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 "util/hash_table.h"
31 #include "util/crc32.h"
32 #include "util/u_async_debug.h"
33 #include "util/u_memory.h"
34 #include "util/u_prim.h"
36 #include "util/disk_cache.h"
37 #include "util/mesa-sha1.h"
38 #include "ac_exp_param.h"
39 #include "ac_shader_util.h"
44 * Return the IR key for the shader cache.
46 void si_get_ir_cache_key(struct si_shader_selector
*sel
, bool ngg
, bool es
,
47 unsigned char ir_sha1_cache_key
[20])
49 struct blob blob
= {};
53 if (sel
->nir_binary
) {
54 ir_binary
= sel
->nir_binary
;
55 ir_size
= sel
->nir_size
;
60 nir_serialize(&blob
, sel
->nir
, true);
61 ir_binary
= blob
.data
;
65 /* These settings affect the compilation, but they are not derived
66 * from the input shader IR.
68 unsigned shader_variant_flags
= 0;
71 shader_variant_flags
|= 1 << 0;
73 shader_variant_flags
|= 1 << 1;
74 if (si_get_wave_size(sel
->screen
, sel
->type
, ngg
, es
) == 32)
75 shader_variant_flags
|= 1 << 2;
76 if (sel
->force_correct_derivs_after_kill
)
77 shader_variant_flags
|= 1 << 3;
80 _mesa_sha1_init(&ctx
);
81 _mesa_sha1_update(&ctx
, &shader_variant_flags
, 4);
82 _mesa_sha1_update(&ctx
, ir_binary
, ir_size
);
83 if (sel
->type
== PIPE_SHADER_VERTEX
||
84 sel
->type
== PIPE_SHADER_TESS_EVAL
||
85 sel
->type
== PIPE_SHADER_GEOMETRY
)
86 _mesa_sha1_update(&ctx
, &sel
->so
, sizeof(sel
->so
));
87 _mesa_sha1_final(&ctx
, ir_sha1_cache_key
);
89 if (ir_binary
== blob
.data
)
93 /** Copy "data" to "ptr" and return the next dword following copied data. */
94 static uint32_t *write_data(uint32_t *ptr
, const void *data
, unsigned size
)
96 /* data may be NULL if size == 0 */
98 memcpy(ptr
, data
, size
);
99 ptr
+= DIV_ROUND_UP(size
, 4);
103 /** Read data from "ptr". Return the next dword following the data. */
104 static uint32_t *read_data(uint32_t *ptr
, void *data
, unsigned size
)
106 memcpy(data
, ptr
, size
);
107 ptr
+= DIV_ROUND_UP(size
, 4);
112 * Write the size as uint followed by the data. Return the next dword
113 * following the copied data.
115 static uint32_t *write_chunk(uint32_t *ptr
, const void *data
, unsigned size
)
118 return write_data(ptr
, data
, size
);
122 * Read the size as uint followed by the data. Return both via parameters.
123 * Return the next dword following the data.
125 static uint32_t *read_chunk(uint32_t *ptr
, void **data
, unsigned *size
)
128 assert(*data
== NULL
);
131 *data
= malloc(*size
);
132 return read_data(ptr
, *data
, *size
);
136 * Return the shader binary in a buffer. The first 4 bytes contain its size
139 static void *si_get_shader_binary(struct si_shader
*shader
)
141 /* There is always a size of data followed by the data itself. */
142 unsigned llvm_ir_size
= shader
->binary
.llvm_ir_string
?
143 strlen(shader
->binary
.llvm_ir_string
) + 1 : 0;
145 /* Refuse to allocate overly large buffers and guard against integer
147 if (shader
->binary
.elf_size
> UINT_MAX
/ 4 ||
148 llvm_ir_size
> UINT_MAX
/ 4)
153 4 + /* CRC32 of the data below */
154 align(sizeof(shader
->config
), 4) +
155 align(sizeof(shader
->info
), 4) +
156 4 + align(shader
->binary
.elf_size
, 4) +
157 4 + align(llvm_ir_size
, 4);
158 void *buffer
= CALLOC(1, size
);
159 uint32_t *ptr
= (uint32_t*)buffer
;
165 ptr
++; /* CRC32 is calculated at the end. */
167 ptr
= write_data(ptr
, &shader
->config
, sizeof(shader
->config
));
168 ptr
= write_data(ptr
, &shader
->info
, sizeof(shader
->info
));
169 ptr
= write_chunk(ptr
, shader
->binary
.elf_buffer
, shader
->binary
.elf_size
);
170 ptr
= write_chunk(ptr
, shader
->binary
.llvm_ir_string
, llvm_ir_size
);
171 assert((char *)ptr
- (char *)buffer
== size
);
174 ptr
= (uint32_t*)buffer
;
176 *ptr
= util_hash_crc32(ptr
+ 1, size
- 8);
181 static bool si_load_shader_binary(struct si_shader
*shader
, void *binary
)
183 uint32_t *ptr
= (uint32_t*)binary
;
184 uint32_t size
= *ptr
++;
185 uint32_t crc32
= *ptr
++;
189 if (util_hash_crc32(ptr
, size
- 8) != crc32
) {
190 fprintf(stderr
, "radeonsi: binary shader has invalid CRC32\n");
194 ptr
= read_data(ptr
, &shader
->config
, sizeof(shader
->config
));
195 ptr
= read_data(ptr
, &shader
->info
, sizeof(shader
->info
));
196 ptr
= read_chunk(ptr
, (void**)&shader
->binary
.elf_buffer
,
198 shader
->binary
.elf_size
= elf_size
;
199 ptr
= read_chunk(ptr
, (void**)&shader
->binary
.llvm_ir_string
, &chunk_size
);
205 * Insert a shader into the cache. It's assumed the shader is not in the cache.
206 * Use si_shader_cache_load_shader before calling this.
208 void si_shader_cache_insert_shader(struct si_screen
*sscreen
,
209 unsigned char ir_sha1_cache_key
[20],
210 struct si_shader
*shader
,
211 bool insert_into_disk_cache
)
214 struct hash_entry
*entry
;
215 uint8_t key
[CACHE_KEY_SIZE
];
217 entry
= _mesa_hash_table_search(sscreen
->shader_cache
, ir_sha1_cache_key
);
219 return; /* already added */
221 hw_binary
= si_get_shader_binary(shader
);
225 if (_mesa_hash_table_insert(sscreen
->shader_cache
,
226 mem_dup(ir_sha1_cache_key
, 20),
227 hw_binary
) == NULL
) {
232 if (sscreen
->disk_shader_cache
&& insert_into_disk_cache
) {
233 disk_cache_compute_key(sscreen
->disk_shader_cache
,
234 ir_sha1_cache_key
, 20, key
);
235 disk_cache_put(sscreen
->disk_shader_cache
, key
, hw_binary
,
236 *((uint32_t *) hw_binary
), NULL
);
240 bool si_shader_cache_load_shader(struct si_screen
*sscreen
,
241 unsigned char ir_sha1_cache_key
[20],
242 struct si_shader
*shader
)
244 struct hash_entry
*entry
=
245 _mesa_hash_table_search(sscreen
->shader_cache
, ir_sha1_cache_key
);
247 if (sscreen
->disk_shader_cache
) {
248 unsigned char sha1
[CACHE_KEY_SIZE
];
250 disk_cache_compute_key(sscreen
->disk_shader_cache
,
251 ir_sha1_cache_key
, 20, sha1
);
255 disk_cache_get(sscreen
->disk_shader_cache
,
260 if (binary_size
< sizeof(uint32_t) ||
261 *((uint32_t*)buffer
) != binary_size
) {
262 /* Something has gone wrong discard the item
263 * from the cache and rebuild/link from
266 assert(!"Invalid radeonsi shader disk cache "
269 disk_cache_remove(sscreen
->disk_shader_cache
,
276 if (!si_load_shader_binary(shader
, buffer
)) {
282 si_shader_cache_insert_shader(sscreen
, ir_sha1_cache_key
,
288 if (!si_load_shader_binary(shader
, entry
->data
))
291 p_atomic_inc(&sscreen
->num_shader_cache_hits
);
295 static uint32_t si_shader_cache_key_hash(const void *key
)
297 /* Take the first dword of SHA1. */
298 return *(uint32_t*)key
;
301 static bool si_shader_cache_key_equals(const void *a
, const void *b
)
304 return memcmp(a
, b
, 20) == 0;
307 static void si_destroy_shader_cache_entry(struct hash_entry
*entry
)
309 FREE((void*)entry
->key
);
313 bool si_init_shader_cache(struct si_screen
*sscreen
)
315 (void) simple_mtx_init(&sscreen
->shader_cache_mutex
, mtx_plain
);
316 sscreen
->shader_cache
=
317 _mesa_hash_table_create(NULL
,
318 si_shader_cache_key_hash
,
319 si_shader_cache_key_equals
);
321 return sscreen
->shader_cache
!= NULL
;
324 void si_destroy_shader_cache(struct si_screen
*sscreen
)
326 if (sscreen
->shader_cache
)
327 _mesa_hash_table_destroy(sscreen
->shader_cache
,
328 si_destroy_shader_cache_entry
);
329 simple_mtx_destroy(&sscreen
->shader_cache_mutex
);
334 static void si_set_tesseval_regs(struct si_screen
*sscreen
,
335 const struct si_shader_selector
*tes
,
336 struct si_pm4_state
*pm4
)
338 const struct tgsi_shader_info
*info
= &tes
->info
;
339 unsigned tes_prim_mode
= info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
340 unsigned tes_spacing
= info
->properties
[TGSI_PROPERTY_TES_SPACING
];
341 bool tes_vertex_order_cw
= info
->properties
[TGSI_PROPERTY_TES_VERTEX_ORDER_CW
];
342 bool tes_point_mode
= info
->properties
[TGSI_PROPERTY_TES_POINT_MODE
];
343 unsigned type
, partitioning
, topology
, distribution_mode
;
345 switch (tes_prim_mode
) {
346 case PIPE_PRIM_LINES
:
347 type
= V_028B6C_TESS_ISOLINE
;
349 case PIPE_PRIM_TRIANGLES
:
350 type
= V_028B6C_TESS_TRIANGLE
;
352 case PIPE_PRIM_QUADS
:
353 type
= V_028B6C_TESS_QUAD
;
360 switch (tes_spacing
) {
361 case PIPE_TESS_SPACING_FRACTIONAL_ODD
:
362 partitioning
= V_028B6C_PART_FRAC_ODD
;
364 case PIPE_TESS_SPACING_FRACTIONAL_EVEN
:
365 partitioning
= V_028B6C_PART_FRAC_EVEN
;
367 case PIPE_TESS_SPACING_EQUAL
:
368 partitioning
= V_028B6C_PART_INTEGER
;
376 topology
= V_028B6C_OUTPUT_POINT
;
377 else if (tes_prim_mode
== PIPE_PRIM_LINES
)
378 topology
= V_028B6C_OUTPUT_LINE
;
379 else if (tes_vertex_order_cw
)
380 /* for some reason, this must be the other way around */
381 topology
= V_028B6C_OUTPUT_TRIANGLE_CCW
;
383 topology
= V_028B6C_OUTPUT_TRIANGLE_CW
;
385 if (sscreen
->info
.has_distributed_tess
) {
386 if (sscreen
->info
.family
== CHIP_FIJI
||
387 sscreen
->info
.family
>= CHIP_POLARIS10
)
388 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS
;
390 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_DONUTS
;
392 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_NO_DIST
;
395 pm4
->shader
->vgt_tf_param
= S_028B6C_TYPE(type
) |
396 S_028B6C_PARTITIONING(partitioning
) |
397 S_028B6C_TOPOLOGY(topology
) |
398 S_028B6C_DISTRIBUTION_MODE(distribution_mode
);
401 /* Polaris needs different VTX_REUSE_DEPTH settings depending on
402 * whether the "fractional odd" tessellation spacing is used.
404 * Possible VGT configurations and which state should set the register:
406 * Reg set in | VGT shader configuration | Value
407 * ------------------------------------------------------
409 * VS as ES | ES -> GS -> VS | 30
410 * TES as VS | LS -> HS -> VS | 14 or 30
411 * TES as ES | LS -> HS -> ES -> GS -> VS | 14 or 30
413 * If "shader" is NULL, it's assumed it's not LS or GS copy shader.
415 static void polaris_set_vgt_vertex_reuse(struct si_screen
*sscreen
,
416 struct si_shader_selector
*sel
,
417 struct si_shader
*shader
,
418 struct si_pm4_state
*pm4
)
420 unsigned type
= sel
->type
;
422 if (sscreen
->info
.family
< CHIP_POLARIS10
||
423 sscreen
->info
.chip_class
>= GFX10
)
426 /* VS as VS, or VS as ES: */
427 if ((type
== PIPE_SHADER_VERTEX
&&
429 (!shader
->key
.as_ls
&& !shader
->is_gs_copy_shader
))) ||
430 /* TES as VS, or TES as ES: */
431 type
== PIPE_SHADER_TESS_EVAL
) {
432 unsigned vtx_reuse_depth
= 30;
434 if (type
== PIPE_SHADER_TESS_EVAL
&&
435 sel
->info
.properties
[TGSI_PROPERTY_TES_SPACING
] ==
436 PIPE_TESS_SPACING_FRACTIONAL_ODD
)
437 vtx_reuse_depth
= 14;
440 pm4
->shader
->vgt_vertex_reuse_block_cntl
= vtx_reuse_depth
;
444 static struct si_pm4_state
*si_get_shader_pm4_state(struct si_shader
*shader
)
447 si_pm4_clear_state(shader
->pm4
);
449 shader
->pm4
= CALLOC_STRUCT(si_pm4_state
);
452 shader
->pm4
->shader
= shader
;
455 fprintf(stderr
, "radeonsi: Failed to create pm4 state.\n");
460 static unsigned si_get_num_vs_user_sgprs(unsigned num_always_on_user_sgprs
)
462 /* Add the pointer to VBO descriptors. */
463 return num_always_on_user_sgprs
+ 1;
466 /* Return VGPR_COMP_CNT for the API vertex shader. This can be hw LS, LSHS, ES, ESGS, VS. */
467 static unsigned si_get_vs_vgpr_comp_cnt(struct si_screen
*sscreen
,
468 struct si_shader
*shader
, bool legacy_vs_prim_id
)
470 assert(shader
->selector
->type
== PIPE_SHADER_VERTEX
||
471 (shader
->previous_stage_sel
&&
472 shader
->previous_stage_sel
->type
== PIPE_SHADER_VERTEX
));
474 /* GFX6-9 LS (VertexID, RelAutoindex, InstanceID / StepRate0(==1), ...).
475 * GFX6-9 ES,VS (VertexID, InstanceID / StepRate0(==1), VSPrimID, ...)
476 * GFX10 LS (VertexID, RelAutoindex, UserVGPR1, InstanceID).
477 * GFX10 ES,VS (VertexID, UserVGPR0, UserVGPR1 or VSPrimID, UserVGPR2 or InstanceID)
479 bool is_ls
= shader
->selector
->type
== PIPE_SHADER_TESS_CTRL
|| shader
->key
.as_ls
;
481 if (sscreen
->info
.chip_class
>= GFX10
&& shader
->info
.uses_instanceid
)
483 else if ((is_ls
&& shader
->info
.uses_instanceid
) || legacy_vs_prim_id
)
485 else if (is_ls
|| shader
->info
.uses_instanceid
)
491 static void si_shader_ls(struct si_screen
*sscreen
, struct si_shader
*shader
)
493 struct si_pm4_state
*pm4
;
496 assert(sscreen
->info
.chip_class
<= GFX8
);
498 pm4
= si_get_shader_pm4_state(shader
);
502 va
= shader
->bo
->gpu_address
;
503 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
505 si_pm4_set_reg(pm4
, R_00B520_SPI_SHADER_PGM_LO_LS
, va
>> 8);
506 si_pm4_set_reg(pm4
, R_00B524_SPI_SHADER_PGM_HI_LS
, S_00B524_MEM_BASE(va
>> 40));
508 shader
->config
.rsrc1
= S_00B528_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
509 S_00B528_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
510 S_00B528_VGPR_COMP_CNT(si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false)) |
511 S_00B528_DX10_CLAMP(1) |
512 S_00B528_FLOAT_MODE(shader
->config
.float_mode
);
513 shader
->config
.rsrc2
= S_00B52C_USER_SGPR(si_get_num_vs_user_sgprs(SI_VS_NUM_USER_SGPR
)) |
514 S_00B52C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
517 static void si_shader_hs(struct si_screen
*sscreen
, struct si_shader
*shader
)
519 struct si_pm4_state
*pm4
;
522 pm4
= si_get_shader_pm4_state(shader
);
526 va
= shader
->bo
->gpu_address
;
527 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
529 if (sscreen
->info
.chip_class
>= GFX9
) {
530 if (sscreen
->info
.chip_class
>= GFX10
) {
531 si_pm4_set_reg(pm4
, R_00B520_SPI_SHADER_PGM_LO_LS
, va
>> 8);
532 si_pm4_set_reg(pm4
, R_00B524_SPI_SHADER_PGM_HI_LS
, S_00B524_MEM_BASE(va
>> 40));
534 si_pm4_set_reg(pm4
, R_00B410_SPI_SHADER_PGM_LO_LS
, va
>> 8);
535 si_pm4_set_reg(pm4
, R_00B414_SPI_SHADER_PGM_HI_LS
, S_00B414_MEM_BASE(va
>> 40));
538 unsigned num_user_sgprs
=
539 si_get_num_vs_user_sgprs(GFX9_TCS_NUM_USER_SGPR
);
541 shader
->config
.rsrc2
=
542 S_00B42C_USER_SGPR(num_user_sgprs
) |
543 S_00B42C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
545 if (sscreen
->info
.chip_class
>= GFX10
)
546 shader
->config
.rsrc2
|= S_00B42C_USER_SGPR_MSB_GFX10(num_user_sgprs
>> 5);
548 shader
->config
.rsrc2
|= S_00B42C_USER_SGPR_MSB_GFX9(num_user_sgprs
>> 5);
550 si_pm4_set_reg(pm4
, R_00B420_SPI_SHADER_PGM_LO_HS
, va
>> 8);
551 si_pm4_set_reg(pm4
, R_00B424_SPI_SHADER_PGM_HI_HS
, S_00B424_MEM_BASE(va
>> 40));
553 shader
->config
.rsrc2
=
554 S_00B42C_USER_SGPR(GFX6_TCS_NUM_USER_SGPR
) |
555 S_00B42C_OC_LDS_EN(1) |
556 S_00B42C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
559 si_pm4_set_reg(pm4
, R_00B428_SPI_SHADER_PGM_RSRC1_HS
,
560 S_00B428_VGPRS((shader
->config
.num_vgprs
- 1) /
561 (sscreen
->ge_wave_size
== 32 ? 8 : 4)) |
562 (sscreen
->info
.chip_class
<= GFX9
?
563 S_00B428_SGPRS((shader
->config
.num_sgprs
- 1) / 8) : 0) |
564 S_00B428_DX10_CLAMP(1) |
565 S_00B428_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
566 S_00B428_WGP_MODE(sscreen
->info
.chip_class
>= GFX10
) |
567 S_00B428_FLOAT_MODE(shader
->config
.float_mode
) |
568 S_00B428_LS_VGPR_COMP_CNT(sscreen
->info
.chip_class
>= GFX9
?
569 si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false) : 0));
571 if (sscreen
->info
.chip_class
<= GFX8
) {
572 si_pm4_set_reg(pm4
, R_00B42C_SPI_SHADER_PGM_RSRC2_HS
,
573 shader
->config
.rsrc2
);
577 static void si_emit_shader_es(struct si_context
*sctx
)
579 struct si_shader
*shader
= sctx
->queued
.named
.es
->shader
;
580 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
585 radeon_opt_set_context_reg(sctx
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
586 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE
,
587 shader
->selector
->esgs_itemsize
/ 4);
589 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
590 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
591 SI_TRACKED_VGT_TF_PARAM
,
592 shader
->vgt_tf_param
);
594 if (shader
->vgt_vertex_reuse_block_cntl
)
595 radeon_opt_set_context_reg(sctx
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
596 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL
,
597 shader
->vgt_vertex_reuse_block_cntl
);
599 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
600 sctx
->context_roll
= true;
603 static void si_shader_es(struct si_screen
*sscreen
, struct si_shader
*shader
)
605 struct si_pm4_state
*pm4
;
606 unsigned num_user_sgprs
;
607 unsigned vgpr_comp_cnt
;
611 assert(sscreen
->info
.chip_class
<= GFX8
);
613 pm4
= si_get_shader_pm4_state(shader
);
617 pm4
->atom
.emit
= si_emit_shader_es
;
618 va
= shader
->bo
->gpu_address
;
619 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
621 if (shader
->selector
->type
== PIPE_SHADER_VERTEX
) {
622 vgpr_comp_cnt
= si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false);
623 num_user_sgprs
= si_get_num_vs_user_sgprs(SI_VS_NUM_USER_SGPR
);
624 } else if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
625 vgpr_comp_cnt
= shader
->selector
->info
.uses_primid
? 3 : 2;
626 num_user_sgprs
= SI_TES_NUM_USER_SGPR
;
628 unreachable("invalid shader selector type");
630 oc_lds_en
= shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
? 1 : 0;
632 si_pm4_set_reg(pm4
, R_00B320_SPI_SHADER_PGM_LO_ES
, va
>> 8);
633 si_pm4_set_reg(pm4
, R_00B324_SPI_SHADER_PGM_HI_ES
, S_00B324_MEM_BASE(va
>> 40));
634 si_pm4_set_reg(pm4
, R_00B328_SPI_SHADER_PGM_RSRC1_ES
,
635 S_00B328_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
636 S_00B328_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
637 S_00B328_VGPR_COMP_CNT(vgpr_comp_cnt
) |
638 S_00B328_DX10_CLAMP(1) |
639 S_00B328_FLOAT_MODE(shader
->config
.float_mode
));
640 si_pm4_set_reg(pm4
, R_00B32C_SPI_SHADER_PGM_RSRC2_ES
,
641 S_00B32C_USER_SGPR(num_user_sgprs
) |
642 S_00B32C_OC_LDS_EN(oc_lds_en
) |
643 S_00B32C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
645 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
646 si_set_tesseval_regs(sscreen
, shader
->selector
, pm4
);
648 polaris_set_vgt_vertex_reuse(sscreen
, shader
->selector
, shader
, pm4
);
651 void gfx9_get_gs_info(struct si_shader_selector
*es
,
652 struct si_shader_selector
*gs
,
653 struct gfx9_gs_info
*out
)
655 unsigned gs_num_invocations
= MAX2(gs
->gs_num_invocations
, 1);
656 unsigned input_prim
= gs
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
];
657 bool uses_adjacency
= input_prim
>= PIPE_PRIM_LINES_ADJACENCY
&&
658 input_prim
<= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
;
660 /* All these are in dwords: */
661 /* We can't allow using the whole LDS, because GS waves compete with
662 * other shader stages for LDS space. */
663 const unsigned max_lds_size
= 8 * 1024;
664 const unsigned esgs_itemsize
= es
->esgs_itemsize
/ 4;
665 unsigned esgs_lds_size
;
667 /* All these are per subgroup: */
668 const unsigned max_out_prims
= 32 * 1024;
669 const unsigned max_es_verts
= 255;
670 const unsigned ideal_gs_prims
= 64;
671 unsigned max_gs_prims
, gs_prims
;
672 unsigned min_es_verts
, es_verts
, worst_case_es_verts
;
674 if (uses_adjacency
|| gs_num_invocations
> 1)
675 max_gs_prims
= 127 / gs_num_invocations
;
679 /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
680 * Make sure we don't go over the maximum value.
682 if (gs
->gs_max_out_vertices
> 0) {
683 max_gs_prims
= MIN2(max_gs_prims
,
685 (gs
->gs_max_out_vertices
* gs_num_invocations
));
687 assert(max_gs_prims
> 0);
689 /* If the primitive has adjacency, halve the number of vertices
690 * that will be reused in multiple primitives.
692 min_es_verts
= gs
->gs_input_verts_per_prim
/ (uses_adjacency
? 2 : 1);
694 gs_prims
= MIN2(ideal_gs_prims
, max_gs_prims
);
695 worst_case_es_verts
= MIN2(min_es_verts
* gs_prims
, max_es_verts
);
697 /* Compute ESGS LDS size based on the worst case number of ES vertices
698 * needed to create the target number of GS prims per subgroup.
700 esgs_lds_size
= esgs_itemsize
* worst_case_es_verts
;
702 /* If total LDS usage is too big, refactor partitions based on ratio
703 * of ESGS item sizes.
705 if (esgs_lds_size
> max_lds_size
) {
706 /* Our target GS Prims Per Subgroup was too large. Calculate
707 * the maximum number of GS Prims Per Subgroup that will fit
708 * into LDS, capped by the maximum that the hardware can support.
710 gs_prims
= MIN2((max_lds_size
/ (esgs_itemsize
* min_es_verts
)),
712 assert(gs_prims
> 0);
713 worst_case_es_verts
= MIN2(min_es_verts
* gs_prims
,
716 esgs_lds_size
= esgs_itemsize
* worst_case_es_verts
;
717 assert(esgs_lds_size
<= max_lds_size
);
720 /* Now calculate remaining ESGS information. */
722 es_verts
= MIN2(esgs_lds_size
/ esgs_itemsize
, max_es_verts
);
724 es_verts
= max_es_verts
;
726 /* Vertices for adjacency primitives are not always reused, so restore
727 * it for ES_VERTS_PER_SUBGRP.
729 min_es_verts
= gs
->gs_input_verts_per_prim
;
731 /* For normal primitives, the VGT only checks if they are past the ES
732 * verts per subgroup after allocating a full GS primitive and if they
733 * are, kick off a new subgroup. But if those additional ES verts are
734 * unique (e.g. not reused) we need to make sure there is enough LDS
735 * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
737 es_verts
-= min_es_verts
- 1;
739 out
->es_verts_per_subgroup
= es_verts
;
740 out
->gs_prims_per_subgroup
= gs_prims
;
741 out
->gs_inst_prims_in_subgroup
= gs_prims
* gs_num_invocations
;
742 out
->max_prims_per_subgroup
= out
->gs_inst_prims_in_subgroup
*
743 gs
->gs_max_out_vertices
;
744 out
->esgs_ring_size
= 4 * esgs_lds_size
;
746 assert(out
->max_prims_per_subgroup
<= max_out_prims
);
749 static void si_emit_shader_gs(struct si_context
*sctx
)
751 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
752 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
757 /* R_028A60_VGT_GSVS_RING_OFFSET_1, R_028A64_VGT_GSVS_RING_OFFSET_2
758 * R_028A68_VGT_GSVS_RING_OFFSET_3 */
759 radeon_opt_set_context_reg3(sctx
, R_028A60_VGT_GSVS_RING_OFFSET_1
,
760 SI_TRACKED_VGT_GSVS_RING_OFFSET_1
,
761 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_1
,
762 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_2
,
763 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_3
);
765 /* R_028AB0_VGT_GSVS_RING_ITEMSIZE */
766 radeon_opt_set_context_reg(sctx
, R_028AB0_VGT_GSVS_RING_ITEMSIZE
,
767 SI_TRACKED_VGT_GSVS_RING_ITEMSIZE
,
768 shader
->ctx_reg
.gs
.vgt_gsvs_ring_itemsize
);
770 /* R_028B38_VGT_GS_MAX_VERT_OUT */
771 radeon_opt_set_context_reg(sctx
, R_028B38_VGT_GS_MAX_VERT_OUT
,
772 SI_TRACKED_VGT_GS_MAX_VERT_OUT
,
773 shader
->ctx_reg
.gs
.vgt_gs_max_vert_out
);
775 /* R_028B5C_VGT_GS_VERT_ITEMSIZE, R_028B60_VGT_GS_VERT_ITEMSIZE_1
776 * R_028B64_VGT_GS_VERT_ITEMSIZE_2, R_028B68_VGT_GS_VERT_ITEMSIZE_3 */
777 radeon_opt_set_context_reg4(sctx
, R_028B5C_VGT_GS_VERT_ITEMSIZE
,
778 SI_TRACKED_VGT_GS_VERT_ITEMSIZE
,
779 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize
,
780 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_1
,
781 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_2
,
782 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_3
);
784 /* R_028B90_VGT_GS_INSTANCE_CNT */
785 radeon_opt_set_context_reg(sctx
, R_028B90_VGT_GS_INSTANCE_CNT
,
786 SI_TRACKED_VGT_GS_INSTANCE_CNT
,
787 shader
->ctx_reg
.gs
.vgt_gs_instance_cnt
);
789 if (sctx
->chip_class
>= GFX9
) {
790 /* R_028A44_VGT_GS_ONCHIP_CNTL */
791 radeon_opt_set_context_reg(sctx
, R_028A44_VGT_GS_ONCHIP_CNTL
,
792 SI_TRACKED_VGT_GS_ONCHIP_CNTL
,
793 shader
->ctx_reg
.gs
.vgt_gs_onchip_cntl
);
794 /* R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP */
795 radeon_opt_set_context_reg(sctx
, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP
,
796 SI_TRACKED_VGT_GS_MAX_PRIMS_PER_SUBGROUP
,
797 shader
->ctx_reg
.gs
.vgt_gs_max_prims_per_subgroup
);
798 /* R_028AAC_VGT_ESGS_RING_ITEMSIZE */
799 radeon_opt_set_context_reg(sctx
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
800 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE
,
801 shader
->ctx_reg
.gs
.vgt_esgs_ring_itemsize
);
803 if (shader
->key
.part
.gs
.es
->type
== PIPE_SHADER_TESS_EVAL
)
804 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
805 SI_TRACKED_VGT_TF_PARAM
,
806 shader
->vgt_tf_param
);
807 if (shader
->vgt_vertex_reuse_block_cntl
)
808 radeon_opt_set_context_reg(sctx
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
809 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL
,
810 shader
->vgt_vertex_reuse_block_cntl
);
813 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
814 sctx
->context_roll
= true;
817 static void si_shader_gs(struct si_screen
*sscreen
, struct si_shader
*shader
)
819 struct si_shader_selector
*sel
= shader
->selector
;
820 const ubyte
*num_components
= sel
->info
.num_stream_output_components
;
821 unsigned gs_num_invocations
= sel
->gs_num_invocations
;
822 struct si_pm4_state
*pm4
;
824 unsigned max_stream
= sel
->max_gs_stream
;
827 pm4
= si_get_shader_pm4_state(shader
);
831 pm4
->atom
.emit
= si_emit_shader_gs
;
833 offset
= num_components
[0] * sel
->gs_max_out_vertices
;
834 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_1
= offset
;
837 offset
+= num_components
[1] * sel
->gs_max_out_vertices
;
838 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_2
= offset
;
841 offset
+= num_components
[2] * sel
->gs_max_out_vertices
;
842 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_3
= offset
;
845 offset
+= num_components
[3] * sel
->gs_max_out_vertices
;
846 shader
->ctx_reg
.gs
.vgt_gsvs_ring_itemsize
= offset
;
848 /* The GSVS_RING_ITEMSIZE register takes 15 bits */
849 assert(offset
< (1 << 15));
851 shader
->ctx_reg
.gs
.vgt_gs_max_vert_out
= sel
->gs_max_out_vertices
;
853 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize
= num_components
[0];
854 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_1
= (max_stream
>= 1) ? num_components
[1] : 0;
855 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_2
= (max_stream
>= 2) ? num_components
[2] : 0;
856 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_3
= (max_stream
>= 3) ? num_components
[3] : 0;
858 shader
->ctx_reg
.gs
.vgt_gs_instance_cnt
= S_028B90_CNT(MIN2(gs_num_invocations
, 127)) |
859 S_028B90_ENABLE(gs_num_invocations
> 0);
861 va
= shader
->bo
->gpu_address
;
862 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
864 if (sscreen
->info
.chip_class
>= GFX9
) {
865 unsigned input_prim
= sel
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
];
866 unsigned es_type
= shader
->key
.part
.gs
.es
->type
;
867 unsigned es_vgpr_comp_cnt
, gs_vgpr_comp_cnt
;
869 if (es_type
== PIPE_SHADER_VERTEX
) {
870 es_vgpr_comp_cnt
= si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false);
871 } else if (es_type
== PIPE_SHADER_TESS_EVAL
)
872 es_vgpr_comp_cnt
= shader
->key
.part
.gs
.es
->info
.uses_primid
? 3 : 2;
874 unreachable("invalid shader selector type");
876 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
877 * VGPR[0:4] are always loaded.
879 if (sel
->info
.uses_invocationid
)
880 gs_vgpr_comp_cnt
= 3; /* VGPR3 contains InvocationID. */
881 else if (sel
->info
.uses_primid
)
882 gs_vgpr_comp_cnt
= 2; /* VGPR2 contains PrimitiveID. */
883 else if (input_prim
>= PIPE_PRIM_TRIANGLES
)
884 gs_vgpr_comp_cnt
= 1; /* VGPR1 contains offsets 2, 3 */
886 gs_vgpr_comp_cnt
= 0; /* VGPR0 contains offsets 0, 1 */
888 unsigned num_user_sgprs
;
889 if (es_type
== PIPE_SHADER_VERTEX
)
890 num_user_sgprs
= si_get_num_vs_user_sgprs(GFX9_VSGS_NUM_USER_SGPR
);
892 num_user_sgprs
= GFX9_TESGS_NUM_USER_SGPR
;
894 if (sscreen
->info
.chip_class
>= GFX10
) {
895 si_pm4_set_reg(pm4
, R_00B320_SPI_SHADER_PGM_LO_ES
, va
>> 8);
896 si_pm4_set_reg(pm4
, R_00B324_SPI_SHADER_PGM_HI_ES
, S_00B324_MEM_BASE(va
>> 40));
898 si_pm4_set_reg(pm4
, R_00B210_SPI_SHADER_PGM_LO_ES
, va
>> 8);
899 si_pm4_set_reg(pm4
, R_00B214_SPI_SHADER_PGM_HI_ES
, S_00B214_MEM_BASE(va
>> 40));
903 S_00B228_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
904 S_00B228_DX10_CLAMP(1) |
905 S_00B228_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
906 S_00B228_WGP_MODE(sscreen
->info
.chip_class
>= GFX10
) |
907 S_00B228_FLOAT_MODE(shader
->config
.float_mode
) |
908 S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt
);
910 S_00B22C_USER_SGPR(num_user_sgprs
) |
911 S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt
) |
912 S_00B22C_OC_LDS_EN(es_type
== PIPE_SHADER_TESS_EVAL
) |
913 S_00B22C_LDS_SIZE(shader
->config
.lds_size
) |
914 S_00B22C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
916 if (sscreen
->info
.chip_class
>= GFX10
) {
917 rsrc2
|= S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs
>> 5);
919 rsrc1
|= S_00B228_SGPRS((shader
->config
.num_sgprs
- 1) / 8);
920 rsrc2
|= S_00B22C_USER_SGPR_MSB_GFX9(num_user_sgprs
>> 5);
923 si_pm4_set_reg(pm4
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
, rsrc1
);
924 si_pm4_set_reg(pm4
, R_00B22C_SPI_SHADER_PGM_RSRC2_GS
, rsrc2
);
926 shader
->ctx_reg
.gs
.vgt_gs_onchip_cntl
=
927 S_028A44_ES_VERTS_PER_SUBGRP(shader
->gs_info
.es_verts_per_subgroup
) |
928 S_028A44_GS_PRIMS_PER_SUBGRP(shader
->gs_info
.gs_prims_per_subgroup
) |
929 S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader
->gs_info
.gs_inst_prims_in_subgroup
);
930 shader
->ctx_reg
.gs
.vgt_gs_max_prims_per_subgroup
=
931 S_028A94_MAX_PRIMS_PER_SUBGROUP(shader
->gs_info
.max_prims_per_subgroup
);
932 shader
->ctx_reg
.gs
.vgt_esgs_ring_itemsize
=
933 shader
->key
.part
.gs
.es
->esgs_itemsize
/ 4;
935 if (es_type
== PIPE_SHADER_TESS_EVAL
)
936 si_set_tesseval_regs(sscreen
, shader
->key
.part
.gs
.es
, pm4
);
938 polaris_set_vgt_vertex_reuse(sscreen
, shader
->key
.part
.gs
.es
,
941 si_pm4_set_reg(pm4
, R_00B220_SPI_SHADER_PGM_LO_GS
, va
>> 8);
942 si_pm4_set_reg(pm4
, R_00B224_SPI_SHADER_PGM_HI_GS
, S_00B224_MEM_BASE(va
>> 40));
944 si_pm4_set_reg(pm4
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
,
945 S_00B228_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
946 S_00B228_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
947 S_00B228_DX10_CLAMP(1) |
948 S_00B228_FLOAT_MODE(shader
->config
.float_mode
));
949 si_pm4_set_reg(pm4
, R_00B22C_SPI_SHADER_PGM_RSRC2_GS
,
950 S_00B22C_USER_SGPR(GFX6_GS_NUM_USER_SGPR
) |
951 S_00B22C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
955 /* Common tail code for NGG primitive shaders. */
956 static void gfx10_emit_shader_ngg_tail(struct si_context
*sctx
,
957 struct si_shader
*shader
,
958 unsigned initial_cdw
)
960 radeon_opt_set_context_reg(sctx
, R_0287FC_GE_MAX_OUTPUT_PER_SUBGROUP
,
961 SI_TRACKED_GE_MAX_OUTPUT_PER_SUBGROUP
,
962 shader
->ctx_reg
.ngg
.ge_max_output_per_subgroup
);
963 radeon_opt_set_context_reg(sctx
, R_028B4C_GE_NGG_SUBGRP_CNTL
,
964 SI_TRACKED_GE_NGG_SUBGRP_CNTL
,
965 shader
->ctx_reg
.ngg
.ge_ngg_subgrp_cntl
);
966 radeon_opt_set_context_reg(sctx
, R_028A84_VGT_PRIMITIVEID_EN
,
967 SI_TRACKED_VGT_PRIMITIVEID_EN
,
968 shader
->ctx_reg
.ngg
.vgt_primitiveid_en
);
969 radeon_opt_set_context_reg(sctx
, R_028A44_VGT_GS_ONCHIP_CNTL
,
970 SI_TRACKED_VGT_GS_ONCHIP_CNTL
,
971 shader
->ctx_reg
.ngg
.vgt_gs_onchip_cntl
);
972 radeon_opt_set_context_reg(sctx
, R_028B90_VGT_GS_INSTANCE_CNT
,
973 SI_TRACKED_VGT_GS_INSTANCE_CNT
,
974 shader
->ctx_reg
.ngg
.vgt_gs_instance_cnt
);
975 radeon_opt_set_context_reg(sctx
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
976 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE
,
977 shader
->ctx_reg
.ngg
.vgt_esgs_ring_itemsize
);
978 radeon_opt_set_context_reg(sctx
, R_0286C4_SPI_VS_OUT_CONFIG
,
979 SI_TRACKED_SPI_VS_OUT_CONFIG
,
980 shader
->ctx_reg
.ngg
.spi_vs_out_config
);
981 radeon_opt_set_context_reg2(sctx
, R_028708_SPI_SHADER_IDX_FORMAT
,
982 SI_TRACKED_SPI_SHADER_IDX_FORMAT
,
983 shader
->ctx_reg
.ngg
.spi_shader_idx_format
,
984 shader
->ctx_reg
.ngg
.spi_shader_pos_format
);
985 radeon_opt_set_context_reg(sctx
, R_028818_PA_CL_VTE_CNTL
,
986 SI_TRACKED_PA_CL_VTE_CNTL
,
987 shader
->ctx_reg
.ngg
.pa_cl_vte_cntl
);
988 radeon_opt_set_context_reg(sctx
, R_028838_PA_CL_NGG_CNTL
,
989 SI_TRACKED_PA_CL_NGG_CNTL
,
990 shader
->ctx_reg
.ngg
.pa_cl_ngg_cntl
);
992 radeon_opt_set_context_reg_rmw(sctx
, R_02881C_PA_CL_VS_OUT_CNTL
,
993 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS
,
994 shader
->pa_cl_vs_out_cntl
,
995 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS_MASK
);
997 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
998 sctx
->context_roll
= true;
1001 static void gfx10_emit_shader_ngg_notess_nogs(struct si_context
*sctx
)
1003 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1004 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1009 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1012 static void gfx10_emit_shader_ngg_tess_nogs(struct si_context
*sctx
)
1014 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1015 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1020 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
1021 SI_TRACKED_VGT_TF_PARAM
,
1022 shader
->vgt_tf_param
);
1024 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1027 static void gfx10_emit_shader_ngg_notess_gs(struct si_context
*sctx
)
1029 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1030 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1035 radeon_opt_set_context_reg(sctx
, R_028B38_VGT_GS_MAX_VERT_OUT
,
1036 SI_TRACKED_VGT_GS_MAX_VERT_OUT
,
1037 shader
->ctx_reg
.ngg
.vgt_gs_max_vert_out
);
1039 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1042 static void gfx10_emit_shader_ngg_tess_gs(struct si_context
*sctx
)
1044 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1045 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1050 radeon_opt_set_context_reg(sctx
, R_028B38_VGT_GS_MAX_VERT_OUT
,
1051 SI_TRACKED_VGT_GS_MAX_VERT_OUT
,
1052 shader
->ctx_reg
.ngg
.vgt_gs_max_vert_out
);
1053 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
1054 SI_TRACKED_VGT_TF_PARAM
,
1055 shader
->vgt_tf_param
);
1057 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1060 unsigned si_get_input_prim(const struct si_shader_selector
*gs
)
1062 if (gs
->type
== PIPE_SHADER_GEOMETRY
)
1063 return gs
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
];
1065 if (gs
->type
== PIPE_SHADER_TESS_EVAL
) {
1066 if (gs
->info
.properties
[TGSI_PROPERTY_TES_POINT_MODE
])
1067 return PIPE_PRIM_POINTS
;
1068 if (gs
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] == PIPE_PRIM_LINES
)
1069 return PIPE_PRIM_LINES
;
1070 return PIPE_PRIM_TRIANGLES
;
1073 /* TODO: Set this correctly if the primitive type is set in the shader key. */
1074 return PIPE_PRIM_TRIANGLES
; /* worst case for all callers */
1077 static unsigned si_get_vs_out_cntl(const struct si_shader_selector
*sel
, bool ngg
)
1080 sel
->info
.writes_psize
|| (sel
->info
.writes_edgeflag
&& !ngg
) ||
1081 sel
->info
.writes_layer
|| sel
->info
.writes_viewport_index
;
1082 return S_02881C_USE_VTX_POINT_SIZE(sel
->info
.writes_psize
) |
1083 S_02881C_USE_VTX_EDGE_FLAG(sel
->info
.writes_edgeflag
&& !ngg
) |
1084 S_02881C_USE_VTX_RENDER_TARGET_INDX(sel
->info
.writes_layer
) |
1085 S_02881C_USE_VTX_VIEWPORT_INDX(sel
->info
.writes_viewport_index
) |
1086 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena
) |
1087 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena
);
1091 * Prepare the PM4 image for \p shader, which will run as a merged ESGS shader
1094 static void gfx10_shader_ngg(struct si_screen
*sscreen
, struct si_shader
*shader
)
1096 const struct si_shader_selector
*gs_sel
= shader
->selector
;
1097 const struct tgsi_shader_info
*gs_info
= &gs_sel
->info
;
1098 enum pipe_shader_type gs_type
= shader
->selector
->type
;
1099 const struct si_shader_selector
*es_sel
=
1100 shader
->previous_stage_sel
? shader
->previous_stage_sel
: shader
->selector
;
1101 const struct tgsi_shader_info
*es_info
= &es_sel
->info
;
1102 enum pipe_shader_type es_type
= es_sel
->type
;
1103 unsigned num_user_sgprs
;
1104 unsigned nparams
, es_vgpr_comp_cnt
, gs_vgpr_comp_cnt
;
1106 unsigned window_space
=
1107 gs_info
->properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
];
1108 bool es_enable_prim_id
= shader
->key
.mono
.u
.vs_export_prim_id
|| es_info
->uses_primid
;
1109 unsigned gs_num_invocations
= MAX2(gs_sel
->gs_num_invocations
, 1);
1110 unsigned input_prim
= si_get_input_prim(gs_sel
);
1111 bool break_wave_at_eoi
= false;
1112 struct si_pm4_state
*pm4
= si_get_shader_pm4_state(shader
);
1116 if (es_type
== PIPE_SHADER_TESS_EVAL
) {
1117 pm4
->atom
.emit
= gs_type
== PIPE_SHADER_GEOMETRY
? gfx10_emit_shader_ngg_tess_gs
1118 : gfx10_emit_shader_ngg_tess_nogs
;
1120 pm4
->atom
.emit
= gs_type
== PIPE_SHADER_GEOMETRY
? gfx10_emit_shader_ngg_notess_gs
1121 : gfx10_emit_shader_ngg_notess_nogs
;
1124 va
= shader
->bo
->gpu_address
;
1125 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
1127 if (es_type
== PIPE_SHADER_VERTEX
) {
1128 es_vgpr_comp_cnt
= si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false);
1130 if (es_info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
]) {
1131 num_user_sgprs
= SI_SGPR_VS_BLIT_DATA
+
1132 es_info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
];
1134 num_user_sgprs
= si_get_num_vs_user_sgprs(GFX9_VSGS_NUM_USER_SGPR
);
1137 assert(es_type
== PIPE_SHADER_TESS_EVAL
);
1138 es_vgpr_comp_cnt
= es_enable_prim_id
? 3 : 2;
1139 num_user_sgprs
= GFX9_TESGS_NUM_USER_SGPR
;
1141 if (es_enable_prim_id
|| gs_info
->uses_primid
)
1142 break_wave_at_eoi
= true;
1145 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
1146 * VGPR[0:4] are always loaded.
1148 * Vertex shaders always need to load VGPR3, because they need to
1149 * pass edge flags for decomposed primitives (such as quads) to the PA
1150 * for the GL_LINE polygon mode to skip rendering lines on inner edges.
1152 if (gs_info
->uses_invocationid
|| gs_type
== PIPE_SHADER_VERTEX
)
1153 gs_vgpr_comp_cnt
= 3; /* VGPR3 contains InvocationID, edge flags. */
1154 else if (gs_info
->uses_primid
)
1155 gs_vgpr_comp_cnt
= 2; /* VGPR2 contains PrimitiveID. */
1156 else if (input_prim
>= PIPE_PRIM_TRIANGLES
)
1157 gs_vgpr_comp_cnt
= 1; /* VGPR1 contains offsets 2, 3 */
1159 gs_vgpr_comp_cnt
= 0; /* VGPR0 contains offsets 0, 1 */
1161 si_pm4_set_reg(pm4
, R_00B320_SPI_SHADER_PGM_LO_ES
, va
>> 8);
1162 si_pm4_set_reg(pm4
, R_00B324_SPI_SHADER_PGM_HI_ES
, va
>> 40);
1163 si_pm4_set_reg(pm4
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
,
1164 S_00B228_VGPRS((shader
->config
.num_vgprs
- 1) /
1165 (sscreen
->ge_wave_size
== 32 ? 8 : 4)) |
1166 S_00B228_FLOAT_MODE(shader
->config
.float_mode
) |
1167 S_00B228_DX10_CLAMP(1) |
1168 S_00B228_MEM_ORDERED(1) |
1169 S_00B228_WGP_MODE(1) |
1170 S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt
));
1171 si_pm4_set_reg(pm4
, R_00B22C_SPI_SHADER_PGM_RSRC2_GS
,
1172 S_00B22C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0) |
1173 S_00B22C_USER_SGPR(num_user_sgprs
) |
1174 S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt
) |
1175 S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs
>> 5) |
1176 S_00B22C_OC_LDS_EN(es_type
== PIPE_SHADER_TESS_EVAL
) |
1177 S_00B22C_LDS_SIZE(shader
->config
.lds_size
));
1179 nparams
= MAX2(shader
->info
.nr_param_exports
, 1);
1180 shader
->ctx_reg
.ngg
.spi_vs_out_config
=
1181 S_0286C4_VS_EXPORT_COUNT(nparams
- 1) |
1182 S_0286C4_NO_PC_EXPORT(shader
->info
.nr_param_exports
== 0);
1184 shader
->ctx_reg
.ngg
.spi_shader_idx_format
=
1185 S_028708_IDX0_EXPORT_FORMAT(V_028708_SPI_SHADER_1COMP
);
1186 shader
->ctx_reg
.ngg
.spi_shader_pos_format
=
1187 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP
) |
1188 S_02870C_POS1_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 1 ?
1189 V_02870C_SPI_SHADER_4COMP
:
1190 V_02870C_SPI_SHADER_NONE
) |
1191 S_02870C_POS2_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 2 ?
1192 V_02870C_SPI_SHADER_4COMP
:
1193 V_02870C_SPI_SHADER_NONE
) |
1194 S_02870C_POS3_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 3 ?
1195 V_02870C_SPI_SHADER_4COMP
:
1196 V_02870C_SPI_SHADER_NONE
);
1198 shader
->ctx_reg
.ngg
.vgt_primitiveid_en
=
1199 S_028A84_PRIMITIVEID_EN(es_enable_prim_id
) |
1200 S_028A84_NGG_DISABLE_PROVOK_REUSE(shader
->key
.mono
.u
.vs_export_prim_id
||
1201 gs_sel
->info
.writes_primid
);
1203 if (gs_type
== PIPE_SHADER_GEOMETRY
) {
1204 shader
->ctx_reg
.ngg
.vgt_esgs_ring_itemsize
= es_sel
->esgs_itemsize
/ 4;
1205 shader
->ctx_reg
.ngg
.vgt_gs_max_vert_out
= gs_sel
->gs_max_out_vertices
;
1207 shader
->ctx_reg
.ngg
.vgt_esgs_ring_itemsize
= 1;
1210 if (es_type
== PIPE_SHADER_TESS_EVAL
)
1211 si_set_tesseval_regs(sscreen
, es_sel
, pm4
);
1213 shader
->ctx_reg
.ngg
.vgt_gs_onchip_cntl
=
1214 S_028A44_ES_VERTS_PER_SUBGRP(shader
->ngg
.hw_max_esverts
) |
1215 S_028A44_GS_PRIMS_PER_SUBGRP(shader
->ngg
.max_gsprims
) |
1216 S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader
->ngg
.max_gsprims
* gs_num_invocations
);
1217 shader
->ctx_reg
.ngg
.ge_max_output_per_subgroup
=
1218 S_0287FC_MAX_VERTS_PER_SUBGROUP(shader
->ngg
.max_out_verts
);
1219 shader
->ctx_reg
.ngg
.ge_ngg_subgrp_cntl
=
1220 S_028B4C_PRIM_AMP_FACTOR(shader
->ngg
.prim_amp_factor
) |
1221 S_028B4C_THDS_PER_SUBGRP(0); /* for fast launch */
1222 shader
->ctx_reg
.ngg
.vgt_gs_instance_cnt
=
1223 S_028B90_CNT(gs_num_invocations
) |
1224 S_028B90_ENABLE(gs_num_invocations
> 1) |
1225 S_028B90_EN_MAX_VERT_OUT_PER_GS_INSTANCE(
1226 shader
->ngg
.max_vert_out_per_gs_instance
);
1228 /* Always output hw-generated edge flags and pass them via the prim
1229 * export to prevent drawing lines on internal edges of decomposed
1230 * primitives (such as quads) with polygon mode = lines. Only VS needs
1233 shader
->ctx_reg
.ngg
.pa_cl_ngg_cntl
=
1234 S_028838_INDEX_BUF_EDGE_FLAG_ENA(gs_type
== PIPE_SHADER_VERTEX
);
1235 shader
->pa_cl_vs_out_cntl
= si_get_vs_out_cntl(gs_sel
, true);
1238 S_03096C_PRIM_GRP_SIZE(shader
->ngg
.max_gsprims
) |
1239 S_03096C_VERT_GRP_SIZE(256) | /* 256 = disable vertex grouping */
1240 S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi
);
1242 /* Bug workaround for a possible hang with non-tessellation cases.
1243 * Tessellation always sets GE_CNTL.VERT_GRP_SIZE = 0
1245 * Requirement: GE_CNTL.VERT_GRP_SIZE = VGT_GS_ONCHIP_CNTL.ES_VERTS_PER_SUBGRP - 5
1247 if ((sscreen
->info
.family
== CHIP_NAVI10
||
1248 sscreen
->info
.family
== CHIP_NAVI12
||
1249 sscreen
->info
.family
== CHIP_NAVI14
) &&
1250 (es_type
== PIPE_SHADER_VERTEX
|| gs_type
== PIPE_SHADER_VERTEX
) && /* = no tess */
1251 shader
->ngg
.hw_max_esverts
!= 256) {
1252 shader
->ge_cntl
&= C_03096C_VERT_GRP_SIZE
;
1254 if (shader
->ngg
.hw_max_esverts
> 5) {
1256 S_03096C_VERT_GRP_SIZE(shader
->ngg
.hw_max_esverts
- 5);
1261 shader
->ctx_reg
.ngg
.pa_cl_vte_cntl
=
1262 S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
1264 shader
->ctx_reg
.ngg
.pa_cl_vte_cntl
=
1265 S_028818_VTX_W0_FMT(1) |
1266 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
1267 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
1268 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
1272 static void si_emit_shader_vs(struct si_context
*sctx
)
1274 struct si_shader
*shader
= sctx
->queued
.named
.vs
->shader
;
1275 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1280 radeon_opt_set_context_reg(sctx
, R_028A40_VGT_GS_MODE
,
1281 SI_TRACKED_VGT_GS_MODE
,
1282 shader
->ctx_reg
.vs
.vgt_gs_mode
);
1283 radeon_opt_set_context_reg(sctx
, R_028A84_VGT_PRIMITIVEID_EN
,
1284 SI_TRACKED_VGT_PRIMITIVEID_EN
,
1285 shader
->ctx_reg
.vs
.vgt_primitiveid_en
);
1287 if (sctx
->chip_class
<= GFX8
) {
1288 radeon_opt_set_context_reg(sctx
, R_028AB4_VGT_REUSE_OFF
,
1289 SI_TRACKED_VGT_REUSE_OFF
,
1290 shader
->ctx_reg
.vs
.vgt_reuse_off
);
1293 radeon_opt_set_context_reg(sctx
, R_0286C4_SPI_VS_OUT_CONFIG
,
1294 SI_TRACKED_SPI_VS_OUT_CONFIG
,
1295 shader
->ctx_reg
.vs
.spi_vs_out_config
);
1297 radeon_opt_set_context_reg(sctx
, R_02870C_SPI_SHADER_POS_FORMAT
,
1298 SI_TRACKED_SPI_SHADER_POS_FORMAT
,
1299 shader
->ctx_reg
.vs
.spi_shader_pos_format
);
1301 radeon_opt_set_context_reg(sctx
, R_028818_PA_CL_VTE_CNTL
,
1302 SI_TRACKED_PA_CL_VTE_CNTL
,
1303 shader
->ctx_reg
.vs
.pa_cl_vte_cntl
);
1305 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
1306 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
1307 SI_TRACKED_VGT_TF_PARAM
,
1308 shader
->vgt_tf_param
);
1310 if (shader
->vgt_vertex_reuse_block_cntl
)
1311 radeon_opt_set_context_reg(sctx
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
1312 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL
,
1313 shader
->vgt_vertex_reuse_block_cntl
);
1315 /* Required programming for tessellation. (legacy pipeline only) */
1316 if (sctx
->chip_class
== GFX10
&&
1317 shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
1318 radeon_opt_set_context_reg(sctx
, R_028A44_VGT_GS_ONCHIP_CNTL
,
1319 SI_TRACKED_VGT_GS_ONCHIP_CNTL
,
1320 S_028A44_ES_VERTS_PER_SUBGRP(250) |
1321 S_028A44_GS_PRIMS_PER_SUBGRP(126) |
1322 S_028A44_GS_INST_PRIMS_IN_SUBGRP(126));
1325 if (sctx
->chip_class
>= GFX10
) {
1326 radeon_opt_set_context_reg_rmw(sctx
, R_02881C_PA_CL_VS_OUT_CNTL
,
1327 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS
,
1328 shader
->pa_cl_vs_out_cntl
,
1329 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS_MASK
);
1332 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
1333 sctx
->context_roll
= true;
1337 * Compute the state for \p shader, which will run as a vertex shader on the
1340 * If \p gs is non-NULL, it points to the geometry shader for which this shader
1341 * is the copy shader.
1343 static void si_shader_vs(struct si_screen
*sscreen
, struct si_shader
*shader
,
1344 struct si_shader_selector
*gs
)
1346 const struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1347 struct si_pm4_state
*pm4
;
1348 unsigned num_user_sgprs
, vgpr_comp_cnt
;
1350 unsigned nparams
, oc_lds_en
;
1351 unsigned window_space
=
1352 info
->properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
];
1353 bool enable_prim_id
= shader
->key
.mono
.u
.vs_export_prim_id
|| info
->uses_primid
;
1355 pm4
= si_get_shader_pm4_state(shader
);
1359 pm4
->atom
.emit
= si_emit_shader_vs
;
1361 /* We always write VGT_GS_MODE in the VS state, because every switch
1362 * between different shader pipelines involving a different GS or no
1363 * GS at all involves a switch of the VS (different GS use different
1364 * copy shaders). On the other hand, when the API switches from a GS to
1365 * no GS and then back to the same GS used originally, the GS state is
1369 unsigned mode
= V_028A40_GS_OFF
;
1371 /* PrimID needs GS scenario A. */
1373 mode
= V_028A40_GS_SCENARIO_A
;
1375 shader
->ctx_reg
.vs
.vgt_gs_mode
= S_028A40_MODE(mode
);
1376 shader
->ctx_reg
.vs
.vgt_primitiveid_en
= enable_prim_id
;
1378 shader
->ctx_reg
.vs
.vgt_gs_mode
= ac_vgt_gs_mode(gs
->gs_max_out_vertices
,
1379 sscreen
->info
.chip_class
);
1380 shader
->ctx_reg
.vs
.vgt_primitiveid_en
= 0;
1383 if (sscreen
->info
.chip_class
<= GFX8
) {
1384 /* Reuse needs to be set off if we write oViewport. */
1385 shader
->ctx_reg
.vs
.vgt_reuse_off
=
1386 S_028AB4_REUSE_OFF(info
->writes_viewport_index
);
1389 va
= shader
->bo
->gpu_address
;
1390 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
1393 vgpr_comp_cnt
= 0; /* only VertexID is needed for GS-COPY. */
1394 num_user_sgprs
= SI_GSCOPY_NUM_USER_SGPR
;
1395 } else if (shader
->selector
->type
== PIPE_SHADER_VERTEX
) {
1396 vgpr_comp_cnt
= si_get_vs_vgpr_comp_cnt(sscreen
, shader
, enable_prim_id
);
1398 if (info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
]) {
1399 num_user_sgprs
= SI_SGPR_VS_BLIT_DATA
+
1400 info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
];
1402 num_user_sgprs
= si_get_num_vs_user_sgprs(SI_VS_NUM_USER_SGPR
);
1404 } else if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
1405 vgpr_comp_cnt
= enable_prim_id
? 3 : 2;
1406 num_user_sgprs
= SI_TES_NUM_USER_SGPR
;
1408 unreachable("invalid shader selector type");
1410 /* VS is required to export at least one param. */
1411 nparams
= MAX2(shader
->info
.nr_param_exports
, 1);
1412 shader
->ctx_reg
.vs
.spi_vs_out_config
= S_0286C4_VS_EXPORT_COUNT(nparams
- 1);
1414 if (sscreen
->info
.chip_class
>= GFX10
) {
1415 shader
->ctx_reg
.vs
.spi_vs_out_config
|=
1416 S_0286C4_NO_PC_EXPORT(shader
->info
.nr_param_exports
== 0);
1419 shader
->ctx_reg
.vs
.spi_shader_pos_format
=
1420 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP
) |
1421 S_02870C_POS1_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 1 ?
1422 V_02870C_SPI_SHADER_4COMP
:
1423 V_02870C_SPI_SHADER_NONE
) |
1424 S_02870C_POS2_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 2 ?
1425 V_02870C_SPI_SHADER_4COMP
:
1426 V_02870C_SPI_SHADER_NONE
) |
1427 S_02870C_POS3_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 3 ?
1428 V_02870C_SPI_SHADER_4COMP
:
1429 V_02870C_SPI_SHADER_NONE
);
1430 shader
->pa_cl_vs_out_cntl
= si_get_vs_out_cntl(shader
->selector
, false);
1432 oc_lds_en
= shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
? 1 : 0;
1434 si_pm4_set_reg(pm4
, R_00B120_SPI_SHADER_PGM_LO_VS
, va
>> 8);
1435 si_pm4_set_reg(pm4
, R_00B124_SPI_SHADER_PGM_HI_VS
, S_00B124_MEM_BASE(va
>> 40));
1437 uint32_t rsrc1
= S_00B128_VGPRS((shader
->config
.num_vgprs
- 1) /
1438 (sscreen
->ge_wave_size
== 32 ? 8 : 4)) |
1439 S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt
) |
1440 S_00B128_DX10_CLAMP(1) |
1441 S_00B128_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
1442 S_00B128_FLOAT_MODE(shader
->config
.float_mode
);
1443 uint32_t rsrc2
= S_00B12C_USER_SGPR(num_user_sgprs
) |
1444 S_00B12C_OC_LDS_EN(oc_lds_en
) |
1445 S_00B12C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
1447 if (sscreen
->info
.chip_class
<= GFX9
)
1448 rsrc1
|= S_00B128_SGPRS((shader
->config
.num_sgprs
- 1) / 8);
1450 if (!sscreen
->use_ngg_streamout
) {
1451 rsrc2
|= S_00B12C_SO_BASE0_EN(!!shader
->selector
->so
.stride
[0]) |
1452 S_00B12C_SO_BASE1_EN(!!shader
->selector
->so
.stride
[1]) |
1453 S_00B12C_SO_BASE2_EN(!!shader
->selector
->so
.stride
[2]) |
1454 S_00B12C_SO_BASE3_EN(!!shader
->selector
->so
.stride
[3]) |
1455 S_00B12C_SO_EN(!!shader
->selector
->so
.num_outputs
);
1458 si_pm4_set_reg(pm4
, R_00B128_SPI_SHADER_PGM_RSRC1_VS
, rsrc1
);
1459 si_pm4_set_reg(pm4
, R_00B12C_SPI_SHADER_PGM_RSRC2_VS
, rsrc2
);
1462 shader
->ctx_reg
.vs
.pa_cl_vte_cntl
=
1463 S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
1465 shader
->ctx_reg
.vs
.pa_cl_vte_cntl
=
1466 S_028818_VTX_W0_FMT(1) |
1467 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
1468 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
1469 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
1471 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
1472 si_set_tesseval_regs(sscreen
, shader
->selector
, pm4
);
1474 polaris_set_vgt_vertex_reuse(sscreen
, shader
->selector
, shader
, pm4
);
1477 static unsigned si_get_ps_num_interp(struct si_shader
*ps
)
1479 struct tgsi_shader_info
*info
= &ps
->selector
->info
;
1480 unsigned num_colors
= !!(info
->colors_read
& 0x0f) +
1481 !!(info
->colors_read
& 0xf0);
1482 unsigned num_interp
= ps
->selector
->info
.num_inputs
+
1483 (ps
->key
.part
.ps
.prolog
.color_two_side
? num_colors
: 0);
1485 assert(num_interp
<= 32);
1486 return MIN2(num_interp
, 32);
1489 static unsigned si_get_spi_shader_col_format(struct si_shader
*shader
)
1491 unsigned value
= shader
->key
.part
.ps
.epilog
.spi_shader_col_format
;
1492 unsigned i
, num_targets
= (util_last_bit(value
) + 3) / 4;
1494 /* If the i-th target format is set, all previous target formats must
1495 * be non-zero to avoid hangs.
1497 for (i
= 0; i
< num_targets
; i
++)
1498 if (!(value
& (0xf << (i
* 4))))
1499 value
|= V_028714_SPI_SHADER_32_R
<< (i
* 4);
1504 static void si_emit_shader_ps(struct si_context
*sctx
)
1506 struct si_shader
*shader
= sctx
->queued
.named
.ps
->shader
;
1507 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1512 /* R_0286CC_SPI_PS_INPUT_ENA, R_0286D0_SPI_PS_INPUT_ADDR*/
1513 radeon_opt_set_context_reg2(sctx
, R_0286CC_SPI_PS_INPUT_ENA
,
1514 SI_TRACKED_SPI_PS_INPUT_ENA
,
1515 shader
->ctx_reg
.ps
.spi_ps_input_ena
,
1516 shader
->ctx_reg
.ps
.spi_ps_input_addr
);
1518 radeon_opt_set_context_reg(sctx
, R_0286E0_SPI_BARYC_CNTL
,
1519 SI_TRACKED_SPI_BARYC_CNTL
,
1520 shader
->ctx_reg
.ps
.spi_baryc_cntl
);
1521 radeon_opt_set_context_reg(sctx
, R_0286D8_SPI_PS_IN_CONTROL
,
1522 SI_TRACKED_SPI_PS_IN_CONTROL
,
1523 shader
->ctx_reg
.ps
.spi_ps_in_control
);
1525 /* R_028710_SPI_SHADER_Z_FORMAT, R_028714_SPI_SHADER_COL_FORMAT */
1526 radeon_opt_set_context_reg2(sctx
, R_028710_SPI_SHADER_Z_FORMAT
,
1527 SI_TRACKED_SPI_SHADER_Z_FORMAT
,
1528 shader
->ctx_reg
.ps
.spi_shader_z_format
,
1529 shader
->ctx_reg
.ps
.spi_shader_col_format
);
1531 radeon_opt_set_context_reg(sctx
, R_02823C_CB_SHADER_MASK
,
1532 SI_TRACKED_CB_SHADER_MASK
,
1533 shader
->ctx_reg
.ps
.cb_shader_mask
);
1535 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
1536 sctx
->context_roll
= true;
1539 static void si_shader_ps(struct si_screen
*sscreen
, struct si_shader
*shader
)
1541 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
1542 struct si_pm4_state
*pm4
;
1543 unsigned spi_ps_in_control
, spi_shader_col_format
, cb_shader_mask
;
1544 unsigned spi_baryc_cntl
= S_0286E0_FRONT_FACE_ALL_BITS(1);
1546 unsigned input_ena
= shader
->config
.spi_ps_input_ena
;
1548 /* we need to enable at least one of them, otherwise we hang the GPU */
1549 assert(G_0286CC_PERSP_SAMPLE_ENA(input_ena
) ||
1550 G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
1551 G_0286CC_PERSP_CENTROID_ENA(input_ena
) ||
1552 G_0286CC_PERSP_PULL_MODEL_ENA(input_ena
) ||
1553 G_0286CC_LINEAR_SAMPLE_ENA(input_ena
) ||
1554 G_0286CC_LINEAR_CENTER_ENA(input_ena
) ||
1555 G_0286CC_LINEAR_CENTROID_ENA(input_ena
) ||
1556 G_0286CC_LINE_STIPPLE_TEX_ENA(input_ena
));
1557 /* POS_W_FLOAT_ENA requires one of the perspective weights. */
1558 assert(!G_0286CC_POS_W_FLOAT_ENA(input_ena
) ||
1559 G_0286CC_PERSP_SAMPLE_ENA(input_ena
) ||
1560 G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
1561 G_0286CC_PERSP_CENTROID_ENA(input_ena
) ||
1562 G_0286CC_PERSP_PULL_MODEL_ENA(input_ena
));
1564 /* Validate interpolation optimization flags (read as implications). */
1565 assert(!shader
->key
.part
.ps
.prolog
.bc_optimize_for_persp
||
1566 (G_0286CC_PERSP_CENTER_ENA(input_ena
) &&
1567 G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
1568 assert(!shader
->key
.part
.ps
.prolog
.bc_optimize_for_linear
||
1569 (G_0286CC_LINEAR_CENTER_ENA(input_ena
) &&
1570 G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
1571 assert(!shader
->key
.part
.ps
.prolog
.force_persp_center_interp
||
1572 (!G_0286CC_PERSP_SAMPLE_ENA(input_ena
) &&
1573 !G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
1574 assert(!shader
->key
.part
.ps
.prolog
.force_linear_center_interp
||
1575 (!G_0286CC_LINEAR_SAMPLE_ENA(input_ena
) &&
1576 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
1577 assert(!shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
||
1578 (!G_0286CC_PERSP_CENTER_ENA(input_ena
) &&
1579 !G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
1580 assert(!shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
||
1581 (!G_0286CC_LINEAR_CENTER_ENA(input_ena
) &&
1582 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
1584 /* Validate cases when the optimizations are off (read as implications). */
1585 assert(shader
->key
.part
.ps
.prolog
.bc_optimize_for_persp
||
1586 !G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
1587 !G_0286CC_PERSP_CENTROID_ENA(input_ena
));
1588 assert(shader
->key
.part
.ps
.prolog
.bc_optimize_for_linear
||
1589 !G_0286CC_LINEAR_CENTER_ENA(input_ena
) ||
1590 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
));
1592 pm4
= si_get_shader_pm4_state(shader
);
1596 pm4
->atom
.emit
= si_emit_shader_ps
;
1598 /* SPI_BARYC_CNTL.POS_FLOAT_LOCATION
1600 * 0 -> Position = pixel center
1601 * 1 -> Position = pixel centroid
1602 * 2 -> Position = at sample position
1604 * From GLSL 4.5 specification, section 7.1:
1605 * "The variable gl_FragCoord is available as an input variable from
1606 * within fragment shaders and it holds the window relative coordinates
1607 * (x, y, z, 1/w) values for the fragment. If multi-sampling, this
1608 * value can be for any location within the pixel, or one of the
1609 * fragment samples. The use of centroid does not further restrict
1610 * this value to be inside the current primitive."
1612 * Meaning that centroid has no effect and we can return anything within
1613 * the pixel. Thus, return the value at sample position, because that's
1614 * the most accurate one shaders can get.
1616 spi_baryc_cntl
|= S_0286E0_POS_FLOAT_LOCATION(2);
1618 if (info
->properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
] ==
1619 TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)
1620 spi_baryc_cntl
|= S_0286E0_POS_FLOAT_ULC(1);
1622 spi_shader_col_format
= si_get_spi_shader_col_format(shader
);
1623 cb_shader_mask
= ac_get_cb_shader_mask(spi_shader_col_format
);
1625 /* Ensure that some export memory is always allocated, for two reasons:
1627 * 1) Correctness: The hardware ignores the EXEC mask if no export
1628 * memory is allocated, so KILL and alpha test do not work correctly
1630 * 2) Performance: Every shader needs at least a NULL export, even when
1631 * it writes no color/depth output. The NULL export instruction
1632 * stalls without this setting.
1634 * Don't add this to CB_SHADER_MASK.
1636 * GFX10 supports pixel shaders without exports by setting both
1637 * the color and Z formats to SPI_SHADER_ZERO. The hw will skip export
1638 * instructions if any are present.
1640 if ((sscreen
->info
.chip_class
<= GFX9
||
1642 shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
) &&
1643 !spi_shader_col_format
&&
1644 !info
->writes_z
&& !info
->writes_stencil
&& !info
->writes_samplemask
)
1645 spi_shader_col_format
= V_028714_SPI_SHADER_32_R
;
1647 shader
->ctx_reg
.ps
.spi_ps_input_ena
= input_ena
;
1648 shader
->ctx_reg
.ps
.spi_ps_input_addr
= shader
->config
.spi_ps_input_addr
;
1650 /* Set interpolation controls. */
1651 spi_ps_in_control
= S_0286D8_NUM_INTERP(si_get_ps_num_interp(shader
)) |
1652 S_0286D8_PS_W32_EN(sscreen
->ps_wave_size
== 32);
1654 shader
->ctx_reg
.ps
.spi_baryc_cntl
= spi_baryc_cntl
;
1655 shader
->ctx_reg
.ps
.spi_ps_in_control
= spi_ps_in_control
;
1656 shader
->ctx_reg
.ps
.spi_shader_z_format
=
1657 ac_get_spi_shader_z_format(info
->writes_z
,
1658 info
->writes_stencil
,
1659 info
->writes_samplemask
);
1660 shader
->ctx_reg
.ps
.spi_shader_col_format
= spi_shader_col_format
;
1661 shader
->ctx_reg
.ps
.cb_shader_mask
= cb_shader_mask
;
1663 va
= shader
->bo
->gpu_address
;
1664 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
1665 si_pm4_set_reg(pm4
, R_00B020_SPI_SHADER_PGM_LO_PS
, va
>> 8);
1666 si_pm4_set_reg(pm4
, R_00B024_SPI_SHADER_PGM_HI_PS
, S_00B024_MEM_BASE(va
>> 40));
1669 S_00B028_VGPRS((shader
->config
.num_vgprs
- 1) /
1670 (sscreen
->ps_wave_size
== 32 ? 8 : 4)) |
1671 S_00B028_DX10_CLAMP(1) |
1672 S_00B028_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
1673 S_00B028_FLOAT_MODE(shader
->config
.float_mode
);
1675 if (sscreen
->info
.chip_class
< GFX10
) {
1676 rsrc1
|= S_00B028_SGPRS((shader
->config
.num_sgprs
- 1) / 8);
1679 si_pm4_set_reg(pm4
, R_00B028_SPI_SHADER_PGM_RSRC1_PS
, rsrc1
);
1680 si_pm4_set_reg(pm4
, R_00B02C_SPI_SHADER_PGM_RSRC2_PS
,
1681 S_00B02C_EXTRA_LDS_SIZE(shader
->config
.lds_size
) |
1682 S_00B02C_USER_SGPR(SI_PS_NUM_USER_SGPR
) |
1683 S_00B32C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
1686 static void si_shader_init_pm4_state(struct si_screen
*sscreen
,
1687 struct si_shader
*shader
)
1689 switch (shader
->selector
->type
) {
1690 case PIPE_SHADER_VERTEX
:
1691 if (shader
->key
.as_ls
)
1692 si_shader_ls(sscreen
, shader
);
1693 else if (shader
->key
.as_es
)
1694 si_shader_es(sscreen
, shader
);
1695 else if (shader
->key
.as_ngg
)
1696 gfx10_shader_ngg(sscreen
, shader
);
1698 si_shader_vs(sscreen
, shader
, NULL
);
1700 case PIPE_SHADER_TESS_CTRL
:
1701 si_shader_hs(sscreen
, shader
);
1703 case PIPE_SHADER_TESS_EVAL
:
1704 if (shader
->key
.as_es
)
1705 si_shader_es(sscreen
, shader
);
1706 else if (shader
->key
.as_ngg
)
1707 gfx10_shader_ngg(sscreen
, shader
);
1709 si_shader_vs(sscreen
, shader
, NULL
);
1711 case PIPE_SHADER_GEOMETRY
:
1712 if (shader
->key
.as_ngg
)
1713 gfx10_shader_ngg(sscreen
, shader
);
1715 si_shader_gs(sscreen
, shader
);
1717 case PIPE_SHADER_FRAGMENT
:
1718 si_shader_ps(sscreen
, shader
);
1725 static unsigned si_get_alpha_test_func(struct si_context
*sctx
)
1727 /* Alpha-test should be disabled if colorbuffer 0 is integer. */
1728 return sctx
->queued
.named
.dsa
->alpha_func
;
1731 void si_shader_selector_key_vs(struct si_context
*sctx
,
1732 struct si_shader_selector
*vs
,
1733 struct si_shader_key
*key
,
1734 struct si_vs_prolog_bits
*prolog_key
)
1736 if (!sctx
->vertex_elements
||
1737 vs
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
])
1740 struct si_vertex_elements
*elts
= sctx
->vertex_elements
;
1742 prolog_key
->instance_divisor_is_one
= elts
->instance_divisor_is_one
;
1743 prolog_key
->instance_divisor_is_fetched
= elts
->instance_divisor_is_fetched
;
1744 prolog_key
->unpack_instance_id_from_vertex_id
=
1745 sctx
->prim_discard_cs_instancing
;
1747 /* Prefer a monolithic shader to allow scheduling divisions around
1749 if (prolog_key
->instance_divisor_is_fetched
)
1750 key
->opt
.prefer_mono
= 1;
1752 unsigned count
= MIN2(vs
->info
.num_inputs
, elts
->count
);
1753 unsigned count_mask
= (1 << count
) - 1;
1754 unsigned fix
= elts
->fix_fetch_always
& count_mask
;
1755 unsigned opencode
= elts
->fix_fetch_opencode
& count_mask
;
1757 if (sctx
->vertex_buffer_unaligned
& elts
->vb_alignment_check_mask
) {
1758 uint32_t mask
= elts
->fix_fetch_unaligned
& count_mask
;
1760 unsigned i
= u_bit_scan(&mask
);
1761 unsigned log_hw_load_size
= 1 + ((elts
->hw_load_is_dword
>> i
) & 1);
1762 unsigned vbidx
= elts
->vertex_buffer_index
[i
];
1763 struct pipe_vertex_buffer
*vb
= &sctx
->vertex_buffer
[vbidx
];
1764 unsigned align_mask
= (1 << log_hw_load_size
) - 1;
1765 if (vb
->buffer_offset
& align_mask
||
1766 vb
->stride
& align_mask
) {
1774 unsigned i
= u_bit_scan(&fix
);
1775 key
->mono
.vs_fix_fetch
[i
].bits
= elts
->fix_fetch
[i
];
1777 key
->mono
.vs_fetch_opencode
= opencode
;
1780 static void si_shader_selector_key_hw_vs(struct si_context
*sctx
,
1781 struct si_shader_selector
*vs
,
1782 struct si_shader_key
*key
)
1784 struct si_shader_selector
*ps
= sctx
->ps_shader
.cso
;
1786 key
->opt
.clip_disable
=
1787 sctx
->queued
.named
.rasterizer
->clip_plane_enable
== 0 &&
1788 (vs
->info
.clipdist_writemask
||
1789 vs
->info
.writes_clipvertex
) &&
1790 !vs
->info
.culldist_writemask
;
1792 /* Find out if PS is disabled. */
1793 bool ps_disabled
= true;
1795 bool ps_modifies_zs
= ps
->info
.uses_kill
||
1796 ps
->info
.writes_z
||
1797 ps
->info
.writes_stencil
||
1798 ps
->info
.writes_samplemask
||
1799 sctx
->queued
.named
.blend
->alpha_to_coverage
||
1800 si_get_alpha_test_func(sctx
) != PIPE_FUNC_ALWAYS
;
1801 unsigned ps_colormask
= si_get_total_colormask(sctx
);
1803 ps_disabled
= sctx
->queued
.named
.rasterizer
->rasterizer_discard
||
1806 !ps
->info
.writes_memory
);
1809 /* Find out which VS outputs aren't used by the PS. */
1810 uint64_t outputs_written
= vs
->outputs_written_before_ps
;
1811 uint64_t inputs_read
= 0;
1813 /* Ignore outputs that are not passed from VS to PS. */
1814 outputs_written
&= ~((1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_POSITION
, 0, true)) |
1815 (1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_PSIZE
, 0, true)) |
1816 (1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_CLIPVERTEX
, 0, true)));
1819 inputs_read
= ps
->inputs_read
;
1822 uint64_t linked
= outputs_written
& inputs_read
;
1824 key
->opt
.kill_outputs
= ~linked
& outputs_written
;
1827 /* Compute the key for the hw shader variant */
1828 static inline void si_shader_selector_key(struct pipe_context
*ctx
,
1829 struct si_shader_selector
*sel
,
1830 union si_vgt_stages_key stages_key
,
1831 struct si_shader_key
*key
)
1833 struct si_context
*sctx
= (struct si_context
*)ctx
;
1835 memset(key
, 0, sizeof(*key
));
1837 switch (sel
->type
) {
1838 case PIPE_SHADER_VERTEX
:
1839 si_shader_selector_key_vs(sctx
, sel
, key
, &key
->part
.vs
.prolog
);
1841 if (sctx
->tes_shader
.cso
)
1843 else if (sctx
->gs_shader
.cso
) {
1845 key
->as_ngg
= stages_key
.u
.ngg
;
1847 key
->as_ngg
= stages_key
.u
.ngg
;
1848 si_shader_selector_key_hw_vs(sctx
, sel
, key
);
1850 if (sctx
->ps_shader
.cso
&& sctx
->ps_shader
.cso
->info
.uses_primid
)
1851 key
->mono
.u
.vs_export_prim_id
= 1;
1854 case PIPE_SHADER_TESS_CTRL
:
1855 if (sctx
->chip_class
>= GFX9
) {
1856 si_shader_selector_key_vs(sctx
, sctx
->vs_shader
.cso
,
1857 key
, &key
->part
.tcs
.ls_prolog
);
1858 key
->part
.tcs
.ls
= sctx
->vs_shader
.cso
;
1860 /* When the LS VGPR fix is needed, monolithic shaders
1862 * - avoid initializing EXEC in both the LS prolog
1863 * and the LS main part when !vs_needs_prolog
1864 * - remove the fixup for unused input VGPRs
1866 key
->part
.tcs
.ls_prolog
.ls_vgpr_fix
= sctx
->ls_vgpr_fix
;
1868 /* The LS output / HS input layout can be communicated
1869 * directly instead of via user SGPRs for merged LS-HS.
1870 * The LS VGPR fix prefers this too.
1872 key
->opt
.prefer_mono
= 1;
1875 key
->part
.tcs
.epilog
.prim_mode
=
1876 sctx
->tes_shader
.cso
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
1877 key
->part
.tcs
.epilog
.invoc0_tess_factors_are_def
=
1878 sel
->tcs_info
.tessfactors_are_def_in_all_invocs
;
1879 key
->part
.tcs
.epilog
.tes_reads_tess_factors
=
1880 sctx
->tes_shader
.cso
->info
.reads_tess_factors
;
1882 if (sel
== sctx
->fixed_func_tcs_shader
.cso
)
1883 key
->mono
.u
.ff_tcs_inputs_to_copy
= sctx
->vs_shader
.cso
->outputs_written
;
1885 case PIPE_SHADER_TESS_EVAL
:
1886 key
->as_ngg
= stages_key
.u
.ngg
;
1888 if (sctx
->gs_shader
.cso
)
1891 si_shader_selector_key_hw_vs(sctx
, sel
, key
);
1893 if (sctx
->ps_shader
.cso
&& sctx
->ps_shader
.cso
->info
.uses_primid
)
1894 key
->mono
.u
.vs_export_prim_id
= 1;
1897 case PIPE_SHADER_GEOMETRY
:
1898 if (sctx
->chip_class
>= GFX9
) {
1899 if (sctx
->tes_shader
.cso
) {
1900 key
->part
.gs
.es
= sctx
->tes_shader
.cso
;
1902 si_shader_selector_key_vs(sctx
, sctx
->vs_shader
.cso
,
1903 key
, &key
->part
.gs
.vs_prolog
);
1904 key
->part
.gs
.es
= sctx
->vs_shader
.cso
;
1905 key
->part
.gs
.prolog
.gfx9_prev_is_vs
= 1;
1908 key
->as_ngg
= stages_key
.u
.ngg
;
1910 /* Merged ES-GS can have unbalanced wave usage.
1912 * ES threads are per-vertex, while GS threads are
1913 * per-primitive. So without any amplification, there
1914 * are fewer GS threads than ES threads, which can result
1915 * in empty (no-op) GS waves. With too much amplification,
1916 * there are more GS threads than ES threads, which
1917 * can result in empty (no-op) ES waves.
1919 * Non-monolithic shaders are implemented by setting EXEC
1920 * at the beginning of shader parts, and don't jump to
1921 * the end if EXEC is 0.
1923 * Monolithic shaders use conditional blocks, so they can
1924 * jump and skip empty waves of ES or GS. So set this to
1925 * always use optimized variants, which are monolithic.
1927 key
->opt
.prefer_mono
= 1;
1929 key
->part
.gs
.prolog
.tri_strip_adj_fix
= sctx
->gs_tri_strip_adj_fix
;
1931 case PIPE_SHADER_FRAGMENT
: {
1932 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
1933 struct si_state_blend
*blend
= sctx
->queued
.named
.blend
;
1935 if (sel
->info
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
] &&
1936 sel
->info
.colors_written
== 0x1)
1937 key
->part
.ps
.epilog
.last_cbuf
= MAX2(sctx
->framebuffer
.state
.nr_cbufs
, 1) - 1;
1939 /* Select the shader color format based on whether
1940 * blending or alpha are needed.
1942 key
->part
.ps
.epilog
.spi_shader_col_format
=
1943 (blend
->blend_enable_4bit
& blend
->need_src_alpha_4bit
&
1944 sctx
->framebuffer
.spi_shader_col_format_blend_alpha
) |
1945 (blend
->blend_enable_4bit
& ~blend
->need_src_alpha_4bit
&
1946 sctx
->framebuffer
.spi_shader_col_format_blend
) |
1947 (~blend
->blend_enable_4bit
& blend
->need_src_alpha_4bit
&
1948 sctx
->framebuffer
.spi_shader_col_format_alpha
) |
1949 (~blend
->blend_enable_4bit
& ~blend
->need_src_alpha_4bit
&
1950 sctx
->framebuffer
.spi_shader_col_format
);
1951 key
->part
.ps
.epilog
.spi_shader_col_format
&= blend
->cb_target_enabled_4bit
;
1953 /* The output for dual source blending should have
1954 * the same format as the first output.
1956 if (blend
->dual_src_blend
) {
1957 key
->part
.ps
.epilog
.spi_shader_col_format
|=
1958 (key
->part
.ps
.epilog
.spi_shader_col_format
& 0xf) << 4;
1961 /* If alpha-to-coverage is enabled, we have to export alpha
1962 * even if there is no color buffer.
1964 if (!(key
->part
.ps
.epilog
.spi_shader_col_format
& 0xf) &&
1965 blend
->alpha_to_coverage
)
1966 key
->part
.ps
.epilog
.spi_shader_col_format
|= V_028710_SPI_SHADER_32_AR
;
1968 /* On GFX6 and GFX7 except Hawaii, the CB doesn't clamp outputs
1969 * to the range supported by the type if a channel has less
1970 * than 16 bits and the export format is 16_ABGR.
1972 if (sctx
->chip_class
<= GFX7
&& sctx
->family
!= CHIP_HAWAII
) {
1973 key
->part
.ps
.epilog
.color_is_int8
= sctx
->framebuffer
.color_is_int8
;
1974 key
->part
.ps
.epilog
.color_is_int10
= sctx
->framebuffer
.color_is_int10
;
1977 /* Disable unwritten outputs (if WRITE_ALL_CBUFS isn't enabled). */
1978 if (!key
->part
.ps
.epilog
.last_cbuf
) {
1979 key
->part
.ps
.epilog
.spi_shader_col_format
&= sel
->colors_written_4bit
;
1980 key
->part
.ps
.epilog
.color_is_int8
&= sel
->info
.colors_written
;
1981 key
->part
.ps
.epilog
.color_is_int10
&= sel
->info
.colors_written
;
1984 bool is_poly
= !util_prim_is_points_or_lines(sctx
->current_rast_prim
);
1985 bool is_line
= util_prim_is_lines(sctx
->current_rast_prim
);
1987 key
->part
.ps
.prolog
.color_two_side
= rs
->two_side
&& sel
->info
.colors_read
;
1988 key
->part
.ps
.prolog
.flatshade_colors
= rs
->flatshade
&& sel
->info
.colors_read
;
1990 key
->part
.ps
.epilog
.alpha_to_one
= blend
->alpha_to_one
&&
1991 rs
->multisample_enable
;
1993 key
->part
.ps
.prolog
.poly_stipple
= rs
->poly_stipple_enable
&& is_poly
;
1994 key
->part
.ps
.epilog
.poly_line_smoothing
= ((is_poly
&& rs
->poly_smooth
) ||
1995 (is_line
&& rs
->line_smooth
)) &&
1996 sctx
->framebuffer
.nr_samples
<= 1;
1997 key
->part
.ps
.epilog
.clamp_color
= rs
->clamp_fragment_color
;
1999 if (sctx
->ps_iter_samples
> 1 &&
2000 sel
->info
.reads_samplemask
) {
2001 key
->part
.ps
.prolog
.samplemask_log_ps_iter
=
2002 util_logbase2(sctx
->ps_iter_samples
);
2005 if (rs
->force_persample_interp
&&
2006 rs
->multisample_enable
&&
2007 sctx
->framebuffer
.nr_samples
> 1 &&
2008 sctx
->ps_iter_samples
> 1) {
2009 key
->part
.ps
.prolog
.force_persp_sample_interp
=
2010 sel
->info
.uses_persp_center
||
2011 sel
->info
.uses_persp_centroid
;
2013 key
->part
.ps
.prolog
.force_linear_sample_interp
=
2014 sel
->info
.uses_linear_center
||
2015 sel
->info
.uses_linear_centroid
;
2016 } else if (rs
->multisample_enable
&&
2017 sctx
->framebuffer
.nr_samples
> 1) {
2018 key
->part
.ps
.prolog
.bc_optimize_for_persp
=
2019 sel
->info
.uses_persp_center
&&
2020 sel
->info
.uses_persp_centroid
;
2021 key
->part
.ps
.prolog
.bc_optimize_for_linear
=
2022 sel
->info
.uses_linear_center
&&
2023 sel
->info
.uses_linear_centroid
;
2025 /* Make sure SPI doesn't compute more than 1 pair
2026 * of (i,j), which is the optimization here. */
2027 key
->part
.ps
.prolog
.force_persp_center_interp
=
2028 sel
->info
.uses_persp_center
+
2029 sel
->info
.uses_persp_centroid
+
2030 sel
->info
.uses_persp_sample
> 1;
2032 key
->part
.ps
.prolog
.force_linear_center_interp
=
2033 sel
->info
.uses_linear_center
+
2034 sel
->info
.uses_linear_centroid
+
2035 sel
->info
.uses_linear_sample
> 1;
2037 if (sel
->info
.uses_persp_opcode_interp_sample
||
2038 sel
->info
.uses_linear_opcode_interp_sample
)
2039 key
->mono
.u
.ps
.interpolate_at_sample_force_center
= 1;
2042 key
->part
.ps
.epilog
.alpha_func
= si_get_alpha_test_func(sctx
);
2044 /* ps_uses_fbfetch is true only if the color buffer is bound. */
2045 if (sctx
->ps_uses_fbfetch
&& !sctx
->blitter
->running
) {
2046 struct pipe_surface
*cb0
= sctx
->framebuffer
.state
.cbufs
[0];
2047 struct pipe_resource
*tex
= cb0
->texture
;
2049 /* 1D textures are allocated and used as 2D on GFX9. */
2050 key
->mono
.u
.ps
.fbfetch_msaa
= sctx
->framebuffer
.nr_samples
> 1;
2051 key
->mono
.u
.ps
.fbfetch_is_1D
= sctx
->chip_class
!= GFX9
&&
2052 (tex
->target
== PIPE_TEXTURE_1D
||
2053 tex
->target
== PIPE_TEXTURE_1D_ARRAY
);
2054 key
->mono
.u
.ps
.fbfetch_layered
= tex
->target
== PIPE_TEXTURE_1D_ARRAY
||
2055 tex
->target
== PIPE_TEXTURE_2D_ARRAY
||
2056 tex
->target
== PIPE_TEXTURE_CUBE
||
2057 tex
->target
== PIPE_TEXTURE_CUBE_ARRAY
||
2058 tex
->target
== PIPE_TEXTURE_3D
;
2066 if (unlikely(sctx
->screen
->debug_flags
& DBG(NO_OPT_VARIANT
)))
2067 memset(&key
->opt
, 0, sizeof(key
->opt
));
2070 static void si_build_shader_variant(struct si_shader
*shader
,
2074 struct si_shader_selector
*sel
= shader
->selector
;
2075 struct si_screen
*sscreen
= sel
->screen
;
2076 struct ac_llvm_compiler
*compiler
;
2077 struct pipe_debug_callback
*debug
= &shader
->compiler_ctx_state
.debug
;
2079 if (thread_index
>= 0) {
2081 assert(thread_index
< ARRAY_SIZE(sscreen
->compiler_lowp
));
2082 compiler
= &sscreen
->compiler_lowp
[thread_index
];
2084 assert(thread_index
< ARRAY_SIZE(sscreen
->compiler
));
2085 compiler
= &sscreen
->compiler
[thread_index
];
2090 assert(!low_priority
);
2091 compiler
= shader
->compiler_ctx_state
.compiler
;
2094 if (!compiler
->passes
)
2095 si_init_compiler(sscreen
, compiler
);
2097 if (unlikely(!si_shader_create(sscreen
, compiler
, shader
, debug
))) {
2098 PRINT_ERR("Failed to build shader variant (type=%u)\n",
2100 shader
->compilation_failed
= true;
2104 if (shader
->compiler_ctx_state
.is_debug_context
) {
2105 FILE *f
= open_memstream(&shader
->shader_log
,
2106 &shader
->shader_log_size
);
2108 si_shader_dump(sscreen
, shader
, NULL
, f
, false);
2113 si_shader_init_pm4_state(sscreen
, shader
);
2116 static void si_build_shader_variant_low_priority(void *job
, int thread_index
)
2118 struct si_shader
*shader
= (struct si_shader
*)job
;
2120 assert(thread_index
>= 0);
2122 si_build_shader_variant(shader
, thread_index
, true);
2125 static const struct si_shader_key zeroed
;
2127 static bool si_check_missing_main_part(struct si_screen
*sscreen
,
2128 struct si_shader_selector
*sel
,
2129 struct si_compiler_ctx_state
*compiler_state
,
2130 struct si_shader_key
*key
)
2132 struct si_shader
**mainp
= si_get_main_shader_part(sel
, key
);
2135 struct si_shader
*main_part
= CALLOC_STRUCT(si_shader
);
2140 /* We can leave the fence as permanently signaled because the
2141 * main part becomes visible globally only after it has been
2143 util_queue_fence_init(&main_part
->ready
);
2145 main_part
->selector
= sel
;
2146 main_part
->key
.as_es
= key
->as_es
;
2147 main_part
->key
.as_ls
= key
->as_ls
;
2148 main_part
->key
.as_ngg
= key
->as_ngg
;
2149 main_part
->is_monolithic
= false;
2151 if (si_compile_shader(sscreen
, compiler_state
->compiler
,
2152 main_part
, &compiler_state
->debug
) != 0) {
2162 * Select a shader variant according to the shader key.
2164 * \param optimized_or_none If the key describes an optimized shader variant and
2165 * the compilation isn't finished, don't select any
2166 * shader and return an error.
2168 int si_shader_select_with_key(struct si_screen
*sscreen
,
2169 struct si_shader_ctx_state
*state
,
2170 struct si_compiler_ctx_state
*compiler_state
,
2171 struct si_shader_key
*key
,
2173 bool optimized_or_none
)
2175 struct si_shader_selector
*sel
= state
->cso
;
2176 struct si_shader_selector
*previous_stage_sel
= NULL
;
2177 struct si_shader
*current
= state
->current
;
2178 struct si_shader
*iter
, *shader
= NULL
;
2181 /* Check if we don't need to change anything.
2182 * This path is also used for most shaders that don't need multiple
2183 * variants, it will cost just a computation of the key and this
2185 if (likely(current
&&
2186 memcmp(¤t
->key
, key
, sizeof(*key
)) == 0)) {
2187 if (unlikely(!util_queue_fence_is_signalled(¤t
->ready
))) {
2188 if (current
->is_optimized
) {
2189 if (optimized_or_none
)
2192 memset(&key
->opt
, 0, sizeof(key
->opt
));
2193 goto current_not_ready
;
2196 util_queue_fence_wait(¤t
->ready
);
2199 return current
->compilation_failed
? -1 : 0;
2203 /* This must be done before the mutex is locked, because async GS
2204 * compilation calls this function too, and therefore must enter
2207 * Only wait if we are in a draw call. Don't wait if we are
2208 * in a compiler thread.
2210 if (thread_index
< 0)
2211 util_queue_fence_wait(&sel
->ready
);
2213 simple_mtx_lock(&sel
->mutex
);
2215 /* Find the shader variant. */
2216 for (iter
= sel
->first_variant
; iter
; iter
= iter
->next_variant
) {
2217 /* Don't check the "current" shader. We checked it above. */
2218 if (current
!= iter
&&
2219 memcmp(&iter
->key
, key
, sizeof(*key
)) == 0) {
2220 simple_mtx_unlock(&sel
->mutex
);
2222 if (unlikely(!util_queue_fence_is_signalled(&iter
->ready
))) {
2223 /* If it's an optimized shader and its compilation has
2224 * been started but isn't done, use the unoptimized
2225 * shader so as not to cause a stall due to compilation.
2227 if (iter
->is_optimized
) {
2228 if (optimized_or_none
)
2230 memset(&key
->opt
, 0, sizeof(key
->opt
));
2234 util_queue_fence_wait(&iter
->ready
);
2237 if (iter
->compilation_failed
) {
2238 return -1; /* skip the draw call */
2241 state
->current
= iter
;
2246 /* Build a new shader. */
2247 shader
= CALLOC_STRUCT(si_shader
);
2249 simple_mtx_unlock(&sel
->mutex
);
2253 util_queue_fence_init(&shader
->ready
);
2255 shader
->selector
= sel
;
2257 shader
->compiler_ctx_state
= *compiler_state
;
2259 /* If this is a merged shader, get the first shader's selector. */
2260 if (sscreen
->info
.chip_class
>= GFX9
) {
2261 if (sel
->type
== PIPE_SHADER_TESS_CTRL
)
2262 previous_stage_sel
= key
->part
.tcs
.ls
;
2263 else if (sel
->type
== PIPE_SHADER_GEOMETRY
)
2264 previous_stage_sel
= key
->part
.gs
.es
;
2266 /* We need to wait for the previous shader. */
2267 if (previous_stage_sel
&& thread_index
< 0)
2268 util_queue_fence_wait(&previous_stage_sel
->ready
);
2271 bool is_pure_monolithic
=
2272 sscreen
->use_monolithic_shaders
||
2273 memcmp(&key
->mono
, &zeroed
.mono
, sizeof(key
->mono
)) != 0;
2275 /* Compile the main shader part if it doesn't exist. This can happen
2276 * if the initial guess was wrong.
2278 * The prim discard CS doesn't need the main shader part.
2280 if (!is_pure_monolithic
&&
2281 !key
->opt
.vs_as_prim_discard_cs
) {
2284 /* Make sure the main shader part is present. This is needed
2285 * for shaders that can be compiled as VS, LS, or ES, and only
2286 * one of them is compiled at creation.
2288 * It is also needed for GS, which can be compiled as non-NGG
2291 * For merged shaders, check that the starting shader's main
2294 if (previous_stage_sel
) {
2295 struct si_shader_key shader1_key
= zeroed
;
2297 if (sel
->type
== PIPE_SHADER_TESS_CTRL
) {
2298 shader1_key
.as_ls
= 1;
2299 } else if (sel
->type
== PIPE_SHADER_GEOMETRY
) {
2300 shader1_key
.as_es
= 1;
2301 shader1_key
.as_ngg
= key
->as_ngg
; /* for Wave32 vs Wave64 */
2306 simple_mtx_lock(&previous_stage_sel
->mutex
);
2307 ok
= si_check_missing_main_part(sscreen
,
2309 compiler_state
, &shader1_key
);
2310 simple_mtx_unlock(&previous_stage_sel
->mutex
);
2314 ok
= si_check_missing_main_part(sscreen
, sel
,
2315 compiler_state
, key
);
2320 simple_mtx_unlock(&sel
->mutex
);
2321 return -ENOMEM
; /* skip the draw call */
2325 /* Keep the reference to the 1st shader of merged shaders, so that
2326 * Gallium can't destroy it before we destroy the 2nd shader.
2328 * Set sctx = NULL, because it's unused if we're not releasing
2329 * the shader, and we don't have any sctx here.
2331 si_shader_selector_reference(NULL
, &shader
->previous_stage_sel
,
2332 previous_stage_sel
);
2334 /* Monolithic-only shaders don't make a distinction between optimized
2335 * and unoptimized. */
2336 shader
->is_monolithic
=
2337 is_pure_monolithic
||
2338 memcmp(&key
->opt
, &zeroed
.opt
, sizeof(key
->opt
)) != 0;
2340 /* The prim discard CS is always optimized. */
2341 shader
->is_optimized
=
2342 (!is_pure_monolithic
|| key
->opt
.vs_as_prim_discard_cs
) &&
2343 memcmp(&key
->opt
, &zeroed
.opt
, sizeof(key
->opt
)) != 0;
2345 /* If it's an optimized shader, compile it asynchronously. */
2346 if (shader
->is_optimized
&& thread_index
< 0) {
2347 /* Compile it asynchronously. */
2348 util_queue_add_job(&sscreen
->shader_compiler_queue_low_priority
,
2349 shader
, &shader
->ready
,
2350 si_build_shader_variant_low_priority
, NULL
,
2353 /* Add only after the ready fence was reset, to guard against a
2354 * race with si_bind_XX_shader. */
2355 if (!sel
->last_variant
) {
2356 sel
->first_variant
= shader
;
2357 sel
->last_variant
= shader
;
2359 sel
->last_variant
->next_variant
= shader
;
2360 sel
->last_variant
= shader
;
2363 /* Use the default (unoptimized) shader for now. */
2364 memset(&key
->opt
, 0, sizeof(key
->opt
));
2365 simple_mtx_unlock(&sel
->mutex
);
2367 if (sscreen
->options
.sync_compile
)
2368 util_queue_fence_wait(&shader
->ready
);
2370 if (optimized_or_none
)
2375 /* Reset the fence before adding to the variant list. */
2376 util_queue_fence_reset(&shader
->ready
);
2378 if (!sel
->last_variant
) {
2379 sel
->first_variant
= shader
;
2380 sel
->last_variant
= shader
;
2382 sel
->last_variant
->next_variant
= shader
;
2383 sel
->last_variant
= shader
;
2386 simple_mtx_unlock(&sel
->mutex
);
2388 assert(!shader
->is_optimized
);
2389 si_build_shader_variant(shader
, thread_index
, false);
2391 util_queue_fence_signal(&shader
->ready
);
2393 if (!shader
->compilation_failed
)
2394 state
->current
= shader
;
2396 return shader
->compilation_failed
? -1 : 0;
2399 static int si_shader_select(struct pipe_context
*ctx
,
2400 struct si_shader_ctx_state
*state
,
2401 union si_vgt_stages_key stages_key
,
2402 struct si_compiler_ctx_state
*compiler_state
)
2404 struct si_context
*sctx
= (struct si_context
*)ctx
;
2405 struct si_shader_key key
;
2407 si_shader_selector_key(ctx
, state
->cso
, stages_key
, &key
);
2408 return si_shader_select_with_key(sctx
->screen
, state
, compiler_state
,
2412 static void si_parse_next_shader_property(const struct tgsi_shader_info
*info
,
2414 struct si_shader_key
*key
)
2416 unsigned next_shader
= info
->properties
[TGSI_PROPERTY_NEXT_SHADER
];
2418 switch (info
->processor
) {
2419 case PIPE_SHADER_VERTEX
:
2420 switch (next_shader
) {
2421 case PIPE_SHADER_GEOMETRY
:
2424 case PIPE_SHADER_TESS_CTRL
:
2425 case PIPE_SHADER_TESS_EVAL
:
2429 /* If POSITION isn't written, it can only be a HW VS
2430 * if streamout is used. If streamout isn't used,
2431 * assume that it's a HW LS. (the next shader is TCS)
2432 * This heuristic is needed for separate shader objects.
2434 if (!info
->writes_position
&& !streamout
)
2439 case PIPE_SHADER_TESS_EVAL
:
2440 if (next_shader
== PIPE_SHADER_GEOMETRY
||
2441 !info
->writes_position
)
2448 * Compile the main shader part or the monolithic shader as part of
2449 * si_shader_selector initialization. Since it can be done asynchronously,
2450 * there is no way to report compile failures to applications.
2452 static void si_init_shader_selector_async(void *job
, int thread_index
)
2454 struct si_shader_selector
*sel
= (struct si_shader_selector
*)job
;
2455 struct si_screen
*sscreen
= sel
->screen
;
2456 struct ac_llvm_compiler
*compiler
;
2457 struct pipe_debug_callback
*debug
= &sel
->compiler_ctx_state
.debug
;
2459 assert(!debug
->debug_message
|| debug
->async
);
2460 assert(thread_index
>= 0);
2461 assert(thread_index
< ARRAY_SIZE(sscreen
->compiler
));
2462 compiler
= &sscreen
->compiler
[thread_index
];
2464 if (!compiler
->passes
)
2465 si_init_compiler(sscreen
, compiler
);
2467 /* Serialize NIR to save memory. Monolithic shader variants
2468 * have to deserialize NIR before compilation.
2475 /* true = remove optional debugging data to increase
2476 * the likehood of getting more shader cache hits.
2477 * It also drops variable names, so we'll save more memory.
2479 nir_serialize(&blob
, sel
->nir
, true);
2480 blob_finish_get_buffer(&blob
, &sel
->nir_binary
, &size
);
2481 sel
->nir_size
= size
;
2484 /* Compile the main shader part for use with a prolog and/or epilog.
2485 * If this fails, the driver will try to compile a monolithic shader
2488 if (!sscreen
->use_monolithic_shaders
) {
2489 struct si_shader
*shader
= CALLOC_STRUCT(si_shader
);
2490 unsigned char ir_sha1_cache_key
[20];
2493 fprintf(stderr
, "radeonsi: can't allocate a main shader part\n");
2497 /* We can leave the fence signaled because use of the default
2498 * main part is guarded by the selector's ready fence. */
2499 util_queue_fence_init(&shader
->ready
);
2501 shader
->selector
= sel
;
2502 shader
->is_monolithic
= false;
2503 si_parse_next_shader_property(&sel
->info
,
2504 sel
->so
.num_outputs
!= 0,
2507 if (sscreen
->use_ngg
&&
2508 (!sel
->so
.num_outputs
|| sscreen
->use_ngg_streamout
) &&
2509 ((sel
->type
== PIPE_SHADER_VERTEX
&& !shader
->key
.as_ls
) ||
2510 sel
->type
== PIPE_SHADER_TESS_EVAL
||
2511 sel
->type
== PIPE_SHADER_GEOMETRY
))
2512 shader
->key
.as_ngg
= 1;
2515 si_get_ir_cache_key(sel
, shader
->key
.as_ngg
,
2516 shader
->key
.as_es
, ir_sha1_cache_key
);
2519 /* Try to load the shader from the shader cache. */
2520 simple_mtx_lock(&sscreen
->shader_cache_mutex
);
2522 if (si_shader_cache_load_shader(sscreen
, ir_sha1_cache_key
, shader
)) {
2523 simple_mtx_unlock(&sscreen
->shader_cache_mutex
);
2524 si_shader_dump_stats_for_shader_db(sscreen
, shader
, debug
);
2526 simple_mtx_unlock(&sscreen
->shader_cache_mutex
);
2528 /* Compile the shader if it hasn't been loaded from the cache. */
2529 if (si_compile_shader(sscreen
, compiler
, shader
,
2532 fprintf(stderr
, "radeonsi: can't compile a main shader part\n");
2536 simple_mtx_lock(&sscreen
->shader_cache_mutex
);
2537 si_shader_cache_insert_shader(sscreen
, ir_sha1_cache_key
,
2539 simple_mtx_unlock(&sscreen
->shader_cache_mutex
);
2542 *si_get_main_shader_part(sel
, &shader
->key
) = shader
;
2544 /* Unset "outputs_written" flags for outputs converted to
2545 * DEFAULT_VAL, so that later inter-shader optimizations don't
2546 * try to eliminate outputs that don't exist in the final
2549 * This is only done if non-monolithic shaders are enabled.
2551 if ((sel
->type
== PIPE_SHADER_VERTEX
||
2552 sel
->type
== PIPE_SHADER_TESS_EVAL
) &&
2553 !shader
->key
.as_ls
&&
2554 !shader
->key
.as_es
) {
2557 for (i
= 0; i
< sel
->info
.num_outputs
; i
++) {
2558 unsigned offset
= shader
->info
.vs_output_param_offset
[i
];
2560 if (offset
<= AC_EXP_PARAM_OFFSET_31
)
2563 unsigned name
= sel
->info
.output_semantic_name
[i
];
2564 unsigned index
= sel
->info
.output_semantic_index
[i
];
2568 case TGSI_SEMANTIC_GENERIC
:
2569 /* don't process indices the function can't handle */
2570 if (index
>= SI_MAX_IO_GENERIC
)
2574 id
= si_shader_io_get_unique_index(name
, index
, true);
2575 sel
->outputs_written_before_ps
&= ~(1ull << id
);
2577 case TGSI_SEMANTIC_POSITION
: /* ignore these */
2578 case TGSI_SEMANTIC_PSIZE
:
2579 case TGSI_SEMANTIC_CLIPVERTEX
:
2580 case TGSI_SEMANTIC_EDGEFLAG
:
2587 /* The GS copy shader is always pre-compiled. */
2588 if (sel
->type
== PIPE_SHADER_GEOMETRY
&&
2589 (!sscreen
->use_ngg
||
2590 !sscreen
->use_ngg_streamout
|| /* also for PRIMITIVES_GENERATED */
2591 sel
->tess_turns_off_ngg
)) {
2592 sel
->gs_copy_shader
= si_generate_gs_copy_shader(sscreen
, compiler
, sel
, debug
);
2593 if (!sel
->gs_copy_shader
) {
2594 fprintf(stderr
, "radeonsi: can't create GS copy shader\n");
2598 si_shader_vs(sscreen
, sel
->gs_copy_shader
, sel
);
2601 /* Free NIR. We only keep serialized NIR after this point. */
2603 ralloc_free(sel
->nir
);
2608 void si_schedule_initial_compile(struct si_context
*sctx
, unsigned processor
,
2609 struct util_queue_fence
*ready_fence
,
2610 struct si_compiler_ctx_state
*compiler_ctx_state
,
2611 void *job
, util_queue_execute_func execute
)
2613 util_queue_fence_init(ready_fence
);
2615 struct util_async_debug_callback async_debug
;
2617 (sctx
->debug
.debug_message
&& !sctx
->debug
.async
) ||
2619 si_can_dump_shader(sctx
->screen
, processor
);
2622 u_async_debug_init(&async_debug
);
2623 compiler_ctx_state
->debug
= async_debug
.base
;
2626 util_queue_add_job(&sctx
->screen
->shader_compiler_queue
, job
,
2627 ready_fence
, execute
, NULL
, 0);
2630 util_queue_fence_wait(ready_fence
);
2631 u_async_debug_drain(&async_debug
, &sctx
->debug
);
2632 u_async_debug_cleanup(&async_debug
);
2635 if (sctx
->screen
->options
.sync_compile
)
2636 util_queue_fence_wait(ready_fence
);
2639 /* Return descriptor slot usage masks from the given shader info. */
2640 void si_get_active_slot_masks(const struct tgsi_shader_info
*info
,
2641 uint32_t *const_and_shader_buffers
,
2642 uint64_t *samplers_and_images
)
2644 unsigned start
, num_shaderbufs
, num_constbufs
, num_images
, num_msaa_images
, num_samplers
;
2646 num_shaderbufs
= util_last_bit(info
->shader_buffers_declared
);
2647 num_constbufs
= util_last_bit(info
->const_buffers_declared
);
2648 /* two 8-byte images share one 16-byte slot */
2649 num_images
= align(util_last_bit(info
->images_declared
), 2);
2650 num_msaa_images
= align(util_last_bit(info
->msaa_images_declared
), 2);
2651 num_samplers
= util_last_bit(info
->samplers_declared
);
2653 /* The layout is: sb[last] ... sb[0], cb[0] ... cb[last] */
2654 start
= si_get_shaderbuf_slot(num_shaderbufs
- 1);
2655 *const_and_shader_buffers
=
2656 u_bit_consecutive(start
, num_shaderbufs
+ num_constbufs
);
2659 * - fmask[last] ... fmask[0] go to [15-last .. 15]
2660 * - image[last] ... image[0] go to [31-last .. 31]
2661 * - sampler[0] ... sampler[last] go to [32 .. 32+last*2]
2663 * FMASKs for images are placed separately, because MSAA images are rare,
2664 * and so we can benefit from a better cache hit rate if we keep image
2665 * descriptors together.
2667 if (num_msaa_images
)
2668 num_images
= SI_NUM_IMAGES
+ num_msaa_images
; /* add FMASK descriptors */
2670 start
= si_get_image_slot(num_images
- 1) / 2;
2671 *samplers_and_images
=
2672 u_bit_consecutive64(start
, num_images
/ 2 + num_samplers
);
2675 static void *si_create_shader_selector(struct pipe_context
*ctx
,
2676 const struct pipe_shader_state
*state
)
2678 struct si_screen
*sscreen
= (struct si_screen
*)ctx
->screen
;
2679 struct si_context
*sctx
= (struct si_context
*)ctx
;
2680 struct si_shader_selector
*sel
= CALLOC_STRUCT(si_shader_selector
);
2686 pipe_reference_init(&sel
->reference
, 1);
2687 sel
->screen
= sscreen
;
2688 sel
->compiler_ctx_state
.debug
= sctx
->debug
;
2689 sel
->compiler_ctx_state
.is_debug_context
= sctx
->is_debug
;
2691 sel
->so
= state
->stream_output
;
2693 if (state
->type
== PIPE_SHADER_IR_TGSI
) {
2694 sel
->nir
= tgsi_to_nir(state
->tokens
, ctx
->screen
);
2696 assert(state
->type
== PIPE_SHADER_IR_NIR
);
2697 sel
->nir
= state
->ir
.nir
;
2700 si_nir_scan_shader(sel
->nir
, &sel
->info
);
2701 si_nir_scan_tess_ctrl(sel
->nir
, &sel
->tcs_info
);
2702 si_nir_adjust_driver_locations(sel
->nir
);
2704 sel
->type
= sel
->info
.processor
;
2705 p_atomic_inc(&sscreen
->num_shaders_created
);
2706 si_get_active_slot_masks(&sel
->info
,
2707 &sel
->active_const_and_shader_buffers
,
2708 &sel
->active_samplers_and_images
);
2710 /* Record which streamout buffers are enabled. */
2711 for (i
= 0; i
< sel
->so
.num_outputs
; i
++) {
2712 sel
->enabled_streamout_buffer_mask
|=
2713 (1 << sel
->so
.output
[i
].output_buffer
) <<
2714 (sel
->so
.output
[i
].stream
* 4);
2717 /* The prolog is a no-op if there are no inputs. */
2718 sel
->vs_needs_prolog
= sel
->type
== PIPE_SHADER_VERTEX
&&
2719 sel
->info
.num_inputs
&&
2720 !sel
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
];
2722 sel
->force_correct_derivs_after_kill
=
2723 sel
->type
== PIPE_SHADER_FRAGMENT
&&
2724 sel
->info
.uses_derivatives
&&
2725 sel
->info
.uses_kill
&&
2726 sctx
->screen
->debug_flags
& DBG(FS_CORRECT_DERIVS_AFTER_KILL
);
2728 sel
->prim_discard_cs_allowed
=
2729 sel
->type
== PIPE_SHADER_VERTEX
&&
2730 !sel
->info
.uses_bindless_images
&&
2731 !sel
->info
.uses_bindless_samplers
&&
2732 !sel
->info
.writes_memory
&&
2733 !sel
->info
.writes_viewport_index
&&
2734 !sel
->info
.properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
] &&
2735 !sel
->so
.num_outputs
;
2737 switch (sel
->type
) {
2738 case PIPE_SHADER_GEOMETRY
:
2739 sel
->gs_output_prim
=
2740 sel
->info
.properties
[TGSI_PROPERTY_GS_OUTPUT_PRIM
];
2742 /* Only possibilities: POINTS, LINE_STRIP, TRIANGLES */
2743 sel
->rast_prim
= sel
->gs_output_prim
;
2744 if (util_rast_prim_is_triangles(sel
->rast_prim
))
2745 sel
->rast_prim
= PIPE_PRIM_TRIANGLES
;
2747 sel
->gs_max_out_vertices
=
2748 sel
->info
.properties
[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES
];
2749 sel
->gs_num_invocations
=
2750 sel
->info
.properties
[TGSI_PROPERTY_GS_INVOCATIONS
];
2751 sel
->gsvs_vertex_size
= sel
->info
.num_outputs
* 16;
2752 sel
->max_gsvs_emit_size
= sel
->gsvs_vertex_size
*
2753 sel
->gs_max_out_vertices
;
2755 sel
->max_gs_stream
= 0;
2756 for (i
= 0; i
< sel
->so
.num_outputs
; i
++)
2757 sel
->max_gs_stream
= MAX2(sel
->max_gs_stream
,
2758 sel
->so
.output
[i
].stream
);
2760 sel
->gs_input_verts_per_prim
=
2761 u_vertices_per_prim(sel
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
]);
2763 /* EN_MAX_VERT_OUT_PER_GS_INSTANCE does not work with tesselation. */
2764 sel
->tess_turns_off_ngg
=
2765 sscreen
->info
.chip_class
== GFX10
&&
2766 sel
->gs_num_invocations
* sel
->gs_max_out_vertices
> 256;
2769 case PIPE_SHADER_TESS_CTRL
:
2770 /* Always reserve space for these. */
2771 sel
->patch_outputs_written
|=
2772 (1ull << si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSINNER
, 0)) |
2773 (1ull << si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSOUTER
, 0));
2775 case PIPE_SHADER_VERTEX
:
2776 case PIPE_SHADER_TESS_EVAL
:
2777 for (i
= 0; i
< sel
->info
.num_outputs
; i
++) {
2778 unsigned name
= sel
->info
.output_semantic_name
[i
];
2779 unsigned index
= sel
->info
.output_semantic_index
[i
];
2782 case TGSI_SEMANTIC_TESSINNER
:
2783 case TGSI_SEMANTIC_TESSOUTER
:
2784 case TGSI_SEMANTIC_PATCH
:
2785 sel
->patch_outputs_written
|=
2786 1ull << si_shader_io_get_unique_index_patch(name
, index
);
2789 case TGSI_SEMANTIC_GENERIC
:
2790 /* don't process indices the function can't handle */
2791 if (index
>= SI_MAX_IO_GENERIC
)
2795 sel
->outputs_written
|=
2796 1ull << si_shader_io_get_unique_index(name
, index
, false);
2797 sel
->outputs_written_before_ps
|=
2798 1ull << si_shader_io_get_unique_index(name
, index
, true);
2800 case TGSI_SEMANTIC_EDGEFLAG
:
2804 sel
->esgs_itemsize
= util_last_bit64(sel
->outputs_written
) * 16;
2805 sel
->lshs_vertex_stride
= sel
->esgs_itemsize
;
2807 /* Add 1 dword to reduce LDS bank conflicts, so that each vertex
2808 * will start on a different bank. (except for the maximum 32*16).
2810 if (sel
->lshs_vertex_stride
< 32*16)
2811 sel
->lshs_vertex_stride
+= 4;
2813 /* For the ESGS ring in LDS, add 1 dword to reduce LDS bank
2814 * conflicts, i.e. each vertex will start at a different bank.
2816 if (sctx
->chip_class
>= GFX9
)
2817 sel
->esgs_itemsize
+= 4;
2819 assert(((sel
->esgs_itemsize
/ 4) & C_028AAC_ITEMSIZE
) == 0);
2822 if (sel
->info
.properties
[TGSI_PROPERTY_TES_POINT_MODE
])
2823 sel
->rast_prim
= PIPE_PRIM_POINTS
;
2824 else if (sel
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] == PIPE_PRIM_LINES
)
2825 sel
->rast_prim
= PIPE_PRIM_LINE_STRIP
;
2827 sel
->rast_prim
= PIPE_PRIM_TRIANGLES
;
2830 case PIPE_SHADER_FRAGMENT
:
2831 for (i
= 0; i
< sel
->info
.num_inputs
; i
++) {
2832 unsigned name
= sel
->info
.input_semantic_name
[i
];
2833 unsigned index
= sel
->info
.input_semantic_index
[i
];
2836 case TGSI_SEMANTIC_GENERIC
:
2837 /* don't process indices the function can't handle */
2838 if (index
>= SI_MAX_IO_GENERIC
)
2843 1ull << si_shader_io_get_unique_index(name
, index
, true);
2845 case TGSI_SEMANTIC_PCOORD
: /* ignore this */
2850 for (i
= 0; i
< 8; i
++)
2851 if (sel
->info
.colors_written
& (1 << i
))
2852 sel
->colors_written_4bit
|= 0xf << (4 * i
);
2854 for (i
= 0; i
< sel
->info
.num_inputs
; i
++) {
2855 if (sel
->info
.input_semantic_name
[i
] == TGSI_SEMANTIC_COLOR
) {
2856 int index
= sel
->info
.input_semantic_index
[i
];
2857 sel
->color_attr_index
[index
] = i
;
2864 /* PA_CL_VS_OUT_CNTL */
2865 if (sctx
->chip_class
<= GFX9
)
2866 sel
->pa_cl_vs_out_cntl
= si_get_vs_out_cntl(sel
, false);
2868 sel
->clipdist_mask
= sel
->info
.writes_clipvertex
?
2869 SIX_BITS
: sel
->info
.clipdist_writemask
;
2870 sel
->culldist_mask
= sel
->info
.culldist_writemask
<<
2871 sel
->info
.num_written_clipdistance
;
2873 /* DB_SHADER_CONTROL */
2874 sel
->db_shader_control
=
2875 S_02880C_Z_EXPORT_ENABLE(sel
->info
.writes_z
) |
2876 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(sel
->info
.writes_stencil
) |
2877 S_02880C_MASK_EXPORT_ENABLE(sel
->info
.writes_samplemask
) |
2878 S_02880C_KILL_ENABLE(sel
->info
.uses_kill
);
2880 switch (sel
->info
.properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
]) {
2881 case TGSI_FS_DEPTH_LAYOUT_GREATER
:
2882 sel
->db_shader_control
|=
2883 S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_GREATER_THAN_Z
);
2885 case TGSI_FS_DEPTH_LAYOUT_LESS
:
2886 sel
->db_shader_control
|=
2887 S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_LESS_THAN_Z
);
2891 /* Z_ORDER, EXEC_ON_HIER_FAIL and EXEC_ON_NOOP should be set as following:
2893 * | early Z/S | writes_mem | allow_ReZ? | Z_ORDER | EXEC_ON_HIER_FAIL | EXEC_ON_NOOP
2894 * --|-----------|------------|------------|--------------------|-------------------|-------------
2895 * 1a| false | false | true | EarlyZ_Then_ReZ | 0 | 0
2896 * 1b| false | false | false | EarlyZ_Then_LateZ | 0 | 0
2897 * 2 | false | true | n/a | LateZ | 1 | 0
2898 * 3 | true | false | n/a | EarlyZ_Then_LateZ | 0 | 0
2899 * 4 | true | true | n/a | EarlyZ_Then_LateZ | 0 | 1
2901 * In cases 3 and 4, HW will force Z_ORDER to EarlyZ regardless of what's set in the register.
2902 * In case 2, NOOP_CULL is a don't care field. In case 2, 3 and 4, ReZ doesn't make sense.
2904 * Don't use ReZ without profiling !!!
2906 * ReZ decreases performance by 15% in DiRT: Showdown on Ultra settings, which has pretty complex
2909 if (sel
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
]) {
2911 sel
->db_shader_control
|= S_02880C_DEPTH_BEFORE_SHADER(1) |
2912 S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z
) |
2913 S_02880C_EXEC_ON_NOOP(sel
->info
.writes_memory
);
2914 } else if (sel
->info
.writes_memory
) {
2916 sel
->db_shader_control
|= S_02880C_Z_ORDER(V_02880C_LATE_Z
) |
2917 S_02880C_EXEC_ON_HIER_FAIL(1);
2920 sel
->db_shader_control
|= S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z
);
2923 if (sel
->info
.properties
[TGSI_PROPERTY_FS_POST_DEPTH_COVERAGE
])
2924 sel
->db_shader_control
|= S_02880C_PRE_SHADER_DEPTH_COVERAGE_ENABLE(1);
2926 (void) simple_mtx_init(&sel
->mutex
, mtx_plain
);
2928 si_schedule_initial_compile(sctx
, sel
->info
.processor
, &sel
->ready
,
2929 &sel
->compiler_ctx_state
, sel
,
2930 si_init_shader_selector_async
);
2934 static void si_update_streamout_state(struct si_context
*sctx
)
2936 struct si_shader_selector
*shader_with_so
= si_get_vs(sctx
)->cso
;
2938 if (!shader_with_so
)
2941 sctx
->streamout
.enabled_stream_buffers_mask
=
2942 shader_with_so
->enabled_streamout_buffer_mask
;
2943 sctx
->streamout
.stride_in_dw
= shader_with_so
->so
.stride
;
2946 static void si_update_clip_regs(struct si_context
*sctx
,
2947 struct si_shader_selector
*old_hw_vs
,
2948 struct si_shader
*old_hw_vs_variant
,
2949 struct si_shader_selector
*next_hw_vs
,
2950 struct si_shader
*next_hw_vs_variant
)
2954 old_hw_vs
->info
.properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
] !=
2955 next_hw_vs
->info
.properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
] ||
2956 old_hw_vs
->pa_cl_vs_out_cntl
!= next_hw_vs
->pa_cl_vs_out_cntl
||
2957 old_hw_vs
->clipdist_mask
!= next_hw_vs
->clipdist_mask
||
2958 old_hw_vs
->culldist_mask
!= next_hw_vs
->culldist_mask
||
2959 !old_hw_vs_variant
||
2960 !next_hw_vs_variant
||
2961 old_hw_vs_variant
->key
.opt
.clip_disable
!=
2962 next_hw_vs_variant
->key
.opt
.clip_disable
))
2963 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.clip_regs
);
2966 static void si_update_common_shader_state(struct si_context
*sctx
)
2968 sctx
->uses_bindless_samplers
=
2969 si_shader_uses_bindless_samplers(sctx
->vs_shader
.cso
) ||
2970 si_shader_uses_bindless_samplers(sctx
->gs_shader
.cso
) ||
2971 si_shader_uses_bindless_samplers(sctx
->ps_shader
.cso
) ||
2972 si_shader_uses_bindless_samplers(sctx
->tcs_shader
.cso
) ||
2973 si_shader_uses_bindless_samplers(sctx
->tes_shader
.cso
);
2974 sctx
->uses_bindless_images
=
2975 si_shader_uses_bindless_images(sctx
->vs_shader
.cso
) ||
2976 si_shader_uses_bindless_images(sctx
->gs_shader
.cso
) ||
2977 si_shader_uses_bindless_images(sctx
->ps_shader
.cso
) ||
2978 si_shader_uses_bindless_images(sctx
->tcs_shader
.cso
) ||
2979 si_shader_uses_bindless_images(sctx
->tes_shader
.cso
);
2980 sctx
->do_update_shaders
= true;
2983 static void si_bind_vs_shader(struct pipe_context
*ctx
, void *state
)
2985 struct si_context
*sctx
= (struct si_context
*)ctx
;
2986 struct si_shader_selector
*old_hw_vs
= si_get_vs(sctx
)->cso
;
2987 struct si_shader
*old_hw_vs_variant
= si_get_vs_state(sctx
);
2988 struct si_shader_selector
*sel
= state
;
2990 if (sctx
->vs_shader
.cso
== sel
)
2993 sctx
->vs_shader
.cso
= sel
;
2994 sctx
->vs_shader
.current
= sel
? sel
->first_variant
: NULL
;
2995 sctx
->num_vs_blit_sgprs
= sel
? sel
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
] : 0;
2997 if (si_update_ngg(sctx
))
2998 si_shader_change_notify(sctx
);
3000 si_update_common_shader_state(sctx
);
3001 si_update_vs_viewport_state(sctx
);
3002 si_set_active_descriptors_for_shader(sctx
, sel
);
3003 si_update_streamout_state(sctx
);
3004 si_update_clip_regs(sctx
, old_hw_vs
, old_hw_vs_variant
,
3005 si_get_vs(sctx
)->cso
, si_get_vs_state(sctx
));
3008 static void si_update_tess_uses_prim_id(struct si_context
*sctx
)
3010 sctx
->ia_multi_vgt_param_key
.u
.tess_uses_prim_id
=
3011 (sctx
->tes_shader
.cso
&&
3012 sctx
->tes_shader
.cso
->info
.uses_primid
) ||
3013 (sctx
->tcs_shader
.cso
&&
3014 sctx
->tcs_shader
.cso
->info
.uses_primid
) ||
3015 (sctx
->gs_shader
.cso
&&
3016 sctx
->gs_shader
.cso
->info
.uses_primid
) ||
3017 (sctx
->ps_shader
.cso
&& !sctx
->gs_shader
.cso
&&
3018 sctx
->ps_shader
.cso
->info
.uses_primid
);
3021 bool si_update_ngg(struct si_context
*sctx
)
3023 if (!sctx
->screen
->use_ngg
) {
3028 bool new_ngg
= true;
3030 if (sctx
->gs_shader
.cso
&& sctx
->tes_shader
.cso
&&
3031 sctx
->gs_shader
.cso
->tess_turns_off_ngg
) {
3033 } else if (!sctx
->screen
->use_ngg_streamout
) {
3034 struct si_shader_selector
*last
= si_get_vs(sctx
)->cso
;
3036 if ((last
&& last
->so
.num_outputs
) ||
3037 sctx
->streamout
.prims_gen_query_enabled
)
3041 if (new_ngg
!= sctx
->ngg
) {
3042 /* Transitioning from NGG to legacy GS requires VGT_FLUSH on Navi10-14.
3043 * VGT_FLUSH is also emitted at the beginning of IBs when legacy GS ring
3046 if ((sctx
->family
== CHIP_NAVI10
||
3047 sctx
->family
== CHIP_NAVI12
||
3048 sctx
->family
== CHIP_NAVI14
) &&
3050 sctx
->flags
|= SI_CONTEXT_VGT_FLUSH
;
3052 sctx
->ngg
= new_ngg
;
3053 sctx
->last_rast_prim
= -1; /* reset this so that it gets updated */
3059 static void si_bind_gs_shader(struct pipe_context
*ctx
, void *state
)
3061 struct si_context
*sctx
= (struct si_context
*)ctx
;
3062 struct si_shader_selector
*old_hw_vs
= si_get_vs(sctx
)->cso
;
3063 struct si_shader
*old_hw_vs_variant
= si_get_vs_state(sctx
);
3064 struct si_shader_selector
*sel
= state
;
3065 bool enable_changed
= !!sctx
->gs_shader
.cso
!= !!sel
;
3068 if (sctx
->gs_shader
.cso
== sel
)
3071 sctx
->gs_shader
.cso
= sel
;
3072 sctx
->gs_shader
.current
= sel
? sel
->first_variant
: NULL
;
3073 sctx
->ia_multi_vgt_param_key
.u
.uses_gs
= sel
!= NULL
;
3075 si_update_common_shader_state(sctx
);
3076 sctx
->last_rast_prim
= -1; /* reset this so that it gets updated */
3078 ngg_changed
= si_update_ngg(sctx
);
3079 if (ngg_changed
|| enable_changed
)
3080 si_shader_change_notify(sctx
);
3081 if (enable_changed
) {
3082 if (sctx
->ia_multi_vgt_param_key
.u
.uses_tess
)
3083 si_update_tess_uses_prim_id(sctx
);
3085 si_update_vs_viewport_state(sctx
);
3086 si_set_active_descriptors_for_shader(sctx
, sel
);
3087 si_update_streamout_state(sctx
);
3088 si_update_clip_regs(sctx
, old_hw_vs
, old_hw_vs_variant
,
3089 si_get_vs(sctx
)->cso
, si_get_vs_state(sctx
));
3092 static void si_bind_tcs_shader(struct pipe_context
*ctx
, void *state
)
3094 struct si_context
*sctx
= (struct si_context
*)ctx
;
3095 struct si_shader_selector
*sel
= state
;
3096 bool enable_changed
= !!sctx
->tcs_shader
.cso
!= !!sel
;
3098 if (sctx
->tcs_shader
.cso
== sel
)
3101 sctx
->tcs_shader
.cso
= sel
;
3102 sctx
->tcs_shader
.current
= sel
? sel
->first_variant
: NULL
;
3103 si_update_tess_uses_prim_id(sctx
);
3105 si_update_common_shader_state(sctx
);
3108 sctx
->last_tcs
= NULL
; /* invalidate derived tess state */
3110 si_set_active_descriptors_for_shader(sctx
, sel
);
3113 static void si_bind_tes_shader(struct pipe_context
*ctx
, void *state
)
3115 struct si_context
*sctx
= (struct si_context
*)ctx
;
3116 struct si_shader_selector
*old_hw_vs
= si_get_vs(sctx
)->cso
;
3117 struct si_shader
*old_hw_vs_variant
= si_get_vs_state(sctx
);
3118 struct si_shader_selector
*sel
= state
;
3119 bool enable_changed
= !!sctx
->tes_shader
.cso
!= !!sel
;
3121 if (sctx
->tes_shader
.cso
== sel
)
3124 sctx
->tes_shader
.cso
= sel
;
3125 sctx
->tes_shader
.current
= sel
? sel
->first_variant
: NULL
;
3126 sctx
->ia_multi_vgt_param_key
.u
.uses_tess
= sel
!= NULL
;
3127 si_update_tess_uses_prim_id(sctx
);
3129 si_update_common_shader_state(sctx
);
3130 sctx
->last_rast_prim
= -1; /* reset this so that it gets updated */
3132 bool ngg_changed
= si_update_ngg(sctx
);
3133 if (ngg_changed
|| enable_changed
)
3134 si_shader_change_notify(sctx
);
3136 sctx
->last_tes_sh_base
= -1; /* invalidate derived tess state */
3137 si_update_vs_viewport_state(sctx
);
3138 si_set_active_descriptors_for_shader(sctx
, sel
);
3139 si_update_streamout_state(sctx
);
3140 si_update_clip_regs(sctx
, old_hw_vs
, old_hw_vs_variant
,
3141 si_get_vs(sctx
)->cso
, si_get_vs_state(sctx
));
3144 static void si_bind_ps_shader(struct pipe_context
*ctx
, void *state
)
3146 struct si_context
*sctx
= (struct si_context
*)ctx
;
3147 struct si_shader_selector
*old_sel
= sctx
->ps_shader
.cso
;
3148 struct si_shader_selector
*sel
= state
;
3150 /* skip if supplied shader is one already in use */
3154 sctx
->ps_shader
.cso
= sel
;
3155 sctx
->ps_shader
.current
= sel
? sel
->first_variant
: NULL
;
3157 si_update_common_shader_state(sctx
);
3159 if (sctx
->ia_multi_vgt_param_key
.u
.uses_tess
)
3160 si_update_tess_uses_prim_id(sctx
);
3163 old_sel
->info
.colors_written
!= sel
->info
.colors_written
)
3164 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.cb_render_state
);
3166 if (sctx
->screen
->has_out_of_order_rast
&&
3168 old_sel
->info
.writes_memory
!= sel
->info
.writes_memory
||
3169 old_sel
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
] !=
3170 sel
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
]))
3171 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.msaa_config
);
3173 si_set_active_descriptors_for_shader(sctx
, sel
);
3174 si_update_ps_colorbuf0_slot(sctx
);
3177 static void si_delete_shader(struct si_context
*sctx
, struct si_shader
*shader
)
3179 if (shader
->is_optimized
) {
3180 util_queue_drop_job(&sctx
->screen
->shader_compiler_queue_low_priority
,
3184 util_queue_fence_destroy(&shader
->ready
);
3187 /* If destroyed shaders were not unbound, the next compiled
3188 * shader variant could get the same pointer address and so
3189 * binding it to the same shader stage would be considered
3190 * a no-op, causing random behavior.
3192 switch (shader
->selector
->type
) {
3193 case PIPE_SHADER_VERTEX
:
3194 if (shader
->key
.as_ls
) {
3195 assert(sctx
->chip_class
<= GFX8
);
3196 si_pm4_delete_state(sctx
, ls
, shader
->pm4
);
3197 } else if (shader
->key
.as_es
) {
3198 assert(sctx
->chip_class
<= GFX8
);
3199 si_pm4_delete_state(sctx
, es
, shader
->pm4
);
3200 } else if (shader
->key
.as_ngg
) {
3201 si_pm4_delete_state(sctx
, gs
, shader
->pm4
);
3203 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
3206 case PIPE_SHADER_TESS_CTRL
:
3207 si_pm4_delete_state(sctx
, hs
, shader
->pm4
);
3209 case PIPE_SHADER_TESS_EVAL
:
3210 if (shader
->key
.as_es
) {
3211 assert(sctx
->chip_class
<= GFX8
);
3212 si_pm4_delete_state(sctx
, es
, shader
->pm4
);
3213 } else if (shader
->key
.as_ngg
) {
3214 si_pm4_delete_state(sctx
, gs
, shader
->pm4
);
3216 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
3219 case PIPE_SHADER_GEOMETRY
:
3220 if (shader
->is_gs_copy_shader
)
3221 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
3223 si_pm4_delete_state(sctx
, gs
, shader
->pm4
);
3225 case PIPE_SHADER_FRAGMENT
:
3226 si_pm4_delete_state(sctx
, ps
, shader
->pm4
);
3232 si_shader_selector_reference(sctx
, &shader
->previous_stage_sel
, NULL
);
3233 si_shader_destroy(shader
);
3237 void si_destroy_shader_selector(struct si_context
*sctx
,
3238 struct si_shader_selector
*sel
)
3240 struct si_shader
*p
= sel
->first_variant
, *c
;
3241 struct si_shader_ctx_state
*current_shader
[SI_NUM_SHADERS
] = {
3242 [PIPE_SHADER_VERTEX
] = &sctx
->vs_shader
,
3243 [PIPE_SHADER_TESS_CTRL
] = &sctx
->tcs_shader
,
3244 [PIPE_SHADER_TESS_EVAL
] = &sctx
->tes_shader
,
3245 [PIPE_SHADER_GEOMETRY
] = &sctx
->gs_shader
,
3246 [PIPE_SHADER_FRAGMENT
] = &sctx
->ps_shader
,
3249 util_queue_drop_job(&sctx
->screen
->shader_compiler_queue
, &sel
->ready
);
3251 if (current_shader
[sel
->type
]->cso
== sel
) {
3252 current_shader
[sel
->type
]->cso
= NULL
;
3253 current_shader
[sel
->type
]->current
= NULL
;
3257 c
= p
->next_variant
;
3258 si_delete_shader(sctx
, p
);
3262 if (sel
->main_shader_part
)
3263 si_delete_shader(sctx
, sel
->main_shader_part
);
3264 if (sel
->main_shader_part_ls
)
3265 si_delete_shader(sctx
, sel
->main_shader_part_ls
);
3266 if (sel
->main_shader_part_es
)
3267 si_delete_shader(sctx
, sel
->main_shader_part_es
);
3268 if (sel
->main_shader_part_ngg
)
3269 si_delete_shader(sctx
, sel
->main_shader_part_ngg
);
3270 if (sel
->gs_copy_shader
)
3271 si_delete_shader(sctx
, sel
->gs_copy_shader
);
3273 util_queue_fence_destroy(&sel
->ready
);
3274 simple_mtx_destroy(&sel
->mutex
);
3275 ralloc_free(sel
->nir
);
3276 free(sel
->nir_binary
);
3280 static void si_delete_shader_selector(struct pipe_context
*ctx
, void *state
)
3282 struct si_context
*sctx
= (struct si_context
*)ctx
;
3283 struct si_shader_selector
*sel
= (struct si_shader_selector
*)state
;
3285 si_shader_selector_reference(sctx
, &sel
, NULL
);
3288 static unsigned si_get_ps_input_cntl(struct si_context
*sctx
,
3289 struct si_shader
*vs
, unsigned name
,
3290 unsigned index
, unsigned interpolate
)
3292 struct tgsi_shader_info
*vsinfo
= &vs
->selector
->info
;
3293 unsigned j
, offset
, ps_input_cntl
= 0;
3295 if (interpolate
== TGSI_INTERPOLATE_CONSTANT
||
3296 (interpolate
== TGSI_INTERPOLATE_COLOR
&& sctx
->flatshade
) ||
3297 name
== TGSI_SEMANTIC_PRIMID
)
3298 ps_input_cntl
|= S_028644_FLAT_SHADE(1);
3300 if (name
== TGSI_SEMANTIC_PCOORD
||
3301 (name
== TGSI_SEMANTIC_TEXCOORD
&&
3302 sctx
->sprite_coord_enable
& (1 << index
))) {
3303 ps_input_cntl
|= S_028644_PT_SPRITE_TEX(1);
3306 for (j
= 0; j
< vsinfo
->num_outputs
; j
++) {
3307 if (name
== vsinfo
->output_semantic_name
[j
] &&
3308 index
== vsinfo
->output_semantic_index
[j
]) {
3309 offset
= vs
->info
.vs_output_param_offset
[j
];
3311 if (offset
<= AC_EXP_PARAM_OFFSET_31
) {
3312 /* The input is loaded from parameter memory. */
3313 ps_input_cntl
|= S_028644_OFFSET(offset
);
3314 } else if (!G_028644_PT_SPRITE_TEX(ps_input_cntl
)) {
3315 if (offset
== AC_EXP_PARAM_UNDEFINED
) {
3316 /* This can happen with depth-only rendering. */
3319 /* The input is a DEFAULT_VAL constant. */
3320 assert(offset
>= AC_EXP_PARAM_DEFAULT_VAL_0000
&&
3321 offset
<= AC_EXP_PARAM_DEFAULT_VAL_1111
);
3322 offset
-= AC_EXP_PARAM_DEFAULT_VAL_0000
;
3325 ps_input_cntl
= S_028644_OFFSET(0x20) |
3326 S_028644_DEFAULT_VAL(offset
);
3332 if (j
== vsinfo
->num_outputs
&& name
== TGSI_SEMANTIC_PRIMID
)
3333 /* PrimID is written after the last output when HW VS is used. */
3334 ps_input_cntl
|= S_028644_OFFSET(vs
->info
.vs_output_param_offset
[vsinfo
->num_outputs
]);
3335 else if (j
== vsinfo
->num_outputs
&& !G_028644_PT_SPRITE_TEX(ps_input_cntl
)) {
3336 /* No corresponding output found, load defaults into input.
3337 * Don't set any other bits.
3338 * (FLAT_SHADE=1 completely changes behavior) */
3339 ps_input_cntl
= S_028644_OFFSET(0x20);
3340 /* D3D 9 behaviour. GL is undefined */
3341 if (name
== TGSI_SEMANTIC_COLOR
&& index
== 0)
3342 ps_input_cntl
|= S_028644_DEFAULT_VAL(3);
3344 return ps_input_cntl
;
3347 static void si_emit_spi_map(struct si_context
*sctx
)
3349 struct si_shader
*ps
= sctx
->ps_shader
.current
;
3350 struct si_shader
*vs
= si_get_vs_state(sctx
);
3351 struct tgsi_shader_info
*psinfo
= ps
? &ps
->selector
->info
: NULL
;
3352 unsigned i
, num_interp
, num_written
= 0, bcol_interp
[2];
3353 unsigned spi_ps_input_cntl
[32];
3355 if (!ps
|| !ps
->selector
->info
.num_inputs
)
3358 num_interp
= si_get_ps_num_interp(ps
);
3359 assert(num_interp
> 0);
3361 for (i
= 0; i
< psinfo
->num_inputs
; i
++) {
3362 unsigned name
= psinfo
->input_semantic_name
[i
];
3363 unsigned index
= psinfo
->input_semantic_index
[i
];
3364 unsigned interpolate
= psinfo
->input_interpolate
[i
];
3366 spi_ps_input_cntl
[num_written
++] = si_get_ps_input_cntl(sctx
, vs
, name
,
3367 index
, interpolate
);
3369 if (name
== TGSI_SEMANTIC_COLOR
) {
3370 assert(index
< ARRAY_SIZE(bcol_interp
));
3371 bcol_interp
[index
] = interpolate
;
3375 if (ps
->key
.part
.ps
.prolog
.color_two_side
) {
3376 unsigned bcol
= TGSI_SEMANTIC_BCOLOR
;
3378 for (i
= 0; i
< 2; i
++) {
3379 if (!(psinfo
->colors_read
& (0xf << (i
* 4))))
3382 spi_ps_input_cntl
[num_written
++] =
3383 si_get_ps_input_cntl(sctx
, vs
, bcol
, i
, bcol_interp
[i
]);
3387 assert(num_interp
== num_written
);
3389 /* R_028644_SPI_PS_INPUT_CNTL_0 */
3390 /* Dota 2: Only ~16% of SPI map updates set different values. */
3391 /* Talos: Only ~9% of SPI map updates set different values. */
3392 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
3393 radeon_opt_set_context_regn(sctx
, R_028644_SPI_PS_INPUT_CNTL_0
,
3395 sctx
->tracked_regs
.spi_ps_input_cntl
, num_interp
);
3397 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
3398 sctx
->context_roll
= true;
3402 * Writing CONFIG or UCONFIG VGT registers requires VGT_FLUSH before that.
3404 static void si_init_config_add_vgt_flush(struct si_context
*sctx
)
3406 if (sctx
->init_config_has_vgt_flush
)
3409 /* Done by Vulkan before VGT_FLUSH. */
3410 si_pm4_cmd_begin(sctx
->init_config
, PKT3_EVENT_WRITE
);
3411 si_pm4_cmd_add(sctx
->init_config
,
3412 EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
3413 si_pm4_cmd_end(sctx
->init_config
, false);
3415 /* VGT_FLUSH is required even if VGT is idle. It resets VGT pointers. */
3416 si_pm4_cmd_begin(sctx
->init_config
, PKT3_EVENT_WRITE
);
3417 si_pm4_cmd_add(sctx
->init_config
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
3418 si_pm4_cmd_end(sctx
->init_config
, false);
3419 sctx
->init_config_has_vgt_flush
= true;
3422 /* Initialize state related to ESGS / GSVS ring buffers */
3423 static bool si_update_gs_ring_buffers(struct si_context
*sctx
)
3425 struct si_shader_selector
*es
=
3426 sctx
->tes_shader
.cso
? sctx
->tes_shader
.cso
: sctx
->vs_shader
.cso
;
3427 struct si_shader_selector
*gs
= sctx
->gs_shader
.cso
;
3428 struct si_pm4_state
*pm4
;
3430 /* Chip constants. */
3431 unsigned num_se
= sctx
->screen
->info
.max_se
;
3432 unsigned wave_size
= 64;
3433 unsigned max_gs_waves
= 32 * num_se
; /* max 32 per SE on GCN */
3434 /* On GFX6-GFX7, the value comes from VGT_GS_VERTEX_REUSE = 16.
3435 * On GFX8+, the value comes from VGT_VERTEX_REUSE_BLOCK_CNTL = 30 (+2).
3437 unsigned gs_vertex_reuse
= (sctx
->chip_class
>= GFX8
? 32 : 16) * num_se
;
3438 unsigned alignment
= 256 * num_se
;
3439 /* The maximum size is 63.999 MB per SE. */
3440 unsigned max_size
= ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se
;
3442 /* Calculate the minimum size. */
3443 unsigned min_esgs_ring_size
= align(es
->esgs_itemsize
* gs_vertex_reuse
*
3444 wave_size
, alignment
);
3446 /* These are recommended sizes, not minimum sizes. */
3447 unsigned esgs_ring_size
= max_gs_waves
* 2 * wave_size
*
3448 es
->esgs_itemsize
* gs
->gs_input_verts_per_prim
;
3449 unsigned gsvs_ring_size
= max_gs_waves
* 2 * wave_size
*
3450 gs
->max_gsvs_emit_size
;
3452 min_esgs_ring_size
= align(min_esgs_ring_size
, alignment
);
3453 esgs_ring_size
= align(esgs_ring_size
, alignment
);
3454 gsvs_ring_size
= align(gsvs_ring_size
, alignment
);
3456 esgs_ring_size
= CLAMP(esgs_ring_size
, min_esgs_ring_size
, max_size
);
3457 gsvs_ring_size
= MIN2(gsvs_ring_size
, max_size
);
3459 /* Some rings don't have to be allocated if shaders don't use them.
3460 * (e.g. no varyings between ES and GS or GS and VS)
3462 * GFX9 doesn't have the ESGS ring.
3464 bool update_esgs
= sctx
->chip_class
<= GFX8
&&
3466 (!sctx
->esgs_ring
||
3467 sctx
->esgs_ring
->width0
< esgs_ring_size
);
3468 bool update_gsvs
= gsvs_ring_size
&&
3469 (!sctx
->gsvs_ring
||
3470 sctx
->gsvs_ring
->width0
< gsvs_ring_size
);
3472 if (!update_esgs
&& !update_gsvs
)
3476 pipe_resource_reference(&sctx
->esgs_ring
, NULL
);
3478 pipe_aligned_buffer_create(sctx
->b
.screen
,
3479 SI_RESOURCE_FLAG_UNMAPPABLE
,
3482 sctx
->screen
->info
.pte_fragment_size
);
3483 if (!sctx
->esgs_ring
)
3488 pipe_resource_reference(&sctx
->gsvs_ring
, NULL
);
3490 pipe_aligned_buffer_create(sctx
->b
.screen
,
3491 SI_RESOURCE_FLAG_UNMAPPABLE
,
3494 sctx
->screen
->info
.pte_fragment_size
);
3495 if (!sctx
->gsvs_ring
)
3499 /* Create the "init_config_gs_rings" state. */
3500 pm4
= CALLOC_STRUCT(si_pm4_state
);
3504 if (sctx
->chip_class
>= GFX7
) {
3505 if (sctx
->esgs_ring
) {
3506 assert(sctx
->chip_class
<= GFX8
);
3507 si_pm4_set_reg(pm4
, R_030900_VGT_ESGS_RING_SIZE
,
3508 sctx
->esgs_ring
->width0
/ 256);
3510 if (sctx
->gsvs_ring
)
3511 si_pm4_set_reg(pm4
, R_030904_VGT_GSVS_RING_SIZE
,
3512 sctx
->gsvs_ring
->width0
/ 256);
3514 if (sctx
->esgs_ring
)
3515 si_pm4_set_reg(pm4
, R_0088C8_VGT_ESGS_RING_SIZE
,
3516 sctx
->esgs_ring
->width0
/ 256);
3517 if (sctx
->gsvs_ring
)
3518 si_pm4_set_reg(pm4
, R_0088CC_VGT_GSVS_RING_SIZE
,
3519 sctx
->gsvs_ring
->width0
/ 256);
3522 /* Set the state. */
3523 if (sctx
->init_config_gs_rings
)
3524 si_pm4_free_state(sctx
, sctx
->init_config_gs_rings
, ~0);
3525 sctx
->init_config_gs_rings
= pm4
;
3527 if (!sctx
->init_config_has_vgt_flush
) {
3528 si_init_config_add_vgt_flush(sctx
);
3529 si_pm4_upload_indirect_buffer(sctx
, sctx
->init_config
);
3532 /* Flush the context to re-emit both init_config states. */
3533 sctx
->initial_gfx_cs_size
= 0; /* force flush */
3534 si_flush_gfx_cs(sctx
, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW
, NULL
);
3536 /* Set ring bindings. */
3537 if (sctx
->esgs_ring
) {
3538 assert(sctx
->chip_class
<= GFX8
);
3539 si_set_ring_buffer(sctx
, SI_ES_RING_ESGS
,
3540 sctx
->esgs_ring
, 0, sctx
->esgs_ring
->width0
,
3541 true, true, 4, 64, 0);
3542 si_set_ring_buffer(sctx
, SI_GS_RING_ESGS
,
3543 sctx
->esgs_ring
, 0, sctx
->esgs_ring
->width0
,
3544 false, false, 0, 0, 0);
3546 if (sctx
->gsvs_ring
) {
3547 si_set_ring_buffer(sctx
, SI_RING_GSVS
,
3548 sctx
->gsvs_ring
, 0, sctx
->gsvs_ring
->width0
,
3549 false, false, 0, 0, 0);
3555 static void si_shader_lock(struct si_shader
*shader
)
3557 simple_mtx_lock(&shader
->selector
->mutex
);
3558 if (shader
->previous_stage_sel
) {
3559 assert(shader
->previous_stage_sel
!= shader
->selector
);
3560 simple_mtx_lock(&shader
->previous_stage_sel
->mutex
);
3564 static void si_shader_unlock(struct si_shader
*shader
)
3566 if (shader
->previous_stage_sel
)
3567 simple_mtx_unlock(&shader
->previous_stage_sel
->mutex
);
3568 simple_mtx_unlock(&shader
->selector
->mutex
);
3572 * @returns 1 if \p sel has been updated to use a new scratch buffer
3574 * < 0 if there was a failure
3576 static int si_update_scratch_buffer(struct si_context
*sctx
,
3577 struct si_shader
*shader
)
3579 uint64_t scratch_va
= sctx
->scratch_buffer
->gpu_address
;
3584 /* This shader doesn't need a scratch buffer */
3585 if (shader
->config
.scratch_bytes_per_wave
== 0)
3588 /* Prevent race conditions when updating:
3589 * - si_shader::scratch_bo
3590 * - si_shader::binary::code
3591 * - si_shader::previous_stage::binary::code.
3593 si_shader_lock(shader
);
3595 /* This shader is already configured to use the current
3596 * scratch buffer. */
3597 if (shader
->scratch_bo
== sctx
->scratch_buffer
) {
3598 si_shader_unlock(shader
);
3602 assert(sctx
->scratch_buffer
);
3604 /* Replace the shader bo with a new bo that has the relocs applied. */
3605 if (!si_shader_binary_upload(sctx
->screen
, shader
, scratch_va
)) {
3606 si_shader_unlock(shader
);
3610 /* Update the shader state to use the new shader bo. */
3611 si_shader_init_pm4_state(sctx
->screen
, shader
);
3613 si_resource_reference(&shader
->scratch_bo
, sctx
->scratch_buffer
);
3615 si_shader_unlock(shader
);
3619 static unsigned si_get_scratch_buffer_bytes_per_wave(struct si_shader
*shader
)
3621 return shader
? shader
->config
.scratch_bytes_per_wave
: 0;
3624 static struct si_shader
*si_get_tcs_current(struct si_context
*sctx
)
3626 if (!sctx
->tes_shader
.cso
)
3627 return NULL
; /* tessellation disabled */
3629 return sctx
->tcs_shader
.cso
? sctx
->tcs_shader
.current
:
3630 sctx
->fixed_func_tcs_shader
.current
;
3633 static bool si_update_scratch_relocs(struct si_context
*sctx
)
3635 struct si_shader
*tcs
= si_get_tcs_current(sctx
);
3638 /* Update the shaders, so that they are using the latest scratch.
3639 * The scratch buffer may have been changed since these shaders were
3640 * last used, so we still need to try to update them, even if they
3641 * require scratch buffers smaller than the current size.
3643 r
= si_update_scratch_buffer(sctx
, sctx
->ps_shader
.current
);
3647 si_pm4_bind_state(sctx
, ps
, sctx
->ps_shader
.current
->pm4
);
3649 r
= si_update_scratch_buffer(sctx
, sctx
->gs_shader
.current
);
3653 si_pm4_bind_state(sctx
, gs
, sctx
->gs_shader
.current
->pm4
);
3655 r
= si_update_scratch_buffer(sctx
, tcs
);
3659 si_pm4_bind_state(sctx
, hs
, tcs
->pm4
);
3661 /* VS can be bound as LS, ES, or VS. */
3662 r
= si_update_scratch_buffer(sctx
, sctx
->vs_shader
.current
);
3666 if (sctx
->vs_shader
.current
->key
.as_ls
)
3667 si_pm4_bind_state(sctx
, ls
, sctx
->vs_shader
.current
->pm4
);
3668 else if (sctx
->vs_shader
.current
->key
.as_es
)
3669 si_pm4_bind_state(sctx
, es
, sctx
->vs_shader
.current
->pm4
);
3670 else if (sctx
->vs_shader
.current
->key
.as_ngg
)
3671 si_pm4_bind_state(sctx
, gs
, sctx
->vs_shader
.current
->pm4
);
3673 si_pm4_bind_state(sctx
, vs
, sctx
->vs_shader
.current
->pm4
);
3676 /* TES can be bound as ES or VS. */
3677 r
= si_update_scratch_buffer(sctx
, sctx
->tes_shader
.current
);
3681 if (sctx
->tes_shader
.current
->key
.as_es
)
3682 si_pm4_bind_state(sctx
, es
, sctx
->tes_shader
.current
->pm4
);
3683 else if (sctx
->tes_shader
.current
->key
.as_ngg
)
3684 si_pm4_bind_state(sctx
, gs
, sctx
->tes_shader
.current
->pm4
);
3686 si_pm4_bind_state(sctx
, vs
, sctx
->tes_shader
.current
->pm4
);
3692 static bool si_update_spi_tmpring_size(struct si_context
*sctx
)
3694 /* SPI_TMPRING_SIZE.WAVESIZE must be constant for each scratch buffer.
3695 * There are 2 cases to handle:
3697 * - If the current needed size is less than the maximum seen size,
3698 * use the maximum seen size, so that WAVESIZE remains the same.
3700 * - If the current needed size is greater than the maximum seen size,
3701 * the scratch buffer is reallocated, so we can increase WAVESIZE.
3703 * Shaders that set SCRATCH_EN=0 don't allocate scratch space.
3704 * Otherwise, the number of waves that can use scratch is
3705 * SPI_TMPRING_SIZE.WAVES.
3709 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->ps_shader
.current
));
3710 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->gs_shader
.current
));
3711 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->vs_shader
.current
));
3713 if (sctx
->tes_shader
.cso
) {
3714 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->tes_shader
.current
));
3715 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(si_get_tcs_current(sctx
)));
3718 sctx
->max_seen_scratch_bytes_per_wave
=
3719 MAX2(sctx
->max_seen_scratch_bytes_per_wave
, bytes
);
3721 unsigned scratch_needed_size
=
3722 sctx
->max_seen_scratch_bytes_per_wave
* sctx
->scratch_waves
;
3723 unsigned spi_tmpring_size
;
3725 if (scratch_needed_size
> 0) {
3726 if (!sctx
->scratch_buffer
||
3727 scratch_needed_size
> sctx
->scratch_buffer
->b
.b
.width0
) {
3728 /* Create a bigger scratch buffer */
3729 si_resource_reference(&sctx
->scratch_buffer
, NULL
);
3731 sctx
->scratch_buffer
=
3732 si_aligned_buffer_create(&sctx
->screen
->b
,
3733 SI_RESOURCE_FLAG_UNMAPPABLE
,
3735 scratch_needed_size
,
3736 sctx
->screen
->info
.pte_fragment_size
);
3737 if (!sctx
->scratch_buffer
)
3740 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.scratch_state
);
3741 si_context_add_resource_size(sctx
,
3742 &sctx
->scratch_buffer
->b
.b
);
3745 if (!si_update_scratch_relocs(sctx
))
3749 /* The LLVM shader backend should be reporting aligned scratch_sizes. */
3750 assert((scratch_needed_size
& ~0x3FF) == scratch_needed_size
&&
3751 "scratch size should already be aligned correctly.");
3753 spi_tmpring_size
= S_0286E8_WAVES(sctx
->scratch_waves
) |
3754 S_0286E8_WAVESIZE(sctx
->max_seen_scratch_bytes_per_wave
>> 10);
3755 if (spi_tmpring_size
!= sctx
->spi_tmpring_size
) {
3756 sctx
->spi_tmpring_size
= spi_tmpring_size
;
3757 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.scratch_state
);
3762 static void si_init_tess_factor_ring(struct si_context
*sctx
)
3764 assert(!sctx
->tess_rings
);
3765 assert(((sctx
->screen
->tess_factor_ring_size
/ 4) & C_030938_SIZE
) == 0);
3767 /* The address must be aligned to 2^19, because the shader only
3768 * receives the high 13 bits.
3770 sctx
->tess_rings
= pipe_aligned_buffer_create(sctx
->b
.screen
,
3771 SI_RESOURCE_FLAG_32BIT
,
3773 sctx
->screen
->tess_offchip_ring_size
+
3774 sctx
->screen
->tess_factor_ring_size
,
3776 if (!sctx
->tess_rings
)
3779 si_init_config_add_vgt_flush(sctx
);
3781 si_pm4_add_bo(sctx
->init_config
, si_resource(sctx
->tess_rings
),
3782 RADEON_USAGE_READWRITE
, RADEON_PRIO_SHADER_RINGS
);
3784 uint64_t factor_va
= si_resource(sctx
->tess_rings
)->gpu_address
+
3785 sctx
->screen
->tess_offchip_ring_size
;
3787 /* Append these registers to the init config state. */
3788 if (sctx
->chip_class
>= GFX7
) {
3789 si_pm4_set_reg(sctx
->init_config
, R_030938_VGT_TF_RING_SIZE
,
3790 S_030938_SIZE(sctx
->screen
->tess_factor_ring_size
/ 4));
3791 si_pm4_set_reg(sctx
->init_config
, R_030940_VGT_TF_MEMORY_BASE
,
3793 if (sctx
->chip_class
>= GFX10
)
3794 si_pm4_set_reg(sctx
->init_config
, R_030984_VGT_TF_MEMORY_BASE_HI_UMD
,
3795 S_030984_BASE_HI(factor_va
>> 40));
3796 else if (sctx
->chip_class
== GFX9
)
3797 si_pm4_set_reg(sctx
->init_config
, R_030944_VGT_TF_MEMORY_BASE_HI
,
3798 S_030944_BASE_HI(factor_va
>> 40));
3799 si_pm4_set_reg(sctx
->init_config
, R_03093C_VGT_HS_OFFCHIP_PARAM
,
3800 sctx
->screen
->vgt_hs_offchip_param
);
3802 si_pm4_set_reg(sctx
->init_config
, R_008988_VGT_TF_RING_SIZE
,
3803 S_008988_SIZE(sctx
->screen
->tess_factor_ring_size
/ 4));
3804 si_pm4_set_reg(sctx
->init_config
, R_0089B8_VGT_TF_MEMORY_BASE
,
3806 si_pm4_set_reg(sctx
->init_config
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
3807 sctx
->screen
->vgt_hs_offchip_param
);
3810 /* Flush the context to re-emit the init_config state.
3811 * This is done only once in a lifetime of a context.
3813 si_pm4_upload_indirect_buffer(sctx
, sctx
->init_config
);
3814 sctx
->initial_gfx_cs_size
= 0; /* force flush */
3815 si_flush_gfx_cs(sctx
, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW
, NULL
);
3818 static struct si_pm4_state
*si_build_vgt_shader_config(struct si_screen
*screen
,
3819 union si_vgt_stages_key key
)
3821 struct si_pm4_state
*pm4
= CALLOC_STRUCT(si_pm4_state
);
3822 uint32_t stages
= 0;
3825 stages
|= S_028B54_LS_EN(V_028B54_LS_STAGE_ON
) |
3826 S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
3829 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_DS
) |
3832 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_DS
);
3834 stages
|= S_028B54_VS_EN(V_028B54_VS_STAGE_DS
);
3835 } else if (key
.u
.gs
) {
3836 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL
) |
3838 } else if (key
.u
.ngg
) {
3839 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL
);
3843 stages
|= S_028B54_PRIMGEN_EN(1) |
3844 S_028B54_NGG_WAVE_ID_EN(key
.u
.streamout
) |
3845 S_028B54_PRIMGEN_PASSTHRU_EN(key
.u
.ngg_passthrough
);
3846 } else if (key
.u
.gs
)
3847 stages
|= S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
3849 if (screen
->info
.chip_class
>= GFX9
)
3850 stages
|= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
3852 if (screen
->info
.chip_class
>= GFX10
&& screen
->ge_wave_size
== 32) {
3853 stages
|= S_028B54_HS_W32_EN(1) |
3854 S_028B54_GS_W32_EN(key
.u
.ngg
) | /* legacy GS only supports Wave64 */
3855 S_028B54_VS_W32_EN(1);
3858 si_pm4_set_reg(pm4
, R_028B54_VGT_SHADER_STAGES_EN
, stages
);
3862 static void si_update_vgt_shader_config(struct si_context
*sctx
,
3863 union si_vgt_stages_key key
)
3865 struct si_pm4_state
**pm4
= &sctx
->vgt_shader_config
[key
.index
];
3867 if (unlikely(!*pm4
))
3868 *pm4
= si_build_vgt_shader_config(sctx
->screen
, key
);
3869 si_pm4_bind_state(sctx
, vgt_shader_config
, *pm4
);
3872 bool si_update_shaders(struct si_context
*sctx
)
3874 struct pipe_context
*ctx
= (struct pipe_context
*)sctx
;
3875 struct si_compiler_ctx_state compiler_state
;
3876 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
3877 struct si_shader
*old_vs
= si_get_vs_state(sctx
);
3878 bool old_clip_disable
= old_vs
? old_vs
->key
.opt
.clip_disable
: false;
3879 struct si_shader
*old_ps
= sctx
->ps_shader
.current
;
3880 union si_vgt_stages_key key
;
3881 unsigned old_spi_shader_col_format
=
3882 old_ps
? old_ps
->key
.part
.ps
.epilog
.spi_shader_col_format
: 0;
3885 if (!sctx
->compiler
.passes
)
3886 si_init_compiler(sctx
->screen
, &sctx
->compiler
);
3888 compiler_state
.compiler
= &sctx
->compiler
;
3889 compiler_state
.debug
= sctx
->debug
;
3890 compiler_state
.is_debug_context
= sctx
->is_debug
;
3894 if (sctx
->tes_shader
.cso
)
3896 if (sctx
->gs_shader
.cso
)
3901 key
.u
.streamout
= !!si_get_vs(sctx
)->cso
->so
.num_outputs
;
3904 /* Update TCS and TES. */
3905 if (sctx
->tes_shader
.cso
) {
3906 if (!sctx
->tess_rings
) {
3907 si_init_tess_factor_ring(sctx
);
3908 if (!sctx
->tess_rings
)
3912 if (sctx
->tcs_shader
.cso
) {
3913 r
= si_shader_select(ctx
, &sctx
->tcs_shader
, key
,
3917 si_pm4_bind_state(sctx
, hs
, sctx
->tcs_shader
.current
->pm4
);
3919 if (!sctx
->fixed_func_tcs_shader
.cso
) {
3920 sctx
->fixed_func_tcs_shader
.cso
=
3921 si_create_fixed_func_tcs(sctx
);
3922 if (!sctx
->fixed_func_tcs_shader
.cso
)
3926 r
= si_shader_select(ctx
, &sctx
->fixed_func_tcs_shader
,
3927 key
, &compiler_state
);
3930 si_pm4_bind_state(sctx
, hs
,
3931 sctx
->fixed_func_tcs_shader
.current
->pm4
);
3934 if (!sctx
->gs_shader
.cso
|| sctx
->chip_class
<= GFX8
) {
3935 r
= si_shader_select(ctx
, &sctx
->tes_shader
, key
, &compiler_state
);
3939 if (sctx
->gs_shader
.cso
) {
3941 assert(sctx
->chip_class
<= GFX8
);
3942 si_pm4_bind_state(sctx
, es
, sctx
->tes_shader
.current
->pm4
);
3943 } else if (key
.u
.ngg
) {
3944 si_pm4_bind_state(sctx
, gs
, sctx
->tes_shader
.current
->pm4
);
3946 si_pm4_bind_state(sctx
, vs
, sctx
->tes_shader
.current
->pm4
);
3950 if (sctx
->chip_class
<= GFX8
)
3951 si_pm4_bind_state(sctx
, ls
, NULL
);
3952 si_pm4_bind_state(sctx
, hs
, NULL
);
3956 if (sctx
->gs_shader
.cso
) {
3957 r
= si_shader_select(ctx
, &sctx
->gs_shader
, key
, &compiler_state
);
3960 si_pm4_bind_state(sctx
, gs
, sctx
->gs_shader
.current
->pm4
);
3962 si_pm4_bind_state(sctx
, vs
, sctx
->gs_shader
.cso
->gs_copy_shader
->pm4
);
3964 if (!si_update_gs_ring_buffers(sctx
))
3967 si_pm4_bind_state(sctx
, vs
, NULL
);
3971 si_pm4_bind_state(sctx
, gs
, NULL
);
3972 if (sctx
->chip_class
<= GFX8
)
3973 si_pm4_bind_state(sctx
, es
, NULL
);
3978 if ((!key
.u
.tess
&& !key
.u
.gs
) || sctx
->chip_class
<= GFX8
) {
3979 r
= si_shader_select(ctx
, &sctx
->vs_shader
, key
, &compiler_state
);
3983 if (!key
.u
.tess
&& !key
.u
.gs
) {
3985 si_pm4_bind_state(sctx
, gs
, sctx
->vs_shader
.current
->pm4
);
3986 si_pm4_bind_state(sctx
, vs
, NULL
);
3988 si_pm4_bind_state(sctx
, vs
, sctx
->vs_shader
.current
->pm4
);
3990 } else if (sctx
->tes_shader
.cso
) {
3991 si_pm4_bind_state(sctx
, ls
, sctx
->vs_shader
.current
->pm4
);
3993 assert(sctx
->gs_shader
.cso
);
3994 si_pm4_bind_state(sctx
, es
, sctx
->vs_shader
.current
->pm4
);
3998 /* This must be done after the shader variant is selected. */
4000 key
.u
.ngg_passthrough
= gfx10_is_ngg_passthrough(si_get_vs(sctx
)->current
);
4002 si_update_vgt_shader_config(sctx
, key
);
4004 if (old_clip_disable
!= si_get_vs_state(sctx
)->key
.opt
.clip_disable
)
4005 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.clip_regs
);
4007 if (sctx
->ps_shader
.cso
) {
4008 unsigned db_shader_control
;
4010 r
= si_shader_select(ctx
, &sctx
->ps_shader
, key
, &compiler_state
);
4013 si_pm4_bind_state(sctx
, ps
, sctx
->ps_shader
.current
->pm4
);
4016 sctx
->ps_shader
.cso
->db_shader_control
|
4017 S_02880C_KILL_ENABLE(si_get_alpha_test_func(sctx
) != PIPE_FUNC_ALWAYS
);
4019 if (si_pm4_state_changed(sctx
, ps
) ||
4020 si_pm4_state_changed(sctx
, vs
) ||
4021 (key
.u
.ngg
&& si_pm4_state_changed(sctx
, gs
)) ||
4022 sctx
->sprite_coord_enable
!= rs
->sprite_coord_enable
||
4023 sctx
->flatshade
!= rs
->flatshade
) {
4024 sctx
->sprite_coord_enable
= rs
->sprite_coord_enable
;
4025 sctx
->flatshade
= rs
->flatshade
;
4026 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.spi_map
);
4029 if (sctx
->screen
->info
.rbplus_allowed
&&
4030 si_pm4_state_changed(sctx
, ps
) &&
4032 old_spi_shader_col_format
!=
4033 sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.spi_shader_col_format
))
4034 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.cb_render_state
);
4036 if (sctx
->ps_db_shader_control
!= db_shader_control
) {
4037 sctx
->ps_db_shader_control
= db_shader_control
;
4038 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.db_render_state
);
4039 if (sctx
->screen
->dpbb_allowed
)
4040 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.dpbb_state
);
4043 if (sctx
->smoothing_enabled
!= sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.poly_line_smoothing
) {
4044 sctx
->smoothing_enabled
= sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.poly_line_smoothing
;
4045 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.msaa_config
);
4047 if (sctx
->chip_class
== GFX6
)
4048 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.db_render_state
);
4050 if (sctx
->framebuffer
.nr_samples
<= 1)
4051 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.msaa_sample_locs
);
4055 if (si_pm4_state_enabled_and_changed(sctx
, ls
) ||
4056 si_pm4_state_enabled_and_changed(sctx
, hs
) ||
4057 si_pm4_state_enabled_and_changed(sctx
, es
) ||
4058 si_pm4_state_enabled_and_changed(sctx
, gs
) ||
4059 si_pm4_state_enabled_and_changed(sctx
, vs
) ||
4060 si_pm4_state_enabled_and_changed(sctx
, ps
)) {
4061 if (!si_update_spi_tmpring_size(sctx
))
4065 if (sctx
->chip_class
>= GFX7
) {
4066 if (si_pm4_state_enabled_and_changed(sctx
, ls
))
4067 sctx
->prefetch_L2_mask
|= SI_PREFETCH_LS
;
4068 else if (!sctx
->queued
.named
.ls
)
4069 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_LS
;
4071 if (si_pm4_state_enabled_and_changed(sctx
, hs
))
4072 sctx
->prefetch_L2_mask
|= SI_PREFETCH_HS
;
4073 else if (!sctx
->queued
.named
.hs
)
4074 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_HS
;
4076 if (si_pm4_state_enabled_and_changed(sctx
, es
))
4077 sctx
->prefetch_L2_mask
|= SI_PREFETCH_ES
;
4078 else if (!sctx
->queued
.named
.es
)
4079 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_ES
;
4081 if (si_pm4_state_enabled_and_changed(sctx
, gs
))
4082 sctx
->prefetch_L2_mask
|= SI_PREFETCH_GS
;
4083 else if (!sctx
->queued
.named
.gs
)
4084 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_GS
;
4086 if (si_pm4_state_enabled_and_changed(sctx
, vs
))
4087 sctx
->prefetch_L2_mask
|= SI_PREFETCH_VS
;
4088 else if (!sctx
->queued
.named
.vs
)
4089 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_VS
;
4091 if (si_pm4_state_enabled_and_changed(sctx
, ps
))
4092 sctx
->prefetch_L2_mask
|= SI_PREFETCH_PS
;
4093 else if (!sctx
->queued
.named
.ps
)
4094 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_PS
;
4097 sctx
->do_update_shaders
= false;
4101 static void si_emit_scratch_state(struct si_context
*sctx
)
4103 struct radeon_cmdbuf
*cs
= sctx
->gfx_cs
;
4105 radeon_set_context_reg(cs
, R_0286E8_SPI_TMPRING_SIZE
,
4106 sctx
->spi_tmpring_size
);
4108 if (sctx
->scratch_buffer
) {
4109 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
4110 sctx
->scratch_buffer
, RADEON_USAGE_READWRITE
,
4111 RADEON_PRIO_SCRATCH_BUFFER
);
4115 void si_init_shader_functions(struct si_context
*sctx
)
4117 sctx
->atoms
.s
.spi_map
.emit
= si_emit_spi_map
;
4118 sctx
->atoms
.s
.scratch_state
.emit
= si_emit_scratch_state
;
4120 sctx
->b
.create_vs_state
= si_create_shader_selector
;
4121 sctx
->b
.create_tcs_state
= si_create_shader_selector
;
4122 sctx
->b
.create_tes_state
= si_create_shader_selector
;
4123 sctx
->b
.create_gs_state
= si_create_shader_selector
;
4124 sctx
->b
.create_fs_state
= si_create_shader_selector
;
4126 sctx
->b
.bind_vs_state
= si_bind_vs_shader
;
4127 sctx
->b
.bind_tcs_state
= si_bind_tcs_shader
;
4128 sctx
->b
.bind_tes_state
= si_bind_tes_shader
;
4129 sctx
->b
.bind_gs_state
= si_bind_gs_shader
;
4130 sctx
->b
.bind_fs_state
= si_bind_ps_shader
;
4132 sctx
->b
.delete_vs_state
= si_delete_shader_selector
;
4133 sctx
->b
.delete_tcs_state
= si_delete_shader_selector
;
4134 sctx
->b
.delete_tes_state
= si_delete_shader_selector
;
4135 sctx
->b
.delete_gs_state
= si_delete_shader_selector
;
4136 sctx
->b
.delete_fs_state
= si_delete_shader_selector
;