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 si_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(struct si_shader
*shader
,
461 unsigned num_always_on_user_sgprs
)
463 struct si_shader_selector
*vs
= shader
->previous_stage_sel
?
464 shader
->previous_stage_sel
: shader
->selector
;
465 unsigned num_vbos_in_user_sgprs
= vs
->num_vbos_in_user_sgprs
;
467 /* 1 SGPR is reserved for the vertex buffer pointer. */
468 assert(num_always_on_user_sgprs
<= SI_SGPR_VS_VB_DESCRIPTOR_FIRST
- 1);
470 if (num_vbos_in_user_sgprs
)
471 return SI_SGPR_VS_VB_DESCRIPTOR_FIRST
+ num_vbos_in_user_sgprs
* 4;
473 /* Add the pointer to VBO descriptors. */
474 return num_always_on_user_sgprs
+ 1;
477 /* Return VGPR_COMP_CNT for the API vertex shader. This can be hw LS, LSHS, ES, ESGS, VS. */
478 static unsigned si_get_vs_vgpr_comp_cnt(struct si_screen
*sscreen
,
479 struct si_shader
*shader
, bool legacy_vs_prim_id
)
481 assert(shader
->selector
->type
== PIPE_SHADER_VERTEX
||
482 (shader
->previous_stage_sel
&&
483 shader
->previous_stage_sel
->type
== PIPE_SHADER_VERTEX
));
485 /* GFX6-9 LS (VertexID, RelAutoindex, InstanceID / StepRate0(==1), ...).
486 * GFX6-9 ES,VS (VertexID, InstanceID / StepRate0(==1), VSPrimID, ...)
487 * GFX10 LS (VertexID, RelAutoindex, UserVGPR1, InstanceID).
488 * GFX10 ES,VS (VertexID, UserVGPR0, UserVGPR1 or VSPrimID, UserVGPR2 or InstanceID)
490 bool is_ls
= shader
->selector
->type
== PIPE_SHADER_TESS_CTRL
|| shader
->key
.as_ls
;
492 if (sscreen
->info
.chip_class
>= GFX10
&& shader
->info
.uses_instanceid
)
494 else if ((is_ls
&& shader
->info
.uses_instanceid
) || legacy_vs_prim_id
)
496 else if (is_ls
|| shader
->info
.uses_instanceid
)
502 static void si_shader_ls(struct si_screen
*sscreen
, struct si_shader
*shader
)
504 struct si_pm4_state
*pm4
;
507 assert(sscreen
->info
.chip_class
<= GFX8
);
509 pm4
= si_get_shader_pm4_state(shader
);
513 va
= shader
->bo
->gpu_address
;
514 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
516 si_pm4_set_reg(pm4
, R_00B520_SPI_SHADER_PGM_LO_LS
, va
>> 8);
517 si_pm4_set_reg(pm4
, R_00B524_SPI_SHADER_PGM_HI_LS
, S_00B524_MEM_BASE(va
>> 40));
519 shader
->config
.rsrc1
= S_00B528_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
520 S_00B528_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
521 S_00B528_VGPR_COMP_CNT(si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false)) |
522 S_00B528_DX10_CLAMP(1) |
523 S_00B528_FLOAT_MODE(shader
->config
.float_mode
);
524 shader
->config
.rsrc2
= S_00B52C_USER_SGPR(si_get_num_vs_user_sgprs(shader
, SI_VS_NUM_USER_SGPR
)) |
525 S_00B52C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
528 static void si_shader_hs(struct si_screen
*sscreen
, struct si_shader
*shader
)
530 struct si_pm4_state
*pm4
;
533 pm4
= si_get_shader_pm4_state(shader
);
537 va
= shader
->bo
->gpu_address
;
538 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
540 if (sscreen
->info
.chip_class
>= GFX9
) {
541 if (sscreen
->info
.chip_class
>= GFX10
) {
542 si_pm4_set_reg(pm4
, R_00B520_SPI_SHADER_PGM_LO_LS
, va
>> 8);
543 si_pm4_set_reg(pm4
, R_00B524_SPI_SHADER_PGM_HI_LS
, S_00B524_MEM_BASE(va
>> 40));
545 si_pm4_set_reg(pm4
, R_00B410_SPI_SHADER_PGM_LO_LS
, va
>> 8);
546 si_pm4_set_reg(pm4
, R_00B414_SPI_SHADER_PGM_HI_LS
, S_00B414_MEM_BASE(va
>> 40));
549 unsigned num_user_sgprs
=
550 si_get_num_vs_user_sgprs(shader
, GFX9_TCS_NUM_USER_SGPR
);
552 shader
->config
.rsrc2
=
553 S_00B42C_USER_SGPR(num_user_sgprs
) |
554 S_00B42C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
556 if (sscreen
->info
.chip_class
>= GFX10
)
557 shader
->config
.rsrc2
|= S_00B42C_USER_SGPR_MSB_GFX10(num_user_sgprs
>> 5);
559 shader
->config
.rsrc2
|= S_00B42C_USER_SGPR_MSB_GFX9(num_user_sgprs
>> 5);
561 si_pm4_set_reg(pm4
, R_00B420_SPI_SHADER_PGM_LO_HS
, va
>> 8);
562 si_pm4_set_reg(pm4
, R_00B424_SPI_SHADER_PGM_HI_HS
, S_00B424_MEM_BASE(va
>> 40));
564 shader
->config
.rsrc2
=
565 S_00B42C_USER_SGPR(GFX6_TCS_NUM_USER_SGPR
) |
566 S_00B42C_OC_LDS_EN(1) |
567 S_00B42C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
570 si_pm4_set_reg(pm4
, R_00B428_SPI_SHADER_PGM_RSRC1_HS
,
571 S_00B428_VGPRS((shader
->config
.num_vgprs
- 1) /
572 (sscreen
->ge_wave_size
== 32 ? 8 : 4)) |
573 (sscreen
->info
.chip_class
<= GFX9
?
574 S_00B428_SGPRS((shader
->config
.num_sgprs
- 1) / 8) : 0) |
575 S_00B428_DX10_CLAMP(1) |
576 S_00B428_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
577 S_00B428_WGP_MODE(sscreen
->info
.chip_class
>= GFX10
) |
578 S_00B428_FLOAT_MODE(shader
->config
.float_mode
) |
579 S_00B428_LS_VGPR_COMP_CNT(sscreen
->info
.chip_class
>= GFX9
?
580 si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false) : 0));
582 if (sscreen
->info
.chip_class
<= GFX8
) {
583 si_pm4_set_reg(pm4
, R_00B42C_SPI_SHADER_PGM_RSRC2_HS
,
584 shader
->config
.rsrc2
);
588 static void si_emit_shader_es(struct si_context
*sctx
)
590 struct si_shader
*shader
= sctx
->queued
.named
.es
->shader
;
591 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
596 radeon_opt_set_context_reg(sctx
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
597 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE
,
598 shader
->selector
->esgs_itemsize
/ 4);
600 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
601 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
602 SI_TRACKED_VGT_TF_PARAM
,
603 shader
->vgt_tf_param
);
605 if (shader
->vgt_vertex_reuse_block_cntl
)
606 radeon_opt_set_context_reg(sctx
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
607 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL
,
608 shader
->vgt_vertex_reuse_block_cntl
);
610 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
611 sctx
->context_roll
= true;
614 static void si_shader_es(struct si_screen
*sscreen
, struct si_shader
*shader
)
616 struct si_pm4_state
*pm4
;
617 unsigned num_user_sgprs
;
618 unsigned vgpr_comp_cnt
;
622 assert(sscreen
->info
.chip_class
<= GFX8
);
624 pm4
= si_get_shader_pm4_state(shader
);
628 pm4
->atom
.emit
= si_emit_shader_es
;
629 va
= shader
->bo
->gpu_address
;
630 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
632 if (shader
->selector
->type
== PIPE_SHADER_VERTEX
) {
633 vgpr_comp_cnt
= si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false);
634 num_user_sgprs
= si_get_num_vs_user_sgprs(shader
, SI_VS_NUM_USER_SGPR
);
635 } else if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
636 vgpr_comp_cnt
= shader
->selector
->info
.uses_primid
? 3 : 2;
637 num_user_sgprs
= SI_TES_NUM_USER_SGPR
;
639 unreachable("invalid shader selector type");
641 oc_lds_en
= shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
? 1 : 0;
643 si_pm4_set_reg(pm4
, R_00B320_SPI_SHADER_PGM_LO_ES
, va
>> 8);
644 si_pm4_set_reg(pm4
, R_00B324_SPI_SHADER_PGM_HI_ES
, S_00B324_MEM_BASE(va
>> 40));
645 si_pm4_set_reg(pm4
, R_00B328_SPI_SHADER_PGM_RSRC1_ES
,
646 S_00B328_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
647 S_00B328_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
648 S_00B328_VGPR_COMP_CNT(vgpr_comp_cnt
) |
649 S_00B328_DX10_CLAMP(1) |
650 S_00B328_FLOAT_MODE(shader
->config
.float_mode
));
651 si_pm4_set_reg(pm4
, R_00B32C_SPI_SHADER_PGM_RSRC2_ES
,
652 S_00B32C_USER_SGPR(num_user_sgprs
) |
653 S_00B32C_OC_LDS_EN(oc_lds_en
) |
654 S_00B32C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
656 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
657 si_set_tesseval_regs(sscreen
, shader
->selector
, pm4
);
659 polaris_set_vgt_vertex_reuse(sscreen
, shader
->selector
, shader
, pm4
);
662 void gfx9_get_gs_info(struct si_shader_selector
*es
,
663 struct si_shader_selector
*gs
,
664 struct gfx9_gs_info
*out
)
666 unsigned gs_num_invocations
= MAX2(gs
->gs_num_invocations
, 1);
667 unsigned input_prim
= gs
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
];
668 bool uses_adjacency
= input_prim
>= PIPE_PRIM_LINES_ADJACENCY
&&
669 input_prim
<= PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
;
671 /* All these are in dwords: */
672 /* We can't allow using the whole LDS, because GS waves compete with
673 * other shader stages for LDS space. */
674 const unsigned max_lds_size
= 8 * 1024;
675 const unsigned esgs_itemsize
= es
->esgs_itemsize
/ 4;
676 unsigned esgs_lds_size
;
678 /* All these are per subgroup: */
679 const unsigned max_out_prims
= 32 * 1024;
680 const unsigned max_es_verts
= 255;
681 const unsigned ideal_gs_prims
= 64;
682 unsigned max_gs_prims
, gs_prims
;
683 unsigned min_es_verts
, es_verts
, worst_case_es_verts
;
685 if (uses_adjacency
|| gs_num_invocations
> 1)
686 max_gs_prims
= 127 / gs_num_invocations
;
690 /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
691 * Make sure we don't go over the maximum value.
693 if (gs
->gs_max_out_vertices
> 0) {
694 max_gs_prims
= MIN2(max_gs_prims
,
696 (gs
->gs_max_out_vertices
* gs_num_invocations
));
698 assert(max_gs_prims
> 0);
700 /* If the primitive has adjacency, halve the number of vertices
701 * that will be reused in multiple primitives.
703 min_es_verts
= gs
->gs_input_verts_per_prim
/ (uses_adjacency
? 2 : 1);
705 gs_prims
= MIN2(ideal_gs_prims
, max_gs_prims
);
706 worst_case_es_verts
= MIN2(min_es_verts
* gs_prims
, max_es_verts
);
708 /* Compute ESGS LDS size based on the worst case number of ES vertices
709 * needed to create the target number of GS prims per subgroup.
711 esgs_lds_size
= esgs_itemsize
* worst_case_es_verts
;
713 /* If total LDS usage is too big, refactor partitions based on ratio
714 * of ESGS item sizes.
716 if (esgs_lds_size
> max_lds_size
) {
717 /* Our target GS Prims Per Subgroup was too large. Calculate
718 * the maximum number of GS Prims Per Subgroup that will fit
719 * into LDS, capped by the maximum that the hardware can support.
721 gs_prims
= MIN2((max_lds_size
/ (esgs_itemsize
* min_es_verts
)),
723 assert(gs_prims
> 0);
724 worst_case_es_verts
= MIN2(min_es_verts
* gs_prims
,
727 esgs_lds_size
= esgs_itemsize
* worst_case_es_verts
;
728 assert(esgs_lds_size
<= max_lds_size
);
731 /* Now calculate remaining ESGS information. */
733 es_verts
= MIN2(esgs_lds_size
/ esgs_itemsize
, max_es_verts
);
735 es_verts
= max_es_verts
;
737 /* Vertices for adjacency primitives are not always reused, so restore
738 * it for ES_VERTS_PER_SUBGRP.
740 min_es_verts
= gs
->gs_input_verts_per_prim
;
742 /* For normal primitives, the VGT only checks if they are past the ES
743 * verts per subgroup after allocating a full GS primitive and if they
744 * are, kick off a new subgroup. But if those additional ES verts are
745 * unique (e.g. not reused) we need to make sure there is enough LDS
746 * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
748 es_verts
-= min_es_verts
- 1;
750 out
->es_verts_per_subgroup
= es_verts
;
751 out
->gs_prims_per_subgroup
= gs_prims
;
752 out
->gs_inst_prims_in_subgroup
= gs_prims
* gs_num_invocations
;
753 out
->max_prims_per_subgroup
= out
->gs_inst_prims_in_subgroup
*
754 gs
->gs_max_out_vertices
;
755 out
->esgs_ring_size
= 4 * esgs_lds_size
;
757 assert(out
->max_prims_per_subgroup
<= max_out_prims
);
760 static void si_emit_shader_gs(struct si_context
*sctx
)
762 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
763 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
768 /* R_028A60_VGT_GSVS_RING_OFFSET_1, R_028A64_VGT_GSVS_RING_OFFSET_2
769 * R_028A68_VGT_GSVS_RING_OFFSET_3 */
770 radeon_opt_set_context_reg3(sctx
, R_028A60_VGT_GSVS_RING_OFFSET_1
,
771 SI_TRACKED_VGT_GSVS_RING_OFFSET_1
,
772 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_1
,
773 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_2
,
774 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_3
);
776 /* R_028AB0_VGT_GSVS_RING_ITEMSIZE */
777 radeon_opt_set_context_reg(sctx
, R_028AB0_VGT_GSVS_RING_ITEMSIZE
,
778 SI_TRACKED_VGT_GSVS_RING_ITEMSIZE
,
779 shader
->ctx_reg
.gs
.vgt_gsvs_ring_itemsize
);
781 /* R_028B38_VGT_GS_MAX_VERT_OUT */
782 radeon_opt_set_context_reg(sctx
, R_028B38_VGT_GS_MAX_VERT_OUT
,
783 SI_TRACKED_VGT_GS_MAX_VERT_OUT
,
784 shader
->ctx_reg
.gs
.vgt_gs_max_vert_out
);
786 /* R_028B5C_VGT_GS_VERT_ITEMSIZE, R_028B60_VGT_GS_VERT_ITEMSIZE_1
787 * R_028B64_VGT_GS_VERT_ITEMSIZE_2, R_028B68_VGT_GS_VERT_ITEMSIZE_3 */
788 radeon_opt_set_context_reg4(sctx
, R_028B5C_VGT_GS_VERT_ITEMSIZE
,
789 SI_TRACKED_VGT_GS_VERT_ITEMSIZE
,
790 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize
,
791 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_1
,
792 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_2
,
793 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_3
);
795 /* R_028B90_VGT_GS_INSTANCE_CNT */
796 radeon_opt_set_context_reg(sctx
, R_028B90_VGT_GS_INSTANCE_CNT
,
797 SI_TRACKED_VGT_GS_INSTANCE_CNT
,
798 shader
->ctx_reg
.gs
.vgt_gs_instance_cnt
);
800 if (sctx
->chip_class
>= GFX9
) {
801 /* R_028A44_VGT_GS_ONCHIP_CNTL */
802 radeon_opt_set_context_reg(sctx
, R_028A44_VGT_GS_ONCHIP_CNTL
,
803 SI_TRACKED_VGT_GS_ONCHIP_CNTL
,
804 shader
->ctx_reg
.gs
.vgt_gs_onchip_cntl
);
805 /* R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP */
806 radeon_opt_set_context_reg(sctx
, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP
,
807 SI_TRACKED_VGT_GS_MAX_PRIMS_PER_SUBGROUP
,
808 shader
->ctx_reg
.gs
.vgt_gs_max_prims_per_subgroup
);
809 /* R_028AAC_VGT_ESGS_RING_ITEMSIZE */
810 radeon_opt_set_context_reg(sctx
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
811 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE
,
812 shader
->ctx_reg
.gs
.vgt_esgs_ring_itemsize
);
814 if (shader
->key
.part
.gs
.es
->type
== PIPE_SHADER_TESS_EVAL
)
815 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
816 SI_TRACKED_VGT_TF_PARAM
,
817 shader
->vgt_tf_param
);
818 if (shader
->vgt_vertex_reuse_block_cntl
)
819 radeon_opt_set_context_reg(sctx
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
820 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL
,
821 shader
->vgt_vertex_reuse_block_cntl
);
824 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
825 sctx
->context_roll
= true;
828 static void si_shader_gs(struct si_screen
*sscreen
, struct si_shader
*shader
)
830 struct si_shader_selector
*sel
= shader
->selector
;
831 const ubyte
*num_components
= sel
->info
.num_stream_output_components
;
832 unsigned gs_num_invocations
= sel
->gs_num_invocations
;
833 struct si_pm4_state
*pm4
;
835 unsigned max_stream
= sel
->max_gs_stream
;
838 pm4
= si_get_shader_pm4_state(shader
);
842 pm4
->atom
.emit
= si_emit_shader_gs
;
844 offset
= num_components
[0] * sel
->gs_max_out_vertices
;
845 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_1
= offset
;
848 offset
+= num_components
[1] * sel
->gs_max_out_vertices
;
849 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_2
= offset
;
852 offset
+= num_components
[2] * sel
->gs_max_out_vertices
;
853 shader
->ctx_reg
.gs
.vgt_gsvs_ring_offset_3
= offset
;
856 offset
+= num_components
[3] * sel
->gs_max_out_vertices
;
857 shader
->ctx_reg
.gs
.vgt_gsvs_ring_itemsize
= offset
;
859 /* The GSVS_RING_ITEMSIZE register takes 15 bits */
860 assert(offset
< (1 << 15));
862 shader
->ctx_reg
.gs
.vgt_gs_max_vert_out
= sel
->gs_max_out_vertices
;
864 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize
= num_components
[0];
865 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_1
= (max_stream
>= 1) ? num_components
[1] : 0;
866 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_2
= (max_stream
>= 2) ? num_components
[2] : 0;
867 shader
->ctx_reg
.gs
.vgt_gs_vert_itemsize_3
= (max_stream
>= 3) ? num_components
[3] : 0;
869 shader
->ctx_reg
.gs
.vgt_gs_instance_cnt
= S_028B90_CNT(MIN2(gs_num_invocations
, 127)) |
870 S_028B90_ENABLE(gs_num_invocations
> 0);
872 va
= shader
->bo
->gpu_address
;
873 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
875 if (sscreen
->info
.chip_class
>= GFX9
) {
876 unsigned input_prim
= sel
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
];
877 unsigned es_type
= shader
->key
.part
.gs
.es
->type
;
878 unsigned es_vgpr_comp_cnt
, gs_vgpr_comp_cnt
;
880 if (es_type
== PIPE_SHADER_VERTEX
) {
881 es_vgpr_comp_cnt
= si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false);
882 } else if (es_type
== PIPE_SHADER_TESS_EVAL
)
883 es_vgpr_comp_cnt
= shader
->key
.part
.gs
.es
->info
.uses_primid
? 3 : 2;
885 unreachable("invalid shader selector type");
887 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
888 * VGPR[0:4] are always loaded.
890 if (sel
->info
.uses_invocationid
)
891 gs_vgpr_comp_cnt
= 3; /* VGPR3 contains InvocationID. */
892 else if (sel
->info
.uses_primid
)
893 gs_vgpr_comp_cnt
= 2; /* VGPR2 contains PrimitiveID. */
894 else if (input_prim
>= PIPE_PRIM_TRIANGLES
)
895 gs_vgpr_comp_cnt
= 1; /* VGPR1 contains offsets 2, 3 */
897 gs_vgpr_comp_cnt
= 0; /* VGPR0 contains offsets 0, 1 */
899 unsigned num_user_sgprs
;
900 if (es_type
== PIPE_SHADER_VERTEX
)
901 num_user_sgprs
= si_get_num_vs_user_sgprs(shader
, GFX9_VSGS_NUM_USER_SGPR
);
903 num_user_sgprs
= GFX9_TESGS_NUM_USER_SGPR
;
905 if (sscreen
->info
.chip_class
>= GFX10
) {
906 si_pm4_set_reg(pm4
, R_00B320_SPI_SHADER_PGM_LO_ES
, va
>> 8);
907 si_pm4_set_reg(pm4
, R_00B324_SPI_SHADER_PGM_HI_ES
, S_00B324_MEM_BASE(va
>> 40));
909 si_pm4_set_reg(pm4
, R_00B210_SPI_SHADER_PGM_LO_ES
, va
>> 8);
910 si_pm4_set_reg(pm4
, R_00B214_SPI_SHADER_PGM_HI_ES
, S_00B214_MEM_BASE(va
>> 40));
914 S_00B228_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
915 S_00B228_DX10_CLAMP(1) |
916 S_00B228_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
917 S_00B228_WGP_MODE(sscreen
->info
.chip_class
>= GFX10
) |
918 S_00B228_FLOAT_MODE(shader
->config
.float_mode
) |
919 S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt
);
921 S_00B22C_USER_SGPR(num_user_sgprs
) |
922 S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt
) |
923 S_00B22C_OC_LDS_EN(es_type
== PIPE_SHADER_TESS_EVAL
) |
924 S_00B22C_LDS_SIZE(shader
->config
.lds_size
) |
925 S_00B22C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
927 if (sscreen
->info
.chip_class
>= GFX10
) {
928 rsrc2
|= S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs
>> 5);
930 rsrc1
|= S_00B228_SGPRS((shader
->config
.num_sgprs
- 1) / 8);
931 rsrc2
|= S_00B22C_USER_SGPR_MSB_GFX9(num_user_sgprs
>> 5);
934 si_pm4_set_reg(pm4
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
, rsrc1
);
935 si_pm4_set_reg(pm4
, R_00B22C_SPI_SHADER_PGM_RSRC2_GS
, rsrc2
);
937 shader
->ctx_reg
.gs
.vgt_gs_onchip_cntl
=
938 S_028A44_ES_VERTS_PER_SUBGRP(shader
->gs_info
.es_verts_per_subgroup
) |
939 S_028A44_GS_PRIMS_PER_SUBGRP(shader
->gs_info
.gs_prims_per_subgroup
) |
940 S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader
->gs_info
.gs_inst_prims_in_subgroup
);
941 shader
->ctx_reg
.gs
.vgt_gs_max_prims_per_subgroup
=
942 S_028A94_MAX_PRIMS_PER_SUBGROUP(shader
->gs_info
.max_prims_per_subgroup
);
943 shader
->ctx_reg
.gs
.vgt_esgs_ring_itemsize
=
944 shader
->key
.part
.gs
.es
->esgs_itemsize
/ 4;
946 if (es_type
== PIPE_SHADER_TESS_EVAL
)
947 si_set_tesseval_regs(sscreen
, shader
->key
.part
.gs
.es
, pm4
);
949 polaris_set_vgt_vertex_reuse(sscreen
, shader
->key
.part
.gs
.es
,
952 si_pm4_set_reg(pm4
, R_00B220_SPI_SHADER_PGM_LO_GS
, va
>> 8);
953 si_pm4_set_reg(pm4
, R_00B224_SPI_SHADER_PGM_HI_GS
, S_00B224_MEM_BASE(va
>> 40));
955 si_pm4_set_reg(pm4
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
,
956 S_00B228_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
957 S_00B228_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
958 S_00B228_DX10_CLAMP(1) |
959 S_00B228_FLOAT_MODE(shader
->config
.float_mode
));
960 si_pm4_set_reg(pm4
, R_00B22C_SPI_SHADER_PGM_RSRC2_GS
,
961 S_00B22C_USER_SGPR(GFX6_GS_NUM_USER_SGPR
) |
962 S_00B22C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
966 /* Common tail code for NGG primitive shaders. */
967 static void gfx10_emit_shader_ngg_tail(struct si_context
*sctx
,
968 struct si_shader
*shader
,
969 unsigned initial_cdw
)
971 radeon_opt_set_context_reg(sctx
, R_0287FC_GE_MAX_OUTPUT_PER_SUBGROUP
,
972 SI_TRACKED_GE_MAX_OUTPUT_PER_SUBGROUP
,
973 shader
->ctx_reg
.ngg
.ge_max_output_per_subgroup
);
974 radeon_opt_set_context_reg(sctx
, R_028B4C_GE_NGG_SUBGRP_CNTL
,
975 SI_TRACKED_GE_NGG_SUBGRP_CNTL
,
976 shader
->ctx_reg
.ngg
.ge_ngg_subgrp_cntl
);
977 radeon_opt_set_context_reg(sctx
, R_028A84_VGT_PRIMITIVEID_EN
,
978 SI_TRACKED_VGT_PRIMITIVEID_EN
,
979 shader
->ctx_reg
.ngg
.vgt_primitiveid_en
);
980 radeon_opt_set_context_reg(sctx
, R_028A44_VGT_GS_ONCHIP_CNTL
,
981 SI_TRACKED_VGT_GS_ONCHIP_CNTL
,
982 shader
->ctx_reg
.ngg
.vgt_gs_onchip_cntl
);
983 radeon_opt_set_context_reg(sctx
, R_028B90_VGT_GS_INSTANCE_CNT
,
984 SI_TRACKED_VGT_GS_INSTANCE_CNT
,
985 shader
->ctx_reg
.ngg
.vgt_gs_instance_cnt
);
986 radeon_opt_set_context_reg(sctx
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
987 SI_TRACKED_VGT_ESGS_RING_ITEMSIZE
,
988 shader
->ctx_reg
.ngg
.vgt_esgs_ring_itemsize
);
989 radeon_opt_set_context_reg(sctx
, R_0286C4_SPI_VS_OUT_CONFIG
,
990 SI_TRACKED_SPI_VS_OUT_CONFIG
,
991 shader
->ctx_reg
.ngg
.spi_vs_out_config
);
992 radeon_opt_set_context_reg2(sctx
, R_028708_SPI_SHADER_IDX_FORMAT
,
993 SI_TRACKED_SPI_SHADER_IDX_FORMAT
,
994 shader
->ctx_reg
.ngg
.spi_shader_idx_format
,
995 shader
->ctx_reg
.ngg
.spi_shader_pos_format
);
996 radeon_opt_set_context_reg(sctx
, R_028818_PA_CL_VTE_CNTL
,
997 SI_TRACKED_PA_CL_VTE_CNTL
,
998 shader
->ctx_reg
.ngg
.pa_cl_vte_cntl
);
999 radeon_opt_set_context_reg(sctx
, R_028838_PA_CL_NGG_CNTL
,
1000 SI_TRACKED_PA_CL_NGG_CNTL
,
1001 shader
->ctx_reg
.ngg
.pa_cl_ngg_cntl
);
1003 radeon_opt_set_context_reg_rmw(sctx
, R_02881C_PA_CL_VS_OUT_CNTL
,
1004 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS
,
1005 shader
->pa_cl_vs_out_cntl
,
1006 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS_MASK
);
1008 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
1009 sctx
->context_roll
= true;
1012 static void gfx10_emit_shader_ngg_notess_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 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1023 static void gfx10_emit_shader_ngg_tess_nogs(struct si_context
*sctx
)
1025 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1026 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1031 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
1032 SI_TRACKED_VGT_TF_PARAM
,
1033 shader
->vgt_tf_param
);
1035 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1038 static void gfx10_emit_shader_ngg_notess_gs(struct si_context
*sctx
)
1040 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1041 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1046 radeon_opt_set_context_reg(sctx
, R_028B38_VGT_GS_MAX_VERT_OUT
,
1047 SI_TRACKED_VGT_GS_MAX_VERT_OUT
,
1048 shader
->ctx_reg
.ngg
.vgt_gs_max_vert_out
);
1050 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1053 static void gfx10_emit_shader_ngg_tess_gs(struct si_context
*sctx
)
1055 struct si_shader
*shader
= sctx
->queued
.named
.gs
->shader
;
1056 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1061 radeon_opt_set_context_reg(sctx
, R_028B38_VGT_GS_MAX_VERT_OUT
,
1062 SI_TRACKED_VGT_GS_MAX_VERT_OUT
,
1063 shader
->ctx_reg
.ngg
.vgt_gs_max_vert_out
);
1064 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
1065 SI_TRACKED_VGT_TF_PARAM
,
1066 shader
->vgt_tf_param
);
1068 gfx10_emit_shader_ngg_tail(sctx
, shader
, initial_cdw
);
1071 unsigned si_get_input_prim(const struct si_shader_selector
*gs
)
1073 if (gs
->type
== PIPE_SHADER_GEOMETRY
)
1074 return gs
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
];
1076 if (gs
->type
== PIPE_SHADER_TESS_EVAL
) {
1077 if (gs
->info
.properties
[TGSI_PROPERTY_TES_POINT_MODE
])
1078 return PIPE_PRIM_POINTS
;
1079 if (gs
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] == PIPE_PRIM_LINES
)
1080 return PIPE_PRIM_LINES
;
1081 return PIPE_PRIM_TRIANGLES
;
1084 /* TODO: Set this correctly if the primitive type is set in the shader key. */
1085 return PIPE_PRIM_TRIANGLES
; /* worst case for all callers */
1088 static unsigned si_get_vs_out_cntl(const struct si_shader_selector
*sel
, bool ngg
)
1091 sel
->info
.writes_psize
|| (sel
->info
.writes_edgeflag
&& !ngg
) ||
1092 sel
->info
.writes_layer
|| sel
->info
.writes_viewport_index
;
1093 return S_02881C_USE_VTX_POINT_SIZE(sel
->info
.writes_psize
) |
1094 S_02881C_USE_VTX_EDGE_FLAG(sel
->info
.writes_edgeflag
&& !ngg
) |
1095 S_02881C_USE_VTX_RENDER_TARGET_INDX(sel
->info
.writes_layer
) |
1096 S_02881C_USE_VTX_VIEWPORT_INDX(sel
->info
.writes_viewport_index
) |
1097 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena
) |
1098 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena
);
1102 * Prepare the PM4 image for \p shader, which will run as a merged ESGS shader
1105 static void gfx10_shader_ngg(struct si_screen
*sscreen
, struct si_shader
*shader
)
1107 const struct si_shader_selector
*gs_sel
= shader
->selector
;
1108 const struct si_shader_info
*gs_info
= &gs_sel
->info
;
1109 enum pipe_shader_type gs_type
= shader
->selector
->type
;
1110 const struct si_shader_selector
*es_sel
=
1111 shader
->previous_stage_sel
? shader
->previous_stage_sel
: shader
->selector
;
1112 const struct si_shader_info
*es_info
= &es_sel
->info
;
1113 enum pipe_shader_type es_type
= es_sel
->type
;
1114 unsigned num_user_sgprs
;
1115 unsigned nparams
, es_vgpr_comp_cnt
, gs_vgpr_comp_cnt
;
1117 unsigned window_space
=
1118 gs_info
->properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
];
1119 bool es_enable_prim_id
= shader
->key
.mono
.u
.vs_export_prim_id
|| es_info
->uses_primid
;
1120 unsigned gs_num_invocations
= MAX2(gs_sel
->gs_num_invocations
, 1);
1121 unsigned input_prim
= si_get_input_prim(gs_sel
);
1122 bool break_wave_at_eoi
= false;
1123 struct si_pm4_state
*pm4
= si_get_shader_pm4_state(shader
);
1127 if (es_type
== PIPE_SHADER_TESS_EVAL
) {
1128 pm4
->atom
.emit
= gs_type
== PIPE_SHADER_GEOMETRY
? gfx10_emit_shader_ngg_tess_gs
1129 : gfx10_emit_shader_ngg_tess_nogs
;
1131 pm4
->atom
.emit
= gs_type
== PIPE_SHADER_GEOMETRY
? gfx10_emit_shader_ngg_notess_gs
1132 : gfx10_emit_shader_ngg_notess_nogs
;
1135 va
= shader
->bo
->gpu_address
;
1136 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
1138 if (es_type
== PIPE_SHADER_VERTEX
) {
1139 es_vgpr_comp_cnt
= si_get_vs_vgpr_comp_cnt(sscreen
, shader
, false);
1141 if (es_info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
]) {
1142 num_user_sgprs
= SI_SGPR_VS_BLIT_DATA
+
1143 es_info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
];
1145 num_user_sgprs
= si_get_num_vs_user_sgprs(shader
, GFX9_VSGS_NUM_USER_SGPR
);
1148 assert(es_type
== PIPE_SHADER_TESS_EVAL
);
1149 es_vgpr_comp_cnt
= es_enable_prim_id
? 3 : 2;
1150 num_user_sgprs
= GFX9_TESGS_NUM_USER_SGPR
;
1152 if (es_enable_prim_id
|| gs_info
->uses_primid
)
1153 break_wave_at_eoi
= true;
1156 /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
1157 * VGPR[0:4] are always loaded.
1159 * Vertex shaders always need to load VGPR3, because they need to
1160 * pass edge flags for decomposed primitives (such as quads) to the PA
1161 * for the GL_LINE polygon mode to skip rendering lines on inner edges.
1163 if (gs_info
->uses_invocationid
||
1164 (gs_type
== PIPE_SHADER_VERTEX
&& !gfx10_is_ngg_passthrough(shader
)))
1165 gs_vgpr_comp_cnt
= 3; /* VGPR3 contains InvocationID, edge flags. */
1166 else if ((gs_type
== PIPE_SHADER_GEOMETRY
&& gs_info
->uses_primid
) ||
1167 (gs_type
== PIPE_SHADER_VERTEX
&& shader
->key
.mono
.u
.vs_export_prim_id
))
1168 gs_vgpr_comp_cnt
= 2; /* VGPR2 contains PrimitiveID. */
1169 else if (input_prim
>= PIPE_PRIM_TRIANGLES
&& !gfx10_is_ngg_passthrough(shader
))
1170 gs_vgpr_comp_cnt
= 1; /* VGPR1 contains offsets 2, 3 */
1172 gs_vgpr_comp_cnt
= 0; /* VGPR0 contains offsets 0, 1 */
1174 si_pm4_set_reg(pm4
, R_00B320_SPI_SHADER_PGM_LO_ES
, va
>> 8);
1175 si_pm4_set_reg(pm4
, R_00B324_SPI_SHADER_PGM_HI_ES
, va
>> 40);
1176 si_pm4_set_reg(pm4
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
,
1177 S_00B228_VGPRS((shader
->config
.num_vgprs
- 1) /
1178 (sscreen
->ge_wave_size
== 32 ? 8 : 4)) |
1179 S_00B228_FLOAT_MODE(shader
->config
.float_mode
) |
1180 S_00B228_DX10_CLAMP(1) |
1181 S_00B228_MEM_ORDERED(1) |
1182 S_00B228_WGP_MODE(1) |
1183 S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt
));
1184 si_pm4_set_reg(pm4
, R_00B22C_SPI_SHADER_PGM_RSRC2_GS
,
1185 S_00B22C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0) |
1186 S_00B22C_USER_SGPR(num_user_sgprs
) |
1187 S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt
) |
1188 S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs
>> 5) |
1189 S_00B22C_OC_LDS_EN(es_type
== PIPE_SHADER_TESS_EVAL
) |
1190 S_00B22C_LDS_SIZE(shader
->config
.lds_size
));
1192 nparams
= MAX2(shader
->info
.nr_param_exports
, 1);
1193 shader
->ctx_reg
.ngg
.spi_vs_out_config
=
1194 S_0286C4_VS_EXPORT_COUNT(nparams
- 1) |
1195 S_0286C4_NO_PC_EXPORT(shader
->info
.nr_param_exports
== 0);
1197 shader
->ctx_reg
.ngg
.spi_shader_idx_format
=
1198 S_028708_IDX0_EXPORT_FORMAT(V_028708_SPI_SHADER_1COMP
);
1199 shader
->ctx_reg
.ngg
.spi_shader_pos_format
=
1200 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP
) |
1201 S_02870C_POS1_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 1 ?
1202 V_02870C_SPI_SHADER_4COMP
:
1203 V_02870C_SPI_SHADER_NONE
) |
1204 S_02870C_POS2_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 2 ?
1205 V_02870C_SPI_SHADER_4COMP
:
1206 V_02870C_SPI_SHADER_NONE
) |
1207 S_02870C_POS3_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 3 ?
1208 V_02870C_SPI_SHADER_4COMP
:
1209 V_02870C_SPI_SHADER_NONE
);
1211 shader
->ctx_reg
.ngg
.vgt_primitiveid_en
=
1212 S_028A84_PRIMITIVEID_EN(es_enable_prim_id
) |
1213 S_028A84_NGG_DISABLE_PROVOK_REUSE(shader
->key
.mono
.u
.vs_export_prim_id
||
1214 gs_sel
->info
.writes_primid
);
1216 if (gs_type
== PIPE_SHADER_GEOMETRY
) {
1217 shader
->ctx_reg
.ngg
.vgt_esgs_ring_itemsize
= es_sel
->esgs_itemsize
/ 4;
1218 shader
->ctx_reg
.ngg
.vgt_gs_max_vert_out
= gs_sel
->gs_max_out_vertices
;
1220 shader
->ctx_reg
.ngg
.vgt_esgs_ring_itemsize
= 1;
1223 if (es_type
== PIPE_SHADER_TESS_EVAL
)
1224 si_set_tesseval_regs(sscreen
, es_sel
, pm4
);
1226 shader
->ctx_reg
.ngg
.vgt_gs_onchip_cntl
=
1227 S_028A44_ES_VERTS_PER_SUBGRP(shader
->ngg
.hw_max_esverts
) |
1228 S_028A44_GS_PRIMS_PER_SUBGRP(shader
->ngg
.max_gsprims
) |
1229 S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader
->ngg
.max_gsprims
* gs_num_invocations
);
1230 shader
->ctx_reg
.ngg
.ge_max_output_per_subgroup
=
1231 S_0287FC_MAX_VERTS_PER_SUBGROUP(shader
->ngg
.max_out_verts
);
1232 shader
->ctx_reg
.ngg
.ge_ngg_subgrp_cntl
=
1233 S_028B4C_PRIM_AMP_FACTOR(shader
->ngg
.prim_amp_factor
) |
1234 S_028B4C_THDS_PER_SUBGRP(0); /* for fast launch */
1235 shader
->ctx_reg
.ngg
.vgt_gs_instance_cnt
=
1236 S_028B90_CNT(gs_num_invocations
) |
1237 S_028B90_ENABLE(gs_num_invocations
> 1) |
1238 S_028B90_EN_MAX_VERT_OUT_PER_GS_INSTANCE(
1239 shader
->ngg
.max_vert_out_per_gs_instance
);
1241 /* Always output hw-generated edge flags and pass them via the prim
1242 * export to prevent drawing lines on internal edges of decomposed
1243 * primitives (such as quads) with polygon mode = lines. Only VS needs
1246 shader
->ctx_reg
.ngg
.pa_cl_ngg_cntl
=
1247 S_028838_INDEX_BUF_EDGE_FLAG_ENA(gs_type
== PIPE_SHADER_VERTEX
);
1248 shader
->pa_cl_vs_out_cntl
= si_get_vs_out_cntl(gs_sel
, true);
1251 S_03096C_PRIM_GRP_SIZE(shader
->ngg
.max_gsprims
) |
1252 S_03096C_VERT_GRP_SIZE(256) | /* 256 = disable vertex grouping */
1253 S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi
);
1255 /* Bug workaround for a possible hang with non-tessellation cases.
1256 * Tessellation always sets GE_CNTL.VERT_GRP_SIZE = 0
1258 * Requirement: GE_CNTL.VERT_GRP_SIZE = VGT_GS_ONCHIP_CNTL.ES_VERTS_PER_SUBGRP - 5
1260 if ((sscreen
->info
.family
== CHIP_NAVI10
||
1261 sscreen
->info
.family
== CHIP_NAVI12
||
1262 sscreen
->info
.family
== CHIP_NAVI14
) &&
1263 (es_type
== PIPE_SHADER_VERTEX
|| gs_type
== PIPE_SHADER_VERTEX
) && /* = no tess */
1264 shader
->ngg
.hw_max_esverts
!= 256) {
1265 shader
->ge_cntl
&= C_03096C_VERT_GRP_SIZE
;
1267 if (shader
->ngg
.hw_max_esverts
> 5) {
1269 S_03096C_VERT_GRP_SIZE(shader
->ngg
.hw_max_esverts
- 5);
1274 shader
->ctx_reg
.ngg
.pa_cl_vte_cntl
=
1275 S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
1277 shader
->ctx_reg
.ngg
.pa_cl_vte_cntl
=
1278 S_028818_VTX_W0_FMT(1) |
1279 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
1280 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
1281 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
1285 static void si_emit_shader_vs(struct si_context
*sctx
)
1287 struct si_shader
*shader
= sctx
->queued
.named
.vs
->shader
;
1288 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1293 radeon_opt_set_context_reg(sctx
, R_028A40_VGT_GS_MODE
,
1294 SI_TRACKED_VGT_GS_MODE
,
1295 shader
->ctx_reg
.vs
.vgt_gs_mode
);
1296 radeon_opt_set_context_reg(sctx
, R_028A84_VGT_PRIMITIVEID_EN
,
1297 SI_TRACKED_VGT_PRIMITIVEID_EN
,
1298 shader
->ctx_reg
.vs
.vgt_primitiveid_en
);
1300 if (sctx
->chip_class
<= GFX8
) {
1301 radeon_opt_set_context_reg(sctx
, R_028AB4_VGT_REUSE_OFF
,
1302 SI_TRACKED_VGT_REUSE_OFF
,
1303 shader
->ctx_reg
.vs
.vgt_reuse_off
);
1306 radeon_opt_set_context_reg(sctx
, R_0286C4_SPI_VS_OUT_CONFIG
,
1307 SI_TRACKED_SPI_VS_OUT_CONFIG
,
1308 shader
->ctx_reg
.vs
.spi_vs_out_config
);
1310 radeon_opt_set_context_reg(sctx
, R_02870C_SPI_SHADER_POS_FORMAT
,
1311 SI_TRACKED_SPI_SHADER_POS_FORMAT
,
1312 shader
->ctx_reg
.vs
.spi_shader_pos_format
);
1314 radeon_opt_set_context_reg(sctx
, R_028818_PA_CL_VTE_CNTL
,
1315 SI_TRACKED_PA_CL_VTE_CNTL
,
1316 shader
->ctx_reg
.vs
.pa_cl_vte_cntl
);
1318 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
1319 radeon_opt_set_context_reg(sctx
, R_028B6C_VGT_TF_PARAM
,
1320 SI_TRACKED_VGT_TF_PARAM
,
1321 shader
->vgt_tf_param
);
1323 if (shader
->vgt_vertex_reuse_block_cntl
)
1324 radeon_opt_set_context_reg(sctx
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
1325 SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL
,
1326 shader
->vgt_vertex_reuse_block_cntl
);
1328 /* Required programming for tessellation. (legacy pipeline only) */
1329 if (sctx
->chip_class
== GFX10
&&
1330 shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
1331 radeon_opt_set_context_reg(sctx
, R_028A44_VGT_GS_ONCHIP_CNTL
,
1332 SI_TRACKED_VGT_GS_ONCHIP_CNTL
,
1333 S_028A44_ES_VERTS_PER_SUBGRP(250) |
1334 S_028A44_GS_PRIMS_PER_SUBGRP(126) |
1335 S_028A44_GS_INST_PRIMS_IN_SUBGRP(126));
1338 if (sctx
->chip_class
>= GFX10
) {
1339 radeon_opt_set_context_reg_rmw(sctx
, R_02881C_PA_CL_VS_OUT_CNTL
,
1340 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS
,
1341 shader
->pa_cl_vs_out_cntl
,
1342 SI_TRACKED_PA_CL_VS_OUT_CNTL__VS_MASK
);
1345 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
1346 sctx
->context_roll
= true;
1350 * Compute the state for \p shader, which will run as a vertex shader on the
1353 * If \p gs is non-NULL, it points to the geometry shader for which this shader
1354 * is the copy shader.
1356 static void si_shader_vs(struct si_screen
*sscreen
, struct si_shader
*shader
,
1357 struct si_shader_selector
*gs
)
1359 const struct si_shader_info
*info
= &shader
->selector
->info
;
1360 struct si_pm4_state
*pm4
;
1361 unsigned num_user_sgprs
, vgpr_comp_cnt
;
1363 unsigned nparams
, oc_lds_en
;
1364 unsigned window_space
=
1365 info
->properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
];
1366 bool enable_prim_id
= shader
->key
.mono
.u
.vs_export_prim_id
|| info
->uses_primid
;
1368 pm4
= si_get_shader_pm4_state(shader
);
1372 pm4
->atom
.emit
= si_emit_shader_vs
;
1374 /* We always write VGT_GS_MODE in the VS state, because every switch
1375 * between different shader pipelines involving a different GS or no
1376 * GS at all involves a switch of the VS (different GS use different
1377 * copy shaders). On the other hand, when the API switches from a GS to
1378 * no GS and then back to the same GS used originally, the GS state is
1382 unsigned mode
= V_028A40_GS_OFF
;
1384 /* PrimID needs GS scenario A. */
1386 mode
= V_028A40_GS_SCENARIO_A
;
1388 shader
->ctx_reg
.vs
.vgt_gs_mode
= S_028A40_MODE(mode
);
1389 shader
->ctx_reg
.vs
.vgt_primitiveid_en
= enable_prim_id
;
1391 shader
->ctx_reg
.vs
.vgt_gs_mode
= ac_vgt_gs_mode(gs
->gs_max_out_vertices
,
1392 sscreen
->info
.chip_class
);
1393 shader
->ctx_reg
.vs
.vgt_primitiveid_en
= 0;
1396 if (sscreen
->info
.chip_class
<= GFX8
) {
1397 /* Reuse needs to be set off if we write oViewport. */
1398 shader
->ctx_reg
.vs
.vgt_reuse_off
=
1399 S_028AB4_REUSE_OFF(info
->writes_viewport_index
);
1402 va
= shader
->bo
->gpu_address
;
1403 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
1406 vgpr_comp_cnt
= 0; /* only VertexID is needed for GS-COPY. */
1407 num_user_sgprs
= SI_GSCOPY_NUM_USER_SGPR
;
1408 } else if (shader
->selector
->type
== PIPE_SHADER_VERTEX
) {
1409 vgpr_comp_cnt
= si_get_vs_vgpr_comp_cnt(sscreen
, shader
, enable_prim_id
);
1411 if (info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
]) {
1412 num_user_sgprs
= SI_SGPR_VS_BLIT_DATA
+
1413 info
->properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
];
1415 num_user_sgprs
= si_get_num_vs_user_sgprs(shader
, SI_VS_NUM_USER_SGPR
);
1417 } else if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
1418 vgpr_comp_cnt
= enable_prim_id
? 3 : 2;
1419 num_user_sgprs
= SI_TES_NUM_USER_SGPR
;
1421 unreachable("invalid shader selector type");
1423 /* VS is required to export at least one param. */
1424 nparams
= MAX2(shader
->info
.nr_param_exports
, 1);
1425 shader
->ctx_reg
.vs
.spi_vs_out_config
= S_0286C4_VS_EXPORT_COUNT(nparams
- 1);
1427 if (sscreen
->info
.chip_class
>= GFX10
) {
1428 shader
->ctx_reg
.vs
.spi_vs_out_config
|=
1429 S_0286C4_NO_PC_EXPORT(shader
->info
.nr_param_exports
== 0);
1432 shader
->ctx_reg
.vs
.spi_shader_pos_format
=
1433 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP
) |
1434 S_02870C_POS1_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 1 ?
1435 V_02870C_SPI_SHADER_4COMP
:
1436 V_02870C_SPI_SHADER_NONE
) |
1437 S_02870C_POS2_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 2 ?
1438 V_02870C_SPI_SHADER_4COMP
:
1439 V_02870C_SPI_SHADER_NONE
) |
1440 S_02870C_POS3_EXPORT_FORMAT(shader
->info
.nr_pos_exports
> 3 ?
1441 V_02870C_SPI_SHADER_4COMP
:
1442 V_02870C_SPI_SHADER_NONE
);
1443 shader
->pa_cl_vs_out_cntl
= si_get_vs_out_cntl(shader
->selector
, false);
1445 oc_lds_en
= shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
? 1 : 0;
1447 si_pm4_set_reg(pm4
, R_00B120_SPI_SHADER_PGM_LO_VS
, va
>> 8);
1448 si_pm4_set_reg(pm4
, R_00B124_SPI_SHADER_PGM_HI_VS
, S_00B124_MEM_BASE(va
>> 40));
1450 uint32_t rsrc1
= S_00B128_VGPRS((shader
->config
.num_vgprs
- 1) /
1451 (sscreen
->ge_wave_size
== 32 ? 8 : 4)) |
1452 S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt
) |
1453 S_00B128_DX10_CLAMP(1) |
1454 S_00B128_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
1455 S_00B128_FLOAT_MODE(shader
->config
.float_mode
);
1456 uint32_t rsrc2
= S_00B12C_USER_SGPR(num_user_sgprs
) |
1457 S_00B12C_OC_LDS_EN(oc_lds_en
) |
1458 S_00B12C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0);
1460 if (sscreen
->info
.chip_class
>= GFX10
)
1461 rsrc2
|= S_00B12C_USER_SGPR_MSB_GFX10(num_user_sgprs
>> 5);
1462 else if (sscreen
->info
.chip_class
== GFX9
)
1463 rsrc2
|= S_00B12C_USER_SGPR_MSB_GFX9(num_user_sgprs
>> 5);
1465 if (sscreen
->info
.chip_class
<= GFX9
)
1466 rsrc1
|= S_00B128_SGPRS((shader
->config
.num_sgprs
- 1) / 8);
1468 if (!sscreen
->use_ngg_streamout
) {
1469 rsrc2
|= S_00B12C_SO_BASE0_EN(!!shader
->selector
->so
.stride
[0]) |
1470 S_00B12C_SO_BASE1_EN(!!shader
->selector
->so
.stride
[1]) |
1471 S_00B12C_SO_BASE2_EN(!!shader
->selector
->so
.stride
[2]) |
1472 S_00B12C_SO_BASE3_EN(!!shader
->selector
->so
.stride
[3]) |
1473 S_00B12C_SO_EN(!!shader
->selector
->so
.num_outputs
);
1476 si_pm4_set_reg(pm4
, R_00B128_SPI_SHADER_PGM_RSRC1_VS
, rsrc1
);
1477 si_pm4_set_reg(pm4
, R_00B12C_SPI_SHADER_PGM_RSRC2_VS
, rsrc2
);
1480 shader
->ctx_reg
.vs
.pa_cl_vte_cntl
=
1481 S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
1483 shader
->ctx_reg
.vs
.pa_cl_vte_cntl
=
1484 S_028818_VTX_W0_FMT(1) |
1485 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
1486 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
1487 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
1489 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
1490 si_set_tesseval_regs(sscreen
, shader
->selector
, pm4
);
1492 polaris_set_vgt_vertex_reuse(sscreen
, shader
->selector
, shader
, pm4
);
1495 static unsigned si_get_ps_num_interp(struct si_shader
*ps
)
1497 struct si_shader_info
*info
= &ps
->selector
->info
;
1498 unsigned num_colors
= !!(info
->colors_read
& 0x0f) +
1499 !!(info
->colors_read
& 0xf0);
1500 unsigned num_interp
= ps
->selector
->info
.num_inputs
+
1501 (ps
->key
.part
.ps
.prolog
.color_two_side
? num_colors
: 0);
1503 assert(num_interp
<= 32);
1504 return MIN2(num_interp
, 32);
1507 static unsigned si_get_spi_shader_col_format(struct si_shader
*shader
)
1509 unsigned value
= shader
->key
.part
.ps
.epilog
.spi_shader_col_format
;
1510 unsigned i
, num_targets
= (util_last_bit(value
) + 3) / 4;
1512 /* If the i-th target format is set, all previous target formats must
1513 * be non-zero to avoid hangs.
1515 for (i
= 0; i
< num_targets
; i
++)
1516 if (!(value
& (0xf << (i
* 4))))
1517 value
|= V_028714_SPI_SHADER_32_R
<< (i
* 4);
1522 static void si_emit_shader_ps(struct si_context
*sctx
)
1524 struct si_shader
*shader
= sctx
->queued
.named
.ps
->shader
;
1525 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
1530 /* R_0286CC_SPI_PS_INPUT_ENA, R_0286D0_SPI_PS_INPUT_ADDR*/
1531 radeon_opt_set_context_reg2(sctx
, R_0286CC_SPI_PS_INPUT_ENA
,
1532 SI_TRACKED_SPI_PS_INPUT_ENA
,
1533 shader
->ctx_reg
.ps
.spi_ps_input_ena
,
1534 shader
->ctx_reg
.ps
.spi_ps_input_addr
);
1536 radeon_opt_set_context_reg(sctx
, R_0286E0_SPI_BARYC_CNTL
,
1537 SI_TRACKED_SPI_BARYC_CNTL
,
1538 shader
->ctx_reg
.ps
.spi_baryc_cntl
);
1539 radeon_opt_set_context_reg(sctx
, R_0286D8_SPI_PS_IN_CONTROL
,
1540 SI_TRACKED_SPI_PS_IN_CONTROL
,
1541 shader
->ctx_reg
.ps
.spi_ps_in_control
);
1543 /* R_028710_SPI_SHADER_Z_FORMAT, R_028714_SPI_SHADER_COL_FORMAT */
1544 radeon_opt_set_context_reg2(sctx
, R_028710_SPI_SHADER_Z_FORMAT
,
1545 SI_TRACKED_SPI_SHADER_Z_FORMAT
,
1546 shader
->ctx_reg
.ps
.spi_shader_z_format
,
1547 shader
->ctx_reg
.ps
.spi_shader_col_format
);
1549 radeon_opt_set_context_reg(sctx
, R_02823C_CB_SHADER_MASK
,
1550 SI_TRACKED_CB_SHADER_MASK
,
1551 shader
->ctx_reg
.ps
.cb_shader_mask
);
1553 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
1554 sctx
->context_roll
= true;
1557 static void si_shader_ps(struct si_screen
*sscreen
, struct si_shader
*shader
)
1559 struct si_shader_info
*info
= &shader
->selector
->info
;
1560 struct si_pm4_state
*pm4
;
1561 unsigned spi_ps_in_control
, spi_shader_col_format
, cb_shader_mask
;
1562 unsigned spi_baryc_cntl
= S_0286E0_FRONT_FACE_ALL_BITS(1);
1564 unsigned input_ena
= shader
->config
.spi_ps_input_ena
;
1566 /* we need to enable at least one of them, otherwise we hang the GPU */
1567 assert(G_0286CC_PERSP_SAMPLE_ENA(input_ena
) ||
1568 G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
1569 G_0286CC_PERSP_CENTROID_ENA(input_ena
) ||
1570 G_0286CC_PERSP_PULL_MODEL_ENA(input_ena
) ||
1571 G_0286CC_LINEAR_SAMPLE_ENA(input_ena
) ||
1572 G_0286CC_LINEAR_CENTER_ENA(input_ena
) ||
1573 G_0286CC_LINEAR_CENTROID_ENA(input_ena
) ||
1574 G_0286CC_LINE_STIPPLE_TEX_ENA(input_ena
));
1575 /* POS_W_FLOAT_ENA requires one of the perspective weights. */
1576 assert(!G_0286CC_POS_W_FLOAT_ENA(input_ena
) ||
1577 G_0286CC_PERSP_SAMPLE_ENA(input_ena
) ||
1578 G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
1579 G_0286CC_PERSP_CENTROID_ENA(input_ena
) ||
1580 G_0286CC_PERSP_PULL_MODEL_ENA(input_ena
));
1582 /* Validate interpolation optimization flags (read as implications). */
1583 assert(!shader
->key
.part
.ps
.prolog
.bc_optimize_for_persp
||
1584 (G_0286CC_PERSP_CENTER_ENA(input_ena
) &&
1585 G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
1586 assert(!shader
->key
.part
.ps
.prolog
.bc_optimize_for_linear
||
1587 (G_0286CC_LINEAR_CENTER_ENA(input_ena
) &&
1588 G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
1589 assert(!shader
->key
.part
.ps
.prolog
.force_persp_center_interp
||
1590 (!G_0286CC_PERSP_SAMPLE_ENA(input_ena
) &&
1591 !G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
1592 assert(!shader
->key
.part
.ps
.prolog
.force_linear_center_interp
||
1593 (!G_0286CC_LINEAR_SAMPLE_ENA(input_ena
) &&
1594 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
1595 assert(!shader
->key
.part
.ps
.prolog
.force_persp_sample_interp
||
1596 (!G_0286CC_PERSP_CENTER_ENA(input_ena
) &&
1597 !G_0286CC_PERSP_CENTROID_ENA(input_ena
)));
1598 assert(!shader
->key
.part
.ps
.prolog
.force_linear_sample_interp
||
1599 (!G_0286CC_LINEAR_CENTER_ENA(input_ena
) &&
1600 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
)));
1602 /* Validate cases when the optimizations are off (read as implications). */
1603 assert(shader
->key
.part
.ps
.prolog
.bc_optimize_for_persp
||
1604 !G_0286CC_PERSP_CENTER_ENA(input_ena
) ||
1605 !G_0286CC_PERSP_CENTROID_ENA(input_ena
));
1606 assert(shader
->key
.part
.ps
.prolog
.bc_optimize_for_linear
||
1607 !G_0286CC_LINEAR_CENTER_ENA(input_ena
) ||
1608 !G_0286CC_LINEAR_CENTROID_ENA(input_ena
));
1610 pm4
= si_get_shader_pm4_state(shader
);
1614 pm4
->atom
.emit
= si_emit_shader_ps
;
1616 /* SPI_BARYC_CNTL.POS_FLOAT_LOCATION
1618 * 0 -> Position = pixel center
1619 * 1 -> Position = pixel centroid
1620 * 2 -> Position = at sample position
1622 * From GLSL 4.5 specification, section 7.1:
1623 * "The variable gl_FragCoord is available as an input variable from
1624 * within fragment shaders and it holds the window relative coordinates
1625 * (x, y, z, 1/w) values for the fragment. If multi-sampling, this
1626 * value can be for any location within the pixel, or one of the
1627 * fragment samples. The use of centroid does not further restrict
1628 * this value to be inside the current primitive."
1630 * Meaning that centroid has no effect and we can return anything within
1631 * the pixel. Thus, return the value at sample position, because that's
1632 * the most accurate one shaders can get.
1634 spi_baryc_cntl
|= S_0286E0_POS_FLOAT_LOCATION(2);
1636 if (info
->properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
] ==
1637 TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)
1638 spi_baryc_cntl
|= S_0286E0_POS_FLOAT_ULC(1);
1640 spi_shader_col_format
= si_get_spi_shader_col_format(shader
);
1641 cb_shader_mask
= ac_get_cb_shader_mask(spi_shader_col_format
);
1643 /* Ensure that some export memory is always allocated, for two reasons:
1645 * 1) Correctness: The hardware ignores the EXEC mask if no export
1646 * memory is allocated, so KILL and alpha test do not work correctly
1648 * 2) Performance: Every shader needs at least a NULL export, even when
1649 * it writes no color/depth output. The NULL export instruction
1650 * stalls without this setting.
1652 * Don't add this to CB_SHADER_MASK.
1654 * GFX10 supports pixel shaders without exports by setting both
1655 * the color and Z formats to SPI_SHADER_ZERO. The hw will skip export
1656 * instructions if any are present.
1658 if ((sscreen
->info
.chip_class
<= GFX9
||
1660 shader
->key
.part
.ps
.epilog
.alpha_func
!= PIPE_FUNC_ALWAYS
) &&
1661 !spi_shader_col_format
&&
1662 !info
->writes_z
&& !info
->writes_stencil
&& !info
->writes_samplemask
)
1663 spi_shader_col_format
= V_028714_SPI_SHADER_32_R
;
1665 shader
->ctx_reg
.ps
.spi_ps_input_ena
= input_ena
;
1666 shader
->ctx_reg
.ps
.spi_ps_input_addr
= shader
->config
.spi_ps_input_addr
;
1668 /* Set interpolation controls. */
1669 spi_ps_in_control
= S_0286D8_NUM_INTERP(si_get_ps_num_interp(shader
)) |
1670 S_0286D8_PS_W32_EN(sscreen
->ps_wave_size
== 32);
1672 shader
->ctx_reg
.ps
.spi_baryc_cntl
= spi_baryc_cntl
;
1673 shader
->ctx_reg
.ps
.spi_ps_in_control
= spi_ps_in_control
;
1674 shader
->ctx_reg
.ps
.spi_shader_z_format
=
1675 ac_get_spi_shader_z_format(info
->writes_z
,
1676 info
->writes_stencil
,
1677 info
->writes_samplemask
);
1678 shader
->ctx_reg
.ps
.spi_shader_col_format
= spi_shader_col_format
;
1679 shader
->ctx_reg
.ps
.cb_shader_mask
= cb_shader_mask
;
1681 va
= shader
->bo
->gpu_address
;
1682 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_BINARY
);
1683 si_pm4_set_reg(pm4
, R_00B020_SPI_SHADER_PGM_LO_PS
, va
>> 8);
1684 si_pm4_set_reg(pm4
, R_00B024_SPI_SHADER_PGM_HI_PS
, S_00B024_MEM_BASE(va
>> 40));
1687 S_00B028_VGPRS((shader
->config
.num_vgprs
- 1) /
1688 (sscreen
->ps_wave_size
== 32 ? 8 : 4)) |
1689 S_00B028_DX10_CLAMP(1) |
1690 S_00B028_MEM_ORDERED(sscreen
->info
.chip_class
>= GFX10
) |
1691 S_00B028_FLOAT_MODE(shader
->config
.float_mode
);
1693 if (sscreen
->info
.chip_class
< GFX10
) {
1694 rsrc1
|= S_00B028_SGPRS((shader
->config
.num_sgprs
- 1) / 8);
1697 si_pm4_set_reg(pm4
, R_00B028_SPI_SHADER_PGM_RSRC1_PS
, rsrc1
);
1698 si_pm4_set_reg(pm4
, R_00B02C_SPI_SHADER_PGM_RSRC2_PS
,
1699 S_00B02C_EXTRA_LDS_SIZE(shader
->config
.lds_size
) |
1700 S_00B02C_USER_SGPR(SI_PS_NUM_USER_SGPR
) |
1701 S_00B32C_SCRATCH_EN(shader
->config
.scratch_bytes_per_wave
> 0));
1704 static void si_shader_init_pm4_state(struct si_screen
*sscreen
,
1705 struct si_shader
*shader
)
1707 switch (shader
->selector
->type
) {
1708 case PIPE_SHADER_VERTEX
:
1709 if (shader
->key
.as_ls
)
1710 si_shader_ls(sscreen
, shader
);
1711 else if (shader
->key
.as_es
)
1712 si_shader_es(sscreen
, shader
);
1713 else if (shader
->key
.as_ngg
)
1714 gfx10_shader_ngg(sscreen
, shader
);
1716 si_shader_vs(sscreen
, shader
, NULL
);
1718 case PIPE_SHADER_TESS_CTRL
:
1719 si_shader_hs(sscreen
, shader
);
1721 case PIPE_SHADER_TESS_EVAL
:
1722 if (shader
->key
.as_es
)
1723 si_shader_es(sscreen
, shader
);
1724 else if (shader
->key
.as_ngg
)
1725 gfx10_shader_ngg(sscreen
, shader
);
1727 si_shader_vs(sscreen
, shader
, NULL
);
1729 case PIPE_SHADER_GEOMETRY
:
1730 if (shader
->key
.as_ngg
)
1731 gfx10_shader_ngg(sscreen
, shader
);
1733 si_shader_gs(sscreen
, shader
);
1735 case PIPE_SHADER_FRAGMENT
:
1736 si_shader_ps(sscreen
, shader
);
1743 static unsigned si_get_alpha_test_func(struct si_context
*sctx
)
1745 /* Alpha-test should be disabled if colorbuffer 0 is integer. */
1746 return sctx
->queued
.named
.dsa
->alpha_func
;
1749 void si_shader_selector_key_vs(struct si_context
*sctx
,
1750 struct si_shader_selector
*vs
,
1751 struct si_shader_key
*key
,
1752 struct si_vs_prolog_bits
*prolog_key
)
1754 if (!sctx
->vertex_elements
||
1755 vs
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
])
1758 struct si_vertex_elements
*elts
= sctx
->vertex_elements
;
1760 prolog_key
->instance_divisor_is_one
= elts
->instance_divisor_is_one
;
1761 prolog_key
->instance_divisor_is_fetched
= elts
->instance_divisor_is_fetched
;
1762 prolog_key
->unpack_instance_id_from_vertex_id
=
1763 sctx
->prim_discard_cs_instancing
;
1765 /* Prefer a monolithic shader to allow scheduling divisions around
1767 if (prolog_key
->instance_divisor_is_fetched
)
1768 key
->opt
.prefer_mono
= 1;
1770 unsigned count
= MIN2(vs
->info
.num_inputs
, elts
->count
);
1771 unsigned count_mask
= (1 << count
) - 1;
1772 unsigned fix
= elts
->fix_fetch_always
& count_mask
;
1773 unsigned opencode
= elts
->fix_fetch_opencode
& count_mask
;
1775 if (sctx
->vertex_buffer_unaligned
& elts
->vb_alignment_check_mask
) {
1776 uint32_t mask
= elts
->fix_fetch_unaligned
& count_mask
;
1778 unsigned i
= u_bit_scan(&mask
);
1779 unsigned log_hw_load_size
= 1 + ((elts
->hw_load_is_dword
>> i
) & 1);
1780 unsigned vbidx
= elts
->vertex_buffer_index
[i
];
1781 struct pipe_vertex_buffer
*vb
= &sctx
->vertex_buffer
[vbidx
];
1782 unsigned align_mask
= (1 << log_hw_load_size
) - 1;
1783 if (vb
->buffer_offset
& align_mask
||
1784 vb
->stride
& align_mask
) {
1792 unsigned i
= u_bit_scan(&fix
);
1793 key
->mono
.vs_fix_fetch
[i
].bits
= elts
->fix_fetch
[i
];
1795 key
->mono
.vs_fetch_opencode
= opencode
;
1798 static void si_shader_selector_key_hw_vs(struct si_context
*sctx
,
1799 struct si_shader_selector
*vs
,
1800 struct si_shader_key
*key
)
1802 struct si_shader_selector
*ps
= sctx
->ps_shader
.cso
;
1804 key
->opt
.clip_disable
=
1805 sctx
->queued
.named
.rasterizer
->clip_plane_enable
== 0 &&
1806 (vs
->info
.clipdist_writemask
||
1807 vs
->info
.writes_clipvertex
) &&
1808 !vs
->info
.culldist_writemask
;
1810 /* Find out if PS is disabled. */
1811 bool ps_disabled
= true;
1813 bool ps_modifies_zs
= ps
->info
.uses_kill
||
1814 ps
->info
.writes_z
||
1815 ps
->info
.writes_stencil
||
1816 ps
->info
.writes_samplemask
||
1817 sctx
->queued
.named
.blend
->alpha_to_coverage
||
1818 si_get_alpha_test_func(sctx
) != PIPE_FUNC_ALWAYS
;
1819 unsigned ps_colormask
= si_get_total_colormask(sctx
);
1821 ps_disabled
= sctx
->queued
.named
.rasterizer
->rasterizer_discard
||
1824 !ps
->info
.writes_memory
);
1827 /* Find out which VS outputs aren't used by the PS. */
1828 uint64_t outputs_written
= vs
->outputs_written_before_ps
;
1829 uint64_t inputs_read
= 0;
1831 /* Ignore outputs that are not passed from VS to PS. */
1832 outputs_written
&= ~((1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_POSITION
, 0, true)) |
1833 (1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_PSIZE
, 0, true)) |
1834 (1ull << si_shader_io_get_unique_index(TGSI_SEMANTIC_CLIPVERTEX
, 0, true)));
1837 inputs_read
= ps
->inputs_read
;
1840 uint64_t linked
= outputs_written
& inputs_read
;
1842 key
->opt
.kill_outputs
= ~linked
& outputs_written
;
1845 /* Compute the key for the hw shader variant */
1846 static inline void si_shader_selector_key(struct pipe_context
*ctx
,
1847 struct si_shader_selector
*sel
,
1848 union si_vgt_stages_key stages_key
,
1849 struct si_shader_key
*key
)
1851 struct si_context
*sctx
= (struct si_context
*)ctx
;
1853 memset(key
, 0, sizeof(*key
));
1855 switch (sel
->type
) {
1856 case PIPE_SHADER_VERTEX
:
1857 si_shader_selector_key_vs(sctx
, sel
, key
, &key
->part
.vs
.prolog
);
1859 if (sctx
->tes_shader
.cso
)
1861 else if (sctx
->gs_shader
.cso
) {
1863 key
->as_ngg
= stages_key
.u
.ngg
;
1865 key
->as_ngg
= stages_key
.u
.ngg
;
1866 si_shader_selector_key_hw_vs(sctx
, sel
, key
);
1868 if (sctx
->ps_shader
.cso
&& sctx
->ps_shader
.cso
->info
.uses_primid
)
1869 key
->mono
.u
.vs_export_prim_id
= 1;
1872 case PIPE_SHADER_TESS_CTRL
:
1873 if (sctx
->chip_class
>= GFX9
) {
1874 si_shader_selector_key_vs(sctx
, sctx
->vs_shader
.cso
,
1875 key
, &key
->part
.tcs
.ls_prolog
);
1876 key
->part
.tcs
.ls
= sctx
->vs_shader
.cso
;
1878 /* When the LS VGPR fix is needed, monolithic shaders
1880 * - avoid initializing EXEC in both the LS prolog
1881 * and the LS main part when !vs_needs_prolog
1882 * - remove the fixup for unused input VGPRs
1884 key
->part
.tcs
.ls_prolog
.ls_vgpr_fix
= sctx
->ls_vgpr_fix
;
1886 /* The LS output / HS input layout can be communicated
1887 * directly instead of via user SGPRs for merged LS-HS.
1888 * The LS VGPR fix prefers this too.
1890 key
->opt
.prefer_mono
= 1;
1893 key
->part
.tcs
.epilog
.prim_mode
=
1894 sctx
->tes_shader
.cso
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
1895 key
->part
.tcs
.epilog
.invoc0_tess_factors_are_def
=
1896 sel
->info
.tessfactors_are_def_in_all_invocs
;
1897 key
->part
.tcs
.epilog
.tes_reads_tess_factors
=
1898 sctx
->tes_shader
.cso
->info
.reads_tess_factors
;
1900 if (sel
== sctx
->fixed_func_tcs_shader
.cso
)
1901 key
->mono
.u
.ff_tcs_inputs_to_copy
= sctx
->vs_shader
.cso
->outputs_written
;
1903 case PIPE_SHADER_TESS_EVAL
:
1904 key
->as_ngg
= stages_key
.u
.ngg
;
1906 if (sctx
->gs_shader
.cso
)
1909 si_shader_selector_key_hw_vs(sctx
, sel
, key
);
1911 if (sctx
->ps_shader
.cso
&& sctx
->ps_shader
.cso
->info
.uses_primid
)
1912 key
->mono
.u
.vs_export_prim_id
= 1;
1915 case PIPE_SHADER_GEOMETRY
:
1916 if (sctx
->chip_class
>= GFX9
) {
1917 if (sctx
->tes_shader
.cso
) {
1918 key
->part
.gs
.es
= sctx
->tes_shader
.cso
;
1920 si_shader_selector_key_vs(sctx
, sctx
->vs_shader
.cso
,
1921 key
, &key
->part
.gs
.vs_prolog
);
1922 key
->part
.gs
.es
= sctx
->vs_shader
.cso
;
1923 key
->part
.gs
.prolog
.gfx9_prev_is_vs
= 1;
1926 key
->as_ngg
= stages_key
.u
.ngg
;
1928 /* Merged ES-GS can have unbalanced wave usage.
1930 * ES threads are per-vertex, while GS threads are
1931 * per-primitive. So without any amplification, there
1932 * are fewer GS threads than ES threads, which can result
1933 * in empty (no-op) GS waves. With too much amplification,
1934 * there are more GS threads than ES threads, which
1935 * can result in empty (no-op) ES waves.
1937 * Non-monolithic shaders are implemented by setting EXEC
1938 * at the beginning of shader parts, and don't jump to
1939 * the end if EXEC is 0.
1941 * Monolithic shaders use conditional blocks, so they can
1942 * jump and skip empty waves of ES or GS. So set this to
1943 * always use optimized variants, which are monolithic.
1945 key
->opt
.prefer_mono
= 1;
1947 key
->part
.gs
.prolog
.tri_strip_adj_fix
= sctx
->gs_tri_strip_adj_fix
;
1949 case PIPE_SHADER_FRAGMENT
: {
1950 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
1951 struct si_state_blend
*blend
= sctx
->queued
.named
.blend
;
1953 if (sel
->info
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
] &&
1954 sel
->info
.colors_written
== 0x1)
1955 key
->part
.ps
.epilog
.last_cbuf
= MAX2(sctx
->framebuffer
.state
.nr_cbufs
, 1) - 1;
1957 /* Select the shader color format based on whether
1958 * blending or alpha are needed.
1960 key
->part
.ps
.epilog
.spi_shader_col_format
=
1961 (blend
->blend_enable_4bit
& blend
->need_src_alpha_4bit
&
1962 sctx
->framebuffer
.spi_shader_col_format_blend_alpha
) |
1963 (blend
->blend_enable_4bit
& ~blend
->need_src_alpha_4bit
&
1964 sctx
->framebuffer
.spi_shader_col_format_blend
) |
1965 (~blend
->blend_enable_4bit
& blend
->need_src_alpha_4bit
&
1966 sctx
->framebuffer
.spi_shader_col_format_alpha
) |
1967 (~blend
->blend_enable_4bit
& ~blend
->need_src_alpha_4bit
&
1968 sctx
->framebuffer
.spi_shader_col_format
);
1969 key
->part
.ps
.epilog
.spi_shader_col_format
&= blend
->cb_target_enabled_4bit
;
1971 /* The output for dual source blending should have
1972 * the same format as the first output.
1974 if (blend
->dual_src_blend
) {
1975 key
->part
.ps
.epilog
.spi_shader_col_format
|=
1976 (key
->part
.ps
.epilog
.spi_shader_col_format
& 0xf) << 4;
1979 /* If alpha-to-coverage is enabled, we have to export alpha
1980 * even if there is no color buffer.
1982 if (!(key
->part
.ps
.epilog
.spi_shader_col_format
& 0xf) &&
1983 blend
->alpha_to_coverage
)
1984 key
->part
.ps
.epilog
.spi_shader_col_format
|= V_028710_SPI_SHADER_32_AR
;
1986 /* On GFX6 and GFX7 except Hawaii, the CB doesn't clamp outputs
1987 * to the range supported by the type if a channel has less
1988 * than 16 bits and the export format is 16_ABGR.
1990 if (sctx
->chip_class
<= GFX7
&& sctx
->family
!= CHIP_HAWAII
) {
1991 key
->part
.ps
.epilog
.color_is_int8
= sctx
->framebuffer
.color_is_int8
;
1992 key
->part
.ps
.epilog
.color_is_int10
= sctx
->framebuffer
.color_is_int10
;
1995 /* Disable unwritten outputs (if WRITE_ALL_CBUFS isn't enabled). */
1996 if (!key
->part
.ps
.epilog
.last_cbuf
) {
1997 key
->part
.ps
.epilog
.spi_shader_col_format
&= sel
->colors_written_4bit
;
1998 key
->part
.ps
.epilog
.color_is_int8
&= sel
->info
.colors_written
;
1999 key
->part
.ps
.epilog
.color_is_int10
&= sel
->info
.colors_written
;
2002 bool is_poly
= !util_prim_is_points_or_lines(sctx
->current_rast_prim
);
2003 bool is_line
= util_prim_is_lines(sctx
->current_rast_prim
);
2005 key
->part
.ps
.prolog
.color_two_side
= rs
->two_side
&& sel
->info
.colors_read
;
2006 key
->part
.ps
.prolog
.flatshade_colors
= rs
->flatshade
&& sel
->info
.colors_read
;
2008 key
->part
.ps
.epilog
.alpha_to_one
= blend
->alpha_to_one
&&
2009 rs
->multisample_enable
;
2011 key
->part
.ps
.prolog
.poly_stipple
= rs
->poly_stipple_enable
&& is_poly
;
2012 key
->part
.ps
.epilog
.poly_line_smoothing
= ((is_poly
&& rs
->poly_smooth
) ||
2013 (is_line
&& rs
->line_smooth
)) &&
2014 sctx
->framebuffer
.nr_samples
<= 1;
2015 key
->part
.ps
.epilog
.clamp_color
= rs
->clamp_fragment_color
;
2017 if (sctx
->ps_iter_samples
> 1 &&
2018 sel
->info
.reads_samplemask
) {
2019 key
->part
.ps
.prolog
.samplemask_log_ps_iter
=
2020 util_logbase2(sctx
->ps_iter_samples
);
2023 if (rs
->force_persample_interp
&&
2024 rs
->multisample_enable
&&
2025 sctx
->framebuffer
.nr_samples
> 1 &&
2026 sctx
->ps_iter_samples
> 1) {
2027 key
->part
.ps
.prolog
.force_persp_sample_interp
=
2028 sel
->info
.uses_persp_center
||
2029 sel
->info
.uses_persp_centroid
;
2031 key
->part
.ps
.prolog
.force_linear_sample_interp
=
2032 sel
->info
.uses_linear_center
||
2033 sel
->info
.uses_linear_centroid
;
2034 } else if (rs
->multisample_enable
&&
2035 sctx
->framebuffer
.nr_samples
> 1) {
2036 key
->part
.ps
.prolog
.bc_optimize_for_persp
=
2037 sel
->info
.uses_persp_center
&&
2038 sel
->info
.uses_persp_centroid
;
2039 key
->part
.ps
.prolog
.bc_optimize_for_linear
=
2040 sel
->info
.uses_linear_center
&&
2041 sel
->info
.uses_linear_centroid
;
2043 /* Make sure SPI doesn't compute more than 1 pair
2044 * of (i,j), which is the optimization here. */
2045 key
->part
.ps
.prolog
.force_persp_center_interp
=
2046 sel
->info
.uses_persp_center
+
2047 sel
->info
.uses_persp_centroid
+
2048 sel
->info
.uses_persp_sample
> 1;
2050 key
->part
.ps
.prolog
.force_linear_center_interp
=
2051 sel
->info
.uses_linear_center
+
2052 sel
->info
.uses_linear_centroid
+
2053 sel
->info
.uses_linear_sample
> 1;
2055 if (sel
->info
.uses_persp_opcode_interp_sample
||
2056 sel
->info
.uses_linear_opcode_interp_sample
)
2057 key
->mono
.u
.ps
.interpolate_at_sample_force_center
= 1;
2060 key
->part
.ps
.epilog
.alpha_func
= si_get_alpha_test_func(sctx
);
2062 /* ps_uses_fbfetch is true only if the color buffer is bound. */
2063 if (sctx
->ps_uses_fbfetch
&& !sctx
->blitter
->running
) {
2064 struct pipe_surface
*cb0
= sctx
->framebuffer
.state
.cbufs
[0];
2065 struct pipe_resource
*tex
= cb0
->texture
;
2067 /* 1D textures are allocated and used as 2D on GFX9. */
2068 key
->mono
.u
.ps
.fbfetch_msaa
= sctx
->framebuffer
.nr_samples
> 1;
2069 key
->mono
.u
.ps
.fbfetch_is_1D
= sctx
->chip_class
!= GFX9
&&
2070 (tex
->target
== PIPE_TEXTURE_1D
||
2071 tex
->target
== PIPE_TEXTURE_1D_ARRAY
);
2072 key
->mono
.u
.ps
.fbfetch_layered
= tex
->target
== PIPE_TEXTURE_1D_ARRAY
||
2073 tex
->target
== PIPE_TEXTURE_2D_ARRAY
||
2074 tex
->target
== PIPE_TEXTURE_CUBE
||
2075 tex
->target
== PIPE_TEXTURE_CUBE_ARRAY
||
2076 tex
->target
== PIPE_TEXTURE_3D
;
2084 if (unlikely(sctx
->screen
->debug_flags
& DBG(NO_OPT_VARIANT
)))
2085 memset(&key
->opt
, 0, sizeof(key
->opt
));
2088 static void si_build_shader_variant(struct si_shader
*shader
,
2092 struct si_shader_selector
*sel
= shader
->selector
;
2093 struct si_screen
*sscreen
= sel
->screen
;
2094 struct ac_llvm_compiler
*compiler
;
2095 struct pipe_debug_callback
*debug
= &shader
->compiler_ctx_state
.debug
;
2097 if (thread_index
>= 0) {
2099 assert(thread_index
< ARRAY_SIZE(sscreen
->compiler_lowp
));
2100 compiler
= &sscreen
->compiler_lowp
[thread_index
];
2102 assert(thread_index
< ARRAY_SIZE(sscreen
->compiler
));
2103 compiler
= &sscreen
->compiler
[thread_index
];
2108 assert(!low_priority
);
2109 compiler
= shader
->compiler_ctx_state
.compiler
;
2112 if (!compiler
->passes
)
2113 si_init_compiler(sscreen
, compiler
);
2115 if (unlikely(!si_create_shader_variant(sscreen
, compiler
, shader
, debug
))) {
2116 PRINT_ERR("Failed to build shader variant (type=%u)\n",
2118 shader
->compilation_failed
= true;
2122 if (shader
->compiler_ctx_state
.is_debug_context
) {
2123 FILE *f
= open_memstream(&shader
->shader_log
,
2124 &shader
->shader_log_size
);
2126 si_shader_dump(sscreen
, shader
, NULL
, f
, false);
2131 si_shader_init_pm4_state(sscreen
, shader
);
2134 static void si_build_shader_variant_low_priority(void *job
, int thread_index
)
2136 struct si_shader
*shader
= (struct si_shader
*)job
;
2138 assert(thread_index
>= 0);
2140 si_build_shader_variant(shader
, thread_index
, true);
2143 static const struct si_shader_key zeroed
;
2145 static bool si_check_missing_main_part(struct si_screen
*sscreen
,
2146 struct si_shader_selector
*sel
,
2147 struct si_compiler_ctx_state
*compiler_state
,
2148 struct si_shader_key
*key
)
2150 struct si_shader
**mainp
= si_get_main_shader_part(sel
, key
);
2153 struct si_shader
*main_part
= CALLOC_STRUCT(si_shader
);
2158 /* We can leave the fence as permanently signaled because the
2159 * main part becomes visible globally only after it has been
2161 util_queue_fence_init(&main_part
->ready
);
2163 main_part
->selector
= sel
;
2164 main_part
->key
.as_es
= key
->as_es
;
2165 main_part
->key
.as_ls
= key
->as_ls
;
2166 main_part
->key
.as_ngg
= key
->as_ngg
;
2167 main_part
->is_monolithic
= false;
2169 if (si_compile_shader(sscreen
, compiler_state
->compiler
,
2170 main_part
, &compiler_state
->debug
) != 0) {
2180 * Select a shader variant according to the shader key.
2182 * \param optimized_or_none If the key describes an optimized shader variant and
2183 * the compilation isn't finished, don't select any
2184 * shader and return an error.
2186 int si_shader_select_with_key(struct si_screen
*sscreen
,
2187 struct si_shader_ctx_state
*state
,
2188 struct si_compiler_ctx_state
*compiler_state
,
2189 struct si_shader_key
*key
,
2191 bool optimized_or_none
)
2193 struct si_shader_selector
*sel
= state
->cso
;
2194 struct si_shader_selector
*previous_stage_sel
= NULL
;
2195 struct si_shader
*current
= state
->current
;
2196 struct si_shader
*iter
, *shader
= NULL
;
2199 /* Check if we don't need to change anything.
2200 * This path is also used for most shaders that don't need multiple
2201 * variants, it will cost just a computation of the key and this
2203 if (likely(current
&&
2204 memcmp(¤t
->key
, key
, sizeof(*key
)) == 0)) {
2205 if (unlikely(!util_queue_fence_is_signalled(¤t
->ready
))) {
2206 if (current
->is_optimized
) {
2207 if (optimized_or_none
)
2210 memset(&key
->opt
, 0, sizeof(key
->opt
));
2211 goto current_not_ready
;
2214 util_queue_fence_wait(¤t
->ready
);
2217 return current
->compilation_failed
? -1 : 0;
2221 /* This must be done before the mutex is locked, because async GS
2222 * compilation calls this function too, and therefore must enter
2225 * Only wait if we are in a draw call. Don't wait if we are
2226 * in a compiler thread.
2228 if (thread_index
< 0)
2229 util_queue_fence_wait(&sel
->ready
);
2231 simple_mtx_lock(&sel
->mutex
);
2233 /* Find the shader variant. */
2234 for (iter
= sel
->first_variant
; iter
; iter
= iter
->next_variant
) {
2235 /* Don't check the "current" shader. We checked it above. */
2236 if (current
!= iter
&&
2237 memcmp(&iter
->key
, key
, sizeof(*key
)) == 0) {
2238 simple_mtx_unlock(&sel
->mutex
);
2240 if (unlikely(!util_queue_fence_is_signalled(&iter
->ready
))) {
2241 /* If it's an optimized shader and its compilation has
2242 * been started but isn't done, use the unoptimized
2243 * shader so as not to cause a stall due to compilation.
2245 if (iter
->is_optimized
) {
2246 if (optimized_or_none
)
2248 memset(&key
->opt
, 0, sizeof(key
->opt
));
2252 util_queue_fence_wait(&iter
->ready
);
2255 if (iter
->compilation_failed
) {
2256 return -1; /* skip the draw call */
2259 state
->current
= iter
;
2264 /* Build a new shader. */
2265 shader
= CALLOC_STRUCT(si_shader
);
2267 simple_mtx_unlock(&sel
->mutex
);
2271 util_queue_fence_init(&shader
->ready
);
2273 shader
->selector
= sel
;
2275 shader
->compiler_ctx_state
= *compiler_state
;
2277 /* If this is a merged shader, get the first shader's selector. */
2278 if (sscreen
->info
.chip_class
>= GFX9
) {
2279 if (sel
->type
== PIPE_SHADER_TESS_CTRL
)
2280 previous_stage_sel
= key
->part
.tcs
.ls
;
2281 else if (sel
->type
== PIPE_SHADER_GEOMETRY
)
2282 previous_stage_sel
= key
->part
.gs
.es
;
2284 /* We need to wait for the previous shader. */
2285 if (previous_stage_sel
&& thread_index
< 0)
2286 util_queue_fence_wait(&previous_stage_sel
->ready
);
2289 bool is_pure_monolithic
=
2290 sscreen
->use_monolithic_shaders
||
2291 memcmp(&key
->mono
, &zeroed
.mono
, sizeof(key
->mono
)) != 0;
2293 /* Compile the main shader part if it doesn't exist. This can happen
2294 * if the initial guess was wrong.
2296 * The prim discard CS doesn't need the main shader part.
2298 if (!is_pure_monolithic
&&
2299 !key
->opt
.vs_as_prim_discard_cs
) {
2302 /* Make sure the main shader part is present. This is needed
2303 * for shaders that can be compiled as VS, LS, or ES, and only
2304 * one of them is compiled at creation.
2306 * It is also needed for GS, which can be compiled as non-NGG
2309 * For merged shaders, check that the starting shader's main
2312 if (previous_stage_sel
) {
2313 struct si_shader_key shader1_key
= zeroed
;
2315 if (sel
->type
== PIPE_SHADER_TESS_CTRL
) {
2316 shader1_key
.as_ls
= 1;
2317 } else if (sel
->type
== PIPE_SHADER_GEOMETRY
) {
2318 shader1_key
.as_es
= 1;
2319 shader1_key
.as_ngg
= key
->as_ngg
; /* for Wave32 vs Wave64 */
2324 simple_mtx_lock(&previous_stage_sel
->mutex
);
2325 ok
= si_check_missing_main_part(sscreen
,
2327 compiler_state
, &shader1_key
);
2328 simple_mtx_unlock(&previous_stage_sel
->mutex
);
2332 ok
= si_check_missing_main_part(sscreen
, sel
,
2333 compiler_state
, key
);
2338 simple_mtx_unlock(&sel
->mutex
);
2339 return -ENOMEM
; /* skip the draw call */
2343 /* Keep the reference to the 1st shader of merged shaders, so that
2344 * Gallium can't destroy it before we destroy the 2nd shader.
2346 * Set sctx = NULL, because it's unused if we're not releasing
2347 * the shader, and we don't have any sctx here.
2349 si_shader_selector_reference(NULL
, &shader
->previous_stage_sel
,
2350 previous_stage_sel
);
2352 /* Monolithic-only shaders don't make a distinction between optimized
2353 * and unoptimized. */
2354 shader
->is_monolithic
=
2355 is_pure_monolithic
||
2356 memcmp(&key
->opt
, &zeroed
.opt
, sizeof(key
->opt
)) != 0;
2358 /* The prim discard CS is always optimized. */
2359 shader
->is_optimized
=
2360 (!is_pure_monolithic
|| key
->opt
.vs_as_prim_discard_cs
) &&
2361 memcmp(&key
->opt
, &zeroed
.opt
, sizeof(key
->opt
)) != 0;
2363 /* If it's an optimized shader, compile it asynchronously. */
2364 if (shader
->is_optimized
&& thread_index
< 0) {
2365 /* Compile it asynchronously. */
2366 util_queue_add_job(&sscreen
->shader_compiler_queue_low_priority
,
2367 shader
, &shader
->ready
,
2368 si_build_shader_variant_low_priority
, NULL
,
2371 /* Add only after the ready fence was reset, to guard against a
2372 * race with si_bind_XX_shader. */
2373 if (!sel
->last_variant
) {
2374 sel
->first_variant
= shader
;
2375 sel
->last_variant
= shader
;
2377 sel
->last_variant
->next_variant
= shader
;
2378 sel
->last_variant
= shader
;
2381 /* Use the default (unoptimized) shader for now. */
2382 memset(&key
->opt
, 0, sizeof(key
->opt
));
2383 simple_mtx_unlock(&sel
->mutex
);
2385 if (sscreen
->options
.sync_compile
)
2386 util_queue_fence_wait(&shader
->ready
);
2388 if (optimized_or_none
)
2393 /* Reset the fence before adding to the variant list. */
2394 util_queue_fence_reset(&shader
->ready
);
2396 if (!sel
->last_variant
) {
2397 sel
->first_variant
= shader
;
2398 sel
->last_variant
= shader
;
2400 sel
->last_variant
->next_variant
= shader
;
2401 sel
->last_variant
= shader
;
2404 simple_mtx_unlock(&sel
->mutex
);
2406 assert(!shader
->is_optimized
);
2407 si_build_shader_variant(shader
, thread_index
, false);
2409 util_queue_fence_signal(&shader
->ready
);
2411 if (!shader
->compilation_failed
)
2412 state
->current
= shader
;
2414 return shader
->compilation_failed
? -1 : 0;
2417 static int si_shader_select(struct pipe_context
*ctx
,
2418 struct si_shader_ctx_state
*state
,
2419 union si_vgt_stages_key stages_key
,
2420 struct si_compiler_ctx_state
*compiler_state
)
2422 struct si_context
*sctx
= (struct si_context
*)ctx
;
2423 struct si_shader_key key
;
2425 si_shader_selector_key(ctx
, state
->cso
, stages_key
, &key
);
2426 return si_shader_select_with_key(sctx
->screen
, state
, compiler_state
,
2430 static void si_parse_next_shader_property(const struct si_shader_info
*info
,
2432 struct si_shader_key
*key
)
2434 unsigned next_shader
= info
->properties
[TGSI_PROPERTY_NEXT_SHADER
];
2436 switch (info
->processor
) {
2437 case PIPE_SHADER_VERTEX
:
2438 switch (next_shader
) {
2439 case PIPE_SHADER_GEOMETRY
:
2442 case PIPE_SHADER_TESS_CTRL
:
2443 case PIPE_SHADER_TESS_EVAL
:
2447 /* If POSITION isn't written, it can only be a HW VS
2448 * if streamout is used. If streamout isn't used,
2449 * assume that it's a HW LS. (the next shader is TCS)
2450 * This heuristic is needed for separate shader objects.
2452 if (!info
->writes_position
&& !streamout
)
2457 case PIPE_SHADER_TESS_EVAL
:
2458 if (next_shader
== PIPE_SHADER_GEOMETRY
||
2459 !info
->writes_position
)
2466 * Compile the main shader part or the monolithic shader as part of
2467 * si_shader_selector initialization. Since it can be done asynchronously,
2468 * there is no way to report compile failures to applications.
2470 static void si_init_shader_selector_async(void *job
, int thread_index
)
2472 struct si_shader_selector
*sel
= (struct si_shader_selector
*)job
;
2473 struct si_screen
*sscreen
= sel
->screen
;
2474 struct ac_llvm_compiler
*compiler
;
2475 struct pipe_debug_callback
*debug
= &sel
->compiler_ctx_state
.debug
;
2477 assert(!debug
->debug_message
|| debug
->async
);
2478 assert(thread_index
>= 0);
2479 assert(thread_index
< ARRAY_SIZE(sscreen
->compiler
));
2480 compiler
= &sscreen
->compiler
[thread_index
];
2482 if (!compiler
->passes
)
2483 si_init_compiler(sscreen
, compiler
);
2485 /* Serialize NIR to save memory. Monolithic shader variants
2486 * have to deserialize NIR before compilation.
2493 /* true = remove optional debugging data to increase
2494 * the likehood of getting more shader cache hits.
2495 * It also drops variable names, so we'll save more memory.
2497 nir_serialize(&blob
, sel
->nir
, true);
2498 blob_finish_get_buffer(&blob
, &sel
->nir_binary
, &size
);
2499 sel
->nir_size
= size
;
2502 /* Compile the main shader part for use with a prolog and/or epilog.
2503 * If this fails, the driver will try to compile a monolithic shader
2506 if (!sscreen
->use_monolithic_shaders
) {
2507 struct si_shader
*shader
= CALLOC_STRUCT(si_shader
);
2508 unsigned char ir_sha1_cache_key
[20];
2511 fprintf(stderr
, "radeonsi: can't allocate a main shader part\n");
2515 /* We can leave the fence signaled because use of the default
2516 * main part is guarded by the selector's ready fence. */
2517 util_queue_fence_init(&shader
->ready
);
2519 shader
->selector
= sel
;
2520 shader
->is_monolithic
= false;
2521 si_parse_next_shader_property(&sel
->info
,
2522 sel
->so
.num_outputs
!= 0,
2525 if (sscreen
->use_ngg
&&
2526 (!sel
->so
.num_outputs
|| sscreen
->use_ngg_streamout
) &&
2527 ((sel
->type
== PIPE_SHADER_VERTEX
&& !shader
->key
.as_ls
) ||
2528 sel
->type
== PIPE_SHADER_TESS_EVAL
||
2529 sel
->type
== PIPE_SHADER_GEOMETRY
))
2530 shader
->key
.as_ngg
= 1;
2533 si_get_ir_cache_key(sel
, shader
->key
.as_ngg
,
2534 shader
->key
.as_es
, ir_sha1_cache_key
);
2537 /* Try to load the shader from the shader cache. */
2538 simple_mtx_lock(&sscreen
->shader_cache_mutex
);
2540 if (si_shader_cache_load_shader(sscreen
, ir_sha1_cache_key
, shader
)) {
2541 simple_mtx_unlock(&sscreen
->shader_cache_mutex
);
2542 si_shader_dump_stats_for_shader_db(sscreen
, shader
, debug
);
2544 simple_mtx_unlock(&sscreen
->shader_cache_mutex
);
2546 /* Compile the shader if it hasn't been loaded from the cache. */
2547 if (si_compile_shader(sscreen
, compiler
, shader
,
2550 fprintf(stderr
, "radeonsi: can't compile a main shader part\n");
2554 simple_mtx_lock(&sscreen
->shader_cache_mutex
);
2555 si_shader_cache_insert_shader(sscreen
, ir_sha1_cache_key
,
2557 simple_mtx_unlock(&sscreen
->shader_cache_mutex
);
2560 *si_get_main_shader_part(sel
, &shader
->key
) = shader
;
2562 /* Unset "outputs_written" flags for outputs converted to
2563 * DEFAULT_VAL, so that later inter-shader optimizations don't
2564 * try to eliminate outputs that don't exist in the final
2567 * This is only done if non-monolithic shaders are enabled.
2569 if ((sel
->type
== PIPE_SHADER_VERTEX
||
2570 sel
->type
== PIPE_SHADER_TESS_EVAL
) &&
2571 !shader
->key
.as_ls
&&
2572 !shader
->key
.as_es
) {
2575 for (i
= 0; i
< sel
->info
.num_outputs
; i
++) {
2576 unsigned offset
= shader
->info
.vs_output_param_offset
[i
];
2578 if (offset
<= AC_EXP_PARAM_OFFSET_31
)
2581 unsigned name
= sel
->info
.output_semantic_name
[i
];
2582 unsigned index
= sel
->info
.output_semantic_index
[i
];
2586 case TGSI_SEMANTIC_GENERIC
:
2587 /* don't process indices the function can't handle */
2588 if (index
>= SI_MAX_IO_GENERIC
)
2592 id
= si_shader_io_get_unique_index(name
, index
, true);
2593 sel
->outputs_written_before_ps
&= ~(1ull << id
);
2595 case TGSI_SEMANTIC_POSITION
: /* ignore these */
2596 case TGSI_SEMANTIC_PSIZE
:
2597 case TGSI_SEMANTIC_CLIPVERTEX
:
2598 case TGSI_SEMANTIC_EDGEFLAG
:
2605 /* The GS copy shader is always pre-compiled. */
2606 if (sel
->type
== PIPE_SHADER_GEOMETRY
&&
2607 (!sscreen
->use_ngg
||
2608 !sscreen
->use_ngg_streamout
|| /* also for PRIMITIVES_GENERATED */
2609 sel
->tess_turns_off_ngg
)) {
2610 sel
->gs_copy_shader
= si_generate_gs_copy_shader(sscreen
, compiler
, sel
, debug
);
2611 if (!sel
->gs_copy_shader
) {
2612 fprintf(stderr
, "radeonsi: can't create GS copy shader\n");
2616 si_shader_vs(sscreen
, sel
->gs_copy_shader
, sel
);
2619 /* Free NIR. We only keep serialized NIR after this point. */
2621 ralloc_free(sel
->nir
);
2626 void si_schedule_initial_compile(struct si_context
*sctx
, unsigned processor
,
2627 struct util_queue_fence
*ready_fence
,
2628 struct si_compiler_ctx_state
*compiler_ctx_state
,
2629 void *job
, util_queue_execute_func execute
)
2631 util_queue_fence_init(ready_fence
);
2633 struct util_async_debug_callback async_debug
;
2635 (sctx
->debug
.debug_message
&& !sctx
->debug
.async
) ||
2637 si_can_dump_shader(sctx
->screen
, processor
);
2640 u_async_debug_init(&async_debug
);
2641 compiler_ctx_state
->debug
= async_debug
.base
;
2644 util_queue_add_job(&sctx
->screen
->shader_compiler_queue
, job
,
2645 ready_fence
, execute
, NULL
, 0);
2648 util_queue_fence_wait(ready_fence
);
2649 u_async_debug_drain(&async_debug
, &sctx
->debug
);
2650 u_async_debug_cleanup(&async_debug
);
2653 if (sctx
->screen
->options
.sync_compile
)
2654 util_queue_fence_wait(ready_fence
);
2657 /* Return descriptor slot usage masks from the given shader info. */
2658 void si_get_active_slot_masks(const struct si_shader_info
*info
,
2659 uint32_t *const_and_shader_buffers
,
2660 uint64_t *samplers_and_images
)
2662 unsigned start
, num_shaderbufs
, num_constbufs
, num_images
, num_msaa_images
, num_samplers
;
2664 num_shaderbufs
= util_last_bit(info
->shader_buffers_declared
);
2665 num_constbufs
= util_last_bit(info
->const_buffers_declared
);
2666 /* two 8-byte images share one 16-byte slot */
2667 num_images
= align(util_last_bit(info
->images_declared
), 2);
2668 num_msaa_images
= align(util_last_bit(info
->msaa_images_declared
), 2);
2669 num_samplers
= util_last_bit(info
->samplers_declared
);
2671 /* The layout is: sb[last] ... sb[0], cb[0] ... cb[last] */
2672 start
= si_get_shaderbuf_slot(num_shaderbufs
- 1);
2673 *const_and_shader_buffers
=
2674 u_bit_consecutive(start
, num_shaderbufs
+ num_constbufs
);
2677 * - fmask[last] ... fmask[0] go to [15-last .. 15]
2678 * - image[last] ... image[0] go to [31-last .. 31]
2679 * - sampler[0] ... sampler[last] go to [32 .. 32+last*2]
2681 * FMASKs for images are placed separately, because MSAA images are rare,
2682 * and so we can benefit from a better cache hit rate if we keep image
2683 * descriptors together.
2685 if (num_msaa_images
)
2686 num_images
= SI_NUM_IMAGES
+ num_msaa_images
; /* add FMASK descriptors */
2688 start
= si_get_image_slot(num_images
- 1) / 2;
2689 *samplers_and_images
=
2690 u_bit_consecutive64(start
, num_images
/ 2 + num_samplers
);
2693 static void *si_create_shader_selector(struct pipe_context
*ctx
,
2694 const struct pipe_shader_state
*state
)
2696 struct si_screen
*sscreen
= (struct si_screen
*)ctx
->screen
;
2697 struct si_context
*sctx
= (struct si_context
*)ctx
;
2698 struct si_shader_selector
*sel
= CALLOC_STRUCT(si_shader_selector
);
2704 pipe_reference_init(&sel
->reference
, 1);
2705 sel
->screen
= sscreen
;
2706 sel
->compiler_ctx_state
.debug
= sctx
->debug
;
2707 sel
->compiler_ctx_state
.is_debug_context
= sctx
->is_debug
;
2709 sel
->so
= state
->stream_output
;
2711 if (state
->type
== PIPE_SHADER_IR_TGSI
) {
2712 sel
->nir
= tgsi_to_nir(state
->tokens
, ctx
->screen
);
2714 assert(state
->type
== PIPE_SHADER_IR_NIR
);
2715 sel
->nir
= state
->ir
.nir
;
2718 si_nir_scan_shader(sel
->nir
, &sel
->info
);
2719 si_nir_adjust_driver_locations(sel
->nir
);
2721 sel
->type
= sel
->info
.processor
;
2722 p_atomic_inc(&sscreen
->num_shaders_created
);
2723 si_get_active_slot_masks(&sel
->info
,
2724 &sel
->active_const_and_shader_buffers
,
2725 &sel
->active_samplers_and_images
);
2727 /* Record which streamout buffers are enabled. */
2728 for (i
= 0; i
< sel
->so
.num_outputs
; i
++) {
2729 sel
->enabled_streamout_buffer_mask
|=
2730 (1 << sel
->so
.output
[i
].output_buffer
) <<
2731 (sel
->so
.output
[i
].stream
* 4);
2734 sel
->num_vs_inputs
= sel
->type
== PIPE_SHADER_VERTEX
&&
2735 !sel
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
] ?
2736 sel
->info
.num_inputs
: 0;
2737 sel
->num_vbos_in_user_sgprs
=
2738 MIN2(sel
->num_vs_inputs
, sscreen
->num_vbos_in_user_sgprs
);
2740 /* The prolog is a no-op if there are no inputs. */
2741 sel
->vs_needs_prolog
= sel
->type
== PIPE_SHADER_VERTEX
&&
2742 sel
->info
.num_inputs
&&
2743 !sel
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
];
2745 sel
->force_correct_derivs_after_kill
=
2746 sel
->type
== PIPE_SHADER_FRAGMENT
&&
2747 sel
->info
.uses_derivatives
&&
2748 sel
->info
.uses_kill
&&
2749 sctx
->screen
->debug_flags
& DBG(FS_CORRECT_DERIVS_AFTER_KILL
);
2751 sel
->prim_discard_cs_allowed
=
2752 sel
->type
== PIPE_SHADER_VERTEX
&&
2753 !sel
->info
.uses_bindless_images
&&
2754 !sel
->info
.uses_bindless_samplers
&&
2755 !sel
->info
.writes_memory
&&
2756 !sel
->info
.writes_viewport_index
&&
2757 !sel
->info
.properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
] &&
2758 !sel
->so
.num_outputs
;
2760 switch (sel
->type
) {
2761 case PIPE_SHADER_GEOMETRY
:
2762 sel
->gs_output_prim
=
2763 sel
->info
.properties
[TGSI_PROPERTY_GS_OUTPUT_PRIM
];
2765 /* Only possibilities: POINTS, LINE_STRIP, TRIANGLES */
2766 sel
->rast_prim
= sel
->gs_output_prim
;
2767 if (util_rast_prim_is_triangles(sel
->rast_prim
))
2768 sel
->rast_prim
= PIPE_PRIM_TRIANGLES
;
2770 sel
->gs_max_out_vertices
=
2771 sel
->info
.properties
[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES
];
2772 sel
->gs_num_invocations
=
2773 sel
->info
.properties
[TGSI_PROPERTY_GS_INVOCATIONS
];
2774 sel
->gsvs_vertex_size
= sel
->info
.num_outputs
* 16;
2775 sel
->max_gsvs_emit_size
= sel
->gsvs_vertex_size
*
2776 sel
->gs_max_out_vertices
;
2778 sel
->max_gs_stream
= 0;
2779 for (i
= 0; i
< sel
->so
.num_outputs
; i
++)
2780 sel
->max_gs_stream
= MAX2(sel
->max_gs_stream
,
2781 sel
->so
.output
[i
].stream
);
2783 sel
->gs_input_verts_per_prim
=
2784 u_vertices_per_prim(sel
->info
.properties
[TGSI_PROPERTY_GS_INPUT_PRIM
]);
2786 /* EN_MAX_VERT_OUT_PER_GS_INSTANCE does not work with tesselation. */
2787 sel
->tess_turns_off_ngg
=
2788 sscreen
->info
.chip_class
== GFX10
&&
2789 sel
->gs_num_invocations
* sel
->gs_max_out_vertices
> 256;
2792 case PIPE_SHADER_TESS_CTRL
:
2793 /* Always reserve space for these. */
2794 sel
->patch_outputs_written
|=
2795 (1ull << si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSINNER
, 0)) |
2796 (1ull << si_shader_io_get_unique_index_patch(TGSI_SEMANTIC_TESSOUTER
, 0));
2798 case PIPE_SHADER_VERTEX
:
2799 case PIPE_SHADER_TESS_EVAL
:
2800 for (i
= 0; i
< sel
->info
.num_outputs
; i
++) {
2801 unsigned name
= sel
->info
.output_semantic_name
[i
];
2802 unsigned index
= sel
->info
.output_semantic_index
[i
];
2805 case TGSI_SEMANTIC_TESSINNER
:
2806 case TGSI_SEMANTIC_TESSOUTER
:
2807 case TGSI_SEMANTIC_PATCH
:
2808 sel
->patch_outputs_written
|=
2809 1ull << si_shader_io_get_unique_index_patch(name
, index
);
2812 case TGSI_SEMANTIC_GENERIC
:
2813 /* don't process indices the function can't handle */
2814 if (index
>= SI_MAX_IO_GENERIC
)
2818 sel
->outputs_written
|=
2819 1ull << si_shader_io_get_unique_index(name
, index
, false);
2820 sel
->outputs_written_before_ps
|=
2821 1ull << si_shader_io_get_unique_index(name
, index
, true);
2823 case TGSI_SEMANTIC_EDGEFLAG
:
2827 sel
->esgs_itemsize
= util_last_bit64(sel
->outputs_written
) * 16;
2828 sel
->lshs_vertex_stride
= sel
->esgs_itemsize
;
2830 /* Add 1 dword to reduce LDS bank conflicts, so that each vertex
2831 * will start on a different bank. (except for the maximum 32*16).
2833 if (sel
->lshs_vertex_stride
< 32*16)
2834 sel
->lshs_vertex_stride
+= 4;
2836 /* For the ESGS ring in LDS, add 1 dword to reduce LDS bank
2837 * conflicts, i.e. each vertex will start at a different bank.
2839 if (sctx
->chip_class
>= GFX9
)
2840 sel
->esgs_itemsize
+= 4;
2842 assert(((sel
->esgs_itemsize
/ 4) & C_028AAC_ITEMSIZE
) == 0);
2845 if (sel
->info
.properties
[TGSI_PROPERTY_TES_POINT_MODE
])
2846 sel
->rast_prim
= PIPE_PRIM_POINTS
;
2847 else if (sel
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
] == PIPE_PRIM_LINES
)
2848 sel
->rast_prim
= PIPE_PRIM_LINE_STRIP
;
2850 sel
->rast_prim
= PIPE_PRIM_TRIANGLES
;
2853 case PIPE_SHADER_FRAGMENT
:
2854 for (i
= 0; i
< sel
->info
.num_inputs
; i
++) {
2855 unsigned name
= sel
->info
.input_semantic_name
[i
];
2856 unsigned index
= sel
->info
.input_semantic_index
[i
];
2859 case TGSI_SEMANTIC_GENERIC
:
2860 /* don't process indices the function can't handle */
2861 if (index
>= SI_MAX_IO_GENERIC
)
2866 1ull << si_shader_io_get_unique_index(name
, index
, true);
2868 case TGSI_SEMANTIC_PCOORD
: /* ignore this */
2873 for (i
= 0; i
< 8; i
++)
2874 if (sel
->info
.colors_written
& (1 << i
))
2875 sel
->colors_written_4bit
|= 0xf << (4 * i
);
2877 for (i
= 0; i
< sel
->info
.num_inputs
; i
++) {
2878 if (sel
->info
.input_semantic_name
[i
] == TGSI_SEMANTIC_COLOR
) {
2879 int index
= sel
->info
.input_semantic_index
[i
];
2880 sel
->color_attr_index
[index
] = i
;
2887 /* PA_CL_VS_OUT_CNTL */
2888 if (sctx
->chip_class
<= GFX9
)
2889 sel
->pa_cl_vs_out_cntl
= si_get_vs_out_cntl(sel
, false);
2891 sel
->clipdist_mask
= sel
->info
.writes_clipvertex
?
2892 SIX_BITS
: sel
->info
.clipdist_writemask
;
2893 sel
->culldist_mask
= sel
->info
.culldist_writemask
<<
2894 sel
->info
.num_written_clipdistance
;
2896 /* DB_SHADER_CONTROL */
2897 sel
->db_shader_control
=
2898 S_02880C_Z_EXPORT_ENABLE(sel
->info
.writes_z
) |
2899 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(sel
->info
.writes_stencil
) |
2900 S_02880C_MASK_EXPORT_ENABLE(sel
->info
.writes_samplemask
) |
2901 S_02880C_KILL_ENABLE(sel
->info
.uses_kill
);
2903 switch (sel
->info
.properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
]) {
2904 case TGSI_FS_DEPTH_LAYOUT_GREATER
:
2905 sel
->db_shader_control
|=
2906 S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_GREATER_THAN_Z
);
2908 case TGSI_FS_DEPTH_LAYOUT_LESS
:
2909 sel
->db_shader_control
|=
2910 S_02880C_CONSERVATIVE_Z_EXPORT(V_02880C_EXPORT_LESS_THAN_Z
);
2914 /* Z_ORDER, EXEC_ON_HIER_FAIL and EXEC_ON_NOOP should be set as following:
2916 * | early Z/S | writes_mem | allow_ReZ? | Z_ORDER | EXEC_ON_HIER_FAIL | EXEC_ON_NOOP
2917 * --|-----------|------------|------------|--------------------|-------------------|-------------
2918 * 1a| false | false | true | EarlyZ_Then_ReZ | 0 | 0
2919 * 1b| false | false | false | EarlyZ_Then_LateZ | 0 | 0
2920 * 2 | false | true | n/a | LateZ | 1 | 0
2921 * 3 | true | false | n/a | EarlyZ_Then_LateZ | 0 | 0
2922 * 4 | true | true | n/a | EarlyZ_Then_LateZ | 0 | 1
2924 * In cases 3 and 4, HW will force Z_ORDER to EarlyZ regardless of what's set in the register.
2925 * In case 2, NOOP_CULL is a don't care field. In case 2, 3 and 4, ReZ doesn't make sense.
2927 * Don't use ReZ without profiling !!!
2929 * ReZ decreases performance by 15% in DiRT: Showdown on Ultra settings, which has pretty complex
2932 if (sel
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
]) {
2934 sel
->db_shader_control
|= S_02880C_DEPTH_BEFORE_SHADER(1) |
2935 S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z
) |
2936 S_02880C_EXEC_ON_NOOP(sel
->info
.writes_memory
);
2937 } else if (sel
->info
.writes_memory
) {
2939 sel
->db_shader_control
|= S_02880C_Z_ORDER(V_02880C_LATE_Z
) |
2940 S_02880C_EXEC_ON_HIER_FAIL(1);
2943 sel
->db_shader_control
|= S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z
);
2946 if (sel
->info
.properties
[TGSI_PROPERTY_FS_POST_DEPTH_COVERAGE
])
2947 sel
->db_shader_control
|= S_02880C_PRE_SHADER_DEPTH_COVERAGE_ENABLE(1);
2949 (void) simple_mtx_init(&sel
->mutex
, mtx_plain
);
2951 si_schedule_initial_compile(sctx
, sel
->info
.processor
, &sel
->ready
,
2952 &sel
->compiler_ctx_state
, sel
,
2953 si_init_shader_selector_async
);
2957 static void si_update_streamout_state(struct si_context
*sctx
)
2959 struct si_shader_selector
*shader_with_so
= si_get_vs(sctx
)->cso
;
2961 if (!shader_with_so
)
2964 sctx
->streamout
.enabled_stream_buffers_mask
=
2965 shader_with_so
->enabled_streamout_buffer_mask
;
2966 sctx
->streamout
.stride_in_dw
= shader_with_so
->so
.stride
;
2969 static void si_update_clip_regs(struct si_context
*sctx
,
2970 struct si_shader_selector
*old_hw_vs
,
2971 struct si_shader
*old_hw_vs_variant
,
2972 struct si_shader_selector
*next_hw_vs
,
2973 struct si_shader
*next_hw_vs_variant
)
2977 old_hw_vs
->info
.properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
] !=
2978 next_hw_vs
->info
.properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
] ||
2979 old_hw_vs
->pa_cl_vs_out_cntl
!= next_hw_vs
->pa_cl_vs_out_cntl
||
2980 old_hw_vs
->clipdist_mask
!= next_hw_vs
->clipdist_mask
||
2981 old_hw_vs
->culldist_mask
!= next_hw_vs
->culldist_mask
||
2982 !old_hw_vs_variant
||
2983 !next_hw_vs_variant
||
2984 old_hw_vs_variant
->key
.opt
.clip_disable
!=
2985 next_hw_vs_variant
->key
.opt
.clip_disable
))
2986 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.clip_regs
);
2989 static void si_update_common_shader_state(struct si_context
*sctx
)
2991 sctx
->uses_bindless_samplers
=
2992 si_shader_uses_bindless_samplers(sctx
->vs_shader
.cso
) ||
2993 si_shader_uses_bindless_samplers(sctx
->gs_shader
.cso
) ||
2994 si_shader_uses_bindless_samplers(sctx
->ps_shader
.cso
) ||
2995 si_shader_uses_bindless_samplers(sctx
->tcs_shader
.cso
) ||
2996 si_shader_uses_bindless_samplers(sctx
->tes_shader
.cso
);
2997 sctx
->uses_bindless_images
=
2998 si_shader_uses_bindless_images(sctx
->vs_shader
.cso
) ||
2999 si_shader_uses_bindless_images(sctx
->gs_shader
.cso
) ||
3000 si_shader_uses_bindless_images(sctx
->ps_shader
.cso
) ||
3001 si_shader_uses_bindless_images(sctx
->tcs_shader
.cso
) ||
3002 si_shader_uses_bindless_images(sctx
->tes_shader
.cso
);
3003 sctx
->do_update_shaders
= true;
3006 static void si_bind_vs_shader(struct pipe_context
*ctx
, void *state
)
3008 struct si_context
*sctx
= (struct si_context
*)ctx
;
3009 struct si_shader_selector
*old_hw_vs
= si_get_vs(sctx
)->cso
;
3010 struct si_shader
*old_hw_vs_variant
= si_get_vs_state(sctx
);
3011 struct si_shader_selector
*sel
= state
;
3013 if (sctx
->vs_shader
.cso
== sel
)
3016 sctx
->vs_shader
.cso
= sel
;
3017 sctx
->vs_shader
.current
= sel
? sel
->first_variant
: NULL
;
3018 sctx
->num_vs_blit_sgprs
= sel
? sel
->info
.properties
[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD
] : 0;
3020 if (si_update_ngg(sctx
))
3021 si_shader_change_notify(sctx
);
3023 si_update_common_shader_state(sctx
);
3024 si_update_vs_viewport_state(sctx
);
3025 si_set_active_descriptors_for_shader(sctx
, sel
);
3026 si_update_streamout_state(sctx
);
3027 si_update_clip_regs(sctx
, old_hw_vs
, old_hw_vs_variant
,
3028 si_get_vs(sctx
)->cso
, si_get_vs_state(sctx
));
3031 static void si_update_tess_uses_prim_id(struct si_context
*sctx
)
3033 sctx
->ia_multi_vgt_param_key
.u
.tess_uses_prim_id
=
3034 (sctx
->tes_shader
.cso
&&
3035 sctx
->tes_shader
.cso
->info
.uses_primid
) ||
3036 (sctx
->tcs_shader
.cso
&&
3037 sctx
->tcs_shader
.cso
->info
.uses_primid
) ||
3038 (sctx
->gs_shader
.cso
&&
3039 sctx
->gs_shader
.cso
->info
.uses_primid
) ||
3040 (sctx
->ps_shader
.cso
&& !sctx
->gs_shader
.cso
&&
3041 sctx
->ps_shader
.cso
->info
.uses_primid
);
3044 bool si_update_ngg(struct si_context
*sctx
)
3046 if (!sctx
->screen
->use_ngg
) {
3051 bool new_ngg
= true;
3053 if (sctx
->gs_shader
.cso
&& sctx
->tes_shader
.cso
&&
3054 sctx
->gs_shader
.cso
->tess_turns_off_ngg
) {
3056 } else if (!sctx
->screen
->use_ngg_streamout
) {
3057 struct si_shader_selector
*last
= si_get_vs(sctx
)->cso
;
3059 if ((last
&& last
->so
.num_outputs
) ||
3060 sctx
->streamout
.prims_gen_query_enabled
)
3064 if (new_ngg
!= sctx
->ngg
) {
3065 /* Transitioning from NGG to legacy GS requires VGT_FLUSH on Navi10-14.
3066 * VGT_FLUSH is also emitted at the beginning of IBs when legacy GS ring
3069 if ((sctx
->family
== CHIP_NAVI10
||
3070 sctx
->family
== CHIP_NAVI12
||
3071 sctx
->family
== CHIP_NAVI14
) &&
3073 sctx
->flags
|= SI_CONTEXT_VGT_FLUSH
;
3075 sctx
->ngg
= new_ngg
;
3076 sctx
->last_gs_out_prim
= -1; /* reset this so that it gets updated */
3082 static void si_bind_gs_shader(struct pipe_context
*ctx
, void *state
)
3084 struct si_context
*sctx
= (struct si_context
*)ctx
;
3085 struct si_shader_selector
*old_hw_vs
= si_get_vs(sctx
)->cso
;
3086 struct si_shader
*old_hw_vs_variant
= si_get_vs_state(sctx
);
3087 struct si_shader_selector
*sel
= state
;
3088 bool enable_changed
= !!sctx
->gs_shader
.cso
!= !!sel
;
3091 if (sctx
->gs_shader
.cso
== sel
)
3094 sctx
->gs_shader
.cso
= sel
;
3095 sctx
->gs_shader
.current
= sel
? sel
->first_variant
: NULL
;
3096 sctx
->ia_multi_vgt_param_key
.u
.uses_gs
= sel
!= NULL
;
3098 si_update_common_shader_state(sctx
);
3099 sctx
->last_gs_out_prim
= -1; /* reset this so that it gets updated */
3101 ngg_changed
= si_update_ngg(sctx
);
3102 if (ngg_changed
|| enable_changed
)
3103 si_shader_change_notify(sctx
);
3104 if (enable_changed
) {
3105 if (sctx
->ia_multi_vgt_param_key
.u
.uses_tess
)
3106 si_update_tess_uses_prim_id(sctx
);
3108 si_update_vs_viewport_state(sctx
);
3109 si_set_active_descriptors_for_shader(sctx
, sel
);
3110 si_update_streamout_state(sctx
);
3111 si_update_clip_regs(sctx
, old_hw_vs
, old_hw_vs_variant
,
3112 si_get_vs(sctx
)->cso
, si_get_vs_state(sctx
));
3115 static void si_bind_tcs_shader(struct pipe_context
*ctx
, void *state
)
3117 struct si_context
*sctx
= (struct si_context
*)ctx
;
3118 struct si_shader_selector
*sel
= state
;
3119 bool enable_changed
= !!sctx
->tcs_shader
.cso
!= !!sel
;
3121 if (sctx
->tcs_shader
.cso
== sel
)
3124 sctx
->tcs_shader
.cso
= sel
;
3125 sctx
->tcs_shader
.current
= sel
? sel
->first_variant
: NULL
;
3126 si_update_tess_uses_prim_id(sctx
);
3128 si_update_common_shader_state(sctx
);
3131 sctx
->last_tcs
= NULL
; /* invalidate derived tess state */
3133 si_set_active_descriptors_for_shader(sctx
, sel
);
3136 static void si_bind_tes_shader(struct pipe_context
*ctx
, void *state
)
3138 struct si_context
*sctx
= (struct si_context
*)ctx
;
3139 struct si_shader_selector
*old_hw_vs
= si_get_vs(sctx
)->cso
;
3140 struct si_shader
*old_hw_vs_variant
= si_get_vs_state(sctx
);
3141 struct si_shader_selector
*sel
= state
;
3142 bool enable_changed
= !!sctx
->tes_shader
.cso
!= !!sel
;
3144 if (sctx
->tes_shader
.cso
== sel
)
3147 sctx
->tes_shader
.cso
= sel
;
3148 sctx
->tes_shader
.current
= sel
? sel
->first_variant
: NULL
;
3149 sctx
->ia_multi_vgt_param_key
.u
.uses_tess
= sel
!= NULL
;
3150 si_update_tess_uses_prim_id(sctx
);
3152 si_update_common_shader_state(sctx
);
3153 sctx
->last_gs_out_prim
= -1; /* reset this so that it gets updated */
3155 bool ngg_changed
= si_update_ngg(sctx
);
3156 if (ngg_changed
|| enable_changed
)
3157 si_shader_change_notify(sctx
);
3159 sctx
->last_tes_sh_base
= -1; /* invalidate derived tess state */
3160 si_update_vs_viewport_state(sctx
);
3161 si_set_active_descriptors_for_shader(sctx
, sel
);
3162 si_update_streamout_state(sctx
);
3163 si_update_clip_regs(sctx
, old_hw_vs
, old_hw_vs_variant
,
3164 si_get_vs(sctx
)->cso
, si_get_vs_state(sctx
));
3167 static void si_bind_ps_shader(struct pipe_context
*ctx
, void *state
)
3169 struct si_context
*sctx
= (struct si_context
*)ctx
;
3170 struct si_shader_selector
*old_sel
= sctx
->ps_shader
.cso
;
3171 struct si_shader_selector
*sel
= state
;
3173 /* skip if supplied shader is one already in use */
3177 sctx
->ps_shader
.cso
= sel
;
3178 sctx
->ps_shader
.current
= sel
? sel
->first_variant
: NULL
;
3180 si_update_common_shader_state(sctx
);
3182 if (sctx
->ia_multi_vgt_param_key
.u
.uses_tess
)
3183 si_update_tess_uses_prim_id(sctx
);
3186 old_sel
->info
.colors_written
!= sel
->info
.colors_written
)
3187 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.cb_render_state
);
3189 if (sctx
->screen
->has_out_of_order_rast
&&
3191 old_sel
->info
.writes_memory
!= sel
->info
.writes_memory
||
3192 old_sel
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
] !=
3193 sel
->info
.properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
]))
3194 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.msaa_config
);
3196 si_set_active_descriptors_for_shader(sctx
, sel
);
3197 si_update_ps_colorbuf0_slot(sctx
);
3200 static void si_delete_shader(struct si_context
*sctx
, struct si_shader
*shader
)
3202 if (shader
->is_optimized
) {
3203 util_queue_drop_job(&sctx
->screen
->shader_compiler_queue_low_priority
,
3207 util_queue_fence_destroy(&shader
->ready
);
3210 /* If destroyed shaders were not unbound, the next compiled
3211 * shader variant could get the same pointer address and so
3212 * binding it to the same shader stage would be considered
3213 * a no-op, causing random behavior.
3215 switch (shader
->selector
->type
) {
3216 case PIPE_SHADER_VERTEX
:
3217 if (shader
->key
.as_ls
) {
3218 assert(sctx
->chip_class
<= GFX8
);
3219 si_pm4_delete_state(sctx
, ls
, shader
->pm4
);
3220 } else if (shader
->key
.as_es
) {
3221 assert(sctx
->chip_class
<= GFX8
);
3222 si_pm4_delete_state(sctx
, es
, shader
->pm4
);
3223 } else if (shader
->key
.as_ngg
) {
3224 si_pm4_delete_state(sctx
, gs
, shader
->pm4
);
3226 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
3229 case PIPE_SHADER_TESS_CTRL
:
3230 si_pm4_delete_state(sctx
, hs
, shader
->pm4
);
3232 case PIPE_SHADER_TESS_EVAL
:
3233 if (shader
->key
.as_es
) {
3234 assert(sctx
->chip_class
<= GFX8
);
3235 si_pm4_delete_state(sctx
, es
, shader
->pm4
);
3236 } else if (shader
->key
.as_ngg
) {
3237 si_pm4_delete_state(sctx
, gs
, shader
->pm4
);
3239 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
3242 case PIPE_SHADER_GEOMETRY
:
3243 if (shader
->is_gs_copy_shader
)
3244 si_pm4_delete_state(sctx
, vs
, shader
->pm4
);
3246 si_pm4_delete_state(sctx
, gs
, shader
->pm4
);
3248 case PIPE_SHADER_FRAGMENT
:
3249 si_pm4_delete_state(sctx
, ps
, shader
->pm4
);
3255 si_shader_selector_reference(sctx
, &shader
->previous_stage_sel
, NULL
);
3256 si_shader_destroy(shader
);
3260 void si_destroy_shader_selector(struct si_context
*sctx
,
3261 struct si_shader_selector
*sel
)
3263 struct si_shader
*p
= sel
->first_variant
, *c
;
3264 struct si_shader_ctx_state
*current_shader
[SI_NUM_SHADERS
] = {
3265 [PIPE_SHADER_VERTEX
] = &sctx
->vs_shader
,
3266 [PIPE_SHADER_TESS_CTRL
] = &sctx
->tcs_shader
,
3267 [PIPE_SHADER_TESS_EVAL
] = &sctx
->tes_shader
,
3268 [PIPE_SHADER_GEOMETRY
] = &sctx
->gs_shader
,
3269 [PIPE_SHADER_FRAGMENT
] = &sctx
->ps_shader
,
3272 util_queue_drop_job(&sctx
->screen
->shader_compiler_queue
, &sel
->ready
);
3274 if (current_shader
[sel
->type
]->cso
== sel
) {
3275 current_shader
[sel
->type
]->cso
= NULL
;
3276 current_shader
[sel
->type
]->current
= NULL
;
3280 c
= p
->next_variant
;
3281 si_delete_shader(sctx
, p
);
3285 if (sel
->main_shader_part
)
3286 si_delete_shader(sctx
, sel
->main_shader_part
);
3287 if (sel
->main_shader_part_ls
)
3288 si_delete_shader(sctx
, sel
->main_shader_part_ls
);
3289 if (sel
->main_shader_part_es
)
3290 si_delete_shader(sctx
, sel
->main_shader_part_es
);
3291 if (sel
->main_shader_part_ngg
)
3292 si_delete_shader(sctx
, sel
->main_shader_part_ngg
);
3293 if (sel
->gs_copy_shader
)
3294 si_delete_shader(sctx
, sel
->gs_copy_shader
);
3296 util_queue_fence_destroy(&sel
->ready
);
3297 simple_mtx_destroy(&sel
->mutex
);
3298 ralloc_free(sel
->nir
);
3299 free(sel
->nir_binary
);
3303 static void si_delete_shader_selector(struct pipe_context
*ctx
, void *state
)
3305 struct si_context
*sctx
= (struct si_context
*)ctx
;
3306 struct si_shader_selector
*sel
= (struct si_shader_selector
*)state
;
3308 si_shader_selector_reference(sctx
, &sel
, NULL
);
3311 static unsigned si_get_ps_input_cntl(struct si_context
*sctx
,
3312 struct si_shader
*vs
, unsigned name
,
3313 unsigned index
, unsigned interpolate
)
3315 struct si_shader_info
*vsinfo
= &vs
->selector
->info
;
3316 unsigned j
, offset
, ps_input_cntl
= 0;
3318 if (interpolate
== TGSI_INTERPOLATE_CONSTANT
||
3319 (interpolate
== TGSI_INTERPOLATE_COLOR
&& sctx
->flatshade
) ||
3320 name
== TGSI_SEMANTIC_PRIMID
)
3321 ps_input_cntl
|= S_028644_FLAT_SHADE(1);
3323 if (name
== TGSI_SEMANTIC_PCOORD
||
3324 (name
== TGSI_SEMANTIC_TEXCOORD
&&
3325 sctx
->sprite_coord_enable
& (1 << index
))) {
3326 ps_input_cntl
|= S_028644_PT_SPRITE_TEX(1);
3329 for (j
= 0; j
< vsinfo
->num_outputs
; j
++) {
3330 if (name
== vsinfo
->output_semantic_name
[j
] &&
3331 index
== vsinfo
->output_semantic_index
[j
]) {
3332 offset
= vs
->info
.vs_output_param_offset
[j
];
3334 if (offset
<= AC_EXP_PARAM_OFFSET_31
) {
3335 /* The input is loaded from parameter memory. */
3336 ps_input_cntl
|= S_028644_OFFSET(offset
);
3337 } else if (!G_028644_PT_SPRITE_TEX(ps_input_cntl
)) {
3338 if (offset
== AC_EXP_PARAM_UNDEFINED
) {
3339 /* This can happen with depth-only rendering. */
3342 /* The input is a DEFAULT_VAL constant. */
3343 assert(offset
>= AC_EXP_PARAM_DEFAULT_VAL_0000
&&
3344 offset
<= AC_EXP_PARAM_DEFAULT_VAL_1111
);
3345 offset
-= AC_EXP_PARAM_DEFAULT_VAL_0000
;
3348 ps_input_cntl
= S_028644_OFFSET(0x20) |
3349 S_028644_DEFAULT_VAL(offset
);
3355 if (j
== vsinfo
->num_outputs
&& name
== TGSI_SEMANTIC_PRIMID
)
3356 /* PrimID is written after the last output when HW VS is used. */
3357 ps_input_cntl
|= S_028644_OFFSET(vs
->info
.vs_output_param_offset
[vsinfo
->num_outputs
]);
3358 else if (j
== vsinfo
->num_outputs
&& !G_028644_PT_SPRITE_TEX(ps_input_cntl
)) {
3359 /* No corresponding output found, load defaults into input.
3360 * Don't set any other bits.
3361 * (FLAT_SHADE=1 completely changes behavior) */
3362 ps_input_cntl
= S_028644_OFFSET(0x20);
3363 /* D3D 9 behaviour. GL is undefined */
3364 if (name
== TGSI_SEMANTIC_COLOR
&& index
== 0)
3365 ps_input_cntl
|= S_028644_DEFAULT_VAL(3);
3367 return ps_input_cntl
;
3370 static void si_emit_spi_map(struct si_context
*sctx
)
3372 struct si_shader
*ps
= sctx
->ps_shader
.current
;
3373 struct si_shader
*vs
= si_get_vs_state(sctx
);
3374 struct si_shader_info
*psinfo
= ps
? &ps
->selector
->info
: NULL
;
3375 unsigned i
, num_interp
, num_written
= 0, bcol_interp
[2];
3376 unsigned spi_ps_input_cntl
[32];
3378 if (!ps
|| !ps
->selector
->info
.num_inputs
)
3381 num_interp
= si_get_ps_num_interp(ps
);
3382 assert(num_interp
> 0);
3384 for (i
= 0; i
< psinfo
->num_inputs
; i
++) {
3385 unsigned name
= psinfo
->input_semantic_name
[i
];
3386 unsigned index
= psinfo
->input_semantic_index
[i
];
3387 unsigned interpolate
= psinfo
->input_interpolate
[i
];
3389 spi_ps_input_cntl
[num_written
++] = si_get_ps_input_cntl(sctx
, vs
, name
,
3390 index
, interpolate
);
3392 if (name
== TGSI_SEMANTIC_COLOR
) {
3393 assert(index
< ARRAY_SIZE(bcol_interp
));
3394 bcol_interp
[index
] = interpolate
;
3398 if (ps
->key
.part
.ps
.prolog
.color_two_side
) {
3399 unsigned bcol
= TGSI_SEMANTIC_BCOLOR
;
3401 for (i
= 0; i
< 2; i
++) {
3402 if (!(psinfo
->colors_read
& (0xf << (i
* 4))))
3405 spi_ps_input_cntl
[num_written
++] =
3406 si_get_ps_input_cntl(sctx
, vs
, bcol
, i
, bcol_interp
[i
]);
3410 assert(num_interp
== num_written
);
3412 /* R_028644_SPI_PS_INPUT_CNTL_0 */
3413 /* Dota 2: Only ~16% of SPI map updates set different values. */
3414 /* Talos: Only ~9% of SPI map updates set different values. */
3415 unsigned initial_cdw
= sctx
->gfx_cs
->current
.cdw
;
3416 radeon_opt_set_context_regn(sctx
, R_028644_SPI_PS_INPUT_CNTL_0
,
3418 sctx
->tracked_regs
.spi_ps_input_cntl
, num_interp
);
3420 if (initial_cdw
!= sctx
->gfx_cs
->current
.cdw
)
3421 sctx
->context_roll
= true;
3425 * Writing CONFIG or UCONFIG VGT registers requires VGT_FLUSH before that.
3427 static void si_init_config_add_vgt_flush(struct si_context
*sctx
)
3429 if (sctx
->init_config_has_vgt_flush
)
3432 /* Done by Vulkan before VGT_FLUSH. */
3433 si_pm4_cmd_begin(sctx
->init_config
, PKT3_EVENT_WRITE
);
3434 si_pm4_cmd_add(sctx
->init_config
,
3435 EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
3436 si_pm4_cmd_end(sctx
->init_config
, false);
3438 /* VGT_FLUSH is required even if VGT is idle. It resets VGT pointers. */
3439 si_pm4_cmd_begin(sctx
->init_config
, PKT3_EVENT_WRITE
);
3440 si_pm4_cmd_add(sctx
->init_config
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
3441 si_pm4_cmd_end(sctx
->init_config
, false);
3442 sctx
->init_config_has_vgt_flush
= true;
3445 /* Initialize state related to ESGS / GSVS ring buffers */
3446 static bool si_update_gs_ring_buffers(struct si_context
*sctx
)
3448 struct si_shader_selector
*es
=
3449 sctx
->tes_shader
.cso
? sctx
->tes_shader
.cso
: sctx
->vs_shader
.cso
;
3450 struct si_shader_selector
*gs
= sctx
->gs_shader
.cso
;
3451 struct si_pm4_state
*pm4
;
3453 /* Chip constants. */
3454 unsigned num_se
= sctx
->screen
->info
.max_se
;
3455 unsigned wave_size
= 64;
3456 unsigned max_gs_waves
= 32 * num_se
; /* max 32 per SE on GCN */
3457 /* On GFX6-GFX7, the value comes from VGT_GS_VERTEX_REUSE = 16.
3458 * On GFX8+, the value comes from VGT_VERTEX_REUSE_BLOCK_CNTL = 30 (+2).
3460 unsigned gs_vertex_reuse
= (sctx
->chip_class
>= GFX8
? 32 : 16) * num_se
;
3461 unsigned alignment
= 256 * num_se
;
3462 /* The maximum size is 63.999 MB per SE. */
3463 unsigned max_size
= ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se
;
3465 /* Calculate the minimum size. */
3466 unsigned min_esgs_ring_size
= align(es
->esgs_itemsize
* gs_vertex_reuse
*
3467 wave_size
, alignment
);
3469 /* These are recommended sizes, not minimum sizes. */
3470 unsigned esgs_ring_size
= max_gs_waves
* 2 * wave_size
*
3471 es
->esgs_itemsize
* gs
->gs_input_verts_per_prim
;
3472 unsigned gsvs_ring_size
= max_gs_waves
* 2 * wave_size
*
3473 gs
->max_gsvs_emit_size
;
3475 min_esgs_ring_size
= align(min_esgs_ring_size
, alignment
);
3476 esgs_ring_size
= align(esgs_ring_size
, alignment
);
3477 gsvs_ring_size
= align(gsvs_ring_size
, alignment
);
3479 esgs_ring_size
= CLAMP(esgs_ring_size
, min_esgs_ring_size
, max_size
);
3480 gsvs_ring_size
= MIN2(gsvs_ring_size
, max_size
);
3482 /* Some rings don't have to be allocated if shaders don't use them.
3483 * (e.g. no varyings between ES and GS or GS and VS)
3485 * GFX9 doesn't have the ESGS ring.
3487 bool update_esgs
= sctx
->chip_class
<= GFX8
&&
3489 (!sctx
->esgs_ring
||
3490 sctx
->esgs_ring
->width0
< esgs_ring_size
);
3491 bool update_gsvs
= gsvs_ring_size
&&
3492 (!sctx
->gsvs_ring
||
3493 sctx
->gsvs_ring
->width0
< gsvs_ring_size
);
3495 if (!update_esgs
&& !update_gsvs
)
3499 pipe_resource_reference(&sctx
->esgs_ring
, NULL
);
3501 pipe_aligned_buffer_create(sctx
->b
.screen
,
3502 SI_RESOURCE_FLAG_UNMAPPABLE
,
3505 sctx
->screen
->info
.pte_fragment_size
);
3506 if (!sctx
->esgs_ring
)
3511 pipe_resource_reference(&sctx
->gsvs_ring
, NULL
);
3513 pipe_aligned_buffer_create(sctx
->b
.screen
,
3514 SI_RESOURCE_FLAG_UNMAPPABLE
,
3517 sctx
->screen
->info
.pte_fragment_size
);
3518 if (!sctx
->gsvs_ring
)
3522 /* Create the "init_config_gs_rings" state. */
3523 pm4
= CALLOC_STRUCT(si_pm4_state
);
3527 if (sctx
->chip_class
>= GFX7
) {
3528 if (sctx
->esgs_ring
) {
3529 assert(sctx
->chip_class
<= GFX8
);
3530 si_pm4_set_reg(pm4
, R_030900_VGT_ESGS_RING_SIZE
,
3531 sctx
->esgs_ring
->width0
/ 256);
3533 if (sctx
->gsvs_ring
)
3534 si_pm4_set_reg(pm4
, R_030904_VGT_GSVS_RING_SIZE
,
3535 sctx
->gsvs_ring
->width0
/ 256);
3537 if (sctx
->esgs_ring
)
3538 si_pm4_set_reg(pm4
, R_0088C8_VGT_ESGS_RING_SIZE
,
3539 sctx
->esgs_ring
->width0
/ 256);
3540 if (sctx
->gsvs_ring
)
3541 si_pm4_set_reg(pm4
, R_0088CC_VGT_GSVS_RING_SIZE
,
3542 sctx
->gsvs_ring
->width0
/ 256);
3545 /* Set the state. */
3546 if (sctx
->init_config_gs_rings
)
3547 si_pm4_free_state(sctx
, sctx
->init_config_gs_rings
, ~0);
3548 sctx
->init_config_gs_rings
= pm4
;
3550 if (!sctx
->init_config_has_vgt_flush
) {
3551 si_init_config_add_vgt_flush(sctx
);
3552 si_pm4_upload_indirect_buffer(sctx
, sctx
->init_config
);
3555 /* Flush the context to re-emit both init_config states. */
3556 sctx
->initial_gfx_cs_size
= 0; /* force flush */
3557 si_flush_gfx_cs(sctx
, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW
, NULL
);
3559 /* Set ring bindings. */
3560 if (sctx
->esgs_ring
) {
3561 assert(sctx
->chip_class
<= GFX8
);
3562 si_set_ring_buffer(sctx
, SI_ES_RING_ESGS
,
3563 sctx
->esgs_ring
, 0, sctx
->esgs_ring
->width0
,
3564 true, true, 4, 64, 0);
3565 si_set_ring_buffer(sctx
, SI_GS_RING_ESGS
,
3566 sctx
->esgs_ring
, 0, sctx
->esgs_ring
->width0
,
3567 false, false, 0, 0, 0);
3569 if (sctx
->gsvs_ring
) {
3570 si_set_ring_buffer(sctx
, SI_RING_GSVS
,
3571 sctx
->gsvs_ring
, 0, sctx
->gsvs_ring
->width0
,
3572 false, false, 0, 0, 0);
3578 static void si_shader_lock(struct si_shader
*shader
)
3580 simple_mtx_lock(&shader
->selector
->mutex
);
3581 if (shader
->previous_stage_sel
) {
3582 assert(shader
->previous_stage_sel
!= shader
->selector
);
3583 simple_mtx_lock(&shader
->previous_stage_sel
->mutex
);
3587 static void si_shader_unlock(struct si_shader
*shader
)
3589 if (shader
->previous_stage_sel
)
3590 simple_mtx_unlock(&shader
->previous_stage_sel
->mutex
);
3591 simple_mtx_unlock(&shader
->selector
->mutex
);
3595 * @returns 1 if \p sel has been updated to use a new scratch buffer
3597 * < 0 if there was a failure
3599 static int si_update_scratch_buffer(struct si_context
*sctx
,
3600 struct si_shader
*shader
)
3602 uint64_t scratch_va
= sctx
->scratch_buffer
->gpu_address
;
3607 /* This shader doesn't need a scratch buffer */
3608 if (shader
->config
.scratch_bytes_per_wave
== 0)
3611 /* Prevent race conditions when updating:
3612 * - si_shader::scratch_bo
3613 * - si_shader::binary::code
3614 * - si_shader::previous_stage::binary::code.
3616 si_shader_lock(shader
);
3618 /* This shader is already configured to use the current
3619 * scratch buffer. */
3620 if (shader
->scratch_bo
== sctx
->scratch_buffer
) {
3621 si_shader_unlock(shader
);
3625 assert(sctx
->scratch_buffer
);
3627 /* Replace the shader bo with a new bo that has the relocs applied. */
3628 if (!si_shader_binary_upload(sctx
->screen
, shader
, scratch_va
)) {
3629 si_shader_unlock(shader
);
3633 /* Update the shader state to use the new shader bo. */
3634 si_shader_init_pm4_state(sctx
->screen
, shader
);
3636 si_resource_reference(&shader
->scratch_bo
, sctx
->scratch_buffer
);
3638 si_shader_unlock(shader
);
3642 static unsigned si_get_scratch_buffer_bytes_per_wave(struct si_shader
*shader
)
3644 return shader
? shader
->config
.scratch_bytes_per_wave
: 0;
3647 static struct si_shader
*si_get_tcs_current(struct si_context
*sctx
)
3649 if (!sctx
->tes_shader
.cso
)
3650 return NULL
; /* tessellation disabled */
3652 return sctx
->tcs_shader
.cso
? sctx
->tcs_shader
.current
:
3653 sctx
->fixed_func_tcs_shader
.current
;
3656 static bool si_update_scratch_relocs(struct si_context
*sctx
)
3658 struct si_shader
*tcs
= si_get_tcs_current(sctx
);
3661 /* Update the shaders, so that they are using the latest scratch.
3662 * The scratch buffer may have been changed since these shaders were
3663 * last used, so we still need to try to update them, even if they
3664 * require scratch buffers smaller than the current size.
3666 r
= si_update_scratch_buffer(sctx
, sctx
->ps_shader
.current
);
3670 si_pm4_bind_state(sctx
, ps
, sctx
->ps_shader
.current
->pm4
);
3672 r
= si_update_scratch_buffer(sctx
, sctx
->gs_shader
.current
);
3676 si_pm4_bind_state(sctx
, gs
, sctx
->gs_shader
.current
->pm4
);
3678 r
= si_update_scratch_buffer(sctx
, tcs
);
3682 si_pm4_bind_state(sctx
, hs
, tcs
->pm4
);
3684 /* VS can be bound as LS, ES, or VS. */
3685 r
= si_update_scratch_buffer(sctx
, sctx
->vs_shader
.current
);
3689 if (sctx
->vs_shader
.current
->key
.as_ls
)
3690 si_pm4_bind_state(sctx
, ls
, sctx
->vs_shader
.current
->pm4
);
3691 else if (sctx
->vs_shader
.current
->key
.as_es
)
3692 si_pm4_bind_state(sctx
, es
, sctx
->vs_shader
.current
->pm4
);
3693 else if (sctx
->vs_shader
.current
->key
.as_ngg
)
3694 si_pm4_bind_state(sctx
, gs
, sctx
->vs_shader
.current
->pm4
);
3696 si_pm4_bind_state(sctx
, vs
, sctx
->vs_shader
.current
->pm4
);
3699 /* TES can be bound as ES or VS. */
3700 r
= si_update_scratch_buffer(sctx
, sctx
->tes_shader
.current
);
3704 if (sctx
->tes_shader
.current
->key
.as_es
)
3705 si_pm4_bind_state(sctx
, es
, sctx
->tes_shader
.current
->pm4
);
3706 else if (sctx
->tes_shader
.current
->key
.as_ngg
)
3707 si_pm4_bind_state(sctx
, gs
, sctx
->tes_shader
.current
->pm4
);
3709 si_pm4_bind_state(sctx
, vs
, sctx
->tes_shader
.current
->pm4
);
3715 static bool si_update_spi_tmpring_size(struct si_context
*sctx
)
3717 /* SPI_TMPRING_SIZE.WAVESIZE must be constant for each scratch buffer.
3718 * There are 2 cases to handle:
3720 * - If the current needed size is less than the maximum seen size,
3721 * use the maximum seen size, so that WAVESIZE remains the same.
3723 * - If the current needed size is greater than the maximum seen size,
3724 * the scratch buffer is reallocated, so we can increase WAVESIZE.
3726 * Shaders that set SCRATCH_EN=0 don't allocate scratch space.
3727 * Otherwise, the number of waves that can use scratch is
3728 * SPI_TMPRING_SIZE.WAVES.
3732 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->ps_shader
.current
));
3733 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->gs_shader
.current
));
3734 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->vs_shader
.current
));
3736 if (sctx
->tes_shader
.cso
) {
3737 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
->tes_shader
.current
));
3738 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(si_get_tcs_current(sctx
)));
3741 sctx
->max_seen_scratch_bytes_per_wave
=
3742 MAX2(sctx
->max_seen_scratch_bytes_per_wave
, bytes
);
3744 unsigned scratch_needed_size
=
3745 sctx
->max_seen_scratch_bytes_per_wave
* sctx
->scratch_waves
;
3746 unsigned spi_tmpring_size
;
3748 if (scratch_needed_size
> 0) {
3749 if (!sctx
->scratch_buffer
||
3750 scratch_needed_size
> sctx
->scratch_buffer
->b
.b
.width0
) {
3751 /* Create a bigger scratch buffer */
3752 si_resource_reference(&sctx
->scratch_buffer
, NULL
);
3754 sctx
->scratch_buffer
=
3755 si_aligned_buffer_create(&sctx
->screen
->b
,
3756 SI_RESOURCE_FLAG_UNMAPPABLE
,
3758 scratch_needed_size
,
3759 sctx
->screen
->info
.pte_fragment_size
);
3760 if (!sctx
->scratch_buffer
)
3763 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.scratch_state
);
3764 si_context_add_resource_size(sctx
,
3765 &sctx
->scratch_buffer
->b
.b
);
3768 if (!si_update_scratch_relocs(sctx
))
3772 /* The LLVM shader backend should be reporting aligned scratch_sizes. */
3773 assert((scratch_needed_size
& ~0x3FF) == scratch_needed_size
&&
3774 "scratch size should already be aligned correctly.");
3776 spi_tmpring_size
= S_0286E8_WAVES(sctx
->scratch_waves
) |
3777 S_0286E8_WAVESIZE(sctx
->max_seen_scratch_bytes_per_wave
>> 10);
3778 if (spi_tmpring_size
!= sctx
->spi_tmpring_size
) {
3779 sctx
->spi_tmpring_size
= spi_tmpring_size
;
3780 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.scratch_state
);
3785 static void si_init_tess_factor_ring(struct si_context
*sctx
)
3787 assert(!sctx
->tess_rings
);
3788 assert(((sctx
->screen
->tess_factor_ring_size
/ 4) & C_030938_SIZE
) == 0);
3790 /* The address must be aligned to 2^19, because the shader only
3791 * receives the high 13 bits.
3793 sctx
->tess_rings
= pipe_aligned_buffer_create(sctx
->b
.screen
,
3794 SI_RESOURCE_FLAG_32BIT
,
3796 sctx
->screen
->tess_offchip_ring_size
+
3797 sctx
->screen
->tess_factor_ring_size
,
3799 if (!sctx
->tess_rings
)
3802 si_init_config_add_vgt_flush(sctx
);
3804 si_pm4_add_bo(sctx
->init_config
, si_resource(sctx
->tess_rings
),
3805 RADEON_USAGE_READWRITE
, RADEON_PRIO_SHADER_RINGS
);
3807 uint64_t factor_va
= si_resource(sctx
->tess_rings
)->gpu_address
+
3808 sctx
->screen
->tess_offchip_ring_size
;
3810 /* Append these registers to the init config state. */
3811 if (sctx
->chip_class
>= GFX7
) {
3812 si_pm4_set_reg(sctx
->init_config
, R_030938_VGT_TF_RING_SIZE
,
3813 S_030938_SIZE(sctx
->screen
->tess_factor_ring_size
/ 4));
3814 si_pm4_set_reg(sctx
->init_config
, R_030940_VGT_TF_MEMORY_BASE
,
3816 if (sctx
->chip_class
>= GFX10
)
3817 si_pm4_set_reg(sctx
->init_config
, R_030984_VGT_TF_MEMORY_BASE_HI_UMD
,
3818 S_030984_BASE_HI(factor_va
>> 40));
3819 else if (sctx
->chip_class
== GFX9
)
3820 si_pm4_set_reg(sctx
->init_config
, R_030944_VGT_TF_MEMORY_BASE_HI
,
3821 S_030944_BASE_HI(factor_va
>> 40));
3822 si_pm4_set_reg(sctx
->init_config
, R_03093C_VGT_HS_OFFCHIP_PARAM
,
3823 sctx
->screen
->vgt_hs_offchip_param
);
3825 si_pm4_set_reg(sctx
->init_config
, R_008988_VGT_TF_RING_SIZE
,
3826 S_008988_SIZE(sctx
->screen
->tess_factor_ring_size
/ 4));
3827 si_pm4_set_reg(sctx
->init_config
, R_0089B8_VGT_TF_MEMORY_BASE
,
3829 si_pm4_set_reg(sctx
->init_config
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
3830 sctx
->screen
->vgt_hs_offchip_param
);
3833 /* Flush the context to re-emit the init_config state.
3834 * This is done only once in a lifetime of a context.
3836 si_pm4_upload_indirect_buffer(sctx
, sctx
->init_config
);
3837 sctx
->initial_gfx_cs_size
= 0; /* force flush */
3838 si_flush_gfx_cs(sctx
, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW
, NULL
);
3841 static struct si_pm4_state
*si_build_vgt_shader_config(struct si_screen
*screen
,
3842 union si_vgt_stages_key key
)
3844 struct si_pm4_state
*pm4
= CALLOC_STRUCT(si_pm4_state
);
3845 uint32_t stages
= 0;
3848 stages
|= S_028B54_LS_EN(V_028B54_LS_STAGE_ON
) |
3849 S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
3852 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_DS
) |
3855 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_DS
);
3857 stages
|= S_028B54_VS_EN(V_028B54_VS_STAGE_DS
);
3858 } else if (key
.u
.gs
) {
3859 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL
) |
3861 } else if (key
.u
.ngg
) {
3862 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL
);
3866 stages
|= S_028B54_PRIMGEN_EN(1) |
3867 S_028B54_NGG_WAVE_ID_EN(key
.u
.streamout
) |
3868 S_028B54_PRIMGEN_PASSTHRU_EN(key
.u
.ngg_passthrough
);
3869 } else if (key
.u
.gs
)
3870 stages
|= S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
3872 if (screen
->info
.chip_class
>= GFX9
)
3873 stages
|= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
3875 if (screen
->info
.chip_class
>= GFX10
&& screen
->ge_wave_size
== 32) {
3876 stages
|= S_028B54_HS_W32_EN(1) |
3877 S_028B54_GS_W32_EN(key
.u
.ngg
) | /* legacy GS only supports Wave64 */
3878 S_028B54_VS_W32_EN(1);
3881 si_pm4_set_reg(pm4
, R_028B54_VGT_SHADER_STAGES_EN
, stages
);
3885 static void si_update_vgt_shader_config(struct si_context
*sctx
,
3886 union si_vgt_stages_key key
)
3888 struct si_pm4_state
**pm4
= &sctx
->vgt_shader_config
[key
.index
];
3890 if (unlikely(!*pm4
))
3891 *pm4
= si_build_vgt_shader_config(sctx
->screen
, key
);
3892 si_pm4_bind_state(sctx
, vgt_shader_config
, *pm4
);
3895 bool si_update_shaders(struct si_context
*sctx
)
3897 struct pipe_context
*ctx
= (struct pipe_context
*)sctx
;
3898 struct si_compiler_ctx_state compiler_state
;
3899 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
3900 struct si_shader
*old_vs
= si_get_vs_state(sctx
);
3901 bool old_clip_disable
= old_vs
? old_vs
->key
.opt
.clip_disable
: false;
3902 struct si_shader
*old_ps
= sctx
->ps_shader
.current
;
3903 union si_vgt_stages_key key
;
3904 unsigned old_spi_shader_col_format
=
3905 old_ps
? old_ps
->key
.part
.ps
.epilog
.spi_shader_col_format
: 0;
3908 if (!sctx
->compiler
.passes
)
3909 si_init_compiler(sctx
->screen
, &sctx
->compiler
);
3911 compiler_state
.compiler
= &sctx
->compiler
;
3912 compiler_state
.debug
= sctx
->debug
;
3913 compiler_state
.is_debug_context
= sctx
->is_debug
;
3917 if (sctx
->tes_shader
.cso
)
3919 if (sctx
->gs_shader
.cso
)
3924 key
.u
.streamout
= !!si_get_vs(sctx
)->cso
->so
.num_outputs
;
3927 /* Update TCS and TES. */
3928 if (sctx
->tes_shader
.cso
) {
3929 if (!sctx
->tess_rings
) {
3930 si_init_tess_factor_ring(sctx
);
3931 if (!sctx
->tess_rings
)
3935 if (sctx
->tcs_shader
.cso
) {
3936 r
= si_shader_select(ctx
, &sctx
->tcs_shader
, key
,
3940 si_pm4_bind_state(sctx
, hs
, sctx
->tcs_shader
.current
->pm4
);
3942 if (!sctx
->fixed_func_tcs_shader
.cso
) {
3943 sctx
->fixed_func_tcs_shader
.cso
=
3944 si_create_fixed_func_tcs(sctx
);
3945 if (!sctx
->fixed_func_tcs_shader
.cso
)
3949 r
= si_shader_select(ctx
, &sctx
->fixed_func_tcs_shader
,
3950 key
, &compiler_state
);
3953 si_pm4_bind_state(sctx
, hs
,
3954 sctx
->fixed_func_tcs_shader
.current
->pm4
);
3957 if (!sctx
->gs_shader
.cso
|| sctx
->chip_class
<= GFX8
) {
3958 r
= si_shader_select(ctx
, &sctx
->tes_shader
, key
, &compiler_state
);
3962 if (sctx
->gs_shader
.cso
) {
3964 assert(sctx
->chip_class
<= GFX8
);
3965 si_pm4_bind_state(sctx
, es
, sctx
->tes_shader
.current
->pm4
);
3966 } else if (key
.u
.ngg
) {
3967 si_pm4_bind_state(sctx
, gs
, sctx
->tes_shader
.current
->pm4
);
3969 si_pm4_bind_state(sctx
, vs
, sctx
->tes_shader
.current
->pm4
);
3973 if (sctx
->chip_class
<= GFX8
)
3974 si_pm4_bind_state(sctx
, ls
, NULL
);
3975 si_pm4_bind_state(sctx
, hs
, NULL
);
3979 if (sctx
->gs_shader
.cso
) {
3980 r
= si_shader_select(ctx
, &sctx
->gs_shader
, key
, &compiler_state
);
3983 si_pm4_bind_state(sctx
, gs
, sctx
->gs_shader
.current
->pm4
);
3985 si_pm4_bind_state(sctx
, vs
, sctx
->gs_shader
.cso
->gs_copy_shader
->pm4
);
3987 if (!si_update_gs_ring_buffers(sctx
))
3990 si_pm4_bind_state(sctx
, vs
, NULL
);
3994 si_pm4_bind_state(sctx
, gs
, NULL
);
3995 if (sctx
->chip_class
<= GFX8
)
3996 si_pm4_bind_state(sctx
, es
, NULL
);
4001 if ((!key
.u
.tess
&& !key
.u
.gs
) || sctx
->chip_class
<= GFX8
) {
4002 r
= si_shader_select(ctx
, &sctx
->vs_shader
, key
, &compiler_state
);
4006 if (!key
.u
.tess
&& !key
.u
.gs
) {
4008 si_pm4_bind_state(sctx
, gs
, sctx
->vs_shader
.current
->pm4
);
4009 si_pm4_bind_state(sctx
, vs
, NULL
);
4011 si_pm4_bind_state(sctx
, vs
, sctx
->vs_shader
.current
->pm4
);
4013 } else if (sctx
->tes_shader
.cso
) {
4014 si_pm4_bind_state(sctx
, ls
, sctx
->vs_shader
.current
->pm4
);
4016 assert(sctx
->gs_shader
.cso
);
4017 si_pm4_bind_state(sctx
, es
, sctx
->vs_shader
.current
->pm4
);
4021 /* This must be done after the shader variant is selected. */
4023 key
.u
.ngg_passthrough
= gfx10_is_ngg_passthrough(si_get_vs(sctx
)->current
);
4025 si_update_vgt_shader_config(sctx
, key
);
4027 if (old_clip_disable
!= si_get_vs_state(sctx
)->key
.opt
.clip_disable
)
4028 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.clip_regs
);
4030 if (sctx
->ps_shader
.cso
) {
4031 unsigned db_shader_control
;
4033 r
= si_shader_select(ctx
, &sctx
->ps_shader
, key
, &compiler_state
);
4036 si_pm4_bind_state(sctx
, ps
, sctx
->ps_shader
.current
->pm4
);
4039 sctx
->ps_shader
.cso
->db_shader_control
|
4040 S_02880C_KILL_ENABLE(si_get_alpha_test_func(sctx
) != PIPE_FUNC_ALWAYS
);
4042 if (si_pm4_state_changed(sctx
, ps
) ||
4043 si_pm4_state_changed(sctx
, vs
) ||
4044 (key
.u
.ngg
&& si_pm4_state_changed(sctx
, gs
)) ||
4045 sctx
->sprite_coord_enable
!= rs
->sprite_coord_enable
||
4046 sctx
->flatshade
!= rs
->flatshade
) {
4047 sctx
->sprite_coord_enable
= rs
->sprite_coord_enable
;
4048 sctx
->flatshade
= rs
->flatshade
;
4049 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.spi_map
);
4052 if (sctx
->screen
->info
.rbplus_allowed
&&
4053 si_pm4_state_changed(sctx
, ps
) &&
4055 old_spi_shader_col_format
!=
4056 sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.spi_shader_col_format
))
4057 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.cb_render_state
);
4059 if (sctx
->ps_db_shader_control
!= db_shader_control
) {
4060 sctx
->ps_db_shader_control
= db_shader_control
;
4061 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.db_render_state
);
4062 if (sctx
->screen
->dpbb_allowed
)
4063 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.dpbb_state
);
4066 if (sctx
->smoothing_enabled
!= sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.poly_line_smoothing
) {
4067 sctx
->smoothing_enabled
= sctx
->ps_shader
.current
->key
.part
.ps
.epilog
.poly_line_smoothing
;
4068 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.msaa_config
);
4070 if (sctx
->chip_class
== GFX6
)
4071 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.db_render_state
);
4073 if (sctx
->framebuffer
.nr_samples
<= 1)
4074 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.msaa_sample_locs
);
4078 if (si_pm4_state_enabled_and_changed(sctx
, ls
) ||
4079 si_pm4_state_enabled_and_changed(sctx
, hs
) ||
4080 si_pm4_state_enabled_and_changed(sctx
, es
) ||
4081 si_pm4_state_enabled_and_changed(sctx
, gs
) ||
4082 si_pm4_state_enabled_and_changed(sctx
, vs
) ||
4083 si_pm4_state_enabled_and_changed(sctx
, ps
)) {
4084 if (!si_update_spi_tmpring_size(sctx
))
4088 if (sctx
->chip_class
>= GFX7
) {
4089 if (si_pm4_state_enabled_and_changed(sctx
, ls
))
4090 sctx
->prefetch_L2_mask
|= SI_PREFETCH_LS
;
4091 else if (!sctx
->queued
.named
.ls
)
4092 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_LS
;
4094 if (si_pm4_state_enabled_and_changed(sctx
, hs
))
4095 sctx
->prefetch_L2_mask
|= SI_PREFETCH_HS
;
4096 else if (!sctx
->queued
.named
.hs
)
4097 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_HS
;
4099 if (si_pm4_state_enabled_and_changed(sctx
, es
))
4100 sctx
->prefetch_L2_mask
|= SI_PREFETCH_ES
;
4101 else if (!sctx
->queued
.named
.es
)
4102 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_ES
;
4104 if (si_pm4_state_enabled_and_changed(sctx
, gs
))
4105 sctx
->prefetch_L2_mask
|= SI_PREFETCH_GS
;
4106 else if (!sctx
->queued
.named
.gs
)
4107 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_GS
;
4109 if (si_pm4_state_enabled_and_changed(sctx
, vs
))
4110 sctx
->prefetch_L2_mask
|= SI_PREFETCH_VS
;
4111 else if (!sctx
->queued
.named
.vs
)
4112 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_VS
;
4114 if (si_pm4_state_enabled_and_changed(sctx
, ps
))
4115 sctx
->prefetch_L2_mask
|= SI_PREFETCH_PS
;
4116 else if (!sctx
->queued
.named
.ps
)
4117 sctx
->prefetch_L2_mask
&= ~SI_PREFETCH_PS
;
4120 sctx
->do_update_shaders
= false;
4124 static void si_emit_scratch_state(struct si_context
*sctx
)
4126 struct radeon_cmdbuf
*cs
= sctx
->gfx_cs
;
4128 radeon_set_context_reg(cs
, R_0286E8_SPI_TMPRING_SIZE
,
4129 sctx
->spi_tmpring_size
);
4131 if (sctx
->scratch_buffer
) {
4132 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
4133 sctx
->scratch_buffer
, RADEON_USAGE_READWRITE
,
4134 RADEON_PRIO_SCRATCH_BUFFER
);
4138 void si_init_shader_functions(struct si_context
*sctx
)
4140 sctx
->atoms
.s
.spi_map
.emit
= si_emit_spi_map
;
4141 sctx
->atoms
.s
.scratch_state
.emit
= si_emit_scratch_state
;
4143 sctx
->b
.create_vs_state
= si_create_shader_selector
;
4144 sctx
->b
.create_tcs_state
= si_create_shader_selector
;
4145 sctx
->b
.create_tes_state
= si_create_shader_selector
;
4146 sctx
->b
.create_gs_state
= si_create_shader_selector
;
4147 sctx
->b
.create_fs_state
= si_create_shader_selector
;
4149 sctx
->b
.bind_vs_state
= si_bind_vs_shader
;
4150 sctx
->b
.bind_tcs_state
= si_bind_tcs_shader
;
4151 sctx
->b
.bind_tes_state
= si_bind_tes_shader
;
4152 sctx
->b
.bind_gs_state
= si_bind_gs_shader
;
4153 sctx
->b
.bind_fs_state
= si_bind_ps_shader
;
4155 sctx
->b
.delete_vs_state
= si_delete_shader_selector
;
4156 sctx
->b
.delete_tcs_state
= si_delete_shader_selector
;
4157 sctx
->b
.delete_tes_state
= si_delete_shader_selector
;
4158 sctx
->b
.delete_gs_state
= si_delete_shader_selector
;
4159 sctx
->b
.delete_fs_state
= si_delete_shader_selector
;