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 "nir/tgsi_to_nir.h"
30 #include "tgsi/tgsi_parse.h"
31 #include "util/hash_table.h"
32 #include "util/crc32.h"
33 #include "util/u_async_debug.h"
34 #include "util/u_memory.h"
35 #include "util/u_prim.h"
37 #include "util/disk_cache.h"
38 #include "util/mesa-sha1.h"
39 #include "ac_exp_param.h"
40 #include "ac_shader_util.h"
45 * Return the IR key for the shader cache.
47 void si_get_ir_cache_key(struct si_shader_selector
*sel
, bool ngg
, bool es
,
48 unsigned char ir_sha1_cache_key
[20])
50 struct blob blob
= {};
55 ir_binary
= sel
->tokens
;
56 ir_size
= tgsi_num_tokens(sel
->tokens
) *
57 sizeof(struct tgsi_token
);
58 } else if (sel
->nir_binary
) {
59 ir_binary
= sel
->nir_binary
;
60 ir_size
= sel
->nir_size
;
65 nir_serialize(&blob
, sel
->nir
, true);
66 ir_binary
= blob
.data
;
70 /* These settings affect the compilation, but they are not derived
71 * from the input shader IR.
73 unsigned shader_variant_flags
= 0;
76 shader_variant_flags
|= 1 << 0;
78 shader_variant_flags
|= 1 << 1;
79 if (si_get_wave_size(sel
->screen
, sel
->type
, ngg
, es
) == 32)
80 shader_variant_flags
|= 1 << 2;
81 if (sel
->force_correct_derivs_after_kill
)
82 shader_variant_flags
|= 1 << 3;
85 _mesa_sha1_init(&ctx
);
86 _mesa_sha1_update(&ctx
, &shader_variant_flags
, 4);
87 _mesa_sha1_update(&ctx
, ir_binary
, ir_size
);
88 if (sel
->type
== PIPE_SHADER_VERTEX
||
89 sel
->type
== PIPE_SHADER_TESS_EVAL
||
90 sel
->type
== PIPE_SHADER_GEOMETRY
)
91 _mesa_sha1_update(&ctx
, &sel
->so
, sizeof(sel
->so
));
92 _mesa_sha1_final(&ctx
, ir_sha1_cache_key
);
94 if (ir_binary
== blob
.data
)
98 /** Copy "data" to "ptr" and return the next dword following copied data. */
99 static uint32_t *write_data(uint32_t *ptr
, const void *data
, unsigned size
)
101 /* data may be NULL if size == 0 */
103 memcpy(ptr
, data
, size
);
104 ptr
+= DIV_ROUND_UP(size
, 4);
108 /** Read data from "ptr". Return the next dword following the data. */
109 static uint32_t *read_data(uint32_t *ptr
, void *data
, unsigned size
)
111 memcpy(data
, ptr
, size
);
112 ptr
+= DIV_ROUND_UP(size
, 4);
117 * Write the size as uint followed by the data. Return the next dword
118 * following the copied data.
120 static uint32_t *write_chunk(uint32_t *ptr
, const void *data
, unsigned size
)
123 return write_data(ptr
, data
, size
);
127 * Read the size as uint followed by the data. Return both via parameters.
128 * Return the next dword following the data.
130 static uint32_t *read_chunk(uint32_t *ptr
, void **data
, unsigned *size
)
133 assert(*data
== NULL
);
136 *data
= malloc(*size
);
137 return read_data(ptr
, *data
, *size
);
141 * Return the shader binary in a buffer. The first 4 bytes contain its size
144 static void *si_get_shader_binary(struct si_shader
*shader
)
146 /* There is always a size of data followed by the data itself. */
147 unsigned llvm_ir_size
= shader
->binary
.llvm_ir_string
?
148 strlen(shader
->binary
.llvm_ir_string
) + 1 : 0;
150 /* Refuse to allocate overly large buffers and guard against integer
152 if (shader
->binary
.elf_size
> UINT_MAX
/ 4 ||
153 llvm_ir_size
> UINT_MAX
/ 4)
158 4 + /* CRC32 of the data below */
159 align(sizeof(shader
->config
), 4) +
160 align(sizeof(shader
->info
), 4) +
161 4 + align(shader
->binary
.elf_size
, 4) +
162 4 + align(llvm_ir_size
, 4);
163 void *buffer
= CALLOC(1, size
);
164 uint32_t *ptr
= (uint32_t*)buffer
;
170 ptr
++; /* CRC32 is calculated at the end. */
172 ptr
= write_data(ptr
, &shader
->config
, sizeof(shader
->config
));
173 ptr
= write_data(ptr
, &shader
->info
, sizeof(shader
->info
));
174 ptr
= write_chunk(ptr
, shader
->binary
.elf_buffer
, shader
->binary
.elf_size
);
175 ptr
= write_chunk(ptr
, shader
->binary
.llvm_ir_string
, llvm_ir_size
);
176 assert((char *)ptr
- (char *)buffer
== size
);
179 ptr
= (uint32_t*)buffer
;
181 *ptr
= util_hash_crc32(ptr
+ 1, size
- 8);
186 static bool si_load_shader_binary(struct si_shader
*shader
, void *binary
)
188 uint32_t *ptr
= (uint32_t*)binary
;
189 uint32_t size
= *ptr
++;
190 uint32_t crc32
= *ptr
++;
194 if (util_hash_crc32(ptr
, size
- 8) != crc32
) {
195 fprintf(stderr
, "radeonsi: binary shader has invalid CRC32\n");
199 ptr
= read_data(ptr
, &shader
->config
, sizeof(shader
->config
));
200 ptr
= read_data(ptr
, &shader
->info
, sizeof(shader
->info
));
201 ptr
= read_chunk(ptr
, (void**)&shader
->binary
.elf_buffer
,
203 shader
->binary
.elf_size
= elf_size
;
204 ptr
= read_chunk(ptr
, (void**)&shader
->binary
.llvm_ir_string
, &chunk_size
);
210 * Insert a shader into the cache. It's assumed the shader is not in the cache.
211 * Use si_shader_cache_load_shader before calling this.
213 void si_shader_cache_insert_shader(struct si_screen
*sscreen
,
214 unsigned char ir_sha1_cache_key
[20],
215 struct si_shader
*shader
,
216 bool insert_into_disk_cache
)
219 struct hash_entry
*entry
;
220 uint8_t key
[CACHE_KEY_SIZE
];
222 entry
= _mesa_hash_table_search(sscreen
->shader_cache
, ir_sha1_cache_key
);
224 return; /* already added */
226 hw_binary
= si_get_shader_binary(shader
);
230 if (_mesa_hash_table_insert(sscreen
->shader_cache
,
231 mem_dup(ir_sha1_cache_key
, 20),
232 hw_binary
) == NULL
) {
237 if (sscreen
->disk_shader_cache
&& insert_into_disk_cache
) {
238 disk_cache_compute_key(sscreen
->disk_shader_cache
,
239 ir_sha1_cache_key
, 20, key
);
240 disk_cache_put(sscreen
->disk_shader_cache
, key
, hw_binary
,
241 *((uint32_t *) hw_binary
), NULL
);
245 bool si_shader_cache_load_shader(struct si_screen
*sscreen
,
246 unsigned char ir_sha1_cache_key
[20],
247 struct si_shader
*shader
)
249 struct hash_entry
*entry
=
250 _mesa_hash_table_search(sscreen
->shader_cache
, ir_sha1_cache_key
);
252 if (sscreen
->disk_shader_cache
) {
253 unsigned char sha1
[CACHE_KEY_SIZE
];
255 disk_cache_compute_key(sscreen
->disk_shader_cache
,
256 ir_sha1_cache_key
, 20, sha1
);
260 disk_cache_get(sscreen
->disk_shader_cache
,
265 if (binary_size
< sizeof(uint32_t) ||
266 *((uint32_t*)buffer
) != binary_size
) {
267 /* Something has gone wrong discard the item
268 * from the cache and rebuild/link from
271 assert(!"Invalid radeonsi shader disk cache "
274 disk_cache_remove(sscreen
->disk_shader_cache
,
281 if (!si_load_shader_binary(shader
, buffer
)) {
287 si_shader_cache_insert_shader(sscreen
, ir_sha1_cache_key
,
293 if (!si_load_shader_binary(shader
, entry
->data
))
296 p_atomic_inc(&sscreen
->num_shader_cache_hits
);
300 static uint32_t si_shader_cache_key_hash(const void *key
)
302 /* Take the first dword of SHA1. */
303 return *(uint32_t*)key
;
306 static bool si_shader_cache_key_equals(const void *a
, const void *b
)
309 return memcmp(a
, b
, 20) == 0;
312 static void si_destroy_shader_cache_entry(struct hash_entry
*entry
)
314 FREE((void*)entry
->key
);
318 bool si_init_shader_cache(struct si_screen
*sscreen
)
320 (void) simple_mtx_init(&sscreen
->shader_cache_mutex
, mtx_plain
);
321 sscreen
->shader_cache
=
322 _mesa_hash_table_create(NULL
,
323 si_shader_cache_key_hash
,
324 si_shader_cache_key_equals
);
326 return sscreen
->shader_cache
!= NULL
;
329 void si_destroy_shader_cache(struct si_screen
*sscreen
)
331 if (sscreen
->shader_cache
)
332 _mesa_hash_table_destroy(sscreen
->shader_cache
,
333 si_destroy_shader_cache_entry
);
334 simple_mtx_destroy(&sscreen
->shader_cache_mutex
);
339 static void si_set_tesseval_regs(struct si_screen
*sscreen
,
340 const struct si_shader_selector
*tes
,
341 struct si_pm4_state
*pm4
)
343 const struct tgsi_shader_info
*info
= &tes
->info
;
344 unsigned tes_prim_mode
= info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
345 unsigned tes_spacing
= info
->properties
[TGSI_PROPERTY_TES_SPACING
];
346 bool tes_vertex_order_cw
= info
->properties
[TGSI_PROPERTY_TES_VERTEX_ORDER_CW
];
347 bool tes_point_mode
= info
->properties
[TGSI_PROPERTY_TES_POINT_MODE
];
348 unsigned type
, partitioning
, topology
, distribution_mode
;
350 switch (tes_prim_mode
) {
351 case PIPE_PRIM_LINES
:
352 type
= V_028B6C_TESS_ISOLINE
;
354 case PIPE_PRIM_TRIANGLES
:
355 type
= V_028B6C_TESS_TRIANGLE
;
357 case PIPE_PRIM_QUADS
:
358 type
= V_028B6C_TESS_QUAD
;
365 switch (tes_spacing
) {
366 case PIPE_TESS_SPACING_FRACTIONAL_ODD
:
367 partitioning
= V_028B6C_PART_FRAC_ODD
;
369 case PIPE_TESS_SPACING_FRACTIONAL_EVEN
:
370 partitioning
= V_028B6C_PART_FRAC_EVEN
;
372 case PIPE_TESS_SPACING_EQUAL
:
373 partitioning
= V_028B6C_PART_INTEGER
;
381 topology
= V_028B6C_OUTPUT_POINT
;
382 else if (tes_prim_mode
== PIPE_PRIM_LINES
)
383 topology
= V_028B6C_OUTPUT_LINE
;
384 else if (tes_vertex_order_cw
)
385 /* for some reason, this must be the other way around */
386 topology
= V_028B6C_OUTPUT_TRIANGLE_CCW
;
388 topology
= V_028B6C_OUTPUT_TRIANGLE_CW
;
390 if (sscreen
->info
.has_distributed_tess
) {
391 if (sscreen
->info
.family
== CHIP_FIJI
||
392 sscreen
->info
.family
>= CHIP_POLARIS10
)
393 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS
;
395 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_DONUTS
;
397 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_NO_DIST
;
400 pm4
->shader
->vgt_tf_param
= S_028B6C_TYPE(type
) |
401 S_028B6C_PARTITIONING(partitioning
) |
402 S_028B6C_TOPOLOGY(topology
) |
403 S_028B6C_DISTRIBUTION_MODE(distribution_mode
);
406 /* Polaris needs different VTX_REUSE_DEPTH settings depending on
407 * whether the "fractional odd" tessellation spacing is used.
409 * Possible VGT configurations and which state should set the register:
411 * Reg set in | VGT shader configuration | Value
412 * ------------------------------------------------------
414 * VS as ES | ES -> GS -> VS | 30
415 * TES as VS | LS -> HS -> VS | 14 or 30
416 * TES as ES | LS -> HS -> ES -> GS -> VS | 14 or 30
418 * If "shader" is NULL, it's assumed it's not LS or GS copy shader.
420 static void polaris_set_vgt_vertex_reuse(struct si_screen
*sscreen
,
421 struct si_shader_selector
*sel
,
422 struct si_shader
*shader
,
423 struct si_pm4_state
*pm4
)
425 unsigned type
= sel
->type
;
427 if (sscreen
->info
.family
< CHIP_POLARIS10
||
428 sscreen
->info
.chip_class
>= GFX10
)
431 /* VS as VS, or VS as ES: */
432 if ((type
== PIPE_SHADER_VERTEX
&&
434 (!shader
->key
.as_ls
&& !shader
->is_gs_copy_shader
))) ||
435 /* TES as VS, or TES as ES: */
436 type
== PIPE_SHADER_TESS_EVAL
) {
437 unsigned vtx_reuse_depth
= 30;
439 if (type
== PIPE_SHADER_TESS_EVAL
&&
440 sel
->info
.properties
[TGSI_PROPERTY_TES_SPACING
] ==
441 PIPE_TESS_SPACING_FRACTIONAL_ODD
)
442 vtx_reuse_depth
= 14;
445 pm4
->shader
->vgt_vertex_reuse_block_cntl
= vtx_reuse_depth
;
449 static struct si_pm4_state
*si_get_shader_pm4_state(struct si_shader
*shader
)
452 si_pm4_clear_state(shader
->pm4
);
454 shader
->pm4
= CALLOC_STRUCT(si_pm4_state
);
457 shader
->pm4
->shader
= shader
;
460 fprintf(stderr
, "radeonsi: Failed to create pm4 state.\n");
465 static unsigned si_get_num_vs_user_sgprs(unsigned num_always_on_user_sgprs
)
467 /* Add the pointer to VBO descriptors. */
468 return num_always_on_user_sgprs
+ 1;
471 /* Return VGPR_COMP_CNT for the API vertex shader. This can be hw LS, LSHS, ES, ESGS, VS. */
472 static unsigned si_get_vs_vgpr_comp_cnt(struct si_screen
*sscreen
,
473 struct si_shader
*shader
, bool legacy_vs_prim_id
)
475 assert(shader
->selector
->type
== PIPE_SHADER_VERTEX
||
476 (shader
->previous_stage_sel
&&
477 shader
->previous_stage_sel
->type
== PIPE_SHADER_VERTEX
));
479 /* GFX6-9 LS (VertexID, RelAutoindex, InstanceID / StepRate0(==1), ...).
480 * GFX6-9 ES,VS (VertexID, InstanceID / StepRate0(==1), VSPrimID, ...)
481 * GFX10 LS (VertexID, RelAutoindex, UserVGPR1, InstanceID).
482 * GFX10 ES,VS (VertexID, UserVGPR0, UserVGPR1 or VSPrimID, UserVGPR2 or InstanceID)
484 bool is_ls
= shader
->selector
->type
== PIPE_SHADER_TESS_CTRL
|| shader
->key
.as_ls
;
486 if (sscreen
->info
.chip_class
>= GFX10
&& shader
->info
.uses_instanceid
)
488 else if ((is_ls
&& shader
->info
.uses_instanceid
) || legacy_vs_prim_id
)
490 else if (is_ls
|| shader
->info
.uses_instanceid
)
496 static void si_shader_ls(struct si_screen
*sscreen
, struct si_shader
*shader
)
498 struct si_pm4_state
*pm4
;
501 assert(sscreen
->info
.chip_class
<= GFX8
);
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 si_pm4_set_reg(pm4
, R_00B520_SPI_SHADER_PGM_LO_LS
, va
>> 8);
511 si_pm4_set_reg(pm4
, R_00B524_SPI_SHADER_PGM_HI_LS
, S_00B524_MEM_BASE(va
>> 40));
513 shader
->config
.rsrc1
= S_00B528_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
514 S_00B528_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
515 S_00B528_VGPR_COMP_CNT(si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false)) |
516 S_00B528_DX10_CLAMP(1) |
517 S_00B528_FLOAT_MODE(shader
->config
.float_mode
);
518 shader
->config
.rsrc2
= S_00B52C_USER_SGPR(si_get_num_vs_user_sgprs(SI_VS_NUM_USER_SGPR
)) |
519 S_00B52C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
522 static void si_shader_hs(struct si_screen
*sscreen
, struct si_shader
*shader
)
524 struct si_pm4_state
*pm4
;
527 pm4
= si_get_shader_pm4_state(shader
);
531 va
= shader
->bo
->gpu_address
;
532 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
534 if (sscreen
->info
.chip_class
>= GFX9
) {
535 if (sscreen
->info
.chip_class
>= GFX10
) {
536 si_pm4_set_reg(pm4
, R_00B520_SPI_SHADER_PGM_LO_LS
, va
>> 8);
537 si_pm4_set_reg(pm4
, R_00B524_SPI_SHADER_PGM_HI_LS
, S_00B524_MEM_BASE(va
>> 40));
539 si_pm4_set_reg(pm4
, R_00B410_SPI_SHADER_PGM_LO_LS
, va
>> 8);
540 si_pm4_set_reg(pm4
, R_00B414_SPI_SHADER_PGM_HI_LS
, S_00B414_MEM_BASE(va
>> 40));
543 unsigned num_user_sgprs
=
544 si_get_num_vs_user_sgprs(GFX9_TCS_NUM_USER_SGPR
);
546 shader
->config
.rsrc2
=
547 S_00B42C_USER_SGPR(num_user_sgprs
) |
548 S_00B42C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
550 if (sscreen
->info
.chip_class
>= GFX10
)
551 shader
->config
.rsrc2
|= S_00B42C_USER_SGPR_MSB_GFX10(num_user_sgprs
>> 5);
553 shader
->config
.rsrc2
|= S_00B42C_USER_SGPR_MSB_GFX9(num_user_sgprs
>> 5);
555 si_pm4_set_reg(pm4
, R_00B420_SPI_SHADER_PGM_LO_HS
, va
>> 8);
556 si_pm4_set_reg(pm4
, R_00B424_SPI_SHADER_PGM_HI_HS
, S_00B424_MEM_BASE(va
>> 40));
558 shader
->config
.rsrc2
=
559 S_00B42C_USER_SGPR(GFX6_TCS_NUM_USER_SGPR
) |
560 S_00B42C_OC_LDS_EN(1) |
561 S_00B42C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
564 si_pm4_set_reg(pm4
, R_00B428_SPI_SHADER_PGM_RSRC1_HS
,
565 S_00B428_VGPRS((shader
->config
.num_vgprs
- 1) /
566 (sscreen
->ge_wave_size
== 32 ? 8 : 4)) |
567 (sscreen
->info
.chip_class
<= GFX9
?
568 S_00B428_SGPRS((shader
->config
.num_sgprs
- 1) / 8) : 0) |
569 S_00B428_DX10_CLAMP(1) |
570 S_00B428_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
571 S_00B428_WGP_MODE(sscreen
->info
.chip_class
>= GFX10
) |
572 S_00B428_FLOAT_MODE(shader
->config
.float_mode
) |
573 S_00B428_LS_VGPR_COMP_CNT(sscreen
->info
.chip_class
>= GFX9
?
574 si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false) : 0));
576 if (sscreen
->info
.chip_class
<= GFX8
) {
577 si_pm4_set_reg(pm4
, R_00B42C_SPI_SHADER_PGM_RSRC2_HS
,
578 shader
->config
.rsrc2
);
582 static void si_emit_shader_es(struct si_context
*sctx
)
584 struct si_shader
*shader
= sctx
->queued
.named
.es
->shader
;
585 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
590 radeon_opt_set_context_reg(sctx
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
591 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE
,
592 shader
->selector
->esgs_itemsize
/ 4);
594 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
595 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
596 SI_TRACKED_VGT_TF_PARAM
,
597 shader
->vgt_tf_param
);
599 if (shader
->vgt_vertex_reuse_block_cntl
)
600 radeon_opt_set_context_reg(sctx
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
601 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL
,
602 shader
->vgt_vertex_reuse_block_cntl
);
604 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
605 sctx
->context_roll
= true;
608 static void si_shader_es(struct si_screen
*sscreen
, struct si_shader
*shader
)
610 struct si_pm4_state
*pm4
;
611 unsigned num_user_sgprs
;
612 unsigned vgpr_comp_cnt
;
616 assert(sscreen
->info
.chip_class
<= GFX8
);
618 pm4
= si_get_shader_pm4_state(shader
);
622 pm4
->atom
.emit
= si_emit_shader_es
;
623 va
= shader
->bo
->gpu_address
;
624 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
626 if (shader
->selector
->type
== PIPE_SHADER_VERTEX
) {
627 vgpr_comp_cnt
= si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false);
628 num_user_sgprs
= si_get_num_vs_user_sgprs(SI_VS_NUM_USER_SGPR
);
629 } else if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
630 vgpr_comp_cnt
= shader
->selector
->info
.uses_primid
? 3 : 2;
631 num_user_sgprs
= SI_TES_NUM_USER_SGPR
;
633 unreachable("invalid shader selector type");
635 oc_lds_en
= shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
? 1 : 0;
637 si_pm4_set_reg(pm4
, R_00B320_SPI_SHADER_PGM_LO_ES
, va
>> 8);
638 si_pm4_set_reg(pm4
, R_00B324_SPI_SHADER_PGM_HI_ES
, S_00B324_MEM_BASE(va
>> 40));
639 si_pm4_set_reg(pm4
, R_00B328_SPI_SHADER_PGM_RSRC1_ES
,
640 S_00B328_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
641 S_00B328_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
642 S_00B328_VGPR_COMP_CNT(vgpr_comp_cnt
) |
643 S_00B328_DX10_CLAMP(1) |
644 S_00B328_FLOAT_MODE(shader
->config
.float_mode
));
645 si_pm4_set_reg(pm4
, R_00B32C_SPI_SHADER_PGM_RSRC2_ES
,
646 S_00B32C_USER_SGPR(num_user_sgprs
) |
647 S_00B32C_OC_LDS_EN(oc_lds_en
) |
648 S_00B32C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
650 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
651 si_set_tesseval_regs(sscreen
, shader
->selector
, pm4
);
653 polaris_set_vgt_vertex_reuse(sscreen
, shader
->selector
, shader
, pm4
);
656 void gfx9_get_gs_info(struct si_shader_selector
*es
,
657 struct si_shader_selector
*gs
,
658 struct gfx9_gs_info
*out
)
660 unsigned gs_num_invocations
= MAX2(gs
->gs_num_invocations
, 1);
661 unsigned input_prim
= gs
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
];
662 bool uses_adjacency
= input_prim
>= PIPE_PRIM_LINES_ADJACENCY
&&
663 input_prim
<= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
;
665 /* All these are in dwords: */
666 /* We can't allow using the whole LDS, because GS waves compete with
667 * other shader stages for LDS space. */
668 const unsigned max_lds_size
= 8 * 1024;
669 const unsigned esgs_itemsize
= es
->esgs_itemsize
/ 4;
670 unsigned esgs_lds_size
;
672 /* All these are per subgroup: */
673 const unsigned max_out_prims
= 32 * 1024;
674 const unsigned max_es_verts
= 255;
675 const unsigned ideal_gs_prims
= 64;
676 unsigned max_gs_prims
, gs_prims
;
677 unsigned min_es_verts
, es_verts
, worst_case_es_verts
;
679 if (uses_adjacency
|| gs_num_invocations
> 1)
680 max_gs_prims
= 127 / gs_num_invocations
;
684 /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
685 * Make sure we don't go over the maximum value.
687 if (gs
->gs_max_out_vertices
> 0) {
688 max_gs_prims
= MIN2(max_gs_prims
,
690 (gs
->gs_max_out_vertices
* gs_num_invocations
));
692 assert(max_gs_prims
> 0);
694 /* If the primitive has adjacency, halve the number of vertices
695 * that will be reused in multiple primitives.
697 min_es_verts
= gs
->gs_input_verts_per_prim
/ (uses_adjacency
? 2 : 1);
699 gs_prims
= MIN2(ideal_gs_prims
, max_gs_prims
);
700 worst_case_es_verts
= MIN2(min_es_verts
* gs_prims
, max_es_verts
);
702 /* Compute ESGS LDS size based on the worst case number of ES vertices
703 * needed to create the target number of GS prims per subgroup.
705 esgs_lds_size
= esgs_itemsize
* worst_case_es_verts
;
707 /* If total LDS usage is too big, refactor partitions based on ratio
708 * of ESGS item sizes.
710 if (esgs_lds_size
> max_lds_size
) {
711 /* Our target GS Prims Per Subgroup was too large. Calculate
712 * the maximum number of GS Prims Per Subgroup that will fit
713 * into LDS, capped by the maximum that the hardware can support.
715 gs_prims
= MIN2((max_lds_size
/ (esgs_itemsize
* min_es_verts
)),
717 assert(gs_prims
> 0);
718 worst_case_es_verts
= MIN2(min_es_verts
* gs_prims
,
721 esgs_lds_size
= esgs_itemsize
* worst_case_es_verts
;
722 assert(esgs_lds_size
<= max_lds_size
);
725 /* Now calculate remaining ESGS information. */
727 es_verts
= MIN2(esgs_lds_size
/ esgs_itemsize
, max_es_verts
);
729 es_verts
= max_es_verts
;
731 /* Vertices for adjacency primitives are not always reused, so restore
732 * it for ES_VERTS_PER_SUBGRP.
734 min_es_verts
= gs
->gs_input_verts_per_prim
;
736 /* For normal primitives, the VGT only checks if they are past the ES
737 * verts per subgroup after allocating a full GS primitive and if they
738 * are, kick off a new subgroup. But if those additional ES verts are
739 * unique (e.g. not reused) we need to make sure there is enough LDS
740 * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
742 es_verts
-= min_es_verts
- 1;
744 out
->es_verts_per_subgroup
= es_verts
;
745 out
->gs_prims_per_subgroup
= gs_prims
;
746 out
->gs_inst_prims_in_subgroup
= gs_prims
* gs_num_invocations
;
747 out
->max_prims_per_subgroup
= out
->gs_inst_prims_in_subgroup
*
748 gs
->gs_max_out_vertices
;
749 out
->esgs_ring_size
= 4 * esgs_lds_size
;
751 assert(out
->max_prims_per_subgroup
<= max_out_prims
);
754 static void si_emit_shader_gs(struct si_context
*sctx
)
756 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
757 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
762 /* R_028A60_VGT_GSVS_RING_OFFSET_1, R_028A64_VGT_GSVS_RING_OFFSET_2
763 * R_028A68_VGT_GSVS_RING_OFFSET_3 */
764 radeon_opt_set_context_reg3(sctx
, R_028A60_VGT_GSVS_RING_OFFSET_1
,
765 SI_TRACKED_VGT_GSVS_RING_OFFSET_1
,
766 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_1
,
767 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_2
,
768 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_3
);
770 /* R_028AB0_VGT_GSVS_RING_ITEMSIZE */
771 radeon_opt_set_context_reg(sctx
, R_028AB0_VGT_GSVS_RING_ITEMSIZE
,
772 SI_TRACKED_VGT_GSVS_RING_ITEMSIZE
,
773 shader
->ctx_reg
.gs
.vgt_gsvs_ring_itemsize
);
775 /* R_028B38_VGT_GS_MAX_VERT_OUT */
776 radeon_opt_set_context_reg(sctx
, R_028B38_VGT_GS_MAX_VERT_OUT
,
777 SI_TRACKED_VGT_GS_MAX_VERT_OUT
,
778 shader
->ctx_reg
.gs
.vgt_gs_max_vert_out
);
780 /* R_028B5C_VGT_GS_VERT_ITEMSIZE, R_028B60_VGT_GS_VERT_ITEMSIZE_1
781 * R_028B64_VGT_GS_VERT_ITEMSIZE_2, R_028B68_VGT_GS_VERT_ITEMSIZE_3 */
782 radeon_opt_set_context_reg4(sctx
, R_028B5C_VGT_GS_VERT_ITEMSIZE
,
783 SI_TRACKED_VGT_GS_VERT_ITEMSIZE
,
784 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize
,
785 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_1
,
786 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_2
,
787 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_3
);
789 /* R_028B90_VGT_GS_INSTANCE_CNT */
790 radeon_opt_set_context_reg(sctx
, R_028B90_VGT_GS_INSTANCE_CNT
,
791 SI_TRACKED_VGT_GS_INSTANCE_CNT
,
792 shader
->ctx_reg
.gs
.vgt_gs_instance_cnt
);
794 if (sctx
->chip_class
>= GFX9
) {
795 /* R_028A44_VGT_GS_ONCHIP_CNTL */
796 radeon_opt_set_context_reg(sctx
, R_028A44_VGT_GS_ONCHIP_CNTL
,
797 SI_TRACKED_VGT_GS_ONCHIP_CNTL
,
798 shader
->ctx_reg
.gs
.vgt_gs_onchip_cntl
);
799 /* R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP */
800 radeon_opt_set_context_reg(sctx
, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP
,
801 SI_TRACKED_VGT_GS_MAX_PRIMS_PER_SUBGROUP
,
802 shader
->ctx_reg
.gs
.vgt_gs_max_prims_per_subgroup
);
803 /* R_028AAC_VGT_ESGS_RING_ITEMSIZE */
804 radeon_opt_set_context_reg(sctx
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
805 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE
,
806 shader
->ctx_reg
.gs
.vgt_esgs_ring_itemsize
);
808 if (shader
->key
.part
.gs
.es
->type
== PIPE_SHADER_TESS_EVAL
)
809 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
810 SI_TRACKED_VGT_TF_PARAM
,
811 shader
->vgt_tf_param
);
812 if (shader
->vgt_vertex_reuse_block_cntl
)
813 radeon_opt_set_context_reg(sctx
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
814 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL
,
815 shader
->vgt_vertex_reuse_block_cntl
);
818 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
819 sctx
->context_roll
= true;
822 static void si_shader_gs(struct si_screen
*sscreen
, struct si_shader
*shader
)
824 struct si_shader_selector
*sel
= shader
->selector
;
825 const ubyte
*num_components
= sel
->info
.num_stream_output_components
;
826 unsigned gs_num_invocations
= sel
->gs_num_invocations
;
827 struct si_pm4_state
*pm4
;
829 unsigned max_stream
= sel
->max_gs_stream
;
832 pm4
= si_get_shader_pm4_state(shader
);
836 pm4
->atom
.emit
= si_emit_shader_gs
;
838 offset
= num_components
[0] * sel
->gs_max_out_vertices
;
839 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_1
= offset
;
842 offset
+= num_components
[1] * sel
->gs_max_out_vertices
;
843 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_2
= offset
;
846 offset
+= num_components
[2] * sel
->gs_max_out_vertices
;
847 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_3
= offset
;
850 offset
+= num_components
[3] * sel
->gs_max_out_vertices
;
851 shader
->ctx_reg
.gs
.vgt_gsvs_ring_itemsize
= offset
;
853 /* The GSVS_RING_ITEMSIZE register takes 15 bits */
854 assert(offset
< (1 << 15));
856 shader
->ctx_reg
.gs
.vgt_gs_max_vert_out
= sel
->gs_max_out_vertices
;
858 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize
= num_components
[0];
859 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_1
= (max_stream
>= 1) ? num_components
[1] : 0;
860 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_2
= (max_stream
>= 2) ? num_components
[2] : 0;
861 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_3
= (max_stream
>= 3) ? num_components
[3] : 0;
863 shader
->ctx_reg
.gs
.vgt_gs_instance_cnt
= S_028B90_CNT(MIN2(gs_num_invocations
, 127)) |
864 S_028B90_ENABLE(gs_num_invocations
> 0);
866 va
= shader
->bo
->gpu_address
;
867 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
869 if (sscreen
->info
.chip_class
>= GFX9
) {
870 unsigned input_prim
= sel
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
];
871 unsigned es_type
= shader
->key
.part
.gs
.es
->type
;
872 unsigned es_vgpr_comp_cnt
, gs_vgpr_comp_cnt
;
874 if (es_type
== PIPE_SHADER_VERTEX
) {
875 es_vgpr_comp_cnt
= si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false);
876 } else if (es_type
== PIPE_SHADER_TESS_EVAL
)
877 es_vgpr_comp_cnt
= shader
->key
.part
.gs
.es
->info
.uses_primid
? 3 : 2;
879 unreachable("invalid shader selector type");
881 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
882 * VGPR[0:4] are always loaded.
884 if (sel
->info
.uses_invocationid
)
885 gs_vgpr_comp_cnt
= 3; /* VGPR3 contains InvocationID. */
886 else if (sel
->info
.uses_primid
)
887 gs_vgpr_comp_cnt
= 2; /* VGPR2 contains PrimitiveID. */
888 else if (input_prim
>= PIPE_PRIM_TRIANGLES
)
889 gs_vgpr_comp_cnt
= 1; /* VGPR1 contains offsets 2, 3 */
891 gs_vgpr_comp_cnt
= 0; /* VGPR0 contains offsets 0, 1 */
893 unsigned num_user_sgprs
;
894 if (es_type
== PIPE_SHADER_VERTEX
)
895 num_user_sgprs
= si_get_num_vs_user_sgprs(GFX9_VSGS_NUM_USER_SGPR
);
897 num_user_sgprs
= GFX9_TESGS_NUM_USER_SGPR
;
899 if (sscreen
->info
.chip_class
>= GFX10
) {
900 si_pm4_set_reg(pm4
, R_00B320_SPI_SHADER_PGM_LO_ES
, va
>> 8);
901 si_pm4_set_reg(pm4
, R_00B324_SPI_SHADER_PGM_HI_ES
, S_00B324_MEM_BASE(va
>> 40));
903 si_pm4_set_reg(pm4
, R_00B210_SPI_SHADER_PGM_LO_ES
, va
>> 8);
904 si_pm4_set_reg(pm4
, R_00B214_SPI_SHADER_PGM_HI_ES
, S_00B214_MEM_BASE(va
>> 40));
908 S_00B228_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
909 S_00B228_DX10_CLAMP(1) |
910 S_00B228_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
911 S_00B228_WGP_MODE(sscreen
->info
.chip_class
>= GFX10
) |
912 S_00B228_FLOAT_MODE(shader
->config
.float_mode
) |
913 S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt
);
915 S_00B22C_USER_SGPR(num_user_sgprs
) |
916 S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt
) |
917 S_00B22C_OC_LDS_EN(es_type
== PIPE_SHADER_TESS_EVAL
) |
918 S_00B22C_LDS_SIZE(shader
->config
.lds_size
) |
919 S_00B22C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
921 if (sscreen
->info
.chip_class
>= GFX10
) {
922 rsrc2
|= S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs
>> 5);
924 rsrc1
|= S_00B228_SGPRS((shader
->config
.num_sgprs
- 1) / 8);
925 rsrc2
|= S_00B22C_USER_SGPR_MSB_GFX9(num_user_sgprs
>> 5);
928 si_pm4_set_reg(pm4
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
, rsrc1
);
929 si_pm4_set_reg(pm4
, R_00B22C_SPI_SHADER_PGM_RSRC2_GS
, rsrc2
);
931 shader
->ctx_reg
.gs
.vgt_gs_onchip_cntl
=
932 S_028A44_ES_VERTS_PER_SUBGRP(shader
->gs_info
.es_verts_per_subgroup
) |
933 S_028A44_GS_PRIMS_PER_SUBGRP(shader
->gs_info
.gs_prims_per_subgroup
) |
934 S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader
->gs_info
.gs_inst_prims_in_subgroup
);
935 shader
->ctx_reg
.gs
.vgt_gs_max_prims_per_subgroup
=
936 S_028A94_MAX_PRIMS_PER_SUBGROUP(shader
->gs_info
.max_prims_per_subgroup
);
937 shader
->ctx_reg
.gs
.vgt_esgs_ring_itemsize
=
938 shader
->key
.part
.gs
.es
->esgs_itemsize
/ 4;
940 if (es_type
== PIPE_SHADER_TESS_EVAL
)
941 si_set_tesseval_regs(sscreen
, shader
->key
.part
.gs
.es
, pm4
);
943 polaris_set_vgt_vertex_reuse(sscreen
, shader
->key
.part
.gs
.es
,
946 si_pm4_set_reg(pm4
, R_00B220_SPI_SHADER_PGM_LO_GS
, va
>> 8);
947 si_pm4_set_reg(pm4
, R_00B224_SPI_SHADER_PGM_HI_GS
, S_00B224_MEM_BASE(va
>> 40));
949 si_pm4_set_reg(pm4
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
,
950 S_00B228_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
951 S_00B228_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
952 S_00B228_DX10_CLAMP(1) |
953 S_00B228_FLOAT_MODE(shader
->config
.float_mode
));
954 si_pm4_set_reg(pm4
, R_00B22C_SPI_SHADER_PGM_RSRC2_GS
,
955 S_00B22C_USER_SGPR(GFX6_GS_NUM_USER_SGPR
) |
956 S_00B22C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
960 /* Common tail code for NGG primitive shaders. */
961 static void gfx10_emit_shader_ngg_tail(struct si_context
*sctx
,
962 struct si_shader
*shader
,
963 unsigned initial_cdw
)
965 radeon_opt_set_context_reg(sctx
, R_0287FC_GE_MAX_OUTPUT_PER_SUBGROUP
,
966 SI_TRACKED_GE_MAX_OUTPUT_PER_SUBGROUP
,
967 shader
->ctx_reg
.ngg
.ge_max_output_per_subgroup
);
968 radeon_opt_set_context_reg(sctx
, R_028B4C_GE_NGG_SUBGRP_CNTL
,
969 SI_TRACKED_GE_NGG_SUBGRP_CNTL
,
970 shader
->ctx_reg
.ngg
.ge_ngg_subgrp_cntl
);
971 radeon_opt_set_context_reg(sctx
, R_028A84_VGT_PRIMITIVEID_EN
,
972 SI_TRACKED_VGT_PRIMITIVEID_EN
,
973 shader
->ctx_reg
.ngg
.vgt_primitiveid_en
);
974 radeon_opt_set_context_reg(sctx
, R_028A44_VGT_GS_ONCHIP_CNTL
,
975 SI_TRACKED_VGT_GS_ONCHIP_CNTL
,
976 shader
->ctx_reg
.ngg
.vgt_gs_onchip_cntl
);
977 radeon_opt_set_context_reg(sctx
, R_028B90_VGT_GS_INSTANCE_CNT
,
978 SI_TRACKED_VGT_GS_INSTANCE_CNT
,
979 shader
->ctx_reg
.ngg
.vgt_gs_instance_cnt
);
980 radeon_opt_set_context_reg(sctx
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
981 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE
,
982 shader
->ctx_reg
.ngg
.vgt_esgs_ring_itemsize
);
983 radeon_opt_set_context_reg(sctx
, R_0286C4_SPI_VS_OUT_CONFIG
,
984 SI_TRACKED_SPI_VS_OUT_CONFIG
,
985 shader
->ctx_reg
.ngg
.spi_vs_out_config
);
986 radeon_opt_set_context_reg2(sctx
, R_028708_SPI_SHADER_IDX_FORMAT
,
987 SI_TRACKED_SPI_SHADER_IDX_FORMAT
,
988 shader
->ctx_reg
.ngg
.spi_shader_idx_format
,
989 shader
->ctx_reg
.ngg
.spi_shader_pos_format
);
990 radeon_opt_set_context_reg(sctx
, R_028818_PA_CL_VTE_CNTL
,
991 SI_TRACKED_PA_CL_VTE_CNTL
,
992 shader
->ctx_reg
.ngg
.pa_cl_vte_cntl
);
993 radeon_opt_set_context_reg(sctx
, R_028838_PA_CL_NGG_CNTL
,
994 SI_TRACKED_PA_CL_NGG_CNTL
,
995 shader
->ctx_reg
.ngg
.pa_cl_ngg_cntl
);
997 radeon_opt_set_context_reg_rmw(sctx
, R_02881C_PA_CL_VS_OUT_CNTL
,
998 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS
,
999 shader
->pa_cl_vs_out_cntl
,
1000 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS_MASK
);
1002 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
1003 sctx
->context_roll
= true;
1006 static void gfx10_emit_shader_ngg_notess_nogs(struct si_context
*sctx
)
1008 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1009 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1014 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1017 static void gfx10_emit_shader_ngg_tess_nogs(struct si_context
*sctx
)
1019 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1020 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1025 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
1026 SI_TRACKED_VGT_TF_PARAM
,
1027 shader
->vgt_tf_param
);
1029 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1032 static void gfx10_emit_shader_ngg_notess_gs(struct si_context
*sctx
)
1034 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1035 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1040 radeon_opt_set_context_reg(sctx
, R_028B38_VGT_GS_MAX_VERT_OUT
,
1041 SI_TRACKED_VGT_GS_MAX_VERT_OUT
,
1042 shader
->ctx_reg
.ngg
.vgt_gs_max_vert_out
);
1044 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1047 static void gfx10_emit_shader_ngg_tess_gs(struct si_context
*sctx
)
1049 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1050 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1055 radeon_opt_set_context_reg(sctx
, R_028B38_VGT_GS_MAX_VERT_OUT
,
1056 SI_TRACKED_VGT_GS_MAX_VERT_OUT
,
1057 shader
->ctx_reg
.ngg
.vgt_gs_max_vert_out
);
1058 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
1059 SI_TRACKED_VGT_TF_PARAM
,
1060 shader
->vgt_tf_param
);
1062 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1065 unsigned si_get_input_prim(const struct si_shader_selector
*gs
)
1067 if (gs
->type
== PIPE_SHADER_GEOMETRY
)
1068 return gs
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
];
1070 if (gs
->type
== PIPE_SHADER_TESS_EVAL
) {
1071 if (gs
->info
.properties
[TGSI_PROPERTY_TES_POINT_MODE
])
1072 return PIPE_PRIM_POINTS
;
1073 if (gs
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] == PIPE_PRIM_LINES
)
1074 return PIPE_PRIM_LINES
;
1075 return PIPE_PRIM_TRIANGLES
;
1078 /* TODO: Set this correctly if the primitive type is set in the shader key. */
1079 return PIPE_PRIM_TRIANGLES
; /* worst case for all callers */
1082 static unsigned si_get_vs_out_cntl(const struct si_shader_selector
*sel
, bool ngg
)
1085 sel
->info
.writes_psize
|| (sel
->info
.writes_edgeflag
&& !ngg
) ||
1086 sel
->info
.writes_layer
|| sel
->info
.writes_viewport_index
;
1087 return S_02881C_USE_VTX_POINT_SIZE(sel
->info
.writes_psize
) |
1088 S_02881C_USE_VTX_EDGE_FLAG(sel
->info
.writes_edgeflag
&& !ngg
) |
1089 S_02881C_USE_VTX_RENDER_TARGET_INDX(sel
->info
.writes_layer
) |
1090 S_02881C_USE_VTX_VIEWPORT_INDX(sel
->info
.writes_viewport_index
) |
1091 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena
) |
1092 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena
);
1096 * Prepare the PM4 image for \p shader, which will run as a merged ESGS shader
1099 static void gfx10_shader_ngg(struct si_screen
*sscreen
, struct si_shader
*shader
)
1101 const struct si_shader_selector
*gs_sel
= shader
->selector
;
1102 const struct tgsi_shader_info
*gs_info
= &gs_sel
->info
;
1103 enum pipe_shader_type gs_type
= shader
->selector
->type
;
1104 const struct si_shader_selector
*es_sel
=
1105 shader
->previous_stage_sel
? shader
->previous_stage_sel
: shader
->selector
;
1106 const struct tgsi_shader_info
*es_info
= &es_sel
->info
;
1107 enum pipe_shader_type es_type
= es_sel
->type
;
1108 unsigned num_user_sgprs
;
1109 unsigned nparams
, es_vgpr_comp_cnt
, gs_vgpr_comp_cnt
;
1111 unsigned window_space
=
1112 gs_info
->properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
];
1113 bool es_enable_prim_id
= shader
->key
.mono
.u
.vs_export_prim_id
|| es_info
->uses_primid
;
1114 unsigned gs_num_invocations
= MAX2(gs_sel
->gs_num_invocations
, 1);
1115 unsigned input_prim
= si_get_input_prim(gs_sel
);
1116 bool break_wave_at_eoi
= false;
1117 struct si_pm4_state
*pm4
= si_get_shader_pm4_state(shader
);
1121 if (es_type
== PIPE_SHADER_TESS_EVAL
) {
1122 pm4
->atom
.emit
= gs_type
== PIPE_SHADER_GEOMETRY
? gfx10_emit_shader_ngg_tess_gs
1123 : gfx10_emit_shader_ngg_tess_nogs
;
1125 pm4
->atom
.emit
= gs_type
== PIPE_SHADER_GEOMETRY
? gfx10_emit_shader_ngg_notess_gs
1126 : gfx10_emit_shader_ngg_notess_nogs
;
1129 va
= shader
->bo
->gpu_address
;
1130 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
1132 if (es_type
== PIPE_SHADER_VERTEX
) {
1133 es_vgpr_comp_cnt
= si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false);
1135 if (es_info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
]) {
1136 num_user_sgprs
= SI_SGPR_VS_BLIT_DATA
+
1137 es_info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
];
1139 num_user_sgprs
= si_get_num_vs_user_sgprs(GFX9_VSGS_NUM_USER_SGPR
);
1142 assert(es_type
== PIPE_SHADER_TESS_EVAL
);
1143 es_vgpr_comp_cnt
= es_enable_prim_id
? 3 : 2;
1144 num_user_sgprs
= GFX9_TESGS_NUM_USER_SGPR
;
1146 if (es_enable_prim_id
|| gs_info
->uses_primid
)
1147 break_wave_at_eoi
= true;
1150 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
1151 * VGPR[0:4] are always loaded.
1153 * Vertex shaders always need to load VGPR3, because they need to
1154 * pass edge flags for decomposed primitives (such as quads) to the PA
1155 * for the GL_LINE polygon mode to skip rendering lines on inner edges.
1157 if (gs_info
->uses_invocationid
|| gs_type
== PIPE_SHADER_VERTEX
)
1158 gs_vgpr_comp_cnt
= 3; /* VGPR3 contains InvocationID, edge flags. */
1159 else if (gs_info
->uses_primid
)
1160 gs_vgpr_comp_cnt
= 2; /* VGPR2 contains PrimitiveID. */
1161 else if (input_prim
>= PIPE_PRIM_TRIANGLES
)
1162 gs_vgpr_comp_cnt
= 1; /* VGPR1 contains offsets 2, 3 */
1164 gs_vgpr_comp_cnt
= 0; /* VGPR0 contains offsets 0, 1 */
1166 si_pm4_set_reg(pm4
, R_00B320_SPI_SHADER_PGM_LO_ES
, va
>> 8);
1167 si_pm4_set_reg(pm4
, R_00B324_SPI_SHADER_PGM_HI_ES
, va
>> 40);
1168 si_pm4_set_reg(pm4
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
,
1169 S_00B228_VGPRS((shader
->config
.num_vgprs
- 1) /
1170 (sscreen
->ge_wave_size
== 32 ? 8 : 4)) |
1171 S_00B228_FLOAT_MODE(shader
->config
.float_mode
) |
1172 S_00B228_DX10_CLAMP(1) |
1173 S_00B228_MEM_ORDERED(1) |
1174 S_00B228_WGP_MODE(1) |
1175 S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt
));
1176 si_pm4_set_reg(pm4
, R_00B22C_SPI_SHADER_PGM_RSRC2_GS
,
1177 S_00B22C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0) |
1178 S_00B22C_USER_SGPR(num_user_sgprs
) |
1179 S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt
) |
1180 S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs
>> 5) |
1181 S_00B22C_OC_LDS_EN(es_type
== PIPE_SHADER_TESS_EVAL
) |
1182 S_00B22C_LDS_SIZE(shader
->config
.lds_size
));
1184 nparams
= MAX2(shader
->info
.nr_param_exports
, 1);
1185 shader
->ctx_reg
.ngg
.spi_vs_out_config
=
1186 S_0286C4_VS_EXPORT_COUNT(nparams
- 1) |
1187 S_0286C4_NO_PC_EXPORT(shader
->info
.nr_param_exports
== 0);
1189 shader
->ctx_reg
.ngg
.spi_shader_idx_format
=
1190 S_028708_IDX0_EXPORT_FORMAT(V_028708_SPI_SHADER_1COMP
);
1191 shader
->ctx_reg
.ngg
.spi_shader_pos_format
=
1192 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP
) |
1193 S_02870C_POS1_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 1 ?
1194 V_02870C_SPI_SHADER_4COMP
:
1195 V_02870C_SPI_SHADER_NONE
) |
1196 S_02870C_POS2_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 2 ?
1197 V_02870C_SPI_SHADER_4COMP
:
1198 V_02870C_SPI_SHADER_NONE
) |
1199 S_02870C_POS3_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 3 ?
1200 V_02870C_SPI_SHADER_4COMP
:
1201 V_02870C_SPI_SHADER_NONE
);
1203 shader
->ctx_reg
.ngg
.vgt_primitiveid_en
=
1204 S_028A84_PRIMITIVEID_EN(es_enable_prim_id
) |
1205 S_028A84_NGG_DISABLE_PROVOK_REUSE(shader
->key
.mono
.u
.vs_export_prim_id
||
1206 gs_sel
->info
.writes_primid
);
1208 if (gs_type
== PIPE_SHADER_GEOMETRY
) {
1209 shader
->ctx_reg
.ngg
.vgt_esgs_ring_itemsize
= es_sel
->esgs_itemsize
/ 4;
1210 shader
->ctx_reg
.ngg
.vgt_gs_max_vert_out
= gs_sel
->gs_max_out_vertices
;
1212 shader
->ctx_reg
.ngg
.vgt_esgs_ring_itemsize
= 1;
1215 if (es_type
== PIPE_SHADER_TESS_EVAL
)
1216 si_set_tesseval_regs(sscreen
, es_sel
, pm4
);
1218 shader
->ctx_reg
.ngg
.vgt_gs_onchip_cntl
=
1219 S_028A44_ES_VERTS_PER_SUBGRP(shader
->ngg
.hw_max_esverts
) |
1220 S_028A44_GS_PRIMS_PER_SUBGRP(shader
->ngg
.max_gsprims
) |
1221 S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader
->ngg
.max_gsprims
* gs_num_invocations
);
1222 shader
->ctx_reg
.ngg
.ge_max_output_per_subgroup
=
1223 S_0287FC_MAX_VERTS_PER_SUBGROUP(shader
->ngg
.max_out_verts
);
1224 shader
->ctx_reg
.ngg
.ge_ngg_subgrp_cntl
=
1225 S_028B4C_PRIM_AMP_FACTOR(shader
->ngg
.prim_amp_factor
) |
1226 S_028B4C_THDS_PER_SUBGRP(0); /* for fast launch */
1227 shader
->ctx_reg
.ngg
.vgt_gs_instance_cnt
=
1228 S_028B90_CNT(gs_num_invocations
) |
1229 S_028B90_ENABLE(gs_num_invocations
> 1) |
1230 S_028B90_EN_MAX_VERT_OUT_PER_GS_INSTANCE(
1231 shader
->ngg
.max_vert_out_per_gs_instance
);
1233 /* Always output hw-generated edge flags and pass them via the prim
1234 * export to prevent drawing lines on internal edges of decomposed
1235 * primitives (such as quads) with polygon mode = lines. Only VS needs
1238 shader
->ctx_reg
.ngg
.pa_cl_ngg_cntl
=
1239 S_028838_INDEX_BUF_EDGE_FLAG_ENA(gs_type
== PIPE_SHADER_VERTEX
);
1240 shader
->pa_cl_vs_out_cntl
= si_get_vs_out_cntl(gs_sel
, true);
1243 S_03096C_PRIM_GRP_SIZE(shader
->ngg
.max_gsprims
) |
1244 S_03096C_VERT_GRP_SIZE(256) | /* 256 = disable vertex grouping */
1245 S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi
);
1247 /* Bug workaround for a possible hang with non-tessellation cases.
1248 * Tessellation always sets GE_CNTL.VERT_GRP_SIZE = 0
1250 * Requirement: GE_CNTL.VERT_GRP_SIZE = VGT_GS_ONCHIP_CNTL.ES_VERTS_PER_SUBGRP - 5
1252 if ((sscreen
->info
.family
== CHIP_NAVI10
||
1253 sscreen
->info
.family
== CHIP_NAVI12
||
1254 sscreen
->info
.family
== CHIP_NAVI14
) &&
1255 (es_type
== PIPE_SHADER_VERTEX
|| gs_type
== PIPE_SHADER_VERTEX
) && /* = no tess */
1256 shader
->ngg
.hw_max_esverts
!= 256) {
1257 shader
->ge_cntl
&= C_03096C_VERT_GRP_SIZE
;
1259 if (shader
->ngg
.hw_max_esverts
> 5) {
1261 S_03096C_VERT_GRP_SIZE(shader
->ngg
.hw_max_esverts
- 5);
1266 shader
->ctx_reg
.ngg
.pa_cl_vte_cntl
=
1267 S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
1269 shader
->ctx_reg
.ngg
.pa_cl_vte_cntl
=
1270 S_028818_VTX_W0_FMT(1) |
1271 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
1272 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
1273 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
1277 static void si_emit_shader_vs(struct si_context
*sctx
)
1279 struct si_shader
*shader
= sctx
->queued
.named
.vs
->shader
;
1280 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1285 radeon_opt_set_context_reg(sctx
, R_028A40_VGT_GS_MODE
,
1286 SI_TRACKED_VGT_GS_MODE
,
1287 shader
->ctx_reg
.vs
.vgt_gs_mode
);
1288 radeon_opt_set_context_reg(sctx
, R_028A84_VGT_PRIMITIVEID_EN
,
1289 SI_TRACKED_VGT_PRIMITIVEID_EN
,
1290 shader
->ctx_reg
.vs
.vgt_primitiveid_en
);
1292 if (sctx
->chip_class
<= GFX8
) {
1293 radeon_opt_set_context_reg(sctx
, R_028AB4_VGT_REUSE_OFF
,
1294 SI_TRACKED_VGT_REUSE_OFF
,
1295 shader
->ctx_reg
.vs
.vgt_reuse_off
);
1298 radeon_opt_set_context_reg(sctx
, R_0286C4_SPI_VS_OUT_CONFIG
,
1299 SI_TRACKED_SPI_VS_OUT_CONFIG
,
1300 shader
->ctx_reg
.vs
.spi_vs_out_config
);
1302 radeon_opt_set_context_reg(sctx
, R_02870C_SPI_SHADER_POS_FORMAT
,
1303 SI_TRACKED_SPI_SHADER_POS_FORMAT
,
1304 shader
->ctx_reg
.vs
.spi_shader_pos_format
);
1306 radeon_opt_set_context_reg(sctx
, R_028818_PA_CL_VTE_CNTL
,
1307 SI_TRACKED_PA_CL_VTE_CNTL
,
1308 shader
->ctx_reg
.vs
.pa_cl_vte_cntl
);
1310 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
1311 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
1312 SI_TRACKED_VGT_TF_PARAM
,
1313 shader
->vgt_tf_param
);
1315 if (shader
->vgt_vertex_reuse_block_cntl
)
1316 radeon_opt_set_context_reg(sctx
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
1317 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL
,
1318 shader
->vgt_vertex_reuse_block_cntl
);
1320 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
1321 sctx
->context_roll
= true;
1323 /* Required programming for tessellation. (legacy pipeline only) */
1324 if (sctx
->chip_class
== GFX10
&&
1325 shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
1326 radeon_opt_set_context_reg(sctx
, R_028A44_VGT_GS_ONCHIP_CNTL
,
1327 SI_TRACKED_VGT_GS_ONCHIP_CNTL
,
1328 S_028A44_ES_VERTS_PER_SUBGRP(250) |
1329 S_028A44_GS_PRIMS_PER_SUBGRP(126) |
1330 S_028A44_GS_INST_PRIMS_IN_SUBGRP(126));
1333 if (sctx
->chip_class
>= GFX10
) {
1334 radeon_opt_set_context_reg_rmw(sctx
, R_02881C_PA_CL_VS_OUT_CNTL
,
1335 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS
,
1336 shader
->pa_cl_vs_out_cntl
,
1337 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS_MASK
);
1342 * Compute the state for \p shader, which will run as a vertex shader on the
1345 * If \p gs is non-NULL, it points to the geometry shader for which this shader
1346 * is the copy shader.
1348 static void si_shader_vs(struct si_screen
*sscreen
, struct si_shader
*shader
,
1349 struct si_shader_selector
*gs
)
1351 const struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1352 struct si_pm4_state
*pm4
;
1353 unsigned num_user_sgprs
, vgpr_comp_cnt
;
1355 unsigned nparams
, oc_lds_en
;
1356 unsigned window_space
=
1357 info
->properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
];
1358 bool enable_prim_id
= shader
->key
.mono
.u
.vs_export_prim_id
|| info
->uses_primid
;
1360 pm4
= si_get_shader_pm4_state(shader
);
1364 pm4
->atom
.emit
= si_emit_shader_vs
;
1366 /* We always write VGT_GS_MODE in the VS state, because every switch
1367 * between different shader pipelines involving a different GS or no
1368 * GS at all involves a switch of the VS (different GS use different
1369 * copy shaders). On the other hand, when the API switches from a GS to
1370 * no GS and then back to the same GS used originally, the GS state is
1374 unsigned mode
= V_028A40_GS_OFF
;
1376 /* PrimID needs GS scenario A. */
1378 mode
= V_028A40_GS_SCENARIO_A
;
1380 shader
->ctx_reg
.vs
.vgt_gs_mode
= S_028A40_MODE(mode
);
1381 shader
->ctx_reg
.vs
.vgt_primitiveid_en
= enable_prim_id
;
1383 shader
->ctx_reg
.vs
.vgt_gs_mode
= ac_vgt_gs_mode(gs
->gs_max_out_vertices
,
1384 sscreen
->info
.chip_class
);
1385 shader
->ctx_reg
.vs
.vgt_primitiveid_en
= 0;
1388 if (sscreen
->info
.chip_class
<= GFX8
) {
1389 /* Reuse needs to be set off if we write oViewport. */
1390 shader
->ctx_reg
.vs
.vgt_reuse_off
=
1391 S_028AB4_REUSE_OFF(info
->writes_viewport_index
);
1394 va
= shader
->bo
->gpu_address
;
1395 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
1398 vgpr_comp_cnt
= 0; /* only VertexID is needed for GS-COPY. */
1399 num_user_sgprs
= SI_GSCOPY_NUM_USER_SGPR
;
1400 } else if (shader
->selector
->type
== PIPE_SHADER_VERTEX
) {
1401 vgpr_comp_cnt
= si_get_vs_vgpr_comp_cnt(sscreen
, shader
, enable_prim_id
);
1403 if (info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
]) {
1404 num_user_sgprs
= SI_SGPR_VS_BLIT_DATA
+
1405 info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
];
1407 num_user_sgprs
= si_get_num_vs_user_sgprs(SI_VS_NUM_USER_SGPR
);
1409 } else if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
1410 vgpr_comp_cnt
= enable_prim_id
? 3 : 2;
1411 num_user_sgprs
= SI_TES_NUM_USER_SGPR
;
1413 unreachable("invalid shader selector type");
1415 /* VS is required to export at least one param. */
1416 nparams
= MAX2(shader
->info
.nr_param_exports
, 1);
1417 shader
->ctx_reg
.vs
.spi_vs_out_config
= S_0286C4_VS_EXPORT_COUNT(nparams
- 1);
1419 if (sscreen
->info
.chip_class
>= GFX10
) {
1420 shader
->ctx_reg
.vs
.spi_vs_out_config
|=
1421 S_0286C4_NO_PC_EXPORT(shader
->info
.nr_param_exports
== 0);
1424 shader
->ctx_reg
.vs
.spi_shader_pos_format
=
1425 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP
) |
1426 S_02870C_POS1_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 1 ?
1427 V_02870C_SPI_SHADER_4COMP
:
1428 V_02870C_SPI_SHADER_NONE
) |
1429 S_02870C_POS2_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 2 ?
1430 V_02870C_SPI_SHADER_4COMP
:
1431 V_02870C_SPI_SHADER_NONE
) |
1432 S_02870C_POS3_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 3 ?
1433 V_02870C_SPI_SHADER_4COMP
:
1434 V_02870C_SPI_SHADER_NONE
);
1435 shader
->pa_cl_vs_out_cntl
= si_get_vs_out_cntl(shader
->selector
, false);
1437 oc_lds_en
= shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
? 1 : 0;
1439 si_pm4_set_reg(pm4
, R_00B120_SPI_SHADER_PGM_LO_VS
, va
>> 8);
1440 si_pm4_set_reg(pm4
, R_00B124_SPI_SHADER_PGM_HI_VS
, S_00B124_MEM_BASE(va
>> 40));
1442 uint32_t rsrc1
= S_00B128_VGPRS((shader
->config
.num_vgprs
- 1) /
1443 (sscreen
->ge_wave_size
== 32 ? 8 : 4)) |
1444 S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt
) |
1445 S_00B128_DX10_CLAMP(1) |
1446 S_00B128_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
1447 S_00B128_FLOAT_MODE(shader
->config
.float_mode
);
1448 uint32_t rsrc2
= S_00B12C_USER_SGPR(num_user_sgprs
) |
1449 S_00B12C_OC_LDS_EN(oc_lds_en
) |
1450 S_00B12C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
1452 if (sscreen
->info
.chip_class
<= GFX9
)
1453 rsrc1
|= S_00B128_SGPRS((shader
->config
.num_sgprs
- 1) / 8);
1455 if (!sscreen
->use_ngg_streamout
) {
1456 rsrc2
|= S_00B12C_SO_BASE0_EN(!!shader
->selector
->so
.stride
[0]) |
1457 S_00B12C_SO_BASE1_EN(!!shader
->selector
->so
.stride
[1]) |
1458 S_00B12C_SO_BASE2_EN(!!shader
->selector
->so
.stride
[2]) |
1459 S_00B12C_SO_BASE3_EN(!!shader
->selector
->so
.stride
[3]) |
1460 S_00B12C_SO_EN(!!shader
->selector
->so
.num_outputs
);
1463 si_pm4_set_reg(pm4
, R_00B128_SPI_SHADER_PGM_RSRC1_VS
, rsrc1
);
1464 si_pm4_set_reg(pm4
, R_00B12C_SPI_SHADER_PGM_RSRC2_VS
, rsrc2
);
1467 shader
->ctx_reg
.vs
.pa_cl_vte_cntl
=
1468 S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
1470 shader
->ctx_reg
.vs
.pa_cl_vte_cntl
=
1471 S_028818_VTX_W0_FMT(1) |
1472 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
1473 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
1474 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
1476 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
1477 si_set_tesseval_regs(sscreen
, shader
->selector
, pm4
);
1479 polaris_set_vgt_vertex_reuse(sscreen
, shader
->selector
, shader
, pm4
);
1482 static unsigned si_get_ps_num_interp(struct si_shader
*ps
)
1484 struct tgsi_shader_info
*info
= &ps
->selector
->info
;
1485 unsigned num_colors
= !!(info
->colors_read
& 0x0f) +
1486 !!(info
->colors_read
& 0xf0);
1487 unsigned num_interp
= ps
->selector
->info
.num_inputs
+
1488 (ps
->key
.part
.ps
.prolog
.color_two_side
? num_colors
: 0);
1490 assert(num_interp
<= 32);
1491 return MIN2(num_interp
, 32);
1494 static unsigned si_get_spi_shader_col_format(struct si_shader
*shader
)
1496 unsigned value
= shader
->key
.part
.ps
.epilog
.spi_shader_col_format
;
1497 unsigned i
, num_targets
= (util_last_bit(value
) + 3) / 4;
1499 /* If the i-th target format is set, all previous target formats must
1500 * be non-zero to avoid hangs.
1502 for (i
= 0; i
< num_targets
; i
++)
1503 if (!(value
& (0xf << (i
* 4))))
1504 value
|= V_028714_SPI_SHADER_32_R
<< (i
* 4);
1509 static void si_emit_shader_ps(struct si_context
*sctx
)
1511 struct si_shader
*shader
= sctx
->queued
.named
.ps
->shader
;
1512 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1517 /* R_0286CC_SPI_PS_INPUT_ENA, R_0286D0_SPI_PS_INPUT_ADDR*/
1518 radeon_opt_set_context_reg2(sctx
, R_0286CC_SPI_PS_INPUT_ENA
,
1519 SI_TRACKED_SPI_PS_INPUT_ENA
,
1520 shader
->ctx_reg
.ps
.spi_ps_input_ena
,
1521 shader
->ctx_reg
.ps
.spi_ps_input_addr
);
1523 radeon_opt_set_context_reg(sctx
, R_0286E0_SPI_BARYC_CNTL
,
1524 SI_TRACKED_SPI_BARYC_CNTL
,
1525 shader
->ctx_reg
.ps
.spi_baryc_cntl
);
1526 radeon_opt_set_context_reg(sctx
, R_0286D8_SPI_PS_IN_CONTROL
,
1527 SI_TRACKED_SPI_PS_IN_CONTROL
,
1528 shader
->ctx_reg
.ps
.spi_ps_in_control
);
1530 /* R_028710_SPI_SHADER_Z_FORMAT, R_028714_SPI_SHADER_COL_FORMAT */
1531 radeon_opt_set_context_reg2(sctx
, R_028710_SPI_SHADER_Z_FORMAT
,
1532 SI_TRACKED_SPI_SHADER_Z_FORMAT
,
1533 shader
->ctx_reg
.ps
.spi_shader_z_format
,
1534 shader
->ctx_reg
.ps
.spi_shader_col_format
);
1536 radeon_opt_set_context_reg(sctx
, R_02823C_CB_SHADER_MASK
,
1537 SI_TRACKED_CB_SHADER_MASK
,
1538 shader
->ctx_reg
.ps
.cb_shader_mask
);
1540 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
1541 sctx
->context_roll
= true;
1544 static void si_shader_ps(struct si_screen
*sscreen
, struct si_shader
*shader
)
1546 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1547 struct si_pm4_state
*pm4
;
1548 unsigned spi_ps_in_control
, spi_shader_col_format
, cb_shader_mask
;
1549 unsigned spi_baryc_cntl
= S_0286E0_FRONT_FACE_ALL_BITS(1);
1551 unsigned input_ena
= shader
->config
.spi_ps_input_ena
;
1553 /* we need to enable at least one of them, otherwise we hang the GPU */
1554 assert(G_0286CC_PERSP_SAMPLE_ENA(input_ena
) ||
1555 G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
1556 G_0286CC_PERSP_CENTROID_ENA(input_ena
) ||
1557 G_0286CC_PERSP_PULL_MODEL_ENA(input_ena
) ||
1558 G_0286CC_LINEAR_SAMPLE_ENA(input_ena
) ||
1559 G_0286CC_LINEAR_CENTER_ENA(input_ena
) ||
1560 G_0286CC_LINEAR_CENTROID_ENA(input_ena
) ||
1561 G_0286CC_LINE_STIPPLE_TEX_ENA(input_ena
));
1562 /* POS_W_FLOAT_ENA requires one of the perspective weights. */
1563 assert(!G_0286CC_POS_W_FLOAT_ENA(input_ena
) ||
1564 G_0286CC_PERSP_SAMPLE_ENA(input_ena
) ||
1565 G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
1566 G_0286CC_PERSP_CENTROID_ENA(input_ena
) ||
1567 G_0286CC_PERSP_PULL_MODEL_ENA(input_ena
));
1569 /* Validate interpolation optimization flags (read as implications). */
1570 assert(!shader
->key
.part
.ps
.prolog
.bc_optimize_for_persp
||
1571 (G_0286CC_PERSP_CENTER_ENA(input_ena
) &&
1572 G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
1573 assert(!shader
->key
.part
.ps
.prolog
.bc_optimize_for_linear
||
1574 (G_0286CC_LINEAR_CENTER_ENA(input_ena
) &&
1575 G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
1576 assert(!shader
->key
.part
.ps
.prolog
.force_persp_center_interp
||
1577 (!G_0286CC_PERSP_SAMPLE_ENA(input_ena
) &&
1578 !G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
1579 assert(!shader
->key
.part
.ps
.prolog
.force_linear_center_interp
||
1580 (!G_0286CC_LINEAR_SAMPLE_ENA(input_ena
) &&
1581 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
1582 assert(!shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
||
1583 (!G_0286CC_PERSP_CENTER_ENA(input_ena
) &&
1584 !G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
1585 assert(!shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
||
1586 (!G_0286CC_LINEAR_CENTER_ENA(input_ena
) &&
1587 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
1589 /* Validate cases when the optimizations are off (read as implications). */
1590 assert(shader
->key
.part
.ps
.prolog
.bc_optimize_for_persp
||
1591 !G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
1592 !G_0286CC_PERSP_CENTROID_ENA(input_ena
));
1593 assert(shader
->key
.part
.ps
.prolog
.bc_optimize_for_linear
||
1594 !G_0286CC_LINEAR_CENTER_ENA(input_ena
) ||
1595 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
));
1597 pm4
= si_get_shader_pm4_state(shader
);
1601 pm4
->atom
.emit
= si_emit_shader_ps
;
1603 /* SPI_BARYC_CNTL.POS_FLOAT_LOCATION
1605 * 0 -> Position = pixel center
1606 * 1 -> Position = pixel centroid
1607 * 2 -> Position = at sample position
1609 * From GLSL 4.5 specification, section 7.1:
1610 * "The variable gl_FragCoord is available as an input variable from
1611 * within fragment shaders and it holds the window relative coordinates
1612 * (x, y, z, 1/w) values for the fragment. If multi-sampling, this
1613 * value can be for any location within the pixel, or one of the
1614 * fragment samples. The use of centroid does not further restrict
1615 * this value to be inside the current primitive."
1617 * Meaning that centroid has no effect and we can return anything within
1618 * the pixel. Thus, return the value at sample position, because that's
1619 * the most accurate one shaders can get.
1621 spi_baryc_cntl
|= S_0286E0_POS_FLOAT_LOCATION(2);
1623 if (info
->properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
] ==
1624 TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)
1625 spi_baryc_cntl
|= S_0286E0_POS_FLOAT_ULC(1);
1627 spi_shader_col_format
= si_get_spi_shader_col_format(shader
);
1628 cb_shader_mask
= ac_get_cb_shader_mask(spi_shader_col_format
);
1630 /* Ensure that some export memory is always allocated, for two reasons:
1632 * 1) Correctness: The hardware ignores the EXEC mask if no export
1633 * memory is allocated, so KILL and alpha test do not work correctly
1635 * 2) Performance: Every shader needs at least a NULL export, even when
1636 * it writes no color/depth output. The NULL export instruction
1637 * stalls without this setting.
1639 * Don't add this to CB_SHADER_MASK.
1641 * GFX10 supports pixel shaders without exports by setting both
1642 * the color and Z formats to SPI_SHADER_ZERO. The hw will skip export
1643 * instructions if any are present.
1645 if ((sscreen
->info
.chip_class
<= GFX9
||
1647 shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
) &&
1648 !spi_shader_col_format
&&
1649 !info
->writes_z
&& !info
->writes_stencil
&& !info
->writes_samplemask
)
1650 spi_shader_col_format
= V_028714_SPI_SHADER_32_R
;
1652 shader
->ctx_reg
.ps
.spi_ps_input_ena
= input_ena
;
1653 shader
->ctx_reg
.ps
.spi_ps_input_addr
= shader
->config
.spi_ps_input_addr
;
1655 /* Set interpolation controls. */
1656 spi_ps_in_control
= S_0286D8_NUM_INTERP(si_get_ps_num_interp(shader
)) |
1657 S_0286D8_PS_W32_EN(sscreen
->ps_wave_size
== 32);
1659 shader
->ctx_reg
.ps
.spi_baryc_cntl
= spi_baryc_cntl
;
1660 shader
->ctx_reg
.ps
.spi_ps_in_control
= spi_ps_in_control
;
1661 shader
->ctx_reg
.ps
.spi_shader_z_format
=
1662 ac_get_spi_shader_z_format(info
->writes_z
,
1663 info
->writes_stencil
,
1664 info
->writes_samplemask
);
1665 shader
->ctx_reg
.ps
.spi_shader_col_format
= spi_shader_col_format
;
1666 shader
->ctx_reg
.ps
.cb_shader_mask
= cb_shader_mask
;
1668 va
= shader
->bo
->gpu_address
;
1669 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
1670 si_pm4_set_reg(pm4
, R_00B020_SPI_SHADER_PGM_LO_PS
, va
>> 8);
1671 si_pm4_set_reg(pm4
, R_00B024_SPI_SHADER_PGM_HI_PS
, S_00B024_MEM_BASE(va
>> 40));
1674 S_00B028_VGPRS((shader
->config
.num_vgprs
- 1) /
1675 (sscreen
->ps_wave_size
== 32 ? 8 : 4)) |
1676 S_00B028_DX10_CLAMP(1) |
1677 S_00B028_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
1678 S_00B028_FLOAT_MODE(shader
->config
.float_mode
);
1680 if (sscreen
->info
.chip_class
< GFX10
) {
1681 rsrc1
|= S_00B028_SGPRS((shader
->config
.num_sgprs
- 1) / 8);
1684 si_pm4_set_reg(pm4
, R_00B028_SPI_SHADER_PGM_RSRC1_PS
, rsrc1
);
1685 si_pm4_set_reg(pm4
, R_00B02C_SPI_SHADER_PGM_RSRC2_PS
,
1686 S_00B02C_EXTRA_LDS_SIZE(shader
->config
.lds_size
) |
1687 S_00B02C_USER_SGPR(SI_PS_NUM_USER_SGPR
) |
1688 S_00B32C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
1691 static void si_shader_init_pm4_state(struct si_screen
*sscreen
,
1692 struct si_shader
*shader
)
1694 switch (shader
->selector
->type
) {
1695 case PIPE_SHADER_VERTEX
:
1696 if (shader
->key
.as_ls
)
1697 si_shader_ls(sscreen
, shader
);
1698 else if (shader
->key
.as_es
)
1699 si_shader_es(sscreen
, shader
);
1700 else if (shader
->key
.as_ngg
)
1701 gfx10_shader_ngg(sscreen
, shader
);
1703 si_shader_vs(sscreen
, shader
, NULL
);
1705 case PIPE_SHADER_TESS_CTRL
:
1706 si_shader_hs(sscreen
, shader
);
1708 case PIPE_SHADER_TESS_EVAL
:
1709 if (shader
->key
.as_es
)
1710 si_shader_es(sscreen
, shader
);
1711 else if (shader
->key
.as_ngg
)
1712 gfx10_shader_ngg(sscreen
, shader
);
1714 si_shader_vs(sscreen
, shader
, NULL
);
1716 case PIPE_SHADER_GEOMETRY
:
1717 if (shader
->key
.as_ngg
)
1718 gfx10_shader_ngg(sscreen
, shader
);
1720 si_shader_gs(sscreen
, shader
);
1722 case PIPE_SHADER_FRAGMENT
:
1723 si_shader_ps(sscreen
, shader
);
1730 static unsigned si_get_alpha_test_func(struct si_context
*sctx
)
1732 /* Alpha-test should be disabled if colorbuffer 0 is integer. */
1733 return sctx
->queued
.named
.dsa
->alpha_func
;
1736 void si_shader_selector_key_vs(struct si_context
*sctx
,
1737 struct si_shader_selector
*vs
,
1738 struct si_shader_key
*key
,
1739 struct si_vs_prolog_bits
*prolog_key
)
1741 if (!sctx
->vertex_elements
||
1742 vs
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
])
1745 struct si_vertex_elements
*elts
= sctx
->vertex_elements
;
1747 prolog_key
->instance_divisor_is_one
= elts
->instance_divisor_is_one
;
1748 prolog_key
->instance_divisor_is_fetched
= elts
->instance_divisor_is_fetched
;
1749 prolog_key
->unpack_instance_id_from_vertex_id
=
1750 sctx
->prim_discard_cs_instancing
;
1752 /* Prefer a monolithic shader to allow scheduling divisions around
1754 if (prolog_key
->instance_divisor_is_fetched
)
1755 key
->opt
.prefer_mono
= 1;
1757 unsigned count
= MIN2(vs
->info
.num_inputs
, elts
->count
);
1758 unsigned count_mask
= (1 << count
) - 1;
1759 unsigned fix
= elts
->fix_fetch_always
& count_mask
;
1760 unsigned opencode
= elts
->fix_fetch_opencode
& count_mask
;
1762 if (sctx
->vertex_buffer_unaligned
& elts
->vb_alignment_check_mask
) {
1763 uint32_t mask
= elts
->fix_fetch_unaligned
& count_mask
;
1765 unsigned i
= u_bit_scan(&mask
);
1766 unsigned log_hw_load_size
= 1 + ((elts
->hw_load_is_dword
>> i
) & 1);
1767 unsigned vbidx
= elts
->vertex_buffer_index
[i
];
1768 struct pipe_vertex_buffer
*vb
= &sctx
->vertex_buffer
[vbidx
];
1769 unsigned align_mask
= (1 << log_hw_load_size
) - 1;
1770 if (vb
->buffer_offset
& align_mask
||
1771 vb
->stride
& align_mask
) {
1779 unsigned i
= u_bit_scan(&fix
);
1780 key
->mono
.vs_fix_fetch
[i
].bits
= elts
->fix_fetch
[i
];
1782 key
->mono
.vs_fetch_opencode
= opencode
;
1785 static void si_shader_selector_key_hw_vs(struct si_context
*sctx
,
1786 struct si_shader_selector
*vs
,
1787 struct si_shader_key
*key
)
1789 struct si_shader_selector
*ps
= sctx
->ps_shader
.cso
;
1791 key
->opt
.clip_disable
=
1792 sctx
->queued
.named
.rasterizer
->clip_plane_enable
== 0 &&
1793 (vs
->info
.clipdist_writemask
||
1794 vs
->info
.writes_clipvertex
) &&
1795 !vs
->info
.culldist_writemask
;
1797 /* Find out if PS is disabled. */
1798 bool ps_disabled
= true;
1800 bool ps_modifies_zs
= ps
->info
.uses_kill
||
1801 ps
->info
.writes_z
||
1802 ps
->info
.writes_stencil
||
1803 ps
->info
.writes_samplemask
||
1804 sctx
->queued
.named
.blend
->alpha_to_coverage
||
1805 si_get_alpha_test_func(sctx
) != PIPE_FUNC_ALWAYS
;
1806 unsigned ps_colormask
= si_get_total_colormask(sctx
);
1808 ps_disabled
= sctx
->queued
.named
.rasterizer
->rasterizer_discard
||
1811 !ps
->info
.writes_memory
);
1814 /* Find out which VS outputs aren't used by the PS. */
1815 uint64_t outputs_written
= vs
->outputs_written_before_ps
;
1816 uint64_t inputs_read
= 0;
1818 /* Ignore outputs that are not passed from VS to PS. */
1819 outputs_written
&= ~((1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_POSITION
, 0, true)) |
1820 (1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_PSIZE
, 0, true)) |
1821 (1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_CLIPVERTEX
, 0, true)));
1824 inputs_read
= ps
->inputs_read
;
1827 uint64_t linked
= outputs_written
& inputs_read
;
1829 key
->opt
.kill_outputs
= ~linked
& outputs_written
;
1832 /* Compute the key for the hw shader variant */
1833 static inline void si_shader_selector_key(struct pipe_context
*ctx
,
1834 struct si_shader_selector
*sel
,
1835 union si_vgt_stages_key stages_key
,
1836 struct si_shader_key
*key
)
1838 struct si_context
*sctx
= (struct si_context
*)ctx
;
1840 memset(key
, 0, sizeof(*key
));
1842 switch (sel
->type
) {
1843 case PIPE_SHADER_VERTEX
:
1844 si_shader_selector_key_vs(sctx
, sel
, key
, &key
->part
.vs
.prolog
);
1846 if (sctx
->tes_shader
.cso
)
1848 else if (sctx
->gs_shader
.cso
) {
1850 key
->as_ngg
= stages_key
.u
.ngg
;
1852 key
->as_ngg
= stages_key
.u
.ngg
;
1853 si_shader_selector_key_hw_vs(sctx
, sel
, key
);
1855 if (sctx
->ps_shader
.cso
&& sctx
->ps_shader
.cso
->info
.uses_primid
)
1856 key
->mono
.u
.vs_export_prim_id
= 1;
1859 case PIPE_SHADER_TESS_CTRL
:
1860 if (sctx
->chip_class
>= GFX9
) {
1861 si_shader_selector_key_vs(sctx
, sctx
->vs_shader
.cso
,
1862 key
, &key
->part
.tcs
.ls_prolog
);
1863 key
->part
.tcs
.ls
= sctx
->vs_shader
.cso
;
1865 /* When the LS VGPR fix is needed, monolithic shaders
1867 * - avoid initializing EXEC in both the LS prolog
1868 * and the LS main part when !vs_needs_prolog
1869 * - remove the fixup for unused input VGPRs
1871 key
->part
.tcs
.ls_prolog
.ls_vgpr_fix
= sctx
->ls_vgpr_fix
;
1873 /* The LS output / HS input layout can be communicated
1874 * directly instead of via user SGPRs for merged LS-HS.
1875 * The LS VGPR fix prefers this too.
1877 key
->opt
.prefer_mono
= 1;
1880 key
->part
.tcs
.epilog
.prim_mode
=
1881 sctx
->tes_shader
.cso
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
1882 key
->part
.tcs
.epilog
.invoc0_tess_factors_are_def
=
1883 sel
->tcs_info
.tessfactors_are_def_in_all_invocs
;
1884 key
->part
.tcs
.epilog
.tes_reads_tess_factors
=
1885 sctx
->tes_shader
.cso
->info
.reads_tess_factors
;
1887 if (sel
== sctx
->fixed_func_tcs_shader
.cso
)
1888 key
->mono
.u
.ff_tcs_inputs_to_copy
= sctx
->vs_shader
.cso
->outputs_written
;
1890 case PIPE_SHADER_TESS_EVAL
:
1891 key
->as_ngg
= stages_key
.u
.ngg
;
1893 if (sctx
->gs_shader
.cso
)
1896 si_shader_selector_key_hw_vs(sctx
, sel
, key
);
1898 if (sctx
->ps_shader
.cso
&& sctx
->ps_shader
.cso
->info
.uses_primid
)
1899 key
->mono
.u
.vs_export_prim_id
= 1;
1902 case PIPE_SHADER_GEOMETRY
:
1903 if (sctx
->chip_class
>= GFX9
) {
1904 if (sctx
->tes_shader
.cso
) {
1905 key
->part
.gs
.es
= sctx
->tes_shader
.cso
;
1907 si_shader_selector_key_vs(sctx
, sctx
->vs_shader
.cso
,
1908 key
, &key
->part
.gs
.vs_prolog
);
1909 key
->part
.gs
.es
= sctx
->vs_shader
.cso
;
1910 key
->part
.gs
.prolog
.gfx9_prev_is_vs
= 1;
1913 key
->as_ngg
= stages_key
.u
.ngg
;
1915 /* Merged ES-GS can have unbalanced wave usage.
1917 * ES threads are per-vertex, while GS threads are
1918 * per-primitive. So without any amplification, there
1919 * are fewer GS threads than ES threads, which can result
1920 * in empty (no-op) GS waves. With too much amplification,
1921 * there are more GS threads than ES threads, which
1922 * can result in empty (no-op) ES waves.
1924 * Non-monolithic shaders are implemented by setting EXEC
1925 * at the beginning of shader parts, and don't jump to
1926 * the end if EXEC is 0.
1928 * Monolithic shaders use conditional blocks, so they can
1929 * jump and skip empty waves of ES or GS. So set this to
1930 * always use optimized variants, which are monolithic.
1932 key
->opt
.prefer_mono
= 1;
1934 key
->part
.gs
.prolog
.tri_strip_adj_fix
= sctx
->gs_tri_strip_adj_fix
;
1936 case PIPE_SHADER_FRAGMENT
: {
1937 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
1938 struct si_state_blend
*blend
= sctx
->queued
.named
.blend
;
1940 if (sel
->info
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
] &&
1941 sel
->info
.colors_written
== 0x1)
1942 key
->part
.ps
.epilog
.last_cbuf
= MAX2(sctx
->framebuffer
.state
.nr_cbufs
, 1) - 1;
1944 /* Select the shader color format based on whether
1945 * blending or alpha are needed.
1947 key
->part
.ps
.epilog
.spi_shader_col_format
=
1948 (blend
->blend_enable_4bit
& blend
->need_src_alpha_4bit
&
1949 sctx
->framebuffer
.spi_shader_col_format_blend_alpha
) |
1950 (blend
->blend_enable_4bit
& ~blend
->need_src_alpha_4bit
&
1951 sctx
->framebuffer
.spi_shader_col_format_blend
) |
1952 (~blend
->blend_enable_4bit
& blend
->need_src_alpha_4bit
&
1953 sctx
->framebuffer
.spi_shader_col_format_alpha
) |
1954 (~blend
->blend_enable_4bit
& ~blend
->need_src_alpha_4bit
&
1955 sctx
->framebuffer
.spi_shader_col_format
);
1956 key
->part
.ps
.epilog
.spi_shader_col_format
&= blend
->cb_target_enabled_4bit
;
1958 /* The output for dual source blending should have
1959 * the same format as the first output.
1961 if (blend
->dual_src_blend
) {
1962 key
->part
.ps
.epilog
.spi_shader_col_format
|=
1963 (key
->part
.ps
.epilog
.spi_shader_col_format
& 0xf) << 4;
1966 /* If alpha-to-coverage is enabled, we have to export alpha
1967 * even if there is no color buffer.
1969 if (!(key
->part
.ps
.epilog
.spi_shader_col_format
& 0xf) &&
1970 blend
->alpha_to_coverage
)
1971 key
->part
.ps
.epilog
.spi_shader_col_format
|= V_028710_SPI_SHADER_32_AR
;
1973 /* On GFX6 and GFX7 except Hawaii, the CB doesn't clamp outputs
1974 * to the range supported by the type if a channel has less
1975 * than 16 bits and the export format is 16_ABGR.
1977 if (sctx
->chip_class
<= GFX7
&& sctx
->family
!= CHIP_HAWAII
) {
1978 key
->part
.ps
.epilog
.color_is_int8
= sctx
->framebuffer
.color_is_int8
;
1979 key
->part
.ps
.epilog
.color_is_int10
= sctx
->framebuffer
.color_is_int10
;
1982 /* Disable unwritten outputs (if WRITE_ALL_CBUFS isn't enabled). */
1983 if (!key
->part
.ps
.epilog
.last_cbuf
) {
1984 key
->part
.ps
.epilog
.spi_shader_col_format
&= sel
->colors_written_4bit
;
1985 key
->part
.ps
.epilog
.color_is_int8
&= sel
->info
.colors_written
;
1986 key
->part
.ps
.epilog
.color_is_int10
&= sel
->info
.colors_written
;
1989 bool is_poly
= !util_prim_is_points_or_lines(sctx
->current_rast_prim
);
1990 bool is_line
= util_prim_is_lines(sctx
->current_rast_prim
);
1992 key
->part
.ps
.prolog
.color_two_side
= rs
->two_side
&& sel
->info
.colors_read
;
1993 key
->part
.ps
.prolog
.flatshade_colors
= rs
->flatshade
&& sel
->info
.colors_read
;
1995 key
->part
.ps
.epilog
.alpha_to_one
= blend
->alpha_to_one
&&
1996 rs
->multisample_enable
;
1998 key
->part
.ps
.prolog
.poly_stipple
= rs
->poly_stipple_enable
&& is_poly
;
1999 key
->part
.ps
.epilog
.poly_line_smoothing
= ((is_poly
&& rs
->poly_smooth
) ||
2000 (is_line
&& rs
->line_smooth
)) &&
2001 sctx
->framebuffer
.nr_samples
<= 1;
2002 key
->part
.ps
.epilog
.clamp_color
= rs
->clamp_fragment_color
;
2004 if (sctx
->ps_iter_samples
> 1 &&
2005 sel
->info
.reads_samplemask
) {
2006 key
->part
.ps
.prolog
.samplemask_log_ps_iter
=
2007 util_logbase2(sctx
->ps_iter_samples
);
2010 if (rs
->force_persample_interp
&&
2011 rs
->multisample_enable
&&
2012 sctx
->framebuffer
.nr_samples
> 1 &&
2013 sctx
->ps_iter_samples
> 1) {
2014 key
->part
.ps
.prolog
.force_persp_sample_interp
=
2015 sel
->info
.uses_persp_center
||
2016 sel
->info
.uses_persp_centroid
;
2018 key
->part
.ps
.prolog
.force_linear_sample_interp
=
2019 sel
->info
.uses_linear_center
||
2020 sel
->info
.uses_linear_centroid
;
2021 } else if (rs
->multisample_enable
&&
2022 sctx
->framebuffer
.nr_samples
> 1) {
2023 key
->part
.ps
.prolog
.bc_optimize_for_persp
=
2024 sel
->info
.uses_persp_center
&&
2025 sel
->info
.uses_persp_centroid
;
2026 key
->part
.ps
.prolog
.bc_optimize_for_linear
=
2027 sel
->info
.uses_linear_center
&&
2028 sel
->info
.uses_linear_centroid
;
2030 /* Make sure SPI doesn't compute more than 1 pair
2031 * of (i,j), which is the optimization here. */
2032 key
->part
.ps
.prolog
.force_persp_center_interp
=
2033 sel
->info
.uses_persp_center
+
2034 sel
->info
.uses_persp_centroid
+
2035 sel
->info
.uses_persp_sample
> 1;
2037 key
->part
.ps
.prolog
.force_linear_center_interp
=
2038 sel
->info
.uses_linear_center
+
2039 sel
->info
.uses_linear_centroid
+
2040 sel
->info
.uses_linear_sample
> 1;
2042 if (sel
->info
.uses_persp_opcode_interp_sample
||
2043 sel
->info
.uses_linear_opcode_interp_sample
)
2044 key
->mono
.u
.ps
.interpolate_at_sample_force_center
= 1;
2047 key
->part
.ps
.epilog
.alpha_func
= si_get_alpha_test_func(sctx
);
2049 /* ps_uses_fbfetch is true only if the color buffer is bound. */
2050 if (sctx
->ps_uses_fbfetch
&& !sctx
->blitter
->running
) {
2051 struct pipe_surface
*cb0
= sctx
->framebuffer
.state
.cbufs
[0];
2052 struct pipe_resource
*tex
= cb0
->texture
;
2054 /* 1D textures are allocated and used as 2D on GFX9. */
2055 key
->mono
.u
.ps
.fbfetch_msaa
= sctx
->framebuffer
.nr_samples
> 1;
2056 key
->mono
.u
.ps
.fbfetch_is_1D
= sctx
->chip_class
!= GFX9
&&
2057 (tex
->target
== PIPE_TEXTURE_1D
||
2058 tex
->target
== PIPE_TEXTURE_1D_ARRAY
);
2059 key
->mono
.u
.ps
.fbfetch_layered
= tex
->target
== PIPE_TEXTURE_1D_ARRAY
||
2060 tex
->target
== PIPE_TEXTURE_2D_ARRAY
||
2061 tex
->target
== PIPE_TEXTURE_CUBE
||
2062 tex
->target
== PIPE_TEXTURE_CUBE_ARRAY
||
2063 tex
->target
== PIPE_TEXTURE_3D
;
2071 if (unlikely(sctx
->screen
->debug_flags
& DBG(NO_OPT_VARIANT
)))
2072 memset(&key
->opt
, 0, sizeof(key
->opt
));
2075 static void si_build_shader_variant(struct si_shader
*shader
,
2079 struct si_shader_selector
*sel
= shader
->selector
;
2080 struct si_screen
*sscreen
= sel
->screen
;
2081 struct ac_llvm_compiler
*compiler
;
2082 struct pipe_debug_callback
*debug
= &shader
->compiler_ctx_state
.debug
;
2084 if (thread_index
>= 0) {
2086 assert(thread_index
< ARRAY_SIZE(sscreen
->compiler_lowp
));
2087 compiler
= &sscreen
->compiler_lowp
[thread_index
];
2089 assert(thread_index
< ARRAY_SIZE(sscreen
->compiler
));
2090 compiler
= &sscreen
->compiler
[thread_index
];
2095 assert(!low_priority
);
2096 compiler
= shader
->compiler_ctx_state
.compiler
;
2099 if (!compiler
->passes
)
2100 si_init_compiler(sscreen
, compiler
);
2102 if (unlikely(!si_shader_create(sscreen
, compiler
, shader
, debug
))) {
2103 PRINT_ERR("Failed to build shader variant (type=%u)\n",
2105 shader
->compilation_failed
= true;
2109 if (shader
->compiler_ctx_state
.is_debug_context
) {
2110 FILE *f
= open_memstream(&shader
->shader_log
,
2111 &shader
->shader_log_size
);
2113 si_shader_dump(sscreen
, shader
, NULL
, f
, false);
2118 si_shader_init_pm4_state(sscreen
, shader
);
2121 static void si_build_shader_variant_low_priority(void *job
, int thread_index
)
2123 struct si_shader
*shader
= (struct si_shader
*)job
;
2125 assert(thread_index
>= 0);
2127 si_build_shader_variant(shader
, thread_index
, true);
2130 static const struct si_shader_key zeroed
;
2132 static bool si_check_missing_main_part(struct si_screen
*sscreen
,
2133 struct si_shader_selector
*sel
,
2134 struct si_compiler_ctx_state
*compiler_state
,
2135 struct si_shader_key
*key
)
2137 struct si_shader
**mainp
= si_get_main_shader_part(sel
, key
);
2140 struct si_shader
*main_part
= CALLOC_STRUCT(si_shader
);
2145 /* We can leave the fence as permanently signaled because the
2146 * main part becomes visible globally only after it has been
2148 util_queue_fence_init(&main_part
->ready
);
2150 main_part
->selector
= sel
;
2151 main_part
->key
.as_es
= key
->as_es
;
2152 main_part
->key
.as_ls
= key
->as_ls
;
2153 main_part
->key
.as_ngg
= key
->as_ngg
;
2154 main_part
->is_monolithic
= false;
2156 if (si_compile_tgsi_shader(sscreen
, compiler_state
->compiler
,
2157 main_part
, &compiler_state
->debug
) != 0) {
2167 * Select a shader variant according to the shader key.
2169 * \param optimized_or_none If the key describes an optimized shader variant and
2170 * the compilation isn't finished, don't select any
2171 * shader and return an error.
2173 int si_shader_select_with_key(struct si_screen
*sscreen
,
2174 struct si_shader_ctx_state
*state
,
2175 struct si_compiler_ctx_state
*compiler_state
,
2176 struct si_shader_key
*key
,
2178 bool optimized_or_none
)
2180 struct si_shader_selector
*sel
= state
->cso
;
2181 struct si_shader_selector
*previous_stage_sel
= NULL
;
2182 struct si_shader
*current
= state
->current
;
2183 struct si_shader
*iter
, *shader
= NULL
;
2186 /* Check if we don't need to change anything.
2187 * This path is also used for most shaders that don't need multiple
2188 * variants, it will cost just a computation of the key and this
2190 if (likely(current
&&
2191 memcmp(¤t
->key
, key
, sizeof(*key
)) == 0)) {
2192 if (unlikely(!util_queue_fence_is_signalled(¤t
->ready
))) {
2193 if (current
->is_optimized
) {
2194 if (optimized_or_none
)
2197 memset(&key
->opt
, 0, sizeof(key
->opt
));
2198 goto current_not_ready
;
2201 util_queue_fence_wait(¤t
->ready
);
2204 return current
->compilation_failed
? -1 : 0;
2208 /* This must be done before the mutex is locked, because async GS
2209 * compilation calls this function too, and therefore must enter
2212 * Only wait if we are in a draw call. Don't wait if we are
2213 * in a compiler thread.
2215 if (thread_index
< 0)
2216 util_queue_fence_wait(&sel
->ready
);
2218 simple_mtx_lock(&sel
->mutex
);
2220 /* Find the shader variant. */
2221 for (iter
= sel
->first_variant
; iter
; iter
= iter
->next_variant
) {
2222 /* Don't check the "current" shader. We checked it above. */
2223 if (current
!= iter
&&
2224 memcmp(&iter
->key
, key
, sizeof(*key
)) == 0) {
2225 simple_mtx_unlock(&sel
->mutex
);
2227 if (unlikely(!util_queue_fence_is_signalled(&iter
->ready
))) {
2228 /* If it's an optimized shader and its compilation has
2229 * been started but isn't done, use the unoptimized
2230 * shader so as not to cause a stall due to compilation.
2232 if (iter
->is_optimized
) {
2233 if (optimized_or_none
)
2235 memset(&key
->opt
, 0, sizeof(key
->opt
));
2239 util_queue_fence_wait(&iter
->ready
);
2242 if (iter
->compilation_failed
) {
2243 return -1; /* skip the draw call */
2246 state
->current
= iter
;
2251 /* Build a new shader. */
2252 shader
= CALLOC_STRUCT(si_shader
);
2254 simple_mtx_unlock(&sel
->mutex
);
2258 util_queue_fence_init(&shader
->ready
);
2260 shader
->selector
= sel
;
2262 shader
->compiler_ctx_state
= *compiler_state
;
2264 /* If this is a merged shader, get the first shader's selector. */
2265 if (sscreen
->info
.chip_class
>= GFX9
) {
2266 if (sel
->type
== PIPE_SHADER_TESS_CTRL
)
2267 previous_stage_sel
= key
->part
.tcs
.ls
;
2268 else if (sel
->type
== PIPE_SHADER_GEOMETRY
)
2269 previous_stage_sel
= key
->part
.gs
.es
;
2271 /* We need to wait for the previous shader. */
2272 if (previous_stage_sel
&& thread_index
< 0)
2273 util_queue_fence_wait(&previous_stage_sel
->ready
);
2276 bool is_pure_monolithic
=
2277 sscreen
->use_monolithic_shaders
||
2278 memcmp(&key
->mono
, &zeroed
.mono
, sizeof(key
->mono
)) != 0;
2280 /* Compile the main shader part if it doesn't exist. This can happen
2281 * if the initial guess was wrong.
2283 * The prim discard CS doesn't need the main shader part.
2285 if (!is_pure_monolithic
&&
2286 !key
->opt
.vs_as_prim_discard_cs
) {
2289 /* Make sure the main shader part is present. This is needed
2290 * for shaders that can be compiled as VS, LS, or ES, and only
2291 * one of them is compiled at creation.
2293 * It is also needed for GS, which can be compiled as non-NGG
2296 * For merged shaders, check that the starting shader's main
2299 if (previous_stage_sel
) {
2300 struct si_shader_key shader1_key
= zeroed
;
2302 if (sel
->type
== PIPE_SHADER_TESS_CTRL
) {
2303 shader1_key
.as_ls
= 1;
2304 } else if (sel
->type
== PIPE_SHADER_GEOMETRY
) {
2305 shader1_key
.as_es
= 1;
2306 shader1_key
.as_ngg
= key
->as_ngg
; /* for Wave32 vs Wave64 */
2311 simple_mtx_lock(&previous_stage_sel
->mutex
);
2312 ok
= si_check_missing_main_part(sscreen
,
2314 compiler_state
, &shader1_key
);
2315 simple_mtx_unlock(&previous_stage_sel
->mutex
);
2319 ok
= si_check_missing_main_part(sscreen
, sel
,
2320 compiler_state
, key
);
2325 simple_mtx_unlock(&sel
->mutex
);
2326 return -ENOMEM
; /* skip the draw call */
2330 /* Keep the reference to the 1st shader of merged shaders, so that
2331 * Gallium can't destroy it before we destroy the 2nd shader.
2333 * Set sctx = NULL, because it's unused if we're not releasing
2334 * the shader, and we don't have any sctx here.
2336 si_shader_selector_reference(NULL
, &shader
->previous_stage_sel
,
2337 previous_stage_sel
);
2339 /* Monolithic-only shaders don't make a distinction between optimized
2340 * and unoptimized. */
2341 shader
->is_monolithic
=
2342 is_pure_monolithic
||
2343 memcmp(&key
->opt
, &zeroed
.opt
, sizeof(key
->opt
)) != 0;
2345 /* The prim discard CS is always optimized. */
2346 shader
->is_optimized
=
2347 (!is_pure_monolithic
|| key
->opt
.vs_as_prim_discard_cs
) &&
2348 memcmp(&key
->opt
, &zeroed
.opt
, sizeof(key
->opt
)) != 0;
2350 /* If it's an optimized shader, compile it asynchronously. */
2351 if (shader
->is_optimized
&& thread_index
< 0) {
2352 /* Compile it asynchronously. */
2353 util_queue_add_job(&sscreen
->shader_compiler_queue_low_priority
,
2354 shader
, &shader
->ready
,
2355 si_build_shader_variant_low_priority
, NULL
,
2358 /* Add only after the ready fence was reset, to guard against a
2359 * race with si_bind_XX_shader. */
2360 if (!sel
->last_variant
) {
2361 sel
->first_variant
= shader
;
2362 sel
->last_variant
= shader
;
2364 sel
->last_variant
->next_variant
= shader
;
2365 sel
->last_variant
= shader
;
2368 /* Use the default (unoptimized) shader for now. */
2369 memset(&key
->opt
, 0, sizeof(key
->opt
));
2370 simple_mtx_unlock(&sel
->mutex
);
2372 if (sscreen
->options
.sync_compile
)
2373 util_queue_fence_wait(&shader
->ready
);
2375 if (optimized_or_none
)
2380 /* Reset the fence before adding to the variant list. */
2381 util_queue_fence_reset(&shader
->ready
);
2383 if (!sel
->last_variant
) {
2384 sel
->first_variant
= shader
;
2385 sel
->last_variant
= shader
;
2387 sel
->last_variant
->next_variant
= shader
;
2388 sel
->last_variant
= shader
;
2391 simple_mtx_unlock(&sel
->mutex
);
2393 assert(!shader
->is_optimized
);
2394 si_build_shader_variant(shader
, thread_index
, false);
2396 util_queue_fence_signal(&shader
->ready
);
2398 if (!shader
->compilation_failed
)
2399 state
->current
= shader
;
2401 return shader
->compilation_failed
? -1 : 0;
2404 static int si_shader_select(struct pipe_context
*ctx
,
2405 struct si_shader_ctx_state
*state
,
2406 union si_vgt_stages_key stages_key
,
2407 struct si_compiler_ctx_state
*compiler_state
)
2409 struct si_context
*sctx
= (struct si_context
*)ctx
;
2410 struct si_shader_key key
;
2412 si_shader_selector_key(ctx
, state
->cso
, stages_key
, &key
);
2413 return si_shader_select_with_key(sctx
->screen
, state
, compiler_state
,
2417 static void si_parse_next_shader_property(const struct tgsi_shader_info
*info
,
2419 struct si_shader_key
*key
)
2421 unsigned next_shader
= info
->properties
[TGSI_PROPERTY_NEXT_SHADER
];
2423 switch (info
->processor
) {
2424 case PIPE_SHADER_VERTEX
:
2425 switch (next_shader
) {
2426 case PIPE_SHADER_GEOMETRY
:
2429 case PIPE_SHADER_TESS_CTRL
:
2430 case PIPE_SHADER_TESS_EVAL
:
2434 /* If POSITION isn't written, it can only be a HW VS
2435 * if streamout is used. If streamout isn't used,
2436 * assume that it's a HW LS. (the next shader is TCS)
2437 * This heuristic is needed for separate shader objects.
2439 if (!info
->writes_position
&& !streamout
)
2444 case PIPE_SHADER_TESS_EVAL
:
2445 if (next_shader
== PIPE_SHADER_GEOMETRY
||
2446 !info
->writes_position
)
2453 * Compile the main shader part or the monolithic shader as part of
2454 * si_shader_selector initialization. Since it can be done asynchronously,
2455 * there is no way to report compile failures to applications.
2457 static void si_init_shader_selector_async(void *job
, int thread_index
)
2459 struct si_shader_selector
*sel
= (struct si_shader_selector
*)job
;
2460 struct si_screen
*sscreen
= sel
->screen
;
2461 struct ac_llvm_compiler
*compiler
;
2462 struct pipe_debug_callback
*debug
= &sel
->compiler_ctx_state
.debug
;
2464 assert(!debug
->debug_message
|| debug
->async
);
2465 assert(thread_index
>= 0);
2466 assert(thread_index
< ARRAY_SIZE(sscreen
->compiler
));
2467 compiler
= &sscreen
->compiler
[thread_index
];
2469 if (!compiler
->passes
)
2470 si_init_compiler(sscreen
, compiler
);
2472 /* Serialize NIR to save memory. Monolithic shader variants
2473 * have to deserialize NIR before compilation.
2480 /* true = remove optional debugging data to increase
2481 * the likehood of getting more shader cache hits.
2482 * It also drops variable names, so we'll save more memory.
2484 nir_serialize(&blob
, sel
->nir
, true);
2485 blob_finish_get_buffer(&blob
, &sel
->nir_binary
, &size
);
2486 sel
->nir_size
= size
;
2489 /* Compile the main shader part for use with a prolog and/or epilog.
2490 * If this fails, the driver will try to compile a monolithic shader
2493 if (!sscreen
->use_monolithic_shaders
) {
2494 struct si_shader
*shader
= CALLOC_STRUCT(si_shader
);
2495 unsigned char ir_sha1_cache_key
[20];
2498 fprintf(stderr
, "radeonsi: can't allocate a main shader part\n");
2502 /* We can leave the fence signaled because use of the default
2503 * main part is guarded by the selector's ready fence. */
2504 util_queue_fence_init(&shader
->ready
);
2506 shader
->selector
= sel
;
2507 shader
->is_monolithic
= false;
2508 si_parse_next_shader_property(&sel
->info
,
2509 sel
->so
.num_outputs
!= 0,
2512 if (sscreen
->use_ngg
&&
2513 (!sel
->so
.num_outputs
|| sscreen
->use_ngg_streamout
) &&
2514 ((sel
->type
== PIPE_SHADER_VERTEX
&& !shader
->key
.as_ls
) ||
2515 sel
->type
== PIPE_SHADER_TESS_EVAL
||
2516 sel
->type
== PIPE_SHADER_GEOMETRY
))
2517 shader
->key
.as_ngg
= 1;
2519 if (sel
->tokens
|| sel
->nir
) {
2520 si_get_ir_cache_key(sel
, shader
->key
.as_ngg
,
2521 shader
->key
.as_es
, ir_sha1_cache_key
);
2524 /* Try to load the shader from the shader cache. */
2525 simple_mtx_lock(&sscreen
->shader_cache_mutex
);
2527 if (si_shader_cache_load_shader(sscreen
, ir_sha1_cache_key
, shader
)) {
2528 simple_mtx_unlock(&sscreen
->shader_cache_mutex
);
2529 si_shader_dump_stats_for_shader_db(sscreen
, shader
, debug
);
2531 simple_mtx_unlock(&sscreen
->shader_cache_mutex
);
2533 /* Compile the shader if it hasn't been loaded from the cache. */
2534 if (si_compile_tgsi_shader(sscreen
, compiler
, shader
,
2537 fprintf(stderr
, "radeonsi: can't compile a main shader part\n");
2541 simple_mtx_lock(&sscreen
->shader_cache_mutex
);
2542 si_shader_cache_insert_shader(sscreen
, ir_sha1_cache_key
,
2544 simple_mtx_unlock(&sscreen
->shader_cache_mutex
);
2547 *si_get_main_shader_part(sel
, &shader
->key
) = shader
;
2549 /* Unset "outputs_written" flags for outputs converted to
2550 * DEFAULT_VAL, so that later inter-shader optimizations don't
2551 * try to eliminate outputs that don't exist in the final
2554 * This is only done if non-monolithic shaders are enabled.
2556 if ((sel
->type
== PIPE_SHADER_VERTEX
||
2557 sel
->type
== PIPE_SHADER_TESS_EVAL
) &&
2558 !shader
->key
.as_ls
&&
2559 !shader
->key
.as_es
) {
2562 for (i
= 0; i
< sel
->info
.num_outputs
; i
++) {
2563 unsigned offset
= shader
->info
.vs_output_param_offset
[i
];
2565 if (offset
<= AC_EXP_PARAM_OFFSET_31
)
2568 unsigned name
= sel
->info
.output_semantic_name
[i
];
2569 unsigned index
= sel
->info
.output_semantic_index
[i
];
2573 case TGSI_SEMANTIC_GENERIC
:
2574 /* don't process indices the function can't handle */
2575 if (index
>= SI_MAX_IO_GENERIC
)
2579 id
= si_shader_io_get_unique_index(name
, index
, true);
2580 sel
->outputs_written_before_ps
&= ~(1ull << id
);
2582 case TGSI_SEMANTIC_POSITION
: /* ignore these */
2583 case TGSI_SEMANTIC_PSIZE
:
2584 case TGSI_SEMANTIC_CLIPVERTEX
:
2585 case TGSI_SEMANTIC_EDGEFLAG
:
2592 /* The GS copy shader is always pre-compiled. */
2593 if (sel
->type
== PIPE_SHADER_GEOMETRY
&&
2594 (!sscreen
->use_ngg
||
2595 !sscreen
->use_ngg_streamout
|| /* also for PRIMITIVES_GENERATED */
2596 sel
->tess_turns_off_ngg
)) {
2597 sel
->gs_copy_shader
= si_generate_gs_copy_shader(sscreen
, compiler
, sel
, debug
);
2598 if (!sel
->gs_copy_shader
) {
2599 fprintf(stderr
, "radeonsi: can't create GS copy shader\n");
2603 si_shader_vs(sscreen
, sel
->gs_copy_shader
, sel
);
2606 /* Free NIR. We only keep serialized NIR after this point. */
2608 ralloc_free(sel
->nir
);
2613 void si_schedule_initial_compile(struct si_context
*sctx
, unsigned processor
,
2614 struct util_queue_fence
*ready_fence
,
2615 struct si_compiler_ctx_state
*compiler_ctx_state
,
2616 void *job
, util_queue_execute_func execute
)
2618 util_queue_fence_init(ready_fence
);
2620 struct util_async_debug_callback async_debug
;
2622 (sctx
->debug
.debug_message
&& !sctx
->debug
.async
) ||
2624 si_can_dump_shader(sctx
->screen
, processor
);
2627 u_async_debug_init(&async_debug
);
2628 compiler_ctx_state
->debug
= async_debug
.base
;
2631 util_queue_add_job(&sctx
->screen
->shader_compiler_queue
, job
,
2632 ready_fence
, execute
, NULL
, 0);
2635 util_queue_fence_wait(ready_fence
);
2636 u_async_debug_drain(&async_debug
, &sctx
->debug
);
2637 u_async_debug_cleanup(&async_debug
);
2640 if (sctx
->screen
->options
.sync_compile
)
2641 util_queue_fence_wait(ready_fence
);
2644 /* Return descriptor slot usage masks from the given shader info. */
2645 void si_get_active_slot_masks(const struct tgsi_shader_info
*info
,
2646 uint32_t *const_and_shader_buffers
,
2647 uint64_t *samplers_and_images
)
2649 unsigned start
, num_shaderbufs
, num_constbufs
, num_images
, num_msaa_images
, num_samplers
;
2651 num_shaderbufs
= util_last_bit(info
->shader_buffers_declared
);
2652 num_constbufs
= util_last_bit(info
->const_buffers_declared
);
2653 /* two 8-byte images share one 16-byte slot */
2654 num_images
= align(util_last_bit(info
->images_declared
), 2);
2655 num_msaa_images
= align(util_last_bit(info
->msaa_images_declared
), 2);
2656 num_samplers
= util_last_bit(info
->samplers_declared
);
2658 /* The layout is: sb[last] ... sb[0], cb[0] ... cb[last] */
2659 start
= si_get_shaderbuf_slot(num_shaderbufs
- 1);
2660 *const_and_shader_buffers
=
2661 u_bit_consecutive(start
, num_shaderbufs
+ num_constbufs
);
2664 * - fmask[last] ... fmask[0] go to [15-last .. 15]
2665 * - image[last] ... image[0] go to [31-last .. 31]
2666 * - sampler[0] ... sampler[last] go to [32 .. 32+last*2]
2668 * FMASKs for images are placed separately, because MSAA images are rare,
2669 * and so we can benefit from a better cache hit rate if we keep image
2670 * descriptors together.
2672 if (num_msaa_images
)
2673 num_images
= SI_NUM_IMAGES
+ num_msaa_images
; /* add FMASK descriptors */
2675 start
= si_get_image_slot(num_images
- 1) / 2;
2676 *samplers_and_images
=
2677 u_bit_consecutive64(start
, num_images
/ 2 + num_samplers
);
2680 static void *si_create_shader_selector(struct pipe_context
*ctx
,
2681 const struct pipe_shader_state
*state
)
2683 struct si_screen
*sscreen
= (struct si_screen
*)ctx
->screen
;
2684 struct si_context
*sctx
= (struct si_context
*)ctx
;
2685 struct si_shader_selector
*sel
= CALLOC_STRUCT(si_shader_selector
);
2691 pipe_reference_init(&sel
->reference
, 1);
2692 sel
->screen
= sscreen
;
2693 sel
->compiler_ctx_state
.debug
= sctx
->debug
;
2694 sel
->compiler_ctx_state
.is_debug_context
= sctx
->is_debug
;
2696 sel
->so
= state
->stream_output
;
2698 if (state
->type
== PIPE_SHADER_IR_TGSI
&&
2699 !sscreen
->options
.enable_nir
) {
2700 sel
->tokens
= tgsi_dup_tokens(state
->tokens
);
2706 tgsi_scan_shader(state
->tokens
, &sel
->info
);
2707 tgsi_scan_tess_ctrl(state
->tokens
, &sel
->info
, &sel
->tcs_info
);
2709 /* Fixup for TGSI: Set which opcode uses which (i,j) pair. */
2710 if (sel
->info
.uses_persp_opcode_interp_centroid
)
2711 sel
->info
.uses_persp_centroid
= true;
2713 if (sel
->info
.uses_linear_opcode_interp_centroid
)
2714 sel
->info
.uses_linear_centroid
= true;
2716 if (sel
->info
.uses_persp_opcode_interp_offset
||
2717 sel
->info
.uses_persp_opcode_interp_sample
)
2718 sel
->info
.uses_persp_center
= true;
2720 if (sel
->info
.uses_linear_opcode_interp_offset
||
2721 sel
->info
.uses_linear_opcode_interp_sample
)
2722 sel
->info
.uses_linear_center
= true;
2724 if (state
->type
== PIPE_SHADER_IR_TGSI
) {
2725 sel
->nir
= tgsi_to_nir(state
->tokens
, ctx
->screen
);
2727 assert(state
->type
== PIPE_SHADER_IR_NIR
);
2728 sel
->nir
= state
->ir
.nir
;
2731 si_nir_scan_shader(sel
->nir
, &sel
->info
);
2732 si_nir_scan_tess_ctrl(sel
->nir
, &sel
->tcs_info
);
2733 si_nir_adjust_driver_locations(sel
->nir
);
2736 sel
->type
= sel
->info
.processor
;
2737 p_atomic_inc(&sscreen
->num_shaders_created
);
2738 si_get_active_slot_masks(&sel
->info
,
2739 &sel
->active_const_and_shader_buffers
,
2740 &sel
->active_samplers_and_images
);
2742 /* Record which streamout buffers are enabled. */
2743 for (i
= 0; i
< sel
->so
.num_outputs
; i
++) {
2744 sel
->enabled_streamout_buffer_mask
|=
2745 (1 << sel
->so
.output
[i
].output_buffer
) <<
2746 (sel
->so
.output
[i
].stream
* 4);
2749 /* The prolog is a no-op if there are no inputs. */
2750 sel
->vs_needs_prolog
= sel
->type
== PIPE_SHADER_VERTEX
&&
2751 sel
->info
.num_inputs
&&
2752 !sel
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
];
2754 sel
->force_correct_derivs_after_kill
=
2755 sel
->type
== PIPE_SHADER_FRAGMENT
&&
2756 sel
->info
.uses_derivatives
&&
2757 sel
->info
.uses_kill
&&
2758 sctx
->screen
->debug_flags
& DBG(FS_CORRECT_DERIVS_AFTER_KILL
);
2760 sel
->prim_discard_cs_allowed
=
2761 sel
->type
== PIPE_SHADER_VERTEX
&&
2762 !sel
->info
.uses_bindless_images
&&
2763 !sel
->info
.uses_bindless_samplers
&&
2764 !sel
->info
.writes_memory
&&
2765 !sel
->info
.writes_viewport_index
&&
2766 !sel
->info
.properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
] &&
2767 !sel
->so
.num_outputs
;
2769 switch (sel
->type
) {
2770 case PIPE_SHADER_GEOMETRY
:
2771 sel
->gs_output_prim
=
2772 sel
->info
.properties
[TGSI_PROPERTY_GS_OUTPUT_PRIM
];
2774 /* Only possibilities: POINTS, LINE_STRIP, TRIANGLES */
2775 sel
->rast_prim
= sel
->gs_output_prim
;
2776 if (util_rast_prim_is_triangles(sel
->rast_prim
))
2777 sel
->rast_prim
= PIPE_PRIM_TRIANGLES
;
2779 sel
->gs_max_out_vertices
=
2780 sel
->info
.properties
[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES
];
2781 sel
->gs_num_invocations
=
2782 sel
->info
.properties
[TGSI_PROPERTY_GS_INVOCATIONS
];
2783 sel
->gsvs_vertex_size
= sel
->info
.num_outputs
* 16;
2784 sel
->max_gsvs_emit_size
= sel
->gsvs_vertex_size
*
2785 sel
->gs_max_out_vertices
;
2787 sel
->max_gs_stream
= 0;
2788 for (i
= 0; i
< sel
->so
.num_outputs
; i
++)
2789 sel
->max_gs_stream
= MAX2(sel
->max_gs_stream
,
2790 sel
->so
.output
[i
].stream
);
2792 sel
->gs_input_verts_per_prim
=
2793 u_vertices_per_prim(sel
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
]);
2795 /* EN_MAX_VERT_OUT_PER_GS_INSTANCE does not work with tesselation. */
2796 sel
->tess_turns_off_ngg
=
2797 sscreen
->info
.chip_class
== GFX10
&&
2798 sel
->gs_num_invocations
* sel
->gs_max_out_vertices
> 256;
2801 case PIPE_SHADER_TESS_CTRL
:
2802 /* Always reserve space for these. */
2803 sel
->patch_outputs_written
|=
2804 (1ull << si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSINNER
, 0)) |
2805 (1ull << si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSOUTER
, 0));
2807 case PIPE_SHADER_VERTEX
:
2808 case PIPE_SHADER_TESS_EVAL
:
2809 for (i
= 0; i
< sel
->info
.num_outputs
; i
++) {
2810 unsigned name
= sel
->info
.output_semantic_name
[i
];
2811 unsigned index
= sel
->info
.output_semantic_index
[i
];
2814 case TGSI_SEMANTIC_TESSINNER
:
2815 case TGSI_SEMANTIC_TESSOUTER
:
2816 case TGSI_SEMANTIC_PATCH
:
2817 sel
->patch_outputs_written
|=
2818 1ull << si_shader_io_get_unique_index_patch(name
, index
);
2821 case TGSI_SEMANTIC_GENERIC
:
2822 /* don't process indices the function can't handle */
2823 if (index
>= SI_MAX_IO_GENERIC
)
2827 sel
->outputs_written
|=
2828 1ull << si_shader_io_get_unique_index(name
, index
, false);
2829 sel
->outputs_written_before_ps
|=
2830 1ull << si_shader_io_get_unique_index(name
, index
, true);
2832 case TGSI_SEMANTIC_EDGEFLAG
:
2836 sel
->esgs_itemsize
= util_last_bit64(sel
->outputs_written
) * 16;
2837 sel
->lshs_vertex_stride
= sel
->esgs_itemsize
;
2839 /* Add 1 dword to reduce LDS bank conflicts, so that each vertex
2840 * will start on a different bank. (except for the maximum 32*16).
2842 if (sel
->lshs_vertex_stride
< 32*16)
2843 sel
->lshs_vertex_stride
+= 4;
2845 /* For the ESGS ring in LDS, add 1 dword to reduce LDS bank
2846 * conflicts, i.e. each vertex will start at a different bank.
2848 if (sctx
->chip_class
>= GFX9
)
2849 sel
->esgs_itemsize
+= 4;
2851 assert(((sel
->esgs_itemsize
/ 4) & C_028AAC_ITEMSIZE
) == 0);
2854 if (sel
->info
.properties
[TGSI_PROPERTY_TES_POINT_MODE
])
2855 sel
->rast_prim
= PIPE_PRIM_POINTS
;
2856 else if (sel
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] == PIPE_PRIM_LINES
)
2857 sel
->rast_prim
= PIPE_PRIM_LINE_STRIP
;
2859 sel
->rast_prim
= PIPE_PRIM_TRIANGLES
;
2862 case PIPE_SHADER_FRAGMENT
:
2863 for (i
= 0; i
< sel
->info
.num_inputs
; i
++) {
2864 unsigned name
= sel
->info
.input_semantic_name
[i
];
2865 unsigned index
= sel
->info
.input_semantic_index
[i
];
2868 case TGSI_SEMANTIC_GENERIC
:
2869 /* don't process indices the function can't handle */
2870 if (index
>= SI_MAX_IO_GENERIC
)
2875 1ull << si_shader_io_get_unique_index(name
, index
, true);
2877 case TGSI_SEMANTIC_PCOORD
: /* ignore this */
2882 for (i
= 0; i
< 8; i
++)
2883 if (sel
->info
.colors_written
& (1 << i
))
2884 sel
->colors_written_4bit
|= 0xf << (4 * i
);
2886 for (i
= 0; i
< sel
->info
.num_inputs
; i
++) {
2887 if (sel
->info
.input_semantic_name
[i
] == TGSI_SEMANTIC_COLOR
) {
2888 int index
= sel
->info
.input_semantic_index
[i
];
2889 sel
->color_attr_index
[index
] = i
;
2896 /* PA_CL_VS_OUT_CNTL */
2897 if (sctx
->chip_class
<= GFX9
)
2898 sel
->pa_cl_vs_out_cntl
= si_get_vs_out_cntl(sel
, false);
2900 sel
->clipdist_mask
= sel
->info
.writes_clipvertex
?
2901 SIX_BITS
: sel
->info
.clipdist_writemask
;
2902 sel
->culldist_mask
= sel
->info
.culldist_writemask
<<
2903 sel
->info
.num_written_clipdistance
;
2905 /* DB_SHADER_CONTROL */
2906 sel
->db_shader_control
=
2907 S_02880C_Z_EXPORT_ENABLE(sel
->info
.writes_z
) |
2908 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(sel
->info
.writes_stencil
) |
2909 S_02880C_MASK_EXPORT_ENABLE(sel
->info
.writes_samplemask
) |
2910 S_02880C_KILL_ENABLE(sel
->info
.uses_kill
);
2912 switch (sel
->info
.properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
]) {
2913 case TGSI_FS_DEPTH_LAYOUT_GREATER
:
2914 sel
->db_shader_control
|=
2915 S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_GREATER_THAN_Z
);
2917 case TGSI_FS_DEPTH_LAYOUT_LESS
:
2918 sel
->db_shader_control
|=
2919 S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_LESS_THAN_Z
);
2923 /* Z_ORDER, EXEC_ON_HIER_FAIL and EXEC_ON_NOOP should be set as following:
2925 * | early Z/S | writes_mem | allow_ReZ? | Z_ORDER | EXEC_ON_HIER_FAIL | EXEC_ON_NOOP
2926 * --|-----------|------------|------------|--------------------|-------------------|-------------
2927 * 1a| false | false | true | EarlyZ_Then_ReZ | 0 | 0
2928 * 1b| false | false | false | EarlyZ_Then_LateZ | 0 | 0
2929 * 2 | false | true | n/a | LateZ | 1 | 0
2930 * 3 | true | false | n/a | EarlyZ_Then_LateZ | 0 | 0
2931 * 4 | true | true | n/a | EarlyZ_Then_LateZ | 0 | 1
2933 * In cases 3 and 4, HW will force Z_ORDER to EarlyZ regardless of what's set in the register.
2934 * In case 2, NOOP_CULL is a don't care field. In case 2, 3 and 4, ReZ doesn't make sense.
2936 * Don't use ReZ without profiling !!!
2938 * ReZ decreases performance by 15% in DiRT: Showdown on Ultra settings, which has pretty complex
2941 if (sel
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
]) {
2943 sel
->db_shader_control
|= S_02880C_DEPTH_BEFORE_SHADER(1) |
2944 S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z
) |
2945 S_02880C_EXEC_ON_NOOP(sel
->info
.writes_memory
);
2946 } else if (sel
->info
.writes_memory
) {
2948 sel
->db_shader_control
|= S_02880C_Z_ORDER(V_02880C_LATE_Z
) |
2949 S_02880C_EXEC_ON_HIER_FAIL(1);
2952 sel
->db_shader_control
|= S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z
);
2955 if (sel
->info
.properties
[TGSI_PROPERTY_FS_POST_DEPTH_COVERAGE
])
2956 sel
->db_shader_control
|= S_02880C_PRE_SHADER_DEPTH_COVERAGE_ENABLE(1);
2958 (void) simple_mtx_init(&sel
->mutex
, mtx_plain
);
2960 si_schedule_initial_compile(sctx
, sel
->info
.processor
, &sel
->ready
,
2961 &sel
->compiler_ctx_state
, sel
,
2962 si_init_shader_selector_async
);
2966 static void si_update_streamout_state(struct si_context
*sctx
)
2968 struct si_shader_selector
*shader_with_so
= si_get_vs(sctx
)->cso
;
2970 if (!shader_with_so
)
2973 sctx
->streamout
.enabled_stream_buffers_mask
=
2974 shader_with_so
->enabled_streamout_buffer_mask
;
2975 sctx
->streamout
.stride_in_dw
= shader_with_so
->so
.stride
;
2978 static void si_update_clip_regs(struct si_context
*sctx
,
2979 struct si_shader_selector
*old_hw_vs
,
2980 struct si_shader
*old_hw_vs_variant
,
2981 struct si_shader_selector
*next_hw_vs
,
2982 struct si_shader
*next_hw_vs_variant
)
2986 old_hw_vs
->info
.properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
] !=
2987 next_hw_vs
->info
.properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
] ||
2988 old_hw_vs
->pa_cl_vs_out_cntl
!= next_hw_vs
->pa_cl_vs_out_cntl
||
2989 old_hw_vs
->clipdist_mask
!= next_hw_vs
->clipdist_mask
||
2990 old_hw_vs
->culldist_mask
!= next_hw_vs
->culldist_mask
||
2991 !old_hw_vs_variant
||
2992 !next_hw_vs_variant
||
2993 old_hw_vs_variant
->key
.opt
.clip_disable
!=
2994 next_hw_vs_variant
->key
.opt
.clip_disable
))
2995 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.clip_regs
);
2998 static void si_update_common_shader_state(struct si_context
*sctx
)
3000 sctx
->uses_bindless_samplers
=
3001 si_shader_uses_bindless_samplers(sctx
->vs_shader
.cso
) ||
3002 si_shader_uses_bindless_samplers(sctx
->gs_shader
.cso
) ||
3003 si_shader_uses_bindless_samplers(sctx
->ps_shader
.cso
) ||
3004 si_shader_uses_bindless_samplers(sctx
->tcs_shader
.cso
) ||
3005 si_shader_uses_bindless_samplers(sctx
->tes_shader
.cso
);
3006 sctx
->uses_bindless_images
=
3007 si_shader_uses_bindless_images(sctx
->vs_shader
.cso
) ||
3008 si_shader_uses_bindless_images(sctx
->gs_shader
.cso
) ||
3009 si_shader_uses_bindless_images(sctx
->ps_shader
.cso
) ||
3010 si_shader_uses_bindless_images(sctx
->tcs_shader
.cso
) ||
3011 si_shader_uses_bindless_images(sctx
->tes_shader
.cso
);
3012 sctx
->do_update_shaders
= true;
3015 static void si_bind_vs_shader(struct pipe_context
*ctx
, void *state
)
3017 struct si_context
*sctx
= (struct si_context
*)ctx
;
3018 struct si_shader_selector
*old_hw_vs
= si_get_vs(sctx
)->cso
;
3019 struct si_shader
*old_hw_vs_variant
= si_get_vs_state(sctx
);
3020 struct si_shader_selector
*sel
= state
;
3022 if (sctx
->vs_shader
.cso
== sel
)
3025 sctx
->vs_shader
.cso
= sel
;
3026 sctx
->vs_shader
.current
= sel
? sel
->first_variant
: NULL
;
3027 sctx
->num_vs_blit_sgprs
= sel
? sel
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
] : 0;
3029 if (si_update_ngg(sctx
))
3030 si_shader_change_notify(sctx
);
3032 si_update_common_shader_state(sctx
);
3033 si_update_vs_viewport_state(sctx
);
3034 si_set_active_descriptors_for_shader(sctx
, sel
);
3035 si_update_streamout_state(sctx
);
3036 si_update_clip_regs(sctx
, old_hw_vs
, old_hw_vs_variant
,
3037 si_get_vs(sctx
)->cso
, si_get_vs_state(sctx
));
3040 static void si_update_tess_uses_prim_id(struct si_context
*sctx
)
3042 sctx
->ia_multi_vgt_param_key
.u
.tess_uses_prim_id
=
3043 (sctx
->tes_shader
.cso
&&
3044 sctx
->tes_shader
.cso
->info
.uses_primid
) ||
3045 (sctx
->tcs_shader
.cso
&&
3046 sctx
->tcs_shader
.cso
->info
.uses_primid
) ||
3047 (sctx
->gs_shader
.cso
&&
3048 sctx
->gs_shader
.cso
->info
.uses_primid
) ||
3049 (sctx
->ps_shader
.cso
&& !sctx
->gs_shader
.cso
&&
3050 sctx
->ps_shader
.cso
->info
.uses_primid
);
3053 bool si_update_ngg(struct si_context
*sctx
)
3055 if (!sctx
->screen
->use_ngg
) {
3060 bool new_ngg
= true;
3062 if (sctx
->gs_shader
.cso
&& sctx
->tes_shader
.cso
&&
3063 sctx
->gs_shader
.cso
->tess_turns_off_ngg
) {
3065 } else if (!sctx
->screen
->use_ngg_streamout
) {
3066 struct si_shader_selector
*last
= si_get_vs(sctx
)->cso
;
3068 if ((last
&& last
->so
.num_outputs
) ||
3069 sctx
->streamout
.prims_gen_query_enabled
)
3073 if (new_ngg
!= sctx
->ngg
) {
3074 /* Transitioning from NGG to legacy GS requires VGT_FLUSH on Navi10-14.
3075 * VGT_FLUSH is also emitted at the beginning of IBs when legacy GS ring
3078 if ((sctx
->family
== CHIP_NAVI10
||
3079 sctx
->family
== CHIP_NAVI12
||
3080 sctx
->family
== CHIP_NAVI14
) &&
3082 sctx
->flags
|= SI_CONTEXT_VGT_FLUSH
;
3084 sctx
->ngg
= new_ngg
;
3085 sctx
->last_rast_prim
= -1; /* reset this so that it gets updated */
3091 static void si_bind_gs_shader(struct pipe_context
*ctx
, void *state
)
3093 struct si_context
*sctx
= (struct si_context
*)ctx
;
3094 struct si_shader_selector
*old_hw_vs
= si_get_vs(sctx
)->cso
;
3095 struct si_shader
*old_hw_vs_variant
= si_get_vs_state(sctx
);
3096 struct si_shader_selector
*sel
= state
;
3097 bool enable_changed
= !!sctx
->gs_shader
.cso
!= !!sel
;
3100 if (sctx
->gs_shader
.cso
== sel
)
3103 sctx
->gs_shader
.cso
= sel
;
3104 sctx
->gs_shader
.current
= sel
? sel
->first_variant
: NULL
;
3105 sctx
->ia_multi_vgt_param_key
.u
.uses_gs
= sel
!= NULL
;
3107 si_update_common_shader_state(sctx
);
3108 sctx
->last_rast_prim
= -1; /* reset this so that it gets updated */
3110 ngg_changed
= si_update_ngg(sctx
);
3111 if (ngg_changed
|| enable_changed
)
3112 si_shader_change_notify(sctx
);
3113 if (enable_changed
) {
3114 if (sctx
->ia_multi_vgt_param_key
.u
.uses_tess
)
3115 si_update_tess_uses_prim_id(sctx
);
3117 si_update_vs_viewport_state(sctx
);
3118 si_set_active_descriptors_for_shader(sctx
, sel
);
3119 si_update_streamout_state(sctx
);
3120 si_update_clip_regs(sctx
, old_hw_vs
, old_hw_vs_variant
,
3121 si_get_vs(sctx
)->cso
, si_get_vs_state(sctx
));
3124 static void si_bind_tcs_shader(struct pipe_context
*ctx
, void *state
)
3126 struct si_context
*sctx
= (struct si_context
*)ctx
;
3127 struct si_shader_selector
*sel
= state
;
3128 bool enable_changed
= !!sctx
->tcs_shader
.cso
!= !!sel
;
3130 if (sctx
->tcs_shader
.cso
== sel
)
3133 sctx
->tcs_shader
.cso
= sel
;
3134 sctx
->tcs_shader
.current
= sel
? sel
->first_variant
: NULL
;
3135 si_update_tess_uses_prim_id(sctx
);
3137 si_update_common_shader_state(sctx
);
3140 sctx
->last_tcs
= NULL
; /* invalidate derived tess state */
3142 si_set_active_descriptors_for_shader(sctx
, sel
);
3145 static void si_bind_tes_shader(struct pipe_context
*ctx
, void *state
)
3147 struct si_context
*sctx
= (struct si_context
*)ctx
;
3148 struct si_shader_selector
*old_hw_vs
= si_get_vs(sctx
)->cso
;
3149 struct si_shader
*old_hw_vs_variant
= si_get_vs_state(sctx
);
3150 struct si_shader_selector
*sel
= state
;
3151 bool enable_changed
= !!sctx
->tes_shader
.cso
!= !!sel
;
3153 if (sctx
->tes_shader
.cso
== sel
)
3156 sctx
->tes_shader
.cso
= sel
;
3157 sctx
->tes_shader
.current
= sel
? sel
->first_variant
: NULL
;
3158 sctx
->ia_multi_vgt_param_key
.u
.uses_tess
= sel
!= NULL
;
3159 si_update_tess_uses_prim_id(sctx
);
3161 si_update_common_shader_state(sctx
);
3162 sctx
->last_rast_prim
= -1; /* reset this so that it gets updated */
3164 bool ngg_changed
= si_update_ngg(sctx
);
3165 if (ngg_changed
|| enable_changed
)
3166 si_shader_change_notify(sctx
);
3168 sctx
->last_tes_sh_base
= -1; /* invalidate derived tess state */
3169 si_update_vs_viewport_state(sctx
);
3170 si_set_active_descriptors_for_shader(sctx
, sel
);
3171 si_update_streamout_state(sctx
);
3172 si_update_clip_regs(sctx
, old_hw_vs
, old_hw_vs_variant
,
3173 si_get_vs(sctx
)->cso
, si_get_vs_state(sctx
));
3176 static void si_bind_ps_shader(struct pipe_context
*ctx
, void *state
)
3178 struct si_context
*sctx
= (struct si_context
*)ctx
;
3179 struct si_shader_selector
*old_sel
= sctx
->ps_shader
.cso
;
3180 struct si_shader_selector
*sel
= state
;
3182 /* skip if supplied shader is one already in use */
3186 sctx
->ps_shader
.cso
= sel
;
3187 sctx
->ps_shader
.current
= sel
? sel
->first_variant
: NULL
;
3189 si_update_common_shader_state(sctx
);
3191 if (sctx
->ia_multi_vgt_param_key
.u
.uses_tess
)
3192 si_update_tess_uses_prim_id(sctx
);
3195 old_sel
->info
.colors_written
!= sel
->info
.colors_written
)
3196 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.cb_render_state
);
3198 if (sctx
->screen
->has_out_of_order_rast
&&
3200 old_sel
->info
.writes_memory
!= sel
->info
.writes_memory
||
3201 old_sel
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
] !=
3202 sel
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
]))
3203 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.msaa_config
);
3205 si_set_active_descriptors_for_shader(sctx
, sel
);
3206 si_update_ps_colorbuf0_slot(sctx
);
3209 static void si_delete_shader(struct si_context
*sctx
, struct si_shader
*shader
)
3211 if (shader
->is_optimized
) {
3212 util_queue_drop_job(&sctx
->screen
->shader_compiler_queue_low_priority
,
3216 util_queue_fence_destroy(&shader
->ready
);
3219 /* If destroyed shaders were not unbound, the next compiled
3220 * shader variant could get the same pointer address and so
3221 * binding it to the same shader stage would be considered
3222 * a no-op, causing random behavior.
3224 switch (shader
->selector
->type
) {
3225 case PIPE_SHADER_VERTEX
:
3226 if (shader
->key
.as_ls
) {
3227 assert(sctx
->chip_class
<= GFX8
);
3228 si_pm4_delete_state(sctx
, ls
, shader
->pm4
);
3229 } else if (shader
->key
.as_es
) {
3230 assert(sctx
->chip_class
<= GFX8
);
3231 si_pm4_delete_state(sctx
, es
, shader
->pm4
);
3232 } else if (shader
->key
.as_ngg
) {
3233 si_pm4_delete_state(sctx
, gs
, shader
->pm4
);
3235 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
3238 case PIPE_SHADER_TESS_CTRL
:
3239 si_pm4_delete_state(sctx
, hs
, shader
->pm4
);
3241 case PIPE_SHADER_TESS_EVAL
:
3242 if (shader
->key
.as_es
) {
3243 assert(sctx
->chip_class
<= GFX8
);
3244 si_pm4_delete_state(sctx
, es
, shader
->pm4
);
3245 } else if (shader
->key
.as_ngg
) {
3246 si_pm4_delete_state(sctx
, gs
, shader
->pm4
);
3248 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
3251 case PIPE_SHADER_GEOMETRY
:
3252 if (shader
->is_gs_copy_shader
)
3253 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
3255 si_pm4_delete_state(sctx
, gs
, shader
->pm4
);
3257 case PIPE_SHADER_FRAGMENT
:
3258 si_pm4_delete_state(sctx
, ps
, shader
->pm4
);
3264 si_shader_selector_reference(sctx
, &shader
->previous_stage_sel
, NULL
);
3265 si_shader_destroy(shader
);
3269 void si_destroy_shader_selector(struct si_context
*sctx
,
3270 struct si_shader_selector
*sel
)
3272 struct si_shader
*p
= sel
->first_variant
, *c
;
3273 struct si_shader_ctx_state
*current_shader
[SI_NUM_SHADERS
] = {
3274 [PIPE_SHADER_VERTEX
] = &sctx
->vs_shader
,
3275 [PIPE_SHADER_TESS_CTRL
] = &sctx
->tcs_shader
,
3276 [PIPE_SHADER_TESS_EVAL
] = &sctx
->tes_shader
,
3277 [PIPE_SHADER_GEOMETRY
] = &sctx
->gs_shader
,
3278 [PIPE_SHADER_FRAGMENT
] = &sctx
->ps_shader
,
3281 util_queue_drop_job(&sctx
->screen
->shader_compiler_queue
, &sel
->ready
);
3283 if (current_shader
[sel
->type
]->cso
== sel
) {
3284 current_shader
[sel
->type
]->cso
= NULL
;
3285 current_shader
[sel
->type
]->current
= NULL
;
3289 c
= p
->next_variant
;
3290 si_delete_shader(sctx
, p
);
3294 if (sel
->main_shader_part
)
3295 si_delete_shader(sctx
, sel
->main_shader_part
);
3296 if (sel
->main_shader_part_ls
)
3297 si_delete_shader(sctx
, sel
->main_shader_part_ls
);
3298 if (sel
->main_shader_part_es
)
3299 si_delete_shader(sctx
, sel
->main_shader_part_es
);
3300 if (sel
->main_shader_part_ngg
)
3301 si_delete_shader(sctx
, sel
->main_shader_part_ngg
);
3302 if (sel
->gs_copy_shader
)
3303 si_delete_shader(sctx
, sel
->gs_copy_shader
);
3305 util_queue_fence_destroy(&sel
->ready
);
3306 simple_mtx_destroy(&sel
->mutex
);
3308 ralloc_free(sel
->nir
);
3309 free(sel
->nir_binary
);
3313 static void si_delete_shader_selector(struct pipe_context
*ctx
, void *state
)
3315 struct si_context
*sctx
= (struct si_context
*)ctx
;
3316 struct si_shader_selector
*sel
= (struct si_shader_selector
*)state
;
3318 si_shader_selector_reference(sctx
, &sel
, NULL
);
3321 static unsigned si_get_ps_input_cntl(struct si_context
*sctx
,
3322 struct si_shader
*vs
, unsigned name
,
3323 unsigned index
, unsigned interpolate
)
3325 struct tgsi_shader_info
*vsinfo
= &vs
->selector
->info
;
3326 unsigned j
, offset
, ps_input_cntl
= 0;
3328 if (interpolate
== TGSI_INTERPOLATE_CONSTANT
||
3329 (interpolate
== TGSI_INTERPOLATE_COLOR
&& sctx
->flatshade
) ||
3330 name
== TGSI_SEMANTIC_PRIMID
)
3331 ps_input_cntl
|= S_028644_FLAT_SHADE(1);
3333 if (name
== TGSI_SEMANTIC_PCOORD
||
3334 (name
== TGSI_SEMANTIC_TEXCOORD
&&
3335 sctx
->sprite_coord_enable
& (1 << index
))) {
3336 ps_input_cntl
|= S_028644_PT_SPRITE_TEX(1);
3339 for (j
= 0; j
< vsinfo
->num_outputs
; j
++) {
3340 if (name
== vsinfo
->output_semantic_name
[j
] &&
3341 index
== vsinfo
->output_semantic_index
[j
]) {
3342 offset
= vs
->info
.vs_output_param_offset
[j
];
3344 if (offset
<= AC_EXP_PARAM_OFFSET_31
) {
3345 /* The input is loaded from parameter memory. */
3346 ps_input_cntl
|= S_028644_OFFSET(offset
);
3347 } else if (!G_028644_PT_SPRITE_TEX(ps_input_cntl
)) {
3348 if (offset
== AC_EXP_PARAM_UNDEFINED
) {
3349 /* This can happen with depth-only rendering. */
3352 /* The input is a DEFAULT_VAL constant. */
3353 assert(offset
>= AC_EXP_PARAM_DEFAULT_VAL_0000
&&
3354 offset
<= AC_EXP_PARAM_DEFAULT_VAL_1111
);
3355 offset
-= AC_EXP_PARAM_DEFAULT_VAL_0000
;
3358 ps_input_cntl
= S_028644_OFFSET(0x20) |
3359 S_028644_DEFAULT_VAL(offset
);
3365 if (j
== vsinfo
->num_outputs
&& name
== TGSI_SEMANTIC_PRIMID
)
3366 /* PrimID is written after the last output when HW VS is used. */
3367 ps_input_cntl
|= S_028644_OFFSET(vs
->info
.vs_output_param_offset
[vsinfo
->num_outputs
]);
3368 else if (j
== vsinfo
->num_outputs
&& !G_028644_PT_SPRITE_TEX(ps_input_cntl
)) {
3369 /* No corresponding output found, load defaults into input.
3370 * Don't set any other bits.
3371 * (FLAT_SHADE=1 completely changes behavior) */
3372 ps_input_cntl
= S_028644_OFFSET(0x20);
3373 /* D3D 9 behaviour. GL is undefined */
3374 if (name
== TGSI_SEMANTIC_COLOR
&& index
== 0)
3375 ps_input_cntl
|= S_028644_DEFAULT_VAL(3);
3377 return ps_input_cntl
;
3380 static void si_emit_spi_map(struct si_context
*sctx
)
3382 struct si_shader
*ps
= sctx
->ps_shader
.current
;
3383 struct si_shader
*vs
= si_get_vs_state(sctx
);
3384 struct tgsi_shader_info
*psinfo
= ps
? &ps
->selector
->info
: NULL
;
3385 unsigned i
, num_interp
, num_written
= 0, bcol_interp
[2];
3386 unsigned spi_ps_input_cntl
[32];
3388 if (!ps
|| !ps
->selector
->info
.num_inputs
)
3391 num_interp
= si_get_ps_num_interp(ps
);
3392 assert(num_interp
> 0);
3394 for (i
= 0; i
< psinfo
->num_inputs
; i
++) {
3395 unsigned name
= psinfo
->input_semantic_name
[i
];
3396 unsigned index
= psinfo
->input_semantic_index
[i
];
3397 unsigned interpolate
= psinfo
->input_interpolate
[i
];
3399 spi_ps_input_cntl
[num_written
++] = si_get_ps_input_cntl(sctx
, vs
, name
,
3400 index
, interpolate
);
3402 if (name
== TGSI_SEMANTIC_COLOR
) {
3403 assert(index
< ARRAY_SIZE(bcol_interp
));
3404 bcol_interp
[index
] = interpolate
;
3408 if (ps
->key
.part
.ps
.prolog
.color_two_side
) {
3409 unsigned bcol
= TGSI_SEMANTIC_BCOLOR
;
3411 for (i
= 0; i
< 2; i
++) {
3412 if (!(psinfo
->colors_read
& (0xf << (i
* 4))))
3415 spi_ps_input_cntl
[num_written
++] =
3416 si_get_ps_input_cntl(sctx
, vs
, bcol
, i
, bcol_interp
[i
]);
3420 assert(num_interp
== num_written
);
3422 /* R_028644_SPI_PS_INPUT_CNTL_0 */
3423 /* Dota 2: Only ~16% of SPI map updates set different values. */
3424 /* Talos: Only ~9% of SPI map updates set different values. */
3425 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
3426 radeon_opt_set_context_regn(sctx
, R_028644_SPI_PS_INPUT_CNTL_0
,
3428 sctx
->tracked_regs
.spi_ps_input_cntl
, num_interp
);
3430 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
3431 sctx
->context_roll
= true;
3435 * Writing CONFIG or UCONFIG VGT registers requires VGT_FLUSH before that.
3437 static void si_init_config_add_vgt_flush(struct si_context
*sctx
)
3439 if (sctx
->init_config_has_vgt_flush
)
3442 /* Done by Vulkan before VGT_FLUSH. */
3443 si_pm4_cmd_begin(sctx
->init_config
, PKT3_EVENT_WRITE
);
3444 si_pm4_cmd_add(sctx
->init_config
,
3445 EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
3446 si_pm4_cmd_end(sctx
->init_config
, false);
3448 /* VGT_FLUSH is required even if VGT is idle. It resets VGT pointers. */
3449 si_pm4_cmd_begin(sctx
->init_config
, PKT3_EVENT_WRITE
);
3450 si_pm4_cmd_add(sctx
->init_config
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
3451 si_pm4_cmd_end(sctx
->init_config
, false);
3452 sctx
->init_config_has_vgt_flush
= true;
3455 /* Initialize state related to ESGS / GSVS ring buffers */
3456 static bool si_update_gs_ring_buffers(struct si_context
*sctx
)
3458 struct si_shader_selector
*es
=
3459 sctx
->tes_shader
.cso
? sctx
->tes_shader
.cso
: sctx
->vs_shader
.cso
;
3460 struct si_shader_selector
*gs
= sctx
->gs_shader
.cso
;
3461 struct si_pm4_state
*pm4
;
3463 /* Chip constants. */
3464 unsigned num_se
= sctx
->screen
->info
.max_se
;
3465 unsigned wave_size
= 64;
3466 unsigned max_gs_waves
= 32 * num_se
; /* max 32 per SE on GCN */
3467 /* On GFX6-GFX7, the value comes from VGT_GS_VERTEX_REUSE = 16.
3468 * On GFX8+, the value comes from VGT_VERTEX_REUSE_BLOCK_CNTL = 30 (+2).
3470 unsigned gs_vertex_reuse
= (sctx
->chip_class
>= GFX8
? 32 : 16) * num_se
;
3471 unsigned alignment
= 256 * num_se
;
3472 /* The maximum size is 63.999 MB per SE. */
3473 unsigned max_size
= ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se
;
3475 /* Calculate the minimum size. */
3476 unsigned min_esgs_ring_size
= align(es
->esgs_itemsize
* gs_vertex_reuse
*
3477 wave_size
, alignment
);
3479 /* These are recommended sizes, not minimum sizes. */
3480 unsigned esgs_ring_size
= max_gs_waves
* 2 * wave_size
*
3481 es
->esgs_itemsize
* gs
->gs_input_verts_per_prim
;
3482 unsigned gsvs_ring_size
= max_gs_waves
* 2 * wave_size
*
3483 gs
->max_gsvs_emit_size
;
3485 min_esgs_ring_size
= align(min_esgs_ring_size
, alignment
);
3486 esgs_ring_size
= align(esgs_ring_size
, alignment
);
3487 gsvs_ring_size
= align(gsvs_ring_size
, alignment
);
3489 esgs_ring_size
= CLAMP(esgs_ring_size
, min_esgs_ring_size
, max_size
);
3490 gsvs_ring_size
= MIN2(gsvs_ring_size
, max_size
);
3492 /* Some rings don't have to be allocated if shaders don't use them.
3493 * (e.g. no varyings between ES and GS or GS and VS)
3495 * GFX9 doesn't have the ESGS ring.
3497 bool update_esgs
= sctx
->chip_class
<= GFX8
&&
3499 (!sctx
->esgs_ring
||
3500 sctx
->esgs_ring
->width0
< esgs_ring_size
);
3501 bool update_gsvs
= gsvs_ring_size
&&
3502 (!sctx
->gsvs_ring
||
3503 sctx
->gsvs_ring
->width0
< gsvs_ring_size
);
3505 if (!update_esgs
&& !update_gsvs
)
3509 pipe_resource_reference(&sctx
->esgs_ring
, NULL
);
3511 pipe_aligned_buffer_create(sctx
->b
.screen
,
3512 SI_RESOURCE_FLAG_UNMAPPABLE
,
3515 sctx
->screen
->info
.pte_fragment_size
);
3516 if (!sctx
->esgs_ring
)
3521 pipe_resource_reference(&sctx
->gsvs_ring
, NULL
);
3523 pipe_aligned_buffer_create(sctx
->b
.screen
,
3524 SI_RESOURCE_FLAG_UNMAPPABLE
,
3527 sctx
->screen
->info
.pte_fragment_size
);
3528 if (!sctx
->gsvs_ring
)
3532 /* Create the "init_config_gs_rings" state. */
3533 pm4
= CALLOC_STRUCT(si_pm4_state
);
3537 if (sctx
->chip_class
>= GFX7
) {
3538 if (sctx
->esgs_ring
) {
3539 assert(sctx
->chip_class
<= GFX8
);
3540 si_pm4_set_reg(pm4
, R_030900_VGT_ESGS_RING_SIZE
,
3541 sctx
->esgs_ring
->width0
/ 256);
3543 if (sctx
->gsvs_ring
)
3544 si_pm4_set_reg(pm4
, R_030904_VGT_GSVS_RING_SIZE
,
3545 sctx
->gsvs_ring
->width0
/ 256);
3547 if (sctx
->esgs_ring
)
3548 si_pm4_set_reg(pm4
, R_0088C8_VGT_ESGS_RING_SIZE
,
3549 sctx
->esgs_ring
->width0
/ 256);
3550 if (sctx
->gsvs_ring
)
3551 si_pm4_set_reg(pm4
, R_0088CC_VGT_GSVS_RING_SIZE
,
3552 sctx
->gsvs_ring
->width0
/ 256);
3555 /* Set the state. */
3556 if (sctx
->init_config_gs_rings
)
3557 si_pm4_free_state(sctx
, sctx
->init_config_gs_rings
, ~0);
3558 sctx
->init_config_gs_rings
= pm4
;
3560 if (!sctx
->init_config_has_vgt_flush
) {
3561 si_init_config_add_vgt_flush(sctx
);
3562 si_pm4_upload_indirect_buffer(sctx
, sctx
->init_config
);
3565 /* Flush the context to re-emit both init_config states. */
3566 sctx
->initial_gfx_cs_size
= 0; /* force flush */
3567 si_flush_gfx_cs(sctx
, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW
, NULL
);
3569 /* Set ring bindings. */
3570 if (sctx
->esgs_ring
) {
3571 assert(sctx
->chip_class
<= GFX8
);
3572 si_set_ring_buffer(sctx
, SI_ES_RING_ESGS
,
3573 sctx
->esgs_ring
, 0, sctx
->esgs_ring
->width0
,
3574 true, true, 4, 64, 0);
3575 si_set_ring_buffer(sctx
, SI_GS_RING_ESGS
,
3576 sctx
->esgs_ring
, 0, sctx
->esgs_ring
->width0
,
3577 false, false, 0, 0, 0);
3579 if (sctx
->gsvs_ring
) {
3580 si_set_ring_buffer(sctx
, SI_RING_GSVS
,
3581 sctx
->gsvs_ring
, 0, sctx
->gsvs_ring
->width0
,
3582 false, false, 0, 0, 0);
3588 static void si_shader_lock(struct si_shader
*shader
)
3590 simple_mtx_lock(&shader
->selector
->mutex
);
3591 if (shader
->previous_stage_sel
) {
3592 assert(shader
->previous_stage_sel
!= shader
->selector
);
3593 simple_mtx_lock(&shader
->previous_stage_sel
->mutex
);
3597 static void si_shader_unlock(struct si_shader
*shader
)
3599 if (shader
->previous_stage_sel
)
3600 simple_mtx_unlock(&shader
->previous_stage_sel
->mutex
);
3601 simple_mtx_unlock(&shader
->selector
->mutex
);
3605 * @returns 1 if \p sel has been updated to use a new scratch buffer
3607 * < 0 if there was a failure
3609 static int si_update_scratch_buffer(struct si_context
*sctx
,
3610 struct si_shader
*shader
)
3612 uint64_t scratch_va
= sctx
->scratch_buffer
->gpu_address
;
3617 /* This shader doesn't need a scratch buffer */
3618 if (shader
->config
.scratch_bytes_per_wave
== 0)
3621 /* Prevent race conditions when updating:
3622 * - si_shader::scratch_bo
3623 * - si_shader::binary::code
3624 * - si_shader::previous_stage::binary::code.
3626 si_shader_lock(shader
);
3628 /* This shader is already configured to use the current
3629 * scratch buffer. */
3630 if (shader
->scratch_bo
== sctx
->scratch_buffer
) {
3631 si_shader_unlock(shader
);
3635 assert(sctx
->scratch_buffer
);
3637 /* Replace the shader bo with a new bo that has the relocs applied. */
3638 if (!si_shader_binary_upload(sctx
->screen
, shader
, scratch_va
)) {
3639 si_shader_unlock(shader
);
3643 /* Update the shader state to use the new shader bo. */
3644 si_shader_init_pm4_state(sctx
->screen
, shader
);
3646 si_resource_reference(&shader
->scratch_bo
, sctx
->scratch_buffer
);
3648 si_shader_unlock(shader
);
3652 static unsigned si_get_scratch_buffer_bytes_per_wave(struct si_shader
*shader
)
3654 return shader
? shader
->config
.scratch_bytes_per_wave
: 0;
3657 static struct si_shader
*si_get_tcs_current(struct si_context
*sctx
)
3659 if (!sctx
->tes_shader
.cso
)
3660 return NULL
; /* tessellation disabled */
3662 return sctx
->tcs_shader
.cso
? sctx
->tcs_shader
.current
:
3663 sctx
->fixed_func_tcs_shader
.current
;
3666 static bool si_update_scratch_relocs(struct si_context
*sctx
)
3668 struct si_shader
*tcs
= si_get_tcs_current(sctx
);
3671 /* Update the shaders, so that they are using the latest scratch.
3672 * The scratch buffer may have been changed since these shaders were
3673 * last used, so we still need to try to update them, even if they
3674 * require scratch buffers smaller than the current size.
3676 r
= si_update_scratch_buffer(sctx
, sctx
->ps_shader
.current
);
3680 si_pm4_bind_state(sctx
, ps
, sctx
->ps_shader
.current
->pm4
);
3682 r
= si_update_scratch_buffer(sctx
, sctx
->gs_shader
.current
);
3686 si_pm4_bind_state(sctx
, gs
, sctx
->gs_shader
.current
->pm4
);
3688 r
= si_update_scratch_buffer(sctx
, tcs
);
3692 si_pm4_bind_state(sctx
, hs
, tcs
->pm4
);
3694 /* VS can be bound as LS, ES, or VS. */
3695 r
= si_update_scratch_buffer(sctx
, sctx
->vs_shader
.current
);
3699 if (sctx
->vs_shader
.current
->key
.as_ls
)
3700 si_pm4_bind_state(sctx
, ls
, sctx
->vs_shader
.current
->pm4
);
3701 else if (sctx
->vs_shader
.current
->key
.as_es
)
3702 si_pm4_bind_state(sctx
, es
, sctx
->vs_shader
.current
->pm4
);
3703 else if (sctx
->vs_shader
.current
->key
.as_ngg
)
3704 si_pm4_bind_state(sctx
, gs
, sctx
->vs_shader
.current
->pm4
);
3706 si_pm4_bind_state(sctx
, vs
, sctx
->vs_shader
.current
->pm4
);
3709 /* TES can be bound as ES or VS. */
3710 r
= si_update_scratch_buffer(sctx
, sctx
->tes_shader
.current
);
3714 if (sctx
->tes_shader
.current
->key
.as_es
)
3715 si_pm4_bind_state(sctx
, es
, sctx
->tes_shader
.current
->pm4
);
3716 else if (sctx
->tes_shader
.current
->key
.as_ngg
)
3717 si_pm4_bind_state(sctx
, gs
, sctx
->tes_shader
.current
->pm4
);
3719 si_pm4_bind_state(sctx
, vs
, sctx
->tes_shader
.current
->pm4
);
3725 static bool si_update_spi_tmpring_size(struct si_context
*sctx
)
3727 /* SPI_TMPRING_SIZE.WAVESIZE must be constant for each scratch buffer.
3728 * There are 2 cases to handle:
3730 * - If the current needed size is less than the maximum seen size,
3731 * use the maximum seen size, so that WAVESIZE remains the same.
3733 * - If the current needed size is greater than the maximum seen size,
3734 * the scratch buffer is reallocated, so we can increase WAVESIZE.
3736 * Shaders that set SCRATCH_EN=0 don't allocate scratch space.
3737 * Otherwise, the number of waves that can use scratch is
3738 * SPI_TMPRING_SIZE.WAVES.
3742 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->ps_shader
.current
));
3743 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->gs_shader
.current
));
3744 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->vs_shader
.current
));
3746 if (sctx
->tes_shader
.cso
) {
3747 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->tes_shader
.current
));
3748 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(si_get_tcs_current(sctx
)));
3751 sctx
->max_seen_scratch_bytes_per_wave
=
3752 MAX2(sctx
->max_seen_scratch_bytes_per_wave
, bytes
);
3754 unsigned scratch_needed_size
=
3755 sctx
->max_seen_scratch_bytes_per_wave
* sctx
->scratch_waves
;
3756 unsigned spi_tmpring_size
;
3758 if (scratch_needed_size
> 0) {
3759 if (!sctx
->scratch_buffer
||
3760 scratch_needed_size
> sctx
->scratch_buffer
->b
.b
.width0
) {
3761 /* Create a bigger scratch buffer */
3762 si_resource_reference(&sctx
->scratch_buffer
, NULL
);
3764 sctx
->scratch_buffer
=
3765 si_aligned_buffer_create(&sctx
->screen
->b
,
3766 SI_RESOURCE_FLAG_UNMAPPABLE
,
3768 scratch_needed_size
,
3769 sctx
->screen
->info
.pte_fragment_size
);
3770 if (!sctx
->scratch_buffer
)
3773 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.scratch_state
);
3774 si_context_add_resource_size(sctx
,
3775 &sctx
->scratch_buffer
->b
.b
);
3778 if (!si_update_scratch_relocs(sctx
))
3782 /* The LLVM shader backend should be reporting aligned scratch_sizes. */
3783 assert((scratch_needed_size
& ~0x3FF) == scratch_needed_size
&&
3784 "scratch size should already be aligned correctly.");
3786 spi_tmpring_size
= S_0286E8_WAVES(sctx
->scratch_waves
) |
3787 S_0286E8_WAVESIZE(sctx
->max_seen_scratch_bytes_per_wave
>> 10);
3788 if (spi_tmpring_size
!= sctx
->spi_tmpring_size
) {
3789 sctx
->spi_tmpring_size
= spi_tmpring_size
;
3790 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.scratch_state
);
3795 static void si_init_tess_factor_ring(struct si_context
*sctx
)
3797 assert(!sctx
->tess_rings
);
3798 assert(((sctx
->screen
->tess_factor_ring_size
/ 4) & C_030938_SIZE
) == 0);
3800 /* The address must be aligned to 2^19, because the shader only
3801 * receives the high 13 bits.
3803 sctx
->tess_rings
= pipe_aligned_buffer_create(sctx
->b
.screen
,
3804 SI_RESOURCE_FLAG_32BIT
,
3806 sctx
->screen
->tess_offchip_ring_size
+
3807 sctx
->screen
->tess_factor_ring_size
,
3809 if (!sctx
->tess_rings
)
3812 si_init_config_add_vgt_flush(sctx
);
3814 si_pm4_add_bo(sctx
->init_config
, si_resource(sctx
->tess_rings
),
3815 RADEON_USAGE_READWRITE
, RADEON_PRIO_SHADER_RINGS
);
3817 uint64_t factor_va
= si_resource(sctx
->tess_rings
)->gpu_address
+
3818 sctx
->screen
->tess_offchip_ring_size
;
3820 /* Append these registers to the init config state. */
3821 if (sctx
->chip_class
>= GFX7
) {
3822 si_pm4_set_reg(sctx
->init_config
, R_030938_VGT_TF_RING_SIZE
,
3823 S_030938_SIZE(sctx
->screen
->tess_factor_ring_size
/ 4));
3824 si_pm4_set_reg(sctx
->init_config
, R_030940_VGT_TF_MEMORY_BASE
,
3826 if (sctx
->chip_class
>= GFX10
)
3827 si_pm4_set_reg(sctx
->init_config
, R_030984_VGT_TF_MEMORY_BASE_HI_UMD
,
3828 S_030984_BASE_HI(factor_va
>> 40));
3829 else if (sctx
->chip_class
== GFX9
)
3830 si_pm4_set_reg(sctx
->init_config
, R_030944_VGT_TF_MEMORY_BASE_HI
,
3831 S_030944_BASE_HI(factor_va
>> 40));
3832 si_pm4_set_reg(sctx
->init_config
, R_03093C_VGT_HS_OFFCHIP_PARAM
,
3833 sctx
->screen
->vgt_hs_offchip_param
);
3835 si_pm4_set_reg(sctx
->init_config
, R_008988_VGT_TF_RING_SIZE
,
3836 S_008988_SIZE(sctx
->screen
->tess_factor_ring_size
/ 4));
3837 si_pm4_set_reg(sctx
->init_config
, R_0089B8_VGT_TF_MEMORY_BASE
,
3839 si_pm4_set_reg(sctx
->init_config
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
3840 sctx
->screen
->vgt_hs_offchip_param
);
3843 /* Flush the context to re-emit the init_config state.
3844 * This is done only once in a lifetime of a context.
3846 si_pm4_upload_indirect_buffer(sctx
, sctx
->init_config
);
3847 sctx
->initial_gfx_cs_size
= 0; /* force flush */
3848 si_flush_gfx_cs(sctx
, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW
, NULL
);
3851 static struct si_pm4_state
*si_build_vgt_shader_config(struct si_screen
*screen
,
3852 union si_vgt_stages_key key
)
3854 struct si_pm4_state
*pm4
= CALLOC_STRUCT(si_pm4_state
);
3855 uint32_t stages
= 0;
3858 stages
|= S_028B54_LS_EN(V_028B54_LS_STAGE_ON
) |
3859 S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
3862 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_DS
) |
3865 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_DS
);
3867 stages
|= S_028B54_VS_EN(V_028B54_VS_STAGE_DS
);
3868 } else if (key
.u
.gs
) {
3869 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL
) |
3871 } else if (key
.u
.ngg
) {
3872 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL
);
3876 stages
|= S_028B54_PRIMGEN_EN(1) |
3877 S_028B54_NGG_WAVE_ID_EN(key
.u
.streamout
) |
3878 S_028B54_PRIMGEN_PASSTHRU_EN(key
.u
.ngg_passthrough
);
3879 } else if (key
.u
.gs
)
3880 stages
|= S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
3882 if (screen
->info
.chip_class
>= GFX9
)
3883 stages
|= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
3885 if (screen
->info
.chip_class
>= GFX10
&& screen
->ge_wave_size
== 32) {
3886 stages
|= S_028B54_HS_W32_EN(1) |
3887 S_028B54_GS_W32_EN(key
.u
.ngg
) | /* legacy GS only supports Wave64 */
3888 S_028B54_VS_W32_EN(1);
3891 si_pm4_set_reg(pm4
, R_028B54_VGT_SHADER_STAGES_EN
, stages
);
3895 static void si_update_vgt_shader_config(struct si_context
*sctx
,
3896 union si_vgt_stages_key key
)
3898 struct si_pm4_state
**pm4
= &sctx
->vgt_shader_config
[key
.index
];
3900 if (unlikely(!*pm4
))
3901 *pm4
= si_build_vgt_shader_config(sctx
->screen
, key
);
3902 si_pm4_bind_state(sctx
, vgt_shader_config
, *pm4
);
3905 bool si_update_shaders(struct si_context
*sctx
)
3907 struct pipe_context
*ctx
= (struct pipe_context
*)sctx
;
3908 struct si_compiler_ctx_state compiler_state
;
3909 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
3910 struct si_shader
*old_vs
= si_get_vs_state(sctx
);
3911 bool old_clip_disable
= old_vs
? old_vs
->key
.opt
.clip_disable
: false;
3912 struct si_shader
*old_ps
= sctx
->ps_shader
.current
;
3913 union si_vgt_stages_key key
;
3914 unsigned old_spi_shader_col_format
=
3915 old_ps
? old_ps
->key
.part
.ps
.epilog
.spi_shader_col_format
: 0;
3918 if (!sctx
->compiler
.passes
)
3919 si_init_compiler(sctx
->screen
, &sctx
->compiler
);
3921 compiler_state
.compiler
= &sctx
->compiler
;
3922 compiler_state
.debug
= sctx
->debug
;
3923 compiler_state
.is_debug_context
= sctx
->is_debug
;
3927 if (sctx
->tes_shader
.cso
)
3929 if (sctx
->gs_shader
.cso
)
3934 key
.u
.streamout
= !!si_get_vs(sctx
)->cso
->so
.num_outputs
;
3937 /* Update TCS and TES. */
3938 if (sctx
->tes_shader
.cso
) {
3939 if (!sctx
->tess_rings
) {
3940 si_init_tess_factor_ring(sctx
);
3941 if (!sctx
->tess_rings
)
3945 if (sctx
->tcs_shader
.cso
) {
3946 r
= si_shader_select(ctx
, &sctx
->tcs_shader
, key
,
3950 si_pm4_bind_state(sctx
, hs
, sctx
->tcs_shader
.current
->pm4
);
3952 if (!sctx
->fixed_func_tcs_shader
.cso
) {
3953 sctx
->fixed_func_tcs_shader
.cso
=
3954 si_create_fixed_func_tcs(sctx
);
3955 if (!sctx
->fixed_func_tcs_shader
.cso
)
3959 r
= si_shader_select(ctx
, &sctx
->fixed_func_tcs_shader
,
3960 key
, &compiler_state
);
3963 si_pm4_bind_state(sctx
, hs
,
3964 sctx
->fixed_func_tcs_shader
.current
->pm4
);
3967 if (!sctx
->gs_shader
.cso
|| sctx
->chip_class
<= GFX8
) {
3968 r
= si_shader_select(ctx
, &sctx
->tes_shader
, key
, &compiler_state
);
3972 if (sctx
->gs_shader
.cso
) {
3974 assert(sctx
->chip_class
<= GFX8
);
3975 si_pm4_bind_state(sctx
, es
, sctx
->tes_shader
.current
->pm4
);
3976 } else if (key
.u
.ngg
) {
3977 si_pm4_bind_state(sctx
, gs
, sctx
->tes_shader
.current
->pm4
);
3979 si_pm4_bind_state(sctx
, vs
, sctx
->tes_shader
.current
->pm4
);
3983 if (sctx
->chip_class
<= GFX8
)
3984 si_pm4_bind_state(sctx
, ls
, NULL
);
3985 si_pm4_bind_state(sctx
, hs
, NULL
);
3989 if (sctx
->gs_shader
.cso
) {
3990 r
= si_shader_select(ctx
, &sctx
->gs_shader
, key
, &compiler_state
);
3993 si_pm4_bind_state(sctx
, gs
, sctx
->gs_shader
.current
->pm4
);
3995 si_pm4_bind_state(sctx
, vs
, sctx
->gs_shader
.cso
->gs_copy_shader
->pm4
);
3997 if (!si_update_gs_ring_buffers(sctx
))
4000 si_pm4_bind_state(sctx
, vs
, NULL
);
4004 si_pm4_bind_state(sctx
, gs
, NULL
);
4005 if (sctx
->chip_class
<= GFX8
)
4006 si_pm4_bind_state(sctx
, es
, NULL
);
4011 if ((!key
.u
.tess
&& !key
.u
.gs
) || sctx
->chip_class
<= GFX8
) {
4012 r
= si_shader_select(ctx
, &sctx
->vs_shader
, key
, &compiler_state
);
4016 if (!key
.u
.tess
&& !key
.u
.gs
) {
4018 si_pm4_bind_state(sctx
, gs
, sctx
->vs_shader
.current
->pm4
);
4019 si_pm4_bind_state(sctx
, vs
, NULL
);
4021 si_pm4_bind_state(sctx
, vs
, sctx
->vs_shader
.current
->pm4
);
4023 } else if (sctx
->tes_shader
.cso
) {
4024 si_pm4_bind_state(sctx
, ls
, sctx
->vs_shader
.current
->pm4
);
4026 assert(sctx
->gs_shader
.cso
);
4027 si_pm4_bind_state(sctx
, es
, sctx
->vs_shader
.current
->pm4
);
4031 /* This must be done after the shader variant is selected. */
4033 key
.u
.ngg_passthrough
= gfx10_is_ngg_passthrough(si_get_vs(sctx
)->current
);
4035 si_update_vgt_shader_config(sctx
, key
);
4037 if (old_clip_disable
!= si_get_vs_state(sctx
)->key
.opt
.clip_disable
)
4038 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.clip_regs
);
4040 if (sctx
->ps_shader
.cso
) {
4041 unsigned db_shader_control
;
4043 r
= si_shader_select(ctx
, &sctx
->ps_shader
, key
, &compiler_state
);
4046 si_pm4_bind_state(sctx
, ps
, sctx
->ps_shader
.current
->pm4
);
4049 sctx
->ps_shader
.cso
->db_shader_control
|
4050 S_02880C_KILL_ENABLE(si_get_alpha_test_func(sctx
) != PIPE_FUNC_ALWAYS
);
4052 if (si_pm4_state_changed(sctx
, ps
) ||
4053 si_pm4_state_changed(sctx
, vs
) ||
4054 (key
.u
.ngg
&& si_pm4_state_changed(sctx
, gs
)) ||
4055 sctx
->sprite_coord_enable
!= rs
->sprite_coord_enable
||
4056 sctx
->flatshade
!= rs
->flatshade
) {
4057 sctx
->sprite_coord_enable
= rs
->sprite_coord_enable
;
4058 sctx
->flatshade
= rs
->flatshade
;
4059 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.spi_map
);
4062 if (sctx
->screen
->info
.rbplus_allowed
&&
4063 si_pm4_state_changed(sctx
, ps
) &&
4065 old_spi_shader_col_format
!=
4066 sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.spi_shader_col_format
))
4067 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.cb_render_state
);
4069 if (sctx
->ps_db_shader_control
!= db_shader_control
) {
4070 sctx
->ps_db_shader_control
= db_shader_control
;
4071 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.db_render_state
);
4072 if (sctx
->screen
->dpbb_allowed
)
4073 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.dpbb_state
);
4076 if (sctx
->smoothing_enabled
!= sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.poly_line_smoothing
) {
4077 sctx
->smoothing_enabled
= sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.poly_line_smoothing
;
4078 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.msaa_config
);
4080 if (sctx
->chip_class
== GFX6
)
4081 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.db_render_state
);
4083 if (sctx
->framebuffer
.nr_samples
<= 1)
4084 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.msaa_sample_locs
);
4088 if (si_pm4_state_enabled_and_changed(sctx
, ls
) ||
4089 si_pm4_state_enabled_and_changed(sctx
, hs
) ||
4090 si_pm4_state_enabled_and_changed(sctx
, es
) ||
4091 si_pm4_state_enabled_and_changed(sctx
, gs
) ||
4092 si_pm4_state_enabled_and_changed(sctx
, vs
) ||
4093 si_pm4_state_enabled_and_changed(sctx
, ps
)) {
4094 if (!si_update_spi_tmpring_size(sctx
))
4098 if (sctx
->chip_class
>= GFX7
) {
4099 if (si_pm4_state_enabled_and_changed(sctx
, ls
))
4100 sctx
->prefetch_L2_mask
|= SI_PREFETCH_LS
;
4101 else if (!sctx
->queued
.named
.ls
)
4102 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_LS
;
4104 if (si_pm4_state_enabled_and_changed(sctx
, hs
))
4105 sctx
->prefetch_L2_mask
|= SI_PREFETCH_HS
;
4106 else if (!sctx
->queued
.named
.hs
)
4107 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_HS
;
4109 if (si_pm4_state_enabled_and_changed(sctx
, es
))
4110 sctx
->prefetch_L2_mask
|= SI_PREFETCH_ES
;
4111 else if (!sctx
->queued
.named
.es
)
4112 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_ES
;
4114 if (si_pm4_state_enabled_and_changed(sctx
, gs
))
4115 sctx
->prefetch_L2_mask
|= SI_PREFETCH_GS
;
4116 else if (!sctx
->queued
.named
.gs
)
4117 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_GS
;
4119 if (si_pm4_state_enabled_and_changed(sctx
, vs
))
4120 sctx
->prefetch_L2_mask
|= SI_PREFETCH_VS
;
4121 else if (!sctx
->queued
.named
.vs
)
4122 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_VS
;
4124 if (si_pm4_state_enabled_and_changed(sctx
, ps
))
4125 sctx
->prefetch_L2_mask
|= SI_PREFETCH_PS
;
4126 else if (!sctx
->queued
.named
.ps
)
4127 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_PS
;
4130 sctx
->do_update_shaders
= false;
4134 static void si_emit_scratch_state(struct si_context
*sctx
)
4136 struct radeon_cmdbuf
*cs
= sctx
->gfx_cs
;
4138 radeon_set_context_reg(cs
, R_0286E8_SPI_TMPRING_SIZE
,
4139 sctx
->spi_tmpring_size
);
4141 if (sctx
->scratch_buffer
) {
4142 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
4143 sctx
->scratch_buffer
, RADEON_USAGE_READWRITE
,
4144 RADEON_PRIO_SCRATCH_BUFFER
);
4148 void si_init_shader_functions(struct si_context
*sctx
)
4150 sctx
->atoms
.s
.spi_map
.emit
= si_emit_spi_map
;
4151 sctx
->atoms
.s
.scratch_state
.emit
= si_emit_scratch_state
;
4153 sctx
->b
.create_vs_state
= si_create_shader_selector
;
4154 sctx
->b
.create_tcs_state
= si_create_shader_selector
;
4155 sctx
->b
.create_tes_state
= si_create_shader_selector
;
4156 sctx
->b
.create_gs_state
= si_create_shader_selector
;
4157 sctx
->b
.create_fs_state
= si_create_shader_selector
;
4159 sctx
->b
.bind_vs_state
= si_bind_vs_shader
;
4160 sctx
->b
.bind_tcs_state
= si_bind_tcs_shader
;
4161 sctx
->b
.bind_tes_state
= si_bind_tes_shader
;
4162 sctx
->b
.bind_gs_state
= si_bind_gs_shader
;
4163 sctx
->b
.bind_fs_state
= si_bind_ps_shader
;
4165 sctx
->b
.delete_vs_state
= si_delete_shader_selector
;
4166 sctx
->b
.delete_tcs_state
= si_delete_shader_selector
;
4167 sctx
->b
.delete_tes_state
= si_delete_shader_selector
;
4168 sctx
->b
.delete_gs_state
= si_delete_shader_selector
;
4169 sctx
->b
.delete_fs_state
= si_delete_shader_selector
;