.delete_gs_state
= si_delete_shader_selector
;
3997 sctx
->b
.delete_fs_state
= si_delete_shader_selector
;