2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include "util/mesa-sha1.h"
29 #include "util/u_atomic.h"
30 #include "radv_debug.h"
31 #include "radv_private.h"
33 #include "radv_shader.h"
35 #include "nir/nir_builder.h"
36 #include "spirv/nir_spirv.h"
39 #include <llvm-c/Core.h>
40 #include <llvm-c/TargetMachine.h>
44 #include "ac_binary.h"
45 #include "ac_llvm_util.h"
46 #include "ac_nir_to_llvm.h"
47 #include "vk_format.h"
48 #include "util/debug.h"
49 #include "ac_exp_param.h"
50 #include "ac_shader_util.h"
51 #include "main/menums.h"
53 struct radv_blend_state
{
54 uint32_t blend_enable_4bit
;
55 uint32_t need_src_alpha
;
57 uint32_t cb_color_control
;
58 uint32_t cb_target_mask
;
59 uint32_t cb_target_enabled_4bit
;
60 uint32_t sx_mrt_blend_opt
[8];
61 uint32_t cb_blend_control
[8];
63 uint32_t spi_shader_col_format
;
64 uint32_t cb_shader_mask
;
65 uint32_t db_alpha_to_mask
;
67 uint32_t commutative_4bit
;
69 bool single_cb_enable
;
70 bool mrt0_is_dual_src
;
73 struct radv_dsa_order_invariance
{
74 /* Whether the final result in Z/S buffers is guaranteed to be
75 * invariant under changes to the order in which fragments arrive.
79 /* Whether the set of fragments that pass the combined Z/S test is
80 * guaranteed to be invariant under changes to the order in which
86 struct radv_tessellation_state
{
87 uint32_t ls_hs_config
;
93 struct radv_gs_state
{
94 uint32_t vgt_gs_onchip_cntl
;
95 uint32_t vgt_gs_max_prims_per_subgroup
;
96 uint32_t vgt_esgs_ring_itemsize
;
101 radv_pipeline_destroy(struct radv_device
*device
,
102 struct radv_pipeline
*pipeline
,
103 const VkAllocationCallbacks
* allocator
)
105 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; ++i
)
106 if (pipeline
->shaders
[i
])
107 radv_shader_variant_destroy(device
, pipeline
->shaders
[i
]);
109 if (pipeline
->gs_copy_shader
)
110 radv_shader_variant_destroy(device
, pipeline
->gs_copy_shader
);
113 free(pipeline
->cs
.buf
);
114 vk_free2(&device
->alloc
, allocator
, pipeline
);
117 void radv_DestroyPipeline(
119 VkPipeline _pipeline
,
120 const VkAllocationCallbacks
* pAllocator
)
122 RADV_FROM_HANDLE(radv_device
, device
, _device
);
123 RADV_FROM_HANDLE(radv_pipeline
, pipeline
, _pipeline
);
128 radv_pipeline_destroy(device
, pipeline
, pAllocator
);
131 static uint32_t get_hash_flags(struct radv_device
*device
)
133 uint32_t hash_flags
= 0;
135 if (device
->instance
->debug_flags
& RADV_DEBUG_UNSAFE_MATH
)
136 hash_flags
|= RADV_HASH_SHADER_UNSAFE_MATH
;
137 if (device
->instance
->perftest_flags
& RADV_PERFTEST_SISCHED
)
138 hash_flags
|= RADV_HASH_SHADER_SISCHED
;
143 radv_pipeline_scratch_init(struct radv_device
*device
,
144 struct radv_pipeline
*pipeline
)
146 unsigned scratch_bytes_per_wave
= 0;
147 unsigned max_waves
= 0;
148 unsigned min_waves
= 1;
150 for (int i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
151 if (pipeline
->shaders
[i
]) {
152 unsigned max_stage_waves
= device
->scratch_waves
;
154 scratch_bytes_per_wave
= MAX2(scratch_bytes_per_wave
,
155 pipeline
->shaders
[i
]->config
.scratch_bytes_per_wave
);
157 max_stage_waves
= MIN2(max_stage_waves
,
158 4 * device
->physical_device
->rad_info
.num_good_compute_units
*
159 (256 / pipeline
->shaders
[i
]->config
.num_vgprs
));
160 max_waves
= MAX2(max_waves
, max_stage_waves
);
164 if (pipeline
->shaders
[MESA_SHADER_COMPUTE
]) {
165 unsigned group_size
= pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.cs
.block_size
[0] *
166 pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.cs
.block_size
[1] *
167 pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.cs
.block_size
[2];
168 min_waves
= MAX2(min_waves
, round_up_u32(group_size
, 64));
171 if (scratch_bytes_per_wave
)
172 max_waves
= MIN2(max_waves
, 0xffffffffu
/ scratch_bytes_per_wave
);
174 if (scratch_bytes_per_wave
&& max_waves
< min_waves
) {
175 /* Not really true at this moment, but will be true on first
176 * execution. Avoid having hanging shaders. */
177 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
179 pipeline
->scratch_bytes_per_wave
= scratch_bytes_per_wave
;
180 pipeline
->max_waves
= max_waves
;
184 static uint32_t si_translate_blend_function(VkBlendOp op
)
187 case VK_BLEND_OP_ADD
:
188 return V_028780_COMB_DST_PLUS_SRC
;
189 case VK_BLEND_OP_SUBTRACT
:
190 return V_028780_COMB_SRC_MINUS_DST
;
191 case VK_BLEND_OP_REVERSE_SUBTRACT
:
192 return V_028780_COMB_DST_MINUS_SRC
;
193 case VK_BLEND_OP_MIN
:
194 return V_028780_COMB_MIN_DST_SRC
;
195 case VK_BLEND_OP_MAX
:
196 return V_028780_COMB_MAX_DST_SRC
;
202 static uint32_t si_translate_blend_factor(VkBlendFactor factor
)
205 case VK_BLEND_FACTOR_ZERO
:
206 return V_028780_BLEND_ZERO
;
207 case VK_BLEND_FACTOR_ONE
:
208 return V_028780_BLEND_ONE
;
209 case VK_BLEND_FACTOR_SRC_COLOR
:
210 return V_028780_BLEND_SRC_COLOR
;
211 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR
:
212 return V_028780_BLEND_ONE_MINUS_SRC_COLOR
;
213 case VK_BLEND_FACTOR_DST_COLOR
:
214 return V_028780_BLEND_DST_COLOR
;
215 case VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR
:
216 return V_028780_BLEND_ONE_MINUS_DST_COLOR
;
217 case VK_BLEND_FACTOR_SRC_ALPHA
:
218 return V_028780_BLEND_SRC_ALPHA
;
219 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
:
220 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA
;
221 case VK_BLEND_FACTOR_DST_ALPHA
:
222 return V_028780_BLEND_DST_ALPHA
;
223 case VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA
:
224 return V_028780_BLEND_ONE_MINUS_DST_ALPHA
;
225 case VK_BLEND_FACTOR_CONSTANT_COLOR
:
226 return V_028780_BLEND_CONSTANT_COLOR
;
227 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR
:
228 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR
;
229 case VK_BLEND_FACTOR_CONSTANT_ALPHA
:
230 return V_028780_BLEND_CONSTANT_ALPHA
;
231 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA
:
232 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA
;
233 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
:
234 return V_028780_BLEND_SRC_ALPHA_SATURATE
;
235 case VK_BLEND_FACTOR_SRC1_COLOR
:
236 return V_028780_BLEND_SRC1_COLOR
;
237 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
:
238 return V_028780_BLEND_INV_SRC1_COLOR
;
239 case VK_BLEND_FACTOR_SRC1_ALPHA
:
240 return V_028780_BLEND_SRC1_ALPHA
;
241 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
:
242 return V_028780_BLEND_INV_SRC1_ALPHA
;
248 static uint32_t si_translate_blend_opt_function(VkBlendOp op
)
251 case VK_BLEND_OP_ADD
:
252 return V_028760_OPT_COMB_ADD
;
253 case VK_BLEND_OP_SUBTRACT
:
254 return V_028760_OPT_COMB_SUBTRACT
;
255 case VK_BLEND_OP_REVERSE_SUBTRACT
:
256 return V_028760_OPT_COMB_REVSUBTRACT
;
257 case VK_BLEND_OP_MIN
:
258 return V_028760_OPT_COMB_MIN
;
259 case VK_BLEND_OP_MAX
:
260 return V_028760_OPT_COMB_MAX
;
262 return V_028760_OPT_COMB_BLEND_DISABLED
;
266 static uint32_t si_translate_blend_opt_factor(VkBlendFactor factor
, bool is_alpha
)
269 case VK_BLEND_FACTOR_ZERO
:
270 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL
;
271 case VK_BLEND_FACTOR_ONE
:
272 return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
;
273 case VK_BLEND_FACTOR_SRC_COLOR
:
274 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
275 : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0
;
276 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR
:
277 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
278 : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1
;
279 case VK_BLEND_FACTOR_SRC_ALPHA
:
280 return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
;
281 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
:
282 return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
;
283 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
:
284 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
285 : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0
;
287 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
292 * Get rid of DST in the blend factors by commuting the operands:
293 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
295 static void si_blend_remove_dst(unsigned *func
, unsigned *src_factor
,
296 unsigned *dst_factor
, unsigned expected_dst
,
297 unsigned replacement_src
)
299 if (*src_factor
== expected_dst
&&
300 *dst_factor
== VK_BLEND_FACTOR_ZERO
) {
301 *src_factor
= VK_BLEND_FACTOR_ZERO
;
302 *dst_factor
= replacement_src
;
304 /* Commuting the operands requires reversing subtractions. */
305 if (*func
== VK_BLEND_OP_SUBTRACT
)
306 *func
= VK_BLEND_OP_REVERSE_SUBTRACT
;
307 else if (*func
== VK_BLEND_OP_REVERSE_SUBTRACT
)
308 *func
= VK_BLEND_OP_SUBTRACT
;
312 static bool si_blend_factor_uses_dst(unsigned factor
)
314 return factor
== VK_BLEND_FACTOR_DST_COLOR
||
315 factor
== VK_BLEND_FACTOR_DST_ALPHA
||
316 factor
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
317 factor
== VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA
||
318 factor
== VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR
;
321 static bool is_dual_src(VkBlendFactor factor
)
324 case VK_BLEND_FACTOR_SRC1_COLOR
:
325 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
:
326 case VK_BLEND_FACTOR_SRC1_ALPHA
:
327 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
:
334 static unsigned si_choose_spi_color_format(VkFormat vk_format
,
336 bool blend_need_alpha
)
338 const struct vk_format_description
*desc
= vk_format_description(vk_format
);
339 unsigned format
, ntype
, swap
;
341 /* Alpha is needed for alpha-to-coverage.
342 * Blending may be with or without alpha.
344 unsigned normal
= 0; /* most optimal, may not support blending or export alpha */
345 unsigned alpha
= 0; /* exports alpha, but may not support blending */
346 unsigned blend
= 0; /* supports blending, but may not export alpha */
347 unsigned blend_alpha
= 0; /* least optimal, supports blending and exports alpha */
349 format
= radv_translate_colorformat(vk_format
);
350 ntype
= radv_translate_color_numformat(vk_format
, desc
,
351 vk_format_get_first_non_void_channel(vk_format
));
352 swap
= radv_translate_colorswap(vk_format
, false);
354 /* Choose the SPI color formats. These are required values for Stoney/RB+.
355 * Other chips have multiple choices, though they are not necessarily better.
358 case V_028C70_COLOR_5_6_5
:
359 case V_028C70_COLOR_1_5_5_5
:
360 case V_028C70_COLOR_5_5_5_1
:
361 case V_028C70_COLOR_4_4_4_4
:
362 case V_028C70_COLOR_10_11_11
:
363 case V_028C70_COLOR_11_11_10
:
364 case V_028C70_COLOR_8
:
365 case V_028C70_COLOR_8_8
:
366 case V_028C70_COLOR_8_8_8_8
:
367 case V_028C70_COLOR_10_10_10_2
:
368 case V_028C70_COLOR_2_10_10_10
:
369 if (ntype
== V_028C70_NUMBER_UINT
)
370 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
371 else if (ntype
== V_028C70_NUMBER_SINT
)
372 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
374 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
377 case V_028C70_COLOR_16
:
378 case V_028C70_COLOR_16_16
:
379 case V_028C70_COLOR_16_16_16_16
:
380 if (ntype
== V_028C70_NUMBER_UNORM
||
381 ntype
== V_028C70_NUMBER_SNORM
) {
382 /* UNORM16 and SNORM16 don't support blending */
383 if (ntype
== V_028C70_NUMBER_UNORM
)
384 normal
= alpha
= V_028714_SPI_SHADER_UNORM16_ABGR
;
386 normal
= alpha
= V_028714_SPI_SHADER_SNORM16_ABGR
;
388 /* Use 32 bits per channel for blending. */
389 if (format
== V_028C70_COLOR_16
) {
390 if (swap
== V_028C70_SWAP_STD
) { /* R */
391 blend
= V_028714_SPI_SHADER_32_R
;
392 blend_alpha
= V_028714_SPI_SHADER_32_AR
;
393 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
394 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
397 } else if (format
== V_028C70_COLOR_16_16
) {
398 if (swap
== V_028C70_SWAP_STD
) { /* RG */
399 blend
= V_028714_SPI_SHADER_32_GR
;
400 blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
401 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
402 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
405 } else /* 16_16_16_16 */
406 blend
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
407 } else if (ntype
== V_028C70_NUMBER_UINT
)
408 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
409 else if (ntype
== V_028C70_NUMBER_SINT
)
410 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
411 else if (ntype
== V_028C70_NUMBER_FLOAT
)
412 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
417 case V_028C70_COLOR_32
:
418 if (swap
== V_028C70_SWAP_STD
) { /* R */
419 blend
= normal
= V_028714_SPI_SHADER_32_R
;
420 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
421 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
422 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
427 case V_028C70_COLOR_32_32
:
428 if (swap
== V_028C70_SWAP_STD
) { /* RG */
429 blend
= normal
= V_028714_SPI_SHADER_32_GR
;
430 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
431 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
432 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
437 case V_028C70_COLOR_32_32_32_32
:
438 case V_028C70_COLOR_8_24
:
439 case V_028C70_COLOR_24_8
:
440 case V_028C70_COLOR_X24_8_32_FLOAT
:
441 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_ABGR
;
445 unreachable("unhandled blend format");
448 if (blend_enable
&& blend_need_alpha
)
450 else if(blend_need_alpha
)
452 else if(blend_enable
)
459 radv_pipeline_compute_spi_color_formats(struct radv_pipeline
*pipeline
,
460 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
461 struct radv_blend_state
*blend
)
463 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
464 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
465 unsigned col_format
= 0;
467 for (unsigned i
= 0; i
< (blend
->single_cb_enable
? 1 : subpass
->color_count
); ++i
) {
470 if (subpass
->color_attachments
[i
].attachment
== VK_ATTACHMENT_UNUSED
) {
471 cf
= V_028714_SPI_SHADER_ZERO
;
473 struct radv_render_pass_attachment
*attachment
= pass
->attachments
+ subpass
->color_attachments
[i
].attachment
;
475 blend
->blend_enable_4bit
& (0xfu
<< (i
* 4));
477 cf
= si_choose_spi_color_format(attachment
->format
,
479 blend
->need_src_alpha
& (1 << i
));
482 col_format
|= cf
<< (4 * i
);
485 blend
->cb_shader_mask
= ac_get_cb_shader_mask(col_format
);
487 if (blend
->mrt0_is_dual_src
)
488 col_format
|= (col_format
& 0xf) << 4;
489 blend
->spi_shader_col_format
= col_format
;
493 format_is_int8(VkFormat format
)
495 const struct vk_format_description
*desc
= vk_format_description(format
);
496 int channel
= vk_format_get_first_non_void_channel(format
);
498 return channel
>= 0 && desc
->channel
[channel
].pure_integer
&&
499 desc
->channel
[channel
].size
== 8;
503 format_is_int10(VkFormat format
)
505 const struct vk_format_description
*desc
= vk_format_description(format
);
507 if (desc
->nr_channels
!= 4)
509 for (unsigned i
= 0; i
< 4; i
++) {
510 if (desc
->channel
[i
].pure_integer
&& desc
->channel
[i
].size
== 10)
516 unsigned radv_format_meta_fs_key(VkFormat format
)
518 unsigned col_format
= si_choose_spi_color_format(format
, false, false) - 1;
519 bool is_int8
= format_is_int8(format
);
520 bool is_int10
= format_is_int10(format
);
522 return col_format
+ (is_int8
? 3 : is_int10
? 5 : 0);
526 radv_pipeline_compute_get_int_clamp(const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
527 unsigned *is_int8
, unsigned *is_int10
)
529 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
530 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
534 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
535 struct radv_render_pass_attachment
*attachment
;
537 if (subpass
->color_attachments
[i
].attachment
== VK_ATTACHMENT_UNUSED
)
540 attachment
= pass
->attachments
+ subpass
->color_attachments
[i
].attachment
;
542 if (format_is_int8(attachment
->format
))
544 if (format_is_int10(attachment
->format
))
550 radv_blend_check_commutativity(struct radv_blend_state
*blend
,
551 VkBlendOp op
, VkBlendFactor src
,
552 VkBlendFactor dst
, unsigned chanmask
)
554 /* Src factor is allowed when it does not depend on Dst. */
555 static const uint32_t src_allowed
=
556 (1u << VK_BLEND_FACTOR_ONE
) |
557 (1u << VK_BLEND_FACTOR_SRC_COLOR
) |
558 (1u << VK_BLEND_FACTOR_SRC_ALPHA
) |
559 (1u << VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
) |
560 (1u << VK_BLEND_FACTOR_CONSTANT_COLOR
) |
561 (1u << VK_BLEND_FACTOR_CONSTANT_ALPHA
) |
562 (1u << VK_BLEND_FACTOR_SRC1_COLOR
) |
563 (1u << VK_BLEND_FACTOR_SRC1_ALPHA
) |
564 (1u << VK_BLEND_FACTOR_ZERO
) |
565 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR
) |
566 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
) |
567 (1u << VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR
) |
568 (1u << VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA
) |
569 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
) |
570 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
);
572 if (dst
== VK_BLEND_FACTOR_ONE
&&
573 (src_allowed
&& (1u << src
))) {
574 /* Addition is commutative, but floating point addition isn't
575 * associative: subtle changes can be introduced via different
576 * rounding. Be conservative, only enable for min and max.
578 if (op
== VK_BLEND_OP_MAX
|| op
== VK_BLEND_OP_MIN
)
579 blend
->commutative_4bit
|= chanmask
;
583 static struct radv_blend_state
584 radv_pipeline_init_blend_state(struct radv_pipeline
*pipeline
,
585 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
586 const struct radv_graphics_pipeline_create_info
*extra
)
588 const VkPipelineColorBlendStateCreateInfo
*vkblend
= pCreateInfo
->pColorBlendState
;
589 const VkPipelineMultisampleStateCreateInfo
*vkms
= pCreateInfo
->pMultisampleState
;
590 struct radv_blend_state blend
= {0};
591 unsigned mode
= V_028808_CB_NORMAL
;
597 if (extra
&& extra
->custom_blend_mode
) {
598 blend
.single_cb_enable
= true;
599 mode
= extra
->custom_blend_mode
;
601 blend
.cb_color_control
= 0;
602 if (vkblend
->logicOpEnable
)
603 blend
.cb_color_control
|= S_028808_ROP3(vkblend
->logicOp
| (vkblend
->logicOp
<< 4));
605 blend
.cb_color_control
|= S_028808_ROP3(0xcc);
607 blend
.db_alpha_to_mask
= S_028B70_ALPHA_TO_MASK_OFFSET0(2) |
608 S_028B70_ALPHA_TO_MASK_OFFSET1(2) |
609 S_028B70_ALPHA_TO_MASK_OFFSET2(2) |
610 S_028B70_ALPHA_TO_MASK_OFFSET3(2);
612 if (vkms
&& vkms
->alphaToCoverageEnable
) {
613 blend
.db_alpha_to_mask
|= S_028B70_ALPHA_TO_MASK_ENABLE(1);
616 blend
.cb_target_mask
= 0;
617 for (i
= 0; i
< vkblend
->attachmentCount
; i
++) {
618 const VkPipelineColorBlendAttachmentState
*att
= &vkblend
->pAttachments
[i
];
619 unsigned blend_cntl
= 0;
620 unsigned srcRGB_opt
, dstRGB_opt
, srcA_opt
, dstA_opt
;
621 VkBlendOp eqRGB
= att
->colorBlendOp
;
622 VkBlendFactor srcRGB
= att
->srcColorBlendFactor
;
623 VkBlendFactor dstRGB
= att
->dstColorBlendFactor
;
624 VkBlendOp eqA
= att
->alphaBlendOp
;
625 VkBlendFactor srcA
= att
->srcAlphaBlendFactor
;
626 VkBlendFactor dstA
= att
->dstAlphaBlendFactor
;
628 blend
.sx_mrt_blend_opt
[i
] = S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED
) | S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED
);
630 if (!att
->colorWriteMask
)
633 blend
.cb_target_mask
|= (unsigned)att
->colorWriteMask
<< (4 * i
);
634 blend
.cb_target_enabled_4bit
|= 0xf << (4 * i
);
635 if (!att
->blendEnable
) {
636 blend
.cb_blend_control
[i
] = blend_cntl
;
640 if (is_dual_src(srcRGB
) || is_dual_src(dstRGB
) || is_dual_src(srcA
) || is_dual_src(dstA
))
642 blend
.mrt0_is_dual_src
= true;
644 if (eqRGB
== VK_BLEND_OP_MIN
|| eqRGB
== VK_BLEND_OP_MAX
) {
645 srcRGB
= VK_BLEND_FACTOR_ONE
;
646 dstRGB
= VK_BLEND_FACTOR_ONE
;
648 if (eqA
== VK_BLEND_OP_MIN
|| eqA
== VK_BLEND_OP_MAX
) {
649 srcA
= VK_BLEND_FACTOR_ONE
;
650 dstA
= VK_BLEND_FACTOR_ONE
;
653 radv_blend_check_commutativity(&blend
, eqRGB
, srcRGB
, dstRGB
,
655 radv_blend_check_commutativity(&blend
, eqA
, srcA
, dstA
,
658 /* Blending optimizations for RB+.
659 * These transformations don't change the behavior.
661 * First, get rid of DST in the blend factors:
662 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
664 si_blend_remove_dst(&eqRGB
, &srcRGB
, &dstRGB
,
665 VK_BLEND_FACTOR_DST_COLOR
,
666 VK_BLEND_FACTOR_SRC_COLOR
);
668 si_blend_remove_dst(&eqA
, &srcA
, &dstA
,
669 VK_BLEND_FACTOR_DST_COLOR
,
670 VK_BLEND_FACTOR_SRC_COLOR
);
672 si_blend_remove_dst(&eqA
, &srcA
, &dstA
,
673 VK_BLEND_FACTOR_DST_ALPHA
,
674 VK_BLEND_FACTOR_SRC_ALPHA
);
676 /* Look up the ideal settings from tables. */
677 srcRGB_opt
= si_translate_blend_opt_factor(srcRGB
, false);
678 dstRGB_opt
= si_translate_blend_opt_factor(dstRGB
, false);
679 srcA_opt
= si_translate_blend_opt_factor(srcA
, true);
680 dstA_opt
= si_translate_blend_opt_factor(dstA
, true);
682 /* Handle interdependencies. */
683 if (si_blend_factor_uses_dst(srcRGB
))
684 dstRGB_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
685 if (si_blend_factor_uses_dst(srcA
))
686 dstA_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
688 if (srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
&&
689 (dstRGB
== VK_BLEND_FACTOR_ZERO
||
690 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
691 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
))
692 dstRGB_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0
;
694 /* Set the final value. */
695 blend
.sx_mrt_blend_opt
[i
] =
696 S_028760_COLOR_SRC_OPT(srcRGB_opt
) |
697 S_028760_COLOR_DST_OPT(dstRGB_opt
) |
698 S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB
)) |
699 S_028760_ALPHA_SRC_OPT(srcA_opt
) |
700 S_028760_ALPHA_DST_OPT(dstA_opt
) |
701 S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA
));
702 blend_cntl
|= S_028780_ENABLE(1);
704 blend_cntl
|= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB
));
705 blend_cntl
|= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB
));
706 blend_cntl
|= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB
));
707 if (srcA
!= srcRGB
|| dstA
!= dstRGB
|| eqA
!= eqRGB
) {
708 blend_cntl
|= S_028780_SEPARATE_ALPHA_BLEND(1);
709 blend_cntl
|= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA
));
710 blend_cntl
|= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA
));
711 blend_cntl
|= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA
));
713 blend
.cb_blend_control
[i
] = blend_cntl
;
715 blend
.blend_enable_4bit
|= 0xfu
<< (i
* 4);
717 if (srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
718 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
719 srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
720 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
721 srcRGB
== VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
||
722 dstRGB
== VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
)
723 blend
.need_src_alpha
|= 1 << i
;
725 for (i
= vkblend
->attachmentCount
; i
< 8; i
++) {
726 blend
.cb_blend_control
[i
] = 0;
727 blend
.sx_mrt_blend_opt
[i
] = S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED
) | S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED
);
730 if (pipeline
->device
->physical_device
->has_rbplus
) {
731 /* Disable RB+ blend optimizations for dual source blending. */
732 if (blend
.mrt0_is_dual_src
) {
733 for (i
= 0; i
< 8; i
++) {
734 blend
.sx_mrt_blend_opt
[i
] =
735 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_NONE
) |
736 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_NONE
);
740 /* RB+ doesn't work with dual source blending, logic op and
743 if (blend
.mrt0_is_dual_src
|| vkblend
->logicOpEnable
||
744 mode
== V_028808_CB_RESOLVE
)
745 blend
.cb_color_control
|= S_028808_DISABLE_DUAL_QUAD(1);
748 if (blend
.cb_target_mask
)
749 blend
.cb_color_control
|= S_028808_MODE(mode
);
751 blend
.cb_color_control
|= S_028808_MODE(V_028808_CB_DISABLE
);
753 radv_pipeline_compute_spi_color_formats(pipeline
, pCreateInfo
, &blend
);
757 static uint32_t si_translate_stencil_op(enum VkStencilOp op
)
760 case VK_STENCIL_OP_KEEP
:
761 return V_02842C_STENCIL_KEEP
;
762 case VK_STENCIL_OP_ZERO
:
763 return V_02842C_STENCIL_ZERO
;
764 case VK_STENCIL_OP_REPLACE
:
765 return V_02842C_STENCIL_REPLACE_TEST
;
766 case VK_STENCIL_OP_INCREMENT_AND_CLAMP
:
767 return V_02842C_STENCIL_ADD_CLAMP
;
768 case VK_STENCIL_OP_DECREMENT_AND_CLAMP
:
769 return V_02842C_STENCIL_SUB_CLAMP
;
770 case VK_STENCIL_OP_INVERT
:
771 return V_02842C_STENCIL_INVERT
;
772 case VK_STENCIL_OP_INCREMENT_AND_WRAP
:
773 return V_02842C_STENCIL_ADD_WRAP
;
774 case VK_STENCIL_OP_DECREMENT_AND_WRAP
:
775 return V_02842C_STENCIL_SUB_WRAP
;
781 static uint32_t si_translate_fill(VkPolygonMode func
)
784 case VK_POLYGON_MODE_FILL
:
785 return V_028814_X_DRAW_TRIANGLES
;
786 case VK_POLYGON_MODE_LINE
:
787 return V_028814_X_DRAW_LINES
;
788 case VK_POLYGON_MODE_POINT
:
789 return V_028814_X_DRAW_POINTS
;
792 return V_028814_X_DRAW_POINTS
;
796 static uint8_t radv_pipeline_get_ps_iter_samples(const VkPipelineMultisampleStateCreateInfo
*vkms
)
798 uint32_t num_samples
= vkms
->rasterizationSamples
;
799 uint32_t ps_iter_samples
= 1;
801 if (vkms
->sampleShadingEnable
) {
802 ps_iter_samples
= ceil(vkms
->minSampleShading
* num_samples
);
803 ps_iter_samples
= util_next_power_of_two(ps_iter_samples
);
805 return ps_iter_samples
;
809 radv_is_depth_write_enabled(const VkPipelineDepthStencilStateCreateInfo
*pCreateInfo
)
811 return pCreateInfo
->depthTestEnable
&&
812 pCreateInfo
->depthWriteEnable
&&
813 pCreateInfo
->depthCompareOp
!= VK_COMPARE_OP_NEVER
;
817 radv_writes_stencil(const VkStencilOpState
*state
)
819 return state
->writeMask
&&
820 (state
->failOp
!= VK_STENCIL_OP_KEEP
||
821 state
->passOp
!= VK_STENCIL_OP_KEEP
||
822 state
->depthFailOp
!= VK_STENCIL_OP_KEEP
);
826 radv_is_stencil_write_enabled(const VkPipelineDepthStencilStateCreateInfo
*pCreateInfo
)
828 return pCreateInfo
->stencilTestEnable
&&
829 (radv_writes_stencil(&pCreateInfo
->front
) ||
830 radv_writes_stencil(&pCreateInfo
->back
));
834 radv_is_ds_write_enabled(const VkPipelineDepthStencilStateCreateInfo
*pCreateInfo
)
836 return radv_is_depth_write_enabled(pCreateInfo
) ||
837 radv_is_stencil_write_enabled(pCreateInfo
);
841 radv_order_invariant_stencil_op(VkStencilOp op
)
843 /* REPLACE is normally order invariant, except when the stencil
844 * reference value is written by the fragment shader. Tracking this
845 * interaction does not seem worth the effort, so be conservative.
847 return op
!= VK_STENCIL_OP_INCREMENT_AND_CLAMP
&&
848 op
!= VK_STENCIL_OP_DECREMENT_AND_CLAMP
&&
849 op
!= VK_STENCIL_OP_REPLACE
;
853 radv_order_invariant_stencil_state(const VkStencilOpState
*state
)
855 /* Compute whether, assuming Z writes are disabled, this stencil state
856 * is order invariant in the sense that the set of passing fragments as
857 * well as the final stencil buffer result does not depend on the order
860 return !state
->writeMask
||
861 /* The following assumes that Z writes are disabled. */
862 (state
->compareOp
== VK_COMPARE_OP_ALWAYS
&&
863 radv_order_invariant_stencil_op(state
->passOp
) &&
864 radv_order_invariant_stencil_op(state
->depthFailOp
)) ||
865 (state
->compareOp
== VK_COMPARE_OP_NEVER
&&
866 radv_order_invariant_stencil_op(state
->failOp
));
870 radv_pipeline_out_of_order_rast(struct radv_pipeline
*pipeline
,
871 struct radv_blend_state
*blend
,
872 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
874 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
875 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
876 unsigned colormask
= blend
->cb_target_enabled_4bit
;
878 if (!pipeline
->device
->physical_device
->out_of_order_rast_allowed
)
881 /* Be conservative if a logic operation is enabled with color buffers. */
882 if (colormask
&& pCreateInfo
->pColorBlendState
->logicOpEnable
)
885 /* Default depth/stencil invariance when no attachment is bound. */
886 struct radv_dsa_order_invariance dsa_order_invariant
= {
887 .zs
= true, .pass_set
= true
890 if (pCreateInfo
->pDepthStencilState
&&
891 subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
892 const VkPipelineDepthStencilStateCreateInfo
*vkds
=
893 pCreateInfo
->pDepthStencilState
;
894 struct radv_render_pass_attachment
*attachment
=
895 pass
->attachments
+ subpass
->depth_stencil_attachment
.attachment
;
896 bool has_stencil
= vk_format_is_stencil(attachment
->format
);
897 struct radv_dsa_order_invariance order_invariance
[2];
898 struct radv_shader_variant
*ps
=
899 pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
901 /* Compute depth/stencil order invariance in order to know if
902 * it's safe to enable out-of-order.
904 bool zfunc_is_ordered
=
905 vkds
->depthCompareOp
== VK_COMPARE_OP_NEVER
||
906 vkds
->depthCompareOp
== VK_COMPARE_OP_LESS
||
907 vkds
->depthCompareOp
== VK_COMPARE_OP_LESS_OR_EQUAL
||
908 vkds
->depthCompareOp
== VK_COMPARE_OP_GREATER
||
909 vkds
->depthCompareOp
== VK_COMPARE_OP_GREATER_OR_EQUAL
;
911 bool nozwrite_and_order_invariant_stencil
=
912 !radv_is_ds_write_enabled(vkds
) ||
913 (!radv_is_depth_write_enabled(vkds
) &&
914 radv_order_invariant_stencil_state(&vkds
->front
) &&
915 radv_order_invariant_stencil_state(&vkds
->back
));
917 order_invariance
[1].zs
=
918 nozwrite_and_order_invariant_stencil
||
919 (!radv_is_stencil_write_enabled(vkds
) &&
921 order_invariance
[0].zs
=
922 !radv_is_depth_write_enabled(vkds
) || zfunc_is_ordered
;
924 order_invariance
[1].pass_set
=
925 nozwrite_and_order_invariant_stencil
||
926 (!radv_is_stencil_write_enabled(vkds
) &&
927 (vkds
->depthCompareOp
== VK_COMPARE_OP_ALWAYS
||
928 vkds
->depthCompareOp
== VK_COMPARE_OP_NEVER
));
929 order_invariance
[0].pass_set
=
930 !radv_is_depth_write_enabled(vkds
) ||
931 (vkds
->depthCompareOp
== VK_COMPARE_OP_ALWAYS
||
932 vkds
->depthCompareOp
== VK_COMPARE_OP_NEVER
);
934 dsa_order_invariant
= order_invariance
[has_stencil
];
935 if (!dsa_order_invariant
.zs
)
938 /* The set of PS invocations is always order invariant,
939 * except when early Z/S tests are requested.
942 ps
->info
.info
.ps
.writes_memory
&&
943 ps
->info
.fs
.early_fragment_test
&&
944 !dsa_order_invariant
.pass_set
)
947 /* Determine if out-of-order rasterization should be disabled
948 * when occlusion queries are used.
950 pipeline
->graphics
.disable_out_of_order_rast_for_occlusion
=
951 !dsa_order_invariant
.pass_set
;
954 /* No color buffers are enabled for writing. */
958 unsigned blendmask
= colormask
& blend
->blend_enable_4bit
;
961 /* Only commutative blending. */
962 if (blendmask
& ~blend
->commutative_4bit
)
965 if (!dsa_order_invariant
.pass_set
)
969 if (colormask
& ~blendmask
)
976 radv_pipeline_init_multisample_state(struct radv_pipeline
*pipeline
,
977 struct radv_blend_state
*blend
,
978 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
980 const VkPipelineMultisampleStateCreateInfo
*vkms
= pCreateInfo
->pMultisampleState
;
981 struct radv_multisample_state
*ms
= &pipeline
->graphics
.ms
;
982 unsigned num_tile_pipes
= pipeline
->device
->physical_device
->rad_info
.num_tile_pipes
;
983 bool out_of_order_rast
= false;
984 int ps_iter_samples
= 1;
985 uint32_t mask
= 0xffff;
988 ms
->num_samples
= vkms
->rasterizationSamples
;
993 ps_iter_samples
= radv_pipeline_get_ps_iter_samples(vkms
);
994 if (vkms
&& !vkms
->sampleShadingEnable
&& pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.force_persample
) {
995 ps_iter_samples
= ms
->num_samples
;
998 ms
->pa_sc_line_cntl
= S_028BDC_DX10_DIAMOND_TEST_ENA(1);
999 ms
->pa_sc_aa_config
= 0;
1000 ms
->db_eqaa
= S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
1001 S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
1002 ms
->pa_sc_mode_cntl_1
=
1003 S_028A4C_WALK_FENCE_ENABLE(1) | //TODO linear dst fixes
1004 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes
== 2 ? 2 : 3) |
1006 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
1007 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
1008 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
1009 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
1010 S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
1011 S_028A4C_FORCE_EOV_REZ_ENABLE(1);
1012 ms
->pa_sc_mode_cntl_0
= S_028A48_ALTERNATE_RBS_PER_TILE(pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) |
1013 S_028A48_VPORT_SCISSOR_ENABLE(1);
1015 if (ms
->num_samples
> 1) {
1016 unsigned log_samples
= util_logbase2(ms
->num_samples
);
1017 unsigned log_ps_iter_samples
= util_logbase2(ps_iter_samples
);
1018 ms
->pa_sc_mode_cntl_0
|= S_028A48_MSAA_ENABLE(1);
1019 ms
->pa_sc_line_cntl
|= S_028BDC_EXPAND_LINE_WIDTH(1); /* CM_R_028BDC_PA_SC_LINE_CNTL */
1020 ms
->db_eqaa
|= S_028804_MAX_ANCHOR_SAMPLES(log_samples
) |
1021 S_028804_PS_ITER_SAMPLES(log_ps_iter_samples
) |
1022 S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples
) |
1023 S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples
);
1024 ms
->pa_sc_aa_config
|= S_028BE0_MSAA_NUM_SAMPLES(log_samples
) |
1025 S_028BE0_MAX_SAMPLE_DIST(radv_cayman_get_maxdist(log_samples
)) |
1026 S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples
); /* CM_R_028BE0_PA_SC_AA_CONFIG */
1027 ms
->pa_sc_mode_cntl_1
|= S_028A4C_PS_ITER_SAMPLE(ps_iter_samples
> 1);
1028 if (ps_iter_samples
> 1)
1029 pipeline
->graphics
.spi_baryc_cntl
|= S_0286E0_POS_FLOAT_LOCATION(2);
1032 const struct VkPipelineRasterizationStateRasterizationOrderAMD
*raster_order
=
1033 vk_find_struct_const(pCreateInfo
->pRasterizationState
->pNext
, PIPELINE_RASTERIZATION_STATE_RASTERIZATION_ORDER_AMD
);
1034 if (raster_order
&& raster_order
->rasterizationOrder
== VK_RASTERIZATION_ORDER_RELAXED_AMD
) {
1035 /* Out-of-order rasterization is explicitly enabled by the
1038 out_of_order_rast
= true;
1040 /* Determine if the driver can enable out-of-order
1041 * rasterization internally.
1044 radv_pipeline_out_of_order_rast(pipeline
, blend
, pCreateInfo
);
1047 if (out_of_order_rast
) {
1048 ms
->pa_sc_mode_cntl_1
|= S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(1) |
1049 S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7);
1052 if (vkms
&& vkms
->pSampleMask
) {
1053 mask
= vkms
->pSampleMask
[0] & 0xffff;
1056 ms
->pa_sc_aa_mask
[0] = mask
| (mask
<< 16);
1057 ms
->pa_sc_aa_mask
[1] = mask
| (mask
<< 16);
1061 radv_prim_can_use_guardband(enum VkPrimitiveTopology topology
)
1064 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
1065 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
1066 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
1067 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
1068 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
1070 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
1071 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
1072 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
1073 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
1074 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
1075 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
1078 unreachable("unhandled primitive type");
1083 si_translate_prim(enum VkPrimitiveTopology topology
)
1086 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
1087 return V_008958_DI_PT_POINTLIST
;
1088 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
1089 return V_008958_DI_PT_LINELIST
;
1090 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
1091 return V_008958_DI_PT_LINESTRIP
;
1092 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
1093 return V_008958_DI_PT_TRILIST
;
1094 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
1095 return V_008958_DI_PT_TRISTRIP
;
1096 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
1097 return V_008958_DI_PT_TRIFAN
;
1098 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
1099 return V_008958_DI_PT_LINELIST_ADJ
;
1100 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
1101 return V_008958_DI_PT_LINESTRIP_ADJ
;
1102 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
1103 return V_008958_DI_PT_TRILIST_ADJ
;
1104 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
1105 return V_008958_DI_PT_TRISTRIP_ADJ
;
1106 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
1107 return V_008958_DI_PT_PATCH
;
1115 si_conv_gl_prim_to_gs_out(unsigned gl_prim
)
1118 case 0: /* GL_POINTS */
1119 return V_028A6C_OUTPRIM_TYPE_POINTLIST
;
1120 case 1: /* GL_LINES */
1121 case 3: /* GL_LINE_STRIP */
1122 case 0xA: /* GL_LINE_STRIP_ADJACENCY_ARB */
1123 case 0x8E7A: /* GL_ISOLINES */
1124 return V_028A6C_OUTPRIM_TYPE_LINESTRIP
;
1126 case 4: /* GL_TRIANGLES */
1127 case 0xc: /* GL_TRIANGLES_ADJACENCY_ARB */
1128 case 5: /* GL_TRIANGLE_STRIP */
1129 case 7: /* GL_QUADS */
1130 return V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
1138 si_conv_prim_to_gs_out(enum VkPrimitiveTopology topology
)
1141 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
1142 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
1143 return V_028A6C_OUTPRIM_TYPE_POINTLIST
;
1144 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
1145 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
1146 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
1147 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
1148 return V_028A6C_OUTPRIM_TYPE_LINESTRIP
;
1149 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
1150 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
1151 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
1152 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
1153 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
1154 return V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
1161 static unsigned si_map_swizzle(unsigned swizzle
)
1165 return V_008F0C_SQ_SEL_Y
;
1167 return V_008F0C_SQ_SEL_Z
;
1169 return V_008F0C_SQ_SEL_W
;
1171 return V_008F0C_SQ_SEL_0
;
1173 return V_008F0C_SQ_SEL_1
;
1174 default: /* VK_SWIZZLE_X */
1175 return V_008F0C_SQ_SEL_X
;
1180 static unsigned radv_dynamic_state_mask(VkDynamicState state
)
1183 case VK_DYNAMIC_STATE_VIEWPORT
:
1184 return RADV_DYNAMIC_VIEWPORT
;
1185 case VK_DYNAMIC_STATE_SCISSOR
:
1186 return RADV_DYNAMIC_SCISSOR
;
1187 case VK_DYNAMIC_STATE_LINE_WIDTH
:
1188 return RADV_DYNAMIC_LINE_WIDTH
;
1189 case VK_DYNAMIC_STATE_DEPTH_BIAS
:
1190 return RADV_DYNAMIC_DEPTH_BIAS
;
1191 case VK_DYNAMIC_STATE_BLEND_CONSTANTS
:
1192 return RADV_DYNAMIC_BLEND_CONSTANTS
;
1193 case VK_DYNAMIC_STATE_DEPTH_BOUNDS
:
1194 return RADV_DYNAMIC_DEPTH_BOUNDS
;
1195 case VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK
:
1196 return RADV_DYNAMIC_STENCIL_COMPARE_MASK
;
1197 case VK_DYNAMIC_STATE_STENCIL_WRITE_MASK
:
1198 return RADV_DYNAMIC_STENCIL_WRITE_MASK
;
1199 case VK_DYNAMIC_STATE_STENCIL_REFERENCE
:
1200 return RADV_DYNAMIC_STENCIL_REFERENCE
;
1201 case VK_DYNAMIC_STATE_DISCARD_RECTANGLE_EXT
:
1202 return RADV_DYNAMIC_DISCARD_RECTANGLE
;
1204 unreachable("Unhandled dynamic state");
1208 static uint32_t radv_pipeline_needed_dynamic_state(const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1210 uint32_t states
= RADV_DYNAMIC_ALL
;
1212 /* If rasterization is disabled we do not care about any of the dynamic states,
1213 * since they are all rasterization related only. */
1214 if (pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
)
1217 if (!pCreateInfo
->pRasterizationState
->depthBiasEnable
)
1218 states
&= ~RADV_DYNAMIC_DEPTH_BIAS
;
1220 if (!pCreateInfo
->pDepthStencilState
||
1221 !pCreateInfo
->pDepthStencilState
->depthBoundsTestEnable
)
1222 states
&= ~RADV_DYNAMIC_DEPTH_BOUNDS
;
1224 if (!pCreateInfo
->pDepthStencilState
||
1225 !pCreateInfo
->pDepthStencilState
->stencilTestEnable
)
1226 states
&= ~(RADV_DYNAMIC_STENCIL_COMPARE_MASK
|
1227 RADV_DYNAMIC_STENCIL_WRITE_MASK
|
1228 RADV_DYNAMIC_STENCIL_REFERENCE
);
1230 if (!vk_find_struct_const(pCreateInfo
->pNext
, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT
))
1231 states
&= ~RADV_DYNAMIC_DISCARD_RECTANGLE
;
1233 /* TODO: blend constants & line width. */
1240 radv_pipeline_init_dynamic_state(struct radv_pipeline
*pipeline
,
1241 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1243 uint32_t needed_states
= radv_pipeline_needed_dynamic_state(pCreateInfo
);
1244 uint32_t states
= needed_states
;
1245 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
1246 struct radv_subpass
*subpass
= &pass
->subpasses
[pCreateInfo
->subpass
];
1248 pipeline
->dynamic_state
= default_dynamic_state
;
1249 pipeline
->graphics
.needed_dynamic_state
= needed_states
;
1251 if (pCreateInfo
->pDynamicState
) {
1252 /* Remove all of the states that are marked as dynamic */
1253 uint32_t count
= pCreateInfo
->pDynamicState
->dynamicStateCount
;
1254 for (uint32_t s
= 0; s
< count
; s
++)
1255 states
&= ~radv_dynamic_state_mask(pCreateInfo
->pDynamicState
->pDynamicStates
[s
]);
1258 struct radv_dynamic_state
*dynamic
= &pipeline
->dynamic_state
;
1260 if (needed_states
& RADV_DYNAMIC_VIEWPORT
) {
1261 assert(pCreateInfo
->pViewportState
);
1263 dynamic
->viewport
.count
= pCreateInfo
->pViewportState
->viewportCount
;
1264 if (states
& RADV_DYNAMIC_VIEWPORT
) {
1265 typed_memcpy(dynamic
->viewport
.viewports
,
1266 pCreateInfo
->pViewportState
->pViewports
,
1267 pCreateInfo
->pViewportState
->viewportCount
);
1271 if (needed_states
& RADV_DYNAMIC_SCISSOR
) {
1272 dynamic
->scissor
.count
= pCreateInfo
->pViewportState
->scissorCount
;
1273 if (states
& RADV_DYNAMIC_SCISSOR
) {
1274 typed_memcpy(dynamic
->scissor
.scissors
,
1275 pCreateInfo
->pViewportState
->pScissors
,
1276 pCreateInfo
->pViewportState
->scissorCount
);
1280 if (states
& RADV_DYNAMIC_LINE_WIDTH
) {
1281 assert(pCreateInfo
->pRasterizationState
);
1282 dynamic
->line_width
= pCreateInfo
->pRasterizationState
->lineWidth
;
1285 if (states
& RADV_DYNAMIC_DEPTH_BIAS
) {
1286 assert(pCreateInfo
->pRasterizationState
);
1287 dynamic
->depth_bias
.bias
=
1288 pCreateInfo
->pRasterizationState
->depthBiasConstantFactor
;
1289 dynamic
->depth_bias
.clamp
=
1290 pCreateInfo
->pRasterizationState
->depthBiasClamp
;
1291 dynamic
->depth_bias
.slope
=
1292 pCreateInfo
->pRasterizationState
->depthBiasSlopeFactor
;
1295 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1297 * pColorBlendState is [...] NULL if the pipeline has rasterization
1298 * disabled or if the subpass of the render pass the pipeline is
1299 * created against does not use any color attachments.
1301 bool uses_color_att
= false;
1302 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
1303 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
) {
1304 uses_color_att
= true;
1309 if (uses_color_att
&& states
& RADV_DYNAMIC_BLEND_CONSTANTS
) {
1310 assert(pCreateInfo
->pColorBlendState
);
1311 typed_memcpy(dynamic
->blend_constants
,
1312 pCreateInfo
->pColorBlendState
->blendConstants
, 4);
1315 /* If there is no depthstencil attachment, then don't read
1316 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
1317 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
1318 * no need to override the depthstencil defaults in
1319 * radv_pipeline::dynamic_state when there is no depthstencil attachment.
1321 * Section 9.2 of the Vulkan 1.0.15 spec says:
1323 * pDepthStencilState is [...] NULL if the pipeline has rasterization
1324 * disabled or if the subpass of the render pass the pipeline is created
1325 * against does not use a depth/stencil attachment.
1327 if (needed_states
&&
1328 subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
1329 assert(pCreateInfo
->pDepthStencilState
);
1331 if (states
& RADV_DYNAMIC_DEPTH_BOUNDS
) {
1332 dynamic
->depth_bounds
.min
=
1333 pCreateInfo
->pDepthStencilState
->minDepthBounds
;
1334 dynamic
->depth_bounds
.max
=
1335 pCreateInfo
->pDepthStencilState
->maxDepthBounds
;
1338 if (states
& RADV_DYNAMIC_STENCIL_COMPARE_MASK
) {
1339 dynamic
->stencil_compare_mask
.front
=
1340 pCreateInfo
->pDepthStencilState
->front
.compareMask
;
1341 dynamic
->stencil_compare_mask
.back
=
1342 pCreateInfo
->pDepthStencilState
->back
.compareMask
;
1345 if (states
& RADV_DYNAMIC_STENCIL_WRITE_MASK
) {
1346 dynamic
->stencil_write_mask
.front
=
1347 pCreateInfo
->pDepthStencilState
->front
.writeMask
;
1348 dynamic
->stencil_write_mask
.back
=
1349 pCreateInfo
->pDepthStencilState
->back
.writeMask
;
1352 if (states
& RADV_DYNAMIC_STENCIL_REFERENCE
) {
1353 dynamic
->stencil_reference
.front
=
1354 pCreateInfo
->pDepthStencilState
->front
.reference
;
1355 dynamic
->stencil_reference
.back
=
1356 pCreateInfo
->pDepthStencilState
->back
.reference
;
1360 const VkPipelineDiscardRectangleStateCreateInfoEXT
*discard_rectangle_info
=
1361 vk_find_struct_const(pCreateInfo
->pNext
, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT
);
1362 if (states
& RADV_DYNAMIC_DISCARD_RECTANGLE
) {
1363 dynamic
->discard_rectangle
.count
= discard_rectangle_info
->discardRectangleCount
;
1364 typed_memcpy(dynamic
->discard_rectangle
.rectangles
,
1365 discard_rectangle_info
->pDiscardRectangles
,
1366 discard_rectangle_info
->discardRectangleCount
);
1369 pipeline
->dynamic_state
.mask
= states
;
1372 static struct radv_gs_state
1373 calculate_gs_info(const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1374 const struct radv_pipeline
*pipeline
)
1376 struct radv_gs_state gs
= {0};
1377 struct radv_shader_variant_info
*gs_info
= &pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
;
1378 struct radv_es_output_info
*es_info
;
1379 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1380 es_info
= radv_pipeline_has_tess(pipeline
) ? &gs_info
->tes
.es_info
: &gs_info
->vs
.es_info
;
1382 es_info
= radv_pipeline_has_tess(pipeline
) ?
1383 &pipeline
->shaders
[MESA_SHADER_TESS_EVAL
]->info
.tes
.es_info
:
1384 &pipeline
->shaders
[MESA_SHADER_VERTEX
]->info
.vs
.es_info
;
1386 unsigned gs_num_invocations
= MAX2(gs_info
->gs
.invocations
, 1);
1387 bool uses_adjacency
;
1388 switch(pCreateInfo
->pInputAssemblyState
->topology
) {
1389 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
1390 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
1391 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
1392 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
1393 uses_adjacency
= true;
1396 uses_adjacency
= false;
1400 /* All these are in dwords: */
1401 /* We can't allow using the whole LDS, because GS waves compete with
1402 * other shader stages for LDS space. */
1403 const unsigned max_lds_size
= 8 * 1024;
1404 const unsigned esgs_itemsize
= es_info
->esgs_itemsize
/ 4;
1405 unsigned esgs_lds_size
;
1407 /* All these are per subgroup: */
1408 const unsigned max_out_prims
= 32 * 1024;
1409 const unsigned max_es_verts
= 255;
1410 const unsigned ideal_gs_prims
= 64;
1411 unsigned max_gs_prims
, gs_prims
;
1412 unsigned min_es_verts
, es_verts
, worst_case_es_verts
;
1414 if (uses_adjacency
|| gs_num_invocations
> 1)
1415 max_gs_prims
= 127 / gs_num_invocations
;
1419 /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
1420 * Make sure we don't go over the maximum value.
1422 if (gs_info
->gs
.vertices_out
> 0) {
1423 max_gs_prims
= MIN2(max_gs_prims
,
1425 (gs_info
->gs
.vertices_out
* gs_num_invocations
));
1427 assert(max_gs_prims
> 0);
1429 /* If the primitive has adjacency, halve the number of vertices
1430 * that will be reused in multiple primitives.
1432 min_es_verts
= gs_info
->gs
.vertices_in
/ (uses_adjacency
? 2 : 1);
1434 gs_prims
= MIN2(ideal_gs_prims
, max_gs_prims
);
1435 worst_case_es_verts
= MIN2(min_es_verts
* gs_prims
, max_es_verts
);
1437 /* Compute ESGS LDS size based on the worst case number of ES vertices
1438 * needed to create the target number of GS prims per subgroup.
1440 esgs_lds_size
= esgs_itemsize
* worst_case_es_verts
;
1442 /* If total LDS usage is too big, refactor partitions based on ratio
1443 * of ESGS item sizes.
1445 if (esgs_lds_size
> max_lds_size
) {
1446 /* Our target GS Prims Per Subgroup was too large. Calculate
1447 * the maximum number of GS Prims Per Subgroup that will fit
1448 * into LDS, capped by the maximum that the hardware can support.
1450 gs_prims
= MIN2((max_lds_size
/ (esgs_itemsize
* min_es_verts
)),
1452 assert(gs_prims
> 0);
1453 worst_case_es_verts
= MIN2(min_es_verts
* gs_prims
,
1456 esgs_lds_size
= esgs_itemsize
* worst_case_es_verts
;
1457 assert(esgs_lds_size
<= max_lds_size
);
1460 /* Now calculate remaining ESGS information. */
1462 es_verts
= MIN2(esgs_lds_size
/ esgs_itemsize
, max_es_verts
);
1464 es_verts
= max_es_verts
;
1466 /* Vertices for adjacency primitives are not always reused, so restore
1467 * it for ES_VERTS_PER_SUBGRP.
1469 min_es_verts
= gs_info
->gs
.vertices_in
;
1471 /* For normal primitives, the VGT only checks if they are past the ES
1472 * verts per subgroup after allocating a full GS primitive and if they
1473 * are, kick off a new subgroup. But if those additional ES verts are
1474 * unique (e.g. not reused) we need to make sure there is enough LDS
1475 * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
1477 es_verts
-= min_es_verts
- 1;
1479 uint32_t es_verts_per_subgroup
= es_verts
;
1480 uint32_t gs_prims_per_subgroup
= gs_prims
;
1481 uint32_t gs_inst_prims_in_subgroup
= gs_prims
* gs_num_invocations
;
1482 uint32_t max_prims_per_subgroup
= gs_inst_prims_in_subgroup
* gs_info
->gs
.vertices_out
;
1483 gs
.lds_size
= align(esgs_lds_size
, 128) / 128;
1484 gs
.vgt_gs_onchip_cntl
= S_028A44_ES_VERTS_PER_SUBGRP(es_verts_per_subgroup
) |
1485 S_028A44_GS_PRIMS_PER_SUBGRP(gs_prims_per_subgroup
) |
1486 S_028A44_GS_INST_PRIMS_IN_SUBGRP(gs_inst_prims_in_subgroup
);
1487 gs
.vgt_gs_max_prims_per_subgroup
= S_028A94_MAX_PRIMS_PER_SUBGROUP(max_prims_per_subgroup
);
1488 gs
.vgt_esgs_ring_itemsize
= esgs_itemsize
;
1489 assert(max_prims_per_subgroup
<= max_out_prims
);
1495 calculate_gs_ring_sizes(struct radv_pipeline
*pipeline
, const struct radv_gs_state
*gs
)
1497 struct radv_device
*device
= pipeline
->device
;
1498 unsigned num_se
= device
->physical_device
->rad_info
.max_se
;
1499 unsigned wave_size
= 64;
1500 unsigned max_gs_waves
= 32 * num_se
; /* max 32 per SE on GCN */
1501 unsigned gs_vertex_reuse
= 16 * num_se
; /* GS_VERTEX_REUSE register (per SE) */
1502 unsigned alignment
= 256 * num_se
;
1503 /* The maximum size is 63.999 MB per SE. */
1504 unsigned max_size
= ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se
;
1505 struct radv_shader_variant_info
*gs_info
= &pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
;
1507 /* Calculate the minimum size. */
1508 unsigned min_esgs_ring_size
= align(gs
->vgt_esgs_ring_itemsize
* 4 * gs_vertex_reuse
*
1509 wave_size
, alignment
);
1510 /* These are recommended sizes, not minimum sizes. */
1511 unsigned esgs_ring_size
= max_gs_waves
* 2 * wave_size
*
1512 gs
->vgt_esgs_ring_itemsize
* 4 * gs_info
->gs
.vertices_in
;
1513 unsigned gsvs_ring_size
= max_gs_waves
* 2 * wave_size
*
1514 gs_info
->gs
.max_gsvs_emit_size
* 1; // no streams in VK (gs->max_gs_stream + 1);
1516 min_esgs_ring_size
= align(min_esgs_ring_size
, alignment
);
1517 esgs_ring_size
= align(esgs_ring_size
, alignment
);
1518 gsvs_ring_size
= align(gsvs_ring_size
, alignment
);
1520 if (pipeline
->device
->physical_device
->rad_info
.chip_class
<= VI
)
1521 pipeline
->graphics
.esgs_ring_size
= CLAMP(esgs_ring_size
, min_esgs_ring_size
, max_size
);
1523 pipeline
->graphics
.gsvs_ring_size
= MIN2(gsvs_ring_size
, max_size
);
1526 static void si_multiwave_lds_size_workaround(struct radv_device
*device
,
1529 /* If tessellation is all offchip and on-chip GS isn't used, this
1530 * workaround is not needed.
1534 /* SPI barrier management bug:
1535 * Make sure we have at least 4k of LDS in use to avoid the bug.
1536 * It applies to workgroup sizes of more than one wavefront.
1538 if (device
->physical_device
->rad_info
.family
== CHIP_BONAIRE
||
1539 device
->physical_device
->rad_info
.family
== CHIP_KABINI
||
1540 device
->physical_device
->rad_info
.family
== CHIP_MULLINS
)
1541 *lds_size
= MAX2(*lds_size
, 8);
1544 struct radv_shader_variant
*
1545 radv_get_vertex_shader(struct radv_pipeline
*pipeline
)
1547 if (pipeline
->shaders
[MESA_SHADER_VERTEX
])
1548 return pipeline
->shaders
[MESA_SHADER_VERTEX
];
1549 if (pipeline
->shaders
[MESA_SHADER_TESS_CTRL
])
1550 return pipeline
->shaders
[MESA_SHADER_TESS_CTRL
];
1551 return pipeline
->shaders
[MESA_SHADER_GEOMETRY
];
1554 static struct radv_shader_variant
*
1555 radv_get_tess_eval_shader(struct radv_pipeline
*pipeline
)
1557 if (pipeline
->shaders
[MESA_SHADER_TESS_EVAL
])
1558 return pipeline
->shaders
[MESA_SHADER_TESS_EVAL
];
1559 return pipeline
->shaders
[MESA_SHADER_GEOMETRY
];
1562 static struct radv_tessellation_state
1563 calculate_tess_state(struct radv_pipeline
*pipeline
,
1564 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1566 unsigned num_tcs_input_cp
;
1567 unsigned num_tcs_output_cp
;
1569 unsigned num_patches
;
1570 struct radv_tessellation_state tess
= {0};
1572 num_tcs_input_cp
= pCreateInfo
->pTessellationState
->patchControlPoints
;
1573 num_tcs_output_cp
= pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.tcs_vertices_out
; //TCS VERTICES OUT
1574 num_patches
= pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.num_patches
;
1576 lds_size
= pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.lds_size
;
1578 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1579 assert(lds_size
<= 65536);
1580 lds_size
= align(lds_size
, 512) / 512;
1582 assert(lds_size
<= 32768);
1583 lds_size
= align(lds_size
, 256) / 256;
1585 si_multiwave_lds_size_workaround(pipeline
->device
, &lds_size
);
1587 tess
.lds_size
= lds_size
;
1589 tess
.ls_hs_config
= S_028B58_NUM_PATCHES(num_patches
) |
1590 S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp
) |
1591 S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp
);
1592 tess
.num_patches
= num_patches
;
1594 struct radv_shader_variant
*tes
= radv_get_tess_eval_shader(pipeline
);
1595 unsigned type
= 0, partitioning
= 0, topology
= 0, distribution_mode
= 0;
1597 switch (tes
->info
.tes
.primitive_mode
) {
1599 type
= V_028B6C_TESS_TRIANGLE
;
1602 type
= V_028B6C_TESS_QUAD
;
1605 type
= V_028B6C_TESS_ISOLINE
;
1609 switch (tes
->info
.tes
.spacing
) {
1610 case TESS_SPACING_EQUAL
:
1611 partitioning
= V_028B6C_PART_INTEGER
;
1613 case TESS_SPACING_FRACTIONAL_ODD
:
1614 partitioning
= V_028B6C_PART_FRAC_ODD
;
1616 case TESS_SPACING_FRACTIONAL_EVEN
:
1617 partitioning
= V_028B6C_PART_FRAC_EVEN
;
1623 bool ccw
= tes
->info
.tes
.ccw
;
1624 const VkPipelineTessellationDomainOriginStateCreateInfoKHR
*domain_origin_state
=
1625 vk_find_struct_const(pCreateInfo
->pTessellationState
,
1626 PIPELINE_TESSELLATION_DOMAIN_ORIGIN_STATE_CREATE_INFO_KHR
);
1628 if (domain_origin_state
&& domain_origin_state
->domainOrigin
!= VK_TESSELLATION_DOMAIN_ORIGIN_UPPER_LEFT_KHR
)
1631 if (tes
->info
.tes
.point_mode
)
1632 topology
= V_028B6C_OUTPUT_POINT
;
1633 else if (tes
->info
.tes
.primitive_mode
== GL_ISOLINES
)
1634 topology
= V_028B6C_OUTPUT_LINE
;
1636 topology
= V_028B6C_OUTPUT_TRIANGLE_CCW
;
1638 topology
= V_028B6C_OUTPUT_TRIANGLE_CW
;
1640 if (pipeline
->device
->has_distributed_tess
) {
1641 if (pipeline
->device
->physical_device
->rad_info
.family
== CHIP_FIJI
||
1642 pipeline
->device
->physical_device
->rad_info
.family
>= CHIP_POLARIS10
)
1643 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS
;
1645 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_DONUTS
;
1647 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_NO_DIST
;
1649 tess
.tf_param
= S_028B6C_TYPE(type
) |
1650 S_028B6C_PARTITIONING(partitioning
) |
1651 S_028B6C_TOPOLOGY(topology
) |
1652 S_028B6C_DISTRIBUTION_MODE(distribution_mode
);
1657 static const struct radv_prim_vertex_count prim_size_table
[] = {
1658 [V_008958_DI_PT_NONE
] = {0, 0},
1659 [V_008958_DI_PT_POINTLIST
] = {1, 1},
1660 [V_008958_DI_PT_LINELIST
] = {2, 2},
1661 [V_008958_DI_PT_LINESTRIP
] = {2, 1},
1662 [V_008958_DI_PT_TRILIST
] = {3, 3},
1663 [V_008958_DI_PT_TRIFAN
] = {3, 1},
1664 [V_008958_DI_PT_TRISTRIP
] = {3, 1},
1665 [V_008958_DI_PT_LINELIST_ADJ
] = {4, 4},
1666 [V_008958_DI_PT_LINESTRIP_ADJ
] = {4, 1},
1667 [V_008958_DI_PT_TRILIST_ADJ
] = {6, 6},
1668 [V_008958_DI_PT_TRISTRIP_ADJ
] = {6, 2},
1669 [V_008958_DI_PT_RECTLIST
] = {3, 3},
1670 [V_008958_DI_PT_LINELOOP
] = {2, 1},
1671 [V_008958_DI_PT_POLYGON
] = {3, 1},
1672 [V_008958_DI_PT_2D_TRI_STRIP
] = {0, 0},
1675 static const struct radv_vs_output_info
*get_vs_output_info(const struct radv_pipeline
*pipeline
)
1677 if (radv_pipeline_has_gs(pipeline
))
1678 return &pipeline
->gs_copy_shader
->info
.vs
.outinfo
;
1679 else if (radv_pipeline_has_tess(pipeline
))
1680 return &pipeline
->shaders
[MESA_SHADER_TESS_EVAL
]->info
.tes
.outinfo
;
1682 return &pipeline
->shaders
[MESA_SHADER_VERTEX
]->info
.vs
.outinfo
;
1686 radv_link_shaders(struct radv_pipeline
*pipeline
, nir_shader
**shaders
)
1688 nir_shader
* ordered_shaders
[MESA_SHADER_STAGES
];
1689 int shader_count
= 0;
1691 if(shaders
[MESA_SHADER_FRAGMENT
]) {
1692 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_FRAGMENT
];
1694 if(shaders
[MESA_SHADER_GEOMETRY
]) {
1695 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_GEOMETRY
];
1697 if(shaders
[MESA_SHADER_TESS_EVAL
]) {
1698 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_TESS_EVAL
];
1700 if(shaders
[MESA_SHADER_TESS_CTRL
]) {
1701 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_TESS_CTRL
];
1703 if(shaders
[MESA_SHADER_VERTEX
]) {
1704 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_VERTEX
];
1707 for (int i
= 1; i
< shader_count
; ++i
) {
1708 nir_lower_io_arrays_to_elements(ordered_shaders
[i
],
1709 ordered_shaders
[i
- 1]);
1711 nir_remove_dead_variables(ordered_shaders
[i
],
1712 nir_var_shader_out
);
1713 nir_remove_dead_variables(ordered_shaders
[i
- 1],
1716 bool progress
= nir_remove_unused_varyings(ordered_shaders
[i
],
1717 ordered_shaders
[i
- 1]);
1719 nir_compact_varyings(ordered_shaders
[i
],
1720 ordered_shaders
[i
- 1], true);
1723 if (nir_lower_global_vars_to_local(ordered_shaders
[i
])) {
1724 ac_lower_indirect_derefs(ordered_shaders
[i
],
1725 pipeline
->device
->physical_device
->rad_info
.chip_class
);
1727 radv_optimize_nir(ordered_shaders
[i
]);
1729 if (nir_lower_global_vars_to_local(ordered_shaders
[i
- 1])) {
1730 ac_lower_indirect_derefs(ordered_shaders
[i
- 1],
1731 pipeline
->device
->physical_device
->rad_info
.chip_class
);
1733 radv_optimize_nir(ordered_shaders
[i
- 1]);
1739 static struct radv_pipeline_key
1740 radv_generate_graphics_pipeline_key(struct radv_pipeline
*pipeline
,
1741 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1742 const struct radv_blend_state
*blend
,
1743 bool has_view_index
)
1745 const VkPipelineVertexInputStateCreateInfo
*input_state
=
1746 pCreateInfo
->pVertexInputState
;
1747 const VkPipelineVertexInputDivisorStateCreateInfoEXT
*divisor_state
=
1748 vk_find_struct_const(input_state
->pNext
, PIPELINE_VERTEX_INPUT_DIVISOR_STATE_CREATE_INFO_EXT
);
1750 struct radv_pipeline_key key
;
1751 memset(&key
, 0, sizeof(key
));
1753 key
.has_multiview_view_index
= has_view_index
;
1755 uint32_t binding_input_rate
= 0;
1756 uint32_t instance_rate_divisors
[MAX_VERTEX_ATTRIBS
];
1757 for (unsigned i
= 0; i
< input_state
->vertexBindingDescriptionCount
; ++i
) {
1758 if (input_state
->pVertexBindingDescriptions
[i
].inputRate
) {
1759 unsigned binding
= input_state
->pVertexBindingDescriptions
[i
].binding
;
1760 binding_input_rate
|= 1u << binding
;
1761 instance_rate_divisors
[binding
] = 1;
1764 if (divisor_state
) {
1765 for (unsigned i
= 0; i
< divisor_state
->vertexBindingDivisorCount
; ++i
) {
1766 instance_rate_divisors
[divisor_state
->pVertexBindingDivisors
[i
].binding
] =
1767 divisor_state
->pVertexBindingDivisors
[i
].divisor
;
1771 for (unsigned i
= 0; i
< input_state
->vertexAttributeDescriptionCount
; ++i
) {
1773 binding
= input_state
->pVertexAttributeDescriptions
[i
].binding
;
1774 if (binding_input_rate
& (1u << binding
)) {
1775 unsigned location
= input_state
->pVertexAttributeDescriptions
[i
].location
;
1776 key
.instance_rate_inputs
|= 1u << location
;
1777 key
.instance_rate_divisors
[location
] = instance_rate_divisors
[binding
];
1781 if (pCreateInfo
->pTessellationState
)
1782 key
.tess_input_vertices
= pCreateInfo
->pTessellationState
->patchControlPoints
;
1785 if (pCreateInfo
->pMultisampleState
&&
1786 pCreateInfo
->pMultisampleState
->rasterizationSamples
> 1) {
1787 uint32_t num_samples
= pCreateInfo
->pMultisampleState
->rasterizationSamples
;
1788 uint32_t ps_iter_samples
= radv_pipeline_get_ps_iter_samples(pCreateInfo
->pMultisampleState
);
1789 key
.multisample
= true;
1790 key
.log2_num_samples
= util_logbase2(num_samples
);
1791 key
.log2_ps_iter_samples
= util_logbase2(ps_iter_samples
);
1794 key
.col_format
= blend
->spi_shader_col_format
;
1795 if (pipeline
->device
->physical_device
->rad_info
.chip_class
< VI
)
1796 radv_pipeline_compute_get_int_clamp(pCreateInfo
, &key
.is_int8
, &key
.is_int10
);
1802 radv_fill_shader_keys(struct radv_shader_variant_key
*keys
,
1803 const struct radv_pipeline_key
*key
,
1806 keys
[MESA_SHADER_VERTEX
].vs
.instance_rate_inputs
= key
->instance_rate_inputs
;
1807 for (unsigned i
= 0; i
< MAX_VERTEX_ATTRIBS
; ++i
)
1808 keys
[MESA_SHADER_VERTEX
].vs
.instance_rate_divisors
[i
] = key
->instance_rate_divisors
[i
];
1810 if (nir
[MESA_SHADER_TESS_CTRL
]) {
1811 keys
[MESA_SHADER_VERTEX
].vs
.as_ls
= true;
1812 keys
[MESA_SHADER_TESS_CTRL
].tcs
.num_inputs
= 0;
1813 keys
[MESA_SHADER_TESS_CTRL
].tcs
.input_vertices
= key
->tess_input_vertices
;
1814 keys
[MESA_SHADER_TESS_CTRL
].tcs
.primitive_mode
= nir
[MESA_SHADER_TESS_EVAL
]->info
.tess
.primitive_mode
;
1816 keys
[MESA_SHADER_TESS_CTRL
].tcs
.tes_reads_tess_factors
= !!(nir
[MESA_SHADER_TESS_EVAL
]->info
.inputs_read
& (VARYING_BIT_TESS_LEVEL_INNER
| VARYING_BIT_TESS_LEVEL_OUTER
));
1819 if (nir
[MESA_SHADER_GEOMETRY
]) {
1820 if (nir
[MESA_SHADER_TESS_CTRL
])
1821 keys
[MESA_SHADER_TESS_EVAL
].tes
.as_es
= true;
1823 keys
[MESA_SHADER_VERTEX
].vs
.as_es
= true;
1826 for(int i
= 0; i
< MESA_SHADER_STAGES
; ++i
)
1827 keys
[i
].has_multiview_view_index
= key
->has_multiview_view_index
;
1829 keys
[MESA_SHADER_FRAGMENT
].fs
.multisample
= key
->multisample
;
1830 keys
[MESA_SHADER_FRAGMENT
].fs
.col_format
= key
->col_format
;
1831 keys
[MESA_SHADER_FRAGMENT
].fs
.is_int8
= key
->is_int8
;
1832 keys
[MESA_SHADER_FRAGMENT
].fs
.is_int10
= key
->is_int10
;
1833 keys
[MESA_SHADER_FRAGMENT
].fs
.log2_ps_iter_samples
= key
->log2_ps_iter_samples
;
1834 keys
[MESA_SHADER_FRAGMENT
].fs
.log2_num_samples
= key
->log2_num_samples
;
1838 merge_tess_info(struct shader_info
*tes_info
,
1839 const struct shader_info
*tcs_info
)
1841 /* The Vulkan 1.0.38 spec, section 21.1 Tessellator says:
1843 * "PointMode. Controls generation of points rather than triangles
1844 * or lines. This functionality defaults to disabled, and is
1845 * enabled if either shader stage includes the execution mode.
1847 * and about Triangles, Quads, IsoLines, VertexOrderCw, VertexOrderCcw,
1848 * PointMode, SpacingEqual, SpacingFractionalEven, SpacingFractionalOdd,
1849 * and OutputVertices, it says:
1851 * "One mode must be set in at least one of the tessellation
1854 * So, the fields can be set in either the TCS or TES, but they must
1855 * agree if set in both. Our backend looks at TES, so bitwise-or in
1856 * the values from the TCS.
1858 assert(tcs_info
->tess
.tcs_vertices_out
== 0 ||
1859 tes_info
->tess
.tcs_vertices_out
== 0 ||
1860 tcs_info
->tess
.tcs_vertices_out
== tes_info
->tess
.tcs_vertices_out
);
1861 tes_info
->tess
.tcs_vertices_out
|= tcs_info
->tess
.tcs_vertices_out
;
1863 assert(tcs_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
1864 tes_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
1865 tcs_info
->tess
.spacing
== tes_info
->tess
.spacing
);
1866 tes_info
->tess
.spacing
|= tcs_info
->tess
.spacing
;
1868 assert(tcs_info
->tess
.primitive_mode
== 0 ||
1869 tes_info
->tess
.primitive_mode
== 0 ||
1870 tcs_info
->tess
.primitive_mode
== tes_info
->tess
.primitive_mode
);
1871 tes_info
->tess
.primitive_mode
|= tcs_info
->tess
.primitive_mode
;
1872 tes_info
->tess
.ccw
|= tcs_info
->tess
.ccw
;
1873 tes_info
->tess
.point_mode
|= tcs_info
->tess
.point_mode
;
1877 void radv_create_shaders(struct radv_pipeline
*pipeline
,
1878 struct radv_device
*device
,
1879 struct radv_pipeline_cache
*cache
,
1880 struct radv_pipeline_key key
,
1881 const VkPipelineShaderStageCreateInfo
**pStages
)
1883 struct radv_shader_module fs_m
= {0};
1884 struct radv_shader_module
*modules
[MESA_SHADER_STAGES
] = { 0, };
1885 nir_shader
*nir
[MESA_SHADER_STAGES
] = {0};
1886 void *codes
[MESA_SHADER_STAGES
] = {0};
1887 unsigned code_sizes
[MESA_SHADER_STAGES
] = {0};
1888 struct radv_shader_variant_key keys
[MESA_SHADER_STAGES
] = {{{{0}}}};
1889 unsigned char hash
[20], gs_copy_hash
[20];
1891 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1893 modules
[i
] = radv_shader_module_from_handle(pStages
[i
]->module
);
1894 if (modules
[i
]->nir
)
1895 _mesa_sha1_compute(modules
[i
]->nir
->info
.name
,
1896 strlen(modules
[i
]->nir
->info
.name
),
1901 radv_hash_shaders(hash
, pStages
, pipeline
->layout
, &key
, get_hash_flags(device
));
1902 memcpy(gs_copy_hash
, hash
, 20);
1903 gs_copy_hash
[0] ^= 1;
1905 if (modules
[MESA_SHADER_GEOMETRY
]) {
1906 struct radv_shader_variant
*variants
[MESA_SHADER_STAGES
] = {0};
1907 radv_create_shader_variants_from_pipeline_cache(device
, cache
, gs_copy_hash
, variants
);
1908 pipeline
->gs_copy_shader
= variants
[MESA_SHADER_GEOMETRY
];
1911 if (radv_create_shader_variants_from_pipeline_cache(device
, cache
, hash
, pipeline
->shaders
) &&
1912 (!modules
[MESA_SHADER_GEOMETRY
] || pipeline
->gs_copy_shader
)) {
1913 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1914 if (pipeline
->shaders
[i
])
1915 pipeline
->active_stages
|= mesa_to_vk_shader_stage(i
);
1920 if (!modules
[MESA_SHADER_FRAGMENT
] && !modules
[MESA_SHADER_COMPUTE
]) {
1922 nir_builder_init_simple_shader(&fs_b
, NULL
, MESA_SHADER_FRAGMENT
, NULL
);
1923 fs_b
.shader
->info
.name
= ralloc_strdup(fs_b
.shader
, "noop_fs");
1924 fs_m
.nir
= fs_b
.shader
;
1925 modules
[MESA_SHADER_FRAGMENT
] = &fs_m
;
1928 /* Determine first and last stage. */
1929 unsigned first
= MESA_SHADER_STAGES
;
1931 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
1934 if (first
== MESA_SHADER_STAGES
)
1939 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1940 const VkPipelineShaderStageCreateInfo
*stage
= pStages
[i
];
1945 nir
[i
] = radv_shader_compile_to_nir(device
, modules
[i
],
1946 stage
? stage
->pName
: "main", i
,
1947 stage
? stage
->pSpecializationInfo
: NULL
);
1948 pipeline
->active_stages
|= mesa_to_vk_shader_stage(i
);
1950 /* We don't want to alter meta shaders IR directly so clone it
1953 if (nir
[i
]->info
.name
) {
1954 nir
[i
] = nir_shader_clone(NULL
, nir
[i
]);
1957 if (first
!= last
) {
1958 nir_variable_mode mask
= 0;
1961 mask
= mask
| nir_var_shader_in
;
1964 mask
= mask
| nir_var_shader_out
;
1966 nir_lower_io_to_scalar_early(nir
[i
], mask
);
1967 radv_optimize_nir(nir
[i
]);
1971 if (nir
[MESA_SHADER_TESS_CTRL
]) {
1972 nir_lower_tes_patch_vertices(nir
[MESA_SHADER_TESS_EVAL
], nir
[MESA_SHADER_TESS_CTRL
]->info
.tess
.tcs_vertices_out
);
1973 merge_tess_info(&nir
[MESA_SHADER_TESS_EVAL
]->info
, &nir
[MESA_SHADER_TESS_CTRL
]->info
);
1976 radv_link_shaders(pipeline
, nir
);
1978 for (int i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1979 if (modules
[i
] && radv_can_dump_shader(device
, modules
[i
]))
1980 nir_print_shader(nir
[i
], stderr
);
1983 radv_fill_shader_keys(keys
, &key
, nir
);
1985 if (nir
[MESA_SHADER_FRAGMENT
]) {
1986 if (!pipeline
->shaders
[MESA_SHADER_FRAGMENT
]) {
1987 pipeline
->shaders
[MESA_SHADER_FRAGMENT
] =
1988 radv_shader_variant_create(device
, modules
[MESA_SHADER_FRAGMENT
], &nir
[MESA_SHADER_FRAGMENT
], 1,
1989 pipeline
->layout
, keys
+ MESA_SHADER_FRAGMENT
,
1990 &codes
[MESA_SHADER_FRAGMENT
], &code_sizes
[MESA_SHADER_FRAGMENT
]);
1993 /* TODO: These are no longer used as keys we should refactor this */
1994 keys
[MESA_SHADER_VERTEX
].vs
.export_prim_id
=
1995 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.prim_id_input
;
1996 keys
[MESA_SHADER_VERTEX
].vs
.export_layer_id
=
1997 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.layer_input
;
1998 keys
[MESA_SHADER_TESS_EVAL
].tes
.export_prim_id
=
1999 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.prim_id_input
;
2000 keys
[MESA_SHADER_TESS_EVAL
].tes
.export_layer_id
=
2001 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.layer_input
;
2004 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
&& modules
[MESA_SHADER_TESS_CTRL
]) {
2005 if (!pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]) {
2006 struct nir_shader
*combined_nir
[] = {nir
[MESA_SHADER_VERTEX
], nir
[MESA_SHADER_TESS_CTRL
]};
2007 struct radv_shader_variant_key key
= keys
[MESA_SHADER_TESS_CTRL
];
2008 key
.tcs
.vs_key
= keys
[MESA_SHADER_VERTEX
].vs
;
2009 pipeline
->shaders
[MESA_SHADER_TESS_CTRL
] = radv_shader_variant_create(device
, modules
[MESA_SHADER_TESS_CTRL
], combined_nir
, 2,
2011 &key
, &codes
[MESA_SHADER_TESS_CTRL
],
2012 &code_sizes
[MESA_SHADER_TESS_CTRL
]);
2014 modules
[MESA_SHADER_VERTEX
] = NULL
;
2015 keys
[MESA_SHADER_TESS_EVAL
].tes
.num_patches
= pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.num_patches
;
2016 keys
[MESA_SHADER_TESS_EVAL
].tes
.tcs_num_outputs
= util_last_bit64(pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.info
.tcs
.outputs_written
);
2019 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
&& modules
[MESA_SHADER_GEOMETRY
]) {
2020 gl_shader_stage pre_stage
= modules
[MESA_SHADER_TESS_EVAL
] ? MESA_SHADER_TESS_EVAL
: MESA_SHADER_VERTEX
;
2021 if (!pipeline
->shaders
[MESA_SHADER_GEOMETRY
]) {
2022 struct nir_shader
*combined_nir
[] = {nir
[pre_stage
], nir
[MESA_SHADER_GEOMETRY
]};
2023 pipeline
->shaders
[MESA_SHADER_GEOMETRY
] = radv_shader_variant_create(device
, modules
[MESA_SHADER_GEOMETRY
], combined_nir
, 2,
2025 &keys
[pre_stage
] , &codes
[MESA_SHADER_GEOMETRY
],
2026 &code_sizes
[MESA_SHADER_GEOMETRY
]);
2028 modules
[pre_stage
] = NULL
;
2031 for (int i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
2032 if(modules
[i
] && !pipeline
->shaders
[i
]) {
2033 if (i
== MESA_SHADER_TESS_CTRL
) {
2034 keys
[MESA_SHADER_TESS_CTRL
].tcs
.num_inputs
= util_last_bit64(pipeline
->shaders
[MESA_SHADER_VERTEX
]->info
.info
.vs
.ls_outputs_written
);
2036 if (i
== MESA_SHADER_TESS_EVAL
) {
2037 keys
[MESA_SHADER_TESS_EVAL
].tes
.num_patches
= pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.num_patches
;
2038 keys
[MESA_SHADER_TESS_EVAL
].tes
.tcs_num_outputs
= util_last_bit64(pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.info
.tcs
.outputs_written
);
2040 pipeline
->shaders
[i
] = radv_shader_variant_create(device
, modules
[i
], &nir
[i
], 1,
2042 keys
+ i
, &codes
[i
],
2047 if(modules
[MESA_SHADER_GEOMETRY
]) {
2048 void *gs_copy_code
= NULL
;
2049 unsigned gs_copy_code_size
= 0;
2050 if (!pipeline
->gs_copy_shader
) {
2051 pipeline
->gs_copy_shader
= radv_create_gs_copy_shader(
2052 device
, nir
[MESA_SHADER_GEOMETRY
], &gs_copy_code
,
2054 keys
[MESA_SHADER_GEOMETRY
].has_multiview_view_index
);
2057 if (pipeline
->gs_copy_shader
) {
2058 void *code
[MESA_SHADER_STAGES
] = {0};
2059 unsigned code_size
[MESA_SHADER_STAGES
] = {0};
2060 struct radv_shader_variant
*variants
[MESA_SHADER_STAGES
] = {0};
2062 code
[MESA_SHADER_GEOMETRY
] = gs_copy_code
;
2063 code_size
[MESA_SHADER_GEOMETRY
] = gs_copy_code_size
;
2064 variants
[MESA_SHADER_GEOMETRY
] = pipeline
->gs_copy_shader
;
2066 radv_pipeline_cache_insert_shaders(device
, cache
,
2075 radv_pipeline_cache_insert_shaders(device
, cache
, hash
, pipeline
->shaders
,
2076 (const void**)codes
, code_sizes
);
2078 for (int i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
2081 if (!pipeline
->device
->keep_shader_info
)
2082 ralloc_free(nir
[i
]);
2084 if (radv_can_dump_shader_stats(device
, modules
[i
]))
2085 radv_shader_dump_stats(device
,
2086 pipeline
->shaders
[i
],
2092 ralloc_free(fs_m
.nir
);
2096 radv_pipeline_stage_to_user_data_0(struct radv_pipeline
*pipeline
,
2097 gl_shader_stage stage
, enum chip_class chip_class
)
2099 bool has_gs
= radv_pipeline_has_gs(pipeline
);
2100 bool has_tess
= radv_pipeline_has_tess(pipeline
);
2102 case MESA_SHADER_FRAGMENT
:
2103 return R_00B030_SPI_SHADER_USER_DATA_PS_0
;
2104 case MESA_SHADER_VERTEX
:
2105 if (chip_class
>= GFX9
) {
2106 return has_tess
? R_00B430_SPI_SHADER_USER_DATA_LS_0
:
2107 has_gs
? R_00B330_SPI_SHADER_USER_DATA_ES_0
:
2108 R_00B130_SPI_SHADER_USER_DATA_VS_0
;
2111 return R_00B530_SPI_SHADER_USER_DATA_LS_0
;
2113 return has_gs
? R_00B330_SPI_SHADER_USER_DATA_ES_0
: R_00B130_SPI_SHADER_USER_DATA_VS_0
;
2114 case MESA_SHADER_GEOMETRY
:
2115 return chip_class
>= GFX9
? R_00B330_SPI_SHADER_USER_DATA_ES_0
:
2116 R_00B230_SPI_SHADER_USER_DATA_GS_0
;
2117 case MESA_SHADER_COMPUTE
:
2118 return R_00B900_COMPUTE_USER_DATA_0
;
2119 case MESA_SHADER_TESS_CTRL
:
2120 return chip_class
>= GFX9
? R_00B430_SPI_SHADER_USER_DATA_LS_0
:
2121 R_00B430_SPI_SHADER_USER_DATA_HS_0
;
2122 case MESA_SHADER_TESS_EVAL
:
2123 if (chip_class
>= GFX9
) {
2124 return has_gs
? R_00B330_SPI_SHADER_USER_DATA_ES_0
:
2125 R_00B130_SPI_SHADER_USER_DATA_VS_0
;
2128 return R_00B330_SPI_SHADER_USER_DATA_ES_0
;
2130 return R_00B130_SPI_SHADER_USER_DATA_VS_0
;
2132 unreachable("unknown shader");
2136 struct radv_bin_size_entry
{
2142 radv_compute_bin_size(struct radv_pipeline
*pipeline
, const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
2144 static const struct radv_bin_size_entry color_size_table
[][3][9] = {
2148 /* One shader engine */
2154 { UINT_MAX
, { 0, 0}},
2157 /* Two shader engines */
2163 { UINT_MAX
, { 0, 0}},
2166 /* Four shader engines */
2171 { UINT_MAX
, { 0, 0}},
2177 /* One shader engine */
2183 { UINT_MAX
, { 0, 0}},
2186 /* Two shader engines */
2192 { UINT_MAX
, { 0, 0}},
2195 /* Four shader engines */
2202 { UINT_MAX
, { 0, 0}},
2208 /* One shader engine */
2215 { UINT_MAX
, { 0, 0}},
2218 /* Two shader engines */
2226 { UINT_MAX
, { 0, 0}},
2229 /* Four shader engines */
2237 { UINT_MAX
, { 0, 0}},
2241 static const struct radv_bin_size_entry ds_size_table
[][3][9] = {
2245 // One shader engine
2252 { UINT_MAX
, { 0, 0}},
2255 // Two shader engines
2263 { UINT_MAX
, { 0, 0}},
2266 // Four shader engines
2274 { UINT_MAX
, { 0, 0}},
2280 // One shader engine
2288 { UINT_MAX
, { 0, 0}},
2291 // Two shader engines
2300 { UINT_MAX
, { 0, 0}},
2303 // Four shader engines
2312 { UINT_MAX
, { 0, 0}},
2318 // One shader engine
2326 { UINT_MAX
, { 0, 0}},
2329 // Two shader engines
2338 { UINT_MAX
, { 0, 0}},
2341 // Four shader engines
2349 { UINT_MAX
, { 0, 0}},
2354 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
2355 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
2356 VkExtent2D extent
= {512, 512};
2358 unsigned log_num_rb_per_se
=
2359 util_logbase2_ceil(pipeline
->device
->physical_device
->rad_info
.num_render_backends
/
2360 pipeline
->device
->physical_device
->rad_info
.max_se
);
2361 unsigned log_num_se
= util_logbase2_ceil(pipeline
->device
->physical_device
->rad_info
.max_se
);
2363 unsigned total_samples
= 1u << G_028BE0_MSAA_NUM_SAMPLES(pipeline
->graphics
.ms
.pa_sc_mode_cntl_1
);
2364 unsigned ps_iter_samples
= 1u << G_028804_PS_ITER_SAMPLES(pipeline
->graphics
.ms
.db_eqaa
);
2365 unsigned effective_samples
= total_samples
;
2366 unsigned color_bytes_per_pixel
= 0;
2368 const VkPipelineColorBlendStateCreateInfo
*vkblend
= pCreateInfo
->pColorBlendState
;
2370 for (unsigned i
= 0; i
< subpass
->color_count
; i
++) {
2371 if (!vkblend
->pAttachments
[i
].colorWriteMask
)
2374 if (subpass
->color_attachments
[i
].attachment
== VK_ATTACHMENT_UNUSED
)
2377 VkFormat format
= pass
->attachments
[subpass
->color_attachments
[i
].attachment
].format
;
2378 color_bytes_per_pixel
+= vk_format_get_blocksize(format
);
2381 /* MSAA images typically don't use all samples all the time. */
2382 if (effective_samples
>= 2 && ps_iter_samples
<= 1)
2383 effective_samples
= 2;
2384 color_bytes_per_pixel
*= effective_samples
;
2387 const struct radv_bin_size_entry
*color_entry
= color_size_table
[log_num_rb_per_se
][log_num_se
];
2388 while(color_entry
->bpp
<= color_bytes_per_pixel
)
2391 extent
= color_entry
->extent
;
2393 if (subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
2394 struct radv_render_pass_attachment
*attachment
= pass
->attachments
+ subpass
->depth_stencil_attachment
.attachment
;
2396 /* Coefficients taken from AMDVLK */
2397 unsigned depth_coeff
= vk_format_is_depth(attachment
->format
) ? 5 : 0;
2398 unsigned stencil_coeff
= vk_format_is_stencil(attachment
->format
) ? 1 : 0;
2399 unsigned ds_bytes_per_pixel
= 4 * (depth_coeff
+ stencil_coeff
) * total_samples
;
2401 const struct radv_bin_size_entry
*ds_entry
= ds_size_table
[log_num_rb_per_se
][log_num_se
];
2402 while(ds_entry
->bpp
<= ds_bytes_per_pixel
)
2405 extent
.width
= MIN2(extent
.width
, ds_entry
->extent
.width
);
2406 extent
.height
= MIN2(extent
.height
, ds_entry
->extent
.height
);
2413 radv_pipeline_generate_binning_state(struct radeon_winsys_cs
*cs
,
2414 struct radv_pipeline
*pipeline
,
2415 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
2417 if (pipeline
->device
->physical_device
->rad_info
.chip_class
< GFX9
)
2420 uint32_t pa_sc_binner_cntl_0
=
2421 S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_LEGACY_SC
) |
2422 S_028C44_DISABLE_START_OF_PRIM(1);
2423 uint32_t db_dfsm_control
= S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF
);
2425 VkExtent2D bin_size
= radv_compute_bin_size(pipeline
, pCreateInfo
);
2427 unsigned context_states_per_bin
; /* allowed range: [1, 6] */
2428 unsigned persistent_states_per_bin
; /* allowed range: [1, 32] */
2429 unsigned fpovs_per_batch
; /* allowed range: [0, 255], 0 = unlimited */
2431 switch (pipeline
->device
->physical_device
->rad_info
.family
) {
2434 context_states_per_bin
= 1;
2435 persistent_states_per_bin
= 1;
2436 fpovs_per_batch
= 63;
2439 context_states_per_bin
= 6;
2440 persistent_states_per_bin
= 32;
2441 fpovs_per_batch
= 63;
2444 unreachable("unhandled family while determining binning state.");
2447 if (pipeline
->device
->pbb_allowed
&& bin_size
.width
&& bin_size
.height
) {
2448 pa_sc_binner_cntl_0
=
2449 S_028C44_BINNING_MODE(V_028C44_BINNING_ALLOWED
) |
2450 S_028C44_BIN_SIZE_X(bin_size
.width
== 16) |
2451 S_028C44_BIN_SIZE_Y(bin_size
.height
== 16) |
2452 S_028C44_BIN_SIZE_X_EXTEND(util_logbase2(MAX2(bin_size
.width
, 32)) - 5) |
2453 S_028C44_BIN_SIZE_Y_EXTEND(util_logbase2(MAX2(bin_size
.height
, 32)) - 5) |
2454 S_028C44_CONTEXT_STATES_PER_BIN(context_states_per_bin
- 1) |
2455 S_028C44_PERSISTENT_STATES_PER_BIN(persistent_states_per_bin
- 1) |
2456 S_028C44_DISABLE_START_OF_PRIM(1) |
2457 S_028C44_FPOVS_PER_BATCH(fpovs_per_batch
) |
2458 S_028C44_OPTIMAL_BIN_SELECTION(1);
2461 radeon_set_context_reg(cs
, R_028C44_PA_SC_BINNER_CNTL_0
,
2462 pa_sc_binner_cntl_0
);
2463 radeon_set_context_reg(cs
, R_028060_DB_DFSM_CONTROL
,
2469 radv_pipeline_generate_depth_stencil_state(struct radeon_winsys_cs
*cs
,
2470 struct radv_pipeline
*pipeline
,
2471 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
2472 const struct radv_graphics_pipeline_create_info
*extra
)
2474 const VkPipelineDepthStencilStateCreateInfo
*vkds
= pCreateInfo
->pDepthStencilState
;
2475 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
2476 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
2477 struct radv_shader_variant
*ps
= pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
2478 struct radv_render_pass_attachment
*attachment
= NULL
;
2479 uint32_t db_depth_control
= 0, db_stencil_control
= 0;
2480 uint32_t db_render_control
= 0, db_render_override2
= 0;
2481 uint32_t db_render_override
= 0;
2483 if (subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
)
2484 attachment
= pass
->attachments
+ subpass
->depth_stencil_attachment
.attachment
;
2486 bool has_depth_attachment
= attachment
&& vk_format_is_depth(attachment
->format
);
2487 bool has_stencil_attachment
= attachment
&& vk_format_is_stencil(attachment
->format
);
2489 if (vkds
&& has_depth_attachment
) {
2490 db_depth_control
= S_028800_Z_ENABLE(vkds
->depthTestEnable
? 1 : 0) |
2491 S_028800_Z_WRITE_ENABLE(vkds
->depthWriteEnable
? 1 : 0) |
2492 S_028800_ZFUNC(vkds
->depthCompareOp
) |
2493 S_028800_DEPTH_BOUNDS_ENABLE(vkds
->depthBoundsTestEnable
? 1 : 0);
2495 /* from amdvlk: For 4xAA and 8xAA need to decompress on flush for better performance */
2496 db_render_override2
|= S_028010_DECOMPRESS_Z_ON_FLUSH(attachment
->samples
> 2);
2499 if (has_stencil_attachment
&& vkds
&& vkds
->stencilTestEnable
) {
2500 db_depth_control
|= S_028800_STENCIL_ENABLE(1) | S_028800_BACKFACE_ENABLE(1);
2501 db_depth_control
|= S_028800_STENCILFUNC(vkds
->front
.compareOp
);
2502 db_stencil_control
|= S_02842C_STENCILFAIL(si_translate_stencil_op(vkds
->front
.failOp
));
2503 db_stencil_control
|= S_02842C_STENCILZPASS(si_translate_stencil_op(vkds
->front
.passOp
));
2504 db_stencil_control
|= S_02842C_STENCILZFAIL(si_translate_stencil_op(vkds
->front
.depthFailOp
));
2506 db_depth_control
|= S_028800_STENCILFUNC_BF(vkds
->back
.compareOp
);
2507 db_stencil_control
|= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(vkds
->back
.failOp
));
2508 db_stencil_control
|= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(vkds
->back
.passOp
));
2509 db_stencil_control
|= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(vkds
->back
.depthFailOp
));
2512 if (attachment
&& extra
) {
2513 db_render_control
|= S_028000_DEPTH_CLEAR_ENABLE(extra
->db_depth_clear
);
2514 db_render_control
|= S_028000_STENCIL_CLEAR_ENABLE(extra
->db_stencil_clear
);
2516 db_render_control
|= S_028000_RESUMMARIZE_ENABLE(extra
->db_resummarize
);
2517 db_render_control
|= S_028000_DEPTH_COMPRESS_DISABLE(extra
->db_flush_depth_inplace
);
2518 db_render_control
|= S_028000_STENCIL_COMPRESS_DISABLE(extra
->db_flush_stencil_inplace
);
2519 db_render_override2
|= S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(extra
->db_depth_disable_expclear
);
2520 db_render_override2
|= S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(extra
->db_stencil_disable_expclear
);
2523 db_render_override
|= S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE
) |
2524 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE
);
2526 if (pipeline
->device
->enabled_extensions
.EXT_depth_range_unrestricted
&&
2527 !pCreateInfo
->pRasterizationState
->depthClampEnable
&&
2528 ps
->info
.info
.ps
.writes_z
) {
2529 /* From VK_EXT_depth_range_unrestricted spec:
2531 * "The behavior described in Primitive Clipping still applies.
2532 * If depth clamping is disabled the depth values are still
2533 * clipped to 0 ≤ zc ≤ wc before the viewport transform. If
2534 * depth clamping is enabled the above equation is ignored and
2535 * the depth values are instead clamped to the VkViewport
2536 * minDepth and maxDepth values, which in the case of this
2537 * extension can be outside of the 0.0 to 1.0 range."
2539 db_render_override
|= S_02800C_DISABLE_VIEWPORT_CLAMP(1);
2542 radeon_set_context_reg(cs
, R_028800_DB_DEPTH_CONTROL
, db_depth_control
);
2543 radeon_set_context_reg(cs
, R_02842C_DB_STENCIL_CONTROL
, db_stencil_control
);
2545 radeon_set_context_reg(cs
, R_028000_DB_RENDER_CONTROL
, db_render_control
);
2546 radeon_set_context_reg(cs
, R_02800C_DB_RENDER_OVERRIDE
, db_render_override
);
2547 radeon_set_context_reg(cs
, R_028010_DB_RENDER_OVERRIDE2
, db_render_override2
);
2551 radv_pipeline_generate_blend_state(struct radeon_winsys_cs
*cs
,
2552 struct radv_pipeline
*pipeline
,
2553 const struct radv_blend_state
*blend
)
2555 radeon_set_context_reg_seq(cs
, R_028780_CB_BLEND0_CONTROL
, 8);
2556 radeon_emit_array(cs
, blend
->cb_blend_control
,
2558 radeon_set_context_reg(cs
, R_028808_CB_COLOR_CONTROL
, blend
->cb_color_control
);
2559 radeon_set_context_reg(cs
, R_028B70_DB_ALPHA_TO_MASK
, blend
->db_alpha_to_mask
);
2561 if (pipeline
->device
->physical_device
->has_rbplus
) {
2563 radeon_set_context_reg_seq(cs
, R_028760_SX_MRT0_BLEND_OPT
, 8);
2564 radeon_emit_array(cs
, blend
->sx_mrt_blend_opt
, 8);
2567 radeon_set_context_reg(cs
, R_028714_SPI_SHADER_COL_FORMAT
, blend
->spi_shader_col_format
);
2569 radeon_set_context_reg(cs
, R_028238_CB_TARGET_MASK
, blend
->cb_target_mask
);
2570 radeon_set_context_reg(cs
, R_02823C_CB_SHADER_MASK
, blend
->cb_shader_mask
);
2572 pipeline
->graphics
.col_format
= blend
->spi_shader_col_format
;
2573 pipeline
->graphics
.cb_target_mask
= blend
->cb_target_mask
;
2578 radv_pipeline_generate_raster_state(struct radeon_winsys_cs
*cs
,
2579 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
2581 const VkPipelineRasterizationStateCreateInfo
*vkraster
= pCreateInfo
->pRasterizationState
;
2583 radeon_set_context_reg(cs
, R_028810_PA_CL_CLIP_CNTL
,
2584 S_028810_PS_UCP_MODE(3) |
2585 S_028810_DX_CLIP_SPACE_DEF(1) | // vulkan uses DX conventions.
2586 S_028810_ZCLIP_NEAR_DISABLE(vkraster
->depthClampEnable
? 1 : 0) |
2587 S_028810_ZCLIP_FAR_DISABLE(vkraster
->depthClampEnable
? 1 : 0) |
2588 S_028810_DX_RASTERIZATION_KILL(vkraster
->rasterizerDiscardEnable
? 1 : 0) |
2589 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1));
2591 radeon_set_context_reg(cs
, R_0286D4_SPI_INTERP_CONTROL_0
,
2592 S_0286D4_FLAT_SHADE_ENA(1) |
2593 S_0286D4_PNT_SPRITE_ENA(1) |
2594 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S
) |
2595 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T
) |
2596 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0
) |
2597 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1
) |
2598 S_0286D4_PNT_SPRITE_TOP_1(0)); /* vulkan is top to bottom - 1.0 at bottom */
2600 radeon_set_context_reg(cs
, R_028BE4_PA_SU_VTX_CNTL
,
2601 S_028BE4_PIX_CENTER(1) | // TODO verify
2602 S_028BE4_ROUND_MODE(V_028BE4_X_ROUND_TO_EVEN
) |
2603 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH
));
2605 radeon_set_context_reg(cs
, R_028814_PA_SU_SC_MODE_CNTL
,
2606 S_028814_FACE(vkraster
->frontFace
) |
2607 S_028814_CULL_FRONT(!!(vkraster
->cullMode
& VK_CULL_MODE_FRONT_BIT
)) |
2608 S_028814_CULL_BACK(!!(vkraster
->cullMode
& VK_CULL_MODE_BACK_BIT
)) |
2609 S_028814_POLY_MODE(vkraster
->polygonMode
!= VK_POLYGON_MODE_FILL
) |
2610 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(vkraster
->polygonMode
)) |
2611 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(vkraster
->polygonMode
)) |
2612 S_028814_POLY_OFFSET_FRONT_ENABLE(vkraster
->depthBiasEnable
? 1 : 0) |
2613 S_028814_POLY_OFFSET_BACK_ENABLE(vkraster
->depthBiasEnable
? 1 : 0) |
2614 S_028814_POLY_OFFSET_PARA_ENABLE(vkraster
->depthBiasEnable
? 1 : 0));
2619 radv_pipeline_generate_multisample_state(struct radeon_winsys_cs
*cs
,
2620 struct radv_pipeline
*pipeline
)
2622 struct radv_multisample_state
*ms
= &pipeline
->graphics
.ms
;
2624 radeon_set_context_reg_seq(cs
, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0
, 2);
2625 radeon_emit(cs
, ms
->pa_sc_aa_mask
[0]);
2626 radeon_emit(cs
, ms
->pa_sc_aa_mask
[1]);
2628 radeon_set_context_reg(cs
, R_028804_DB_EQAA
, ms
->db_eqaa
);
2629 radeon_set_context_reg(cs
, R_028A4C_PA_SC_MODE_CNTL_1
, ms
->pa_sc_mode_cntl_1
);
2631 if (pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.needs_sample_positions
) {
2633 struct radv_userdata_info
*loc
= radv_lookup_user_sgpr(pipeline
, MESA_SHADER_FRAGMENT
, AC_UD_PS_SAMPLE_POS_OFFSET
);
2634 uint32_t base_reg
= pipeline
->user_data_0
[MESA_SHADER_FRAGMENT
];
2635 if (loc
->sgpr_idx
== -1)
2637 assert(loc
->num_sgprs
== 1);
2638 assert(!loc
->indirect
);
2639 switch (pipeline
->graphics
.ms
.num_samples
) {
2657 radeon_set_sh_reg(cs
, base_reg
+ loc
->sgpr_idx
* 4, offset
);
2662 radv_pipeline_generate_vgt_gs_mode(struct radeon_winsys_cs
*cs
,
2663 const struct radv_pipeline
*pipeline
)
2665 const struct radv_vs_output_info
*outinfo
= get_vs_output_info(pipeline
);
2667 uint32_t vgt_primitiveid_en
= false;
2668 uint32_t vgt_gs_mode
= 0;
2670 if (radv_pipeline_has_gs(pipeline
)) {
2671 const struct radv_shader_variant
*gs
=
2672 pipeline
->shaders
[MESA_SHADER_GEOMETRY
];
2674 vgt_gs_mode
= ac_vgt_gs_mode(gs
->info
.gs
.vertices_out
,
2675 pipeline
->device
->physical_device
->rad_info
.chip_class
);
2676 } else if (outinfo
->export_prim_id
) {
2677 vgt_gs_mode
= S_028A40_MODE(V_028A40_GS_SCENARIO_A
);
2678 vgt_primitiveid_en
= true;
2681 radeon_set_context_reg(cs
, R_028A84_VGT_PRIMITIVEID_EN
, vgt_primitiveid_en
);
2682 radeon_set_context_reg(cs
, R_028A40_VGT_GS_MODE
, vgt_gs_mode
);
2686 radv_pipeline_generate_hw_vs(struct radeon_winsys_cs
*cs
,
2687 struct radv_pipeline
*pipeline
,
2688 struct radv_shader_variant
*shader
)
2690 uint64_t va
= radv_buffer_get_va(shader
->bo
) + shader
->bo_offset
;
2692 radeon_set_sh_reg_seq(cs
, R_00B120_SPI_SHADER_PGM_LO_VS
, 4);
2693 radeon_emit(cs
, va
>> 8);
2694 radeon_emit(cs
, S_00B124_MEM_BASE(va
>> 40));
2695 radeon_emit(cs
, shader
->rsrc1
);
2696 radeon_emit(cs
, shader
->rsrc2
);
2698 const struct radv_vs_output_info
*outinfo
= get_vs_output_info(pipeline
);
2699 unsigned clip_dist_mask
, cull_dist_mask
, total_mask
;
2700 clip_dist_mask
= outinfo
->clip_dist_mask
;
2701 cull_dist_mask
= outinfo
->cull_dist_mask
;
2702 total_mask
= clip_dist_mask
| cull_dist_mask
;
2703 bool misc_vec_ena
= outinfo
->writes_pointsize
||
2704 outinfo
->writes_layer
||
2705 outinfo
->writes_viewport_index
;
2707 radeon_set_context_reg(cs
, R_0286C4_SPI_VS_OUT_CONFIG
,
2708 S_0286C4_VS_EXPORT_COUNT(MAX2(1, outinfo
->param_exports
) - 1));
2710 radeon_set_context_reg(cs
, R_02870C_SPI_SHADER_POS_FORMAT
,
2711 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP
) |
2712 S_02870C_POS1_EXPORT_FORMAT(outinfo
->pos_exports
> 1 ?
2713 V_02870C_SPI_SHADER_4COMP
:
2714 V_02870C_SPI_SHADER_NONE
) |
2715 S_02870C_POS2_EXPORT_FORMAT(outinfo
->pos_exports
> 2 ?
2716 V_02870C_SPI_SHADER_4COMP
:
2717 V_02870C_SPI_SHADER_NONE
) |
2718 S_02870C_POS3_EXPORT_FORMAT(outinfo
->pos_exports
> 3 ?
2719 V_02870C_SPI_SHADER_4COMP
:
2720 V_02870C_SPI_SHADER_NONE
));
2722 radeon_set_context_reg(cs
, R_028818_PA_CL_VTE_CNTL
,
2723 S_028818_VTX_W0_FMT(1) |
2724 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
2725 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
2726 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
2728 radeon_set_context_reg(cs
, R_02881C_PA_CL_VS_OUT_CNTL
,
2729 S_02881C_USE_VTX_POINT_SIZE(outinfo
->writes_pointsize
) |
2730 S_02881C_USE_VTX_RENDER_TARGET_INDX(outinfo
->writes_layer
) |
2731 S_02881C_USE_VTX_VIEWPORT_INDX(outinfo
->writes_viewport_index
) |
2732 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena
) |
2733 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena
) |
2734 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask
& 0x0f) != 0) |
2735 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask
& 0xf0) != 0) |
2736 cull_dist_mask
<< 8 |
2739 if (pipeline
->device
->physical_device
->rad_info
.chip_class
<= VI
)
2740 radeon_set_context_reg(cs
, R_028AB4_VGT_REUSE_OFF
,
2741 outinfo
->writes_viewport_index
);
2745 radv_pipeline_generate_hw_es(struct radeon_winsys_cs
*cs
,
2746 struct radv_pipeline
*pipeline
,
2747 struct radv_shader_variant
*shader
)
2749 uint64_t va
= radv_buffer_get_va(shader
->bo
) + shader
->bo_offset
;
2751 radeon_set_sh_reg_seq(cs
, R_00B320_SPI_SHADER_PGM_LO_ES
, 4);
2752 radeon_emit(cs
, va
>> 8);
2753 radeon_emit(cs
, S_00B324_MEM_BASE(va
>> 40));
2754 radeon_emit(cs
, shader
->rsrc1
);
2755 radeon_emit(cs
, shader
->rsrc2
);
2759 radv_pipeline_generate_hw_ls(struct radeon_winsys_cs
*cs
,
2760 struct radv_pipeline
*pipeline
,
2761 struct radv_shader_variant
*shader
,
2762 const struct radv_tessellation_state
*tess
)
2764 uint64_t va
= radv_buffer_get_va(shader
->bo
) + shader
->bo_offset
;
2765 uint32_t rsrc2
= shader
->rsrc2
;
2767 radeon_set_sh_reg_seq(cs
, R_00B520_SPI_SHADER_PGM_LO_LS
, 2);
2768 radeon_emit(cs
, va
>> 8);
2769 radeon_emit(cs
, S_00B524_MEM_BASE(va
>> 40));
2771 rsrc2
|= S_00B52C_LDS_SIZE(tess
->lds_size
);
2772 if (pipeline
->device
->physical_device
->rad_info
.chip_class
== CIK
&&
2773 pipeline
->device
->physical_device
->rad_info
.family
!= CHIP_HAWAII
)
2774 radeon_set_sh_reg(cs
, R_00B52C_SPI_SHADER_PGM_RSRC2_LS
, rsrc2
);
2776 radeon_set_sh_reg_seq(cs
, R_00B528_SPI_SHADER_PGM_RSRC1_LS
, 2);
2777 radeon_emit(cs
, shader
->rsrc1
);
2778 radeon_emit(cs
, rsrc2
);
2782 radv_pipeline_generate_hw_hs(struct radeon_winsys_cs
*cs
,
2783 struct radv_pipeline
*pipeline
,
2784 struct radv_shader_variant
*shader
,
2785 const struct radv_tessellation_state
*tess
)
2787 uint64_t va
= radv_buffer_get_va(shader
->bo
) + shader
->bo_offset
;
2789 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2790 radeon_set_sh_reg_seq(cs
, R_00B410_SPI_SHADER_PGM_LO_LS
, 2);
2791 radeon_emit(cs
, va
>> 8);
2792 radeon_emit(cs
, S_00B414_MEM_BASE(va
>> 40));
2794 radeon_set_sh_reg_seq(cs
, R_00B428_SPI_SHADER_PGM_RSRC1_HS
, 2);
2795 radeon_emit(cs
, shader
->rsrc1
);
2796 radeon_emit(cs
, shader
->rsrc2
|
2797 S_00B42C_LDS_SIZE(tess
->lds_size
));
2799 radeon_set_sh_reg_seq(cs
, R_00B420_SPI_SHADER_PGM_LO_HS
, 4);
2800 radeon_emit(cs
, va
>> 8);
2801 radeon_emit(cs
, S_00B424_MEM_BASE(va
>> 40));
2802 radeon_emit(cs
, shader
->rsrc1
);
2803 radeon_emit(cs
, shader
->rsrc2
);
2808 radv_pipeline_generate_vertex_shader(struct radeon_winsys_cs
*cs
,
2809 struct radv_pipeline
*pipeline
,
2810 const struct radv_tessellation_state
*tess
)
2812 struct radv_shader_variant
*vs
;
2814 /* Skip shaders merged into HS/GS */
2815 vs
= pipeline
->shaders
[MESA_SHADER_VERTEX
];
2819 if (vs
->info
.vs
.as_ls
)
2820 radv_pipeline_generate_hw_ls(cs
, pipeline
, vs
, tess
);
2821 else if (vs
->info
.vs
.as_es
)
2822 radv_pipeline_generate_hw_es(cs
, pipeline
, vs
);
2824 radv_pipeline_generate_hw_vs(cs
, pipeline
, vs
);
2828 radv_pipeline_generate_tess_shaders(struct radeon_winsys_cs
*cs
,
2829 struct radv_pipeline
*pipeline
,
2830 const struct radv_tessellation_state
*tess
)
2832 if (!radv_pipeline_has_tess(pipeline
))
2835 struct radv_shader_variant
*tes
, *tcs
;
2837 tcs
= pipeline
->shaders
[MESA_SHADER_TESS_CTRL
];
2838 tes
= pipeline
->shaders
[MESA_SHADER_TESS_EVAL
];
2841 if (tes
->info
.tes
.as_es
)
2842 radv_pipeline_generate_hw_es(cs
, pipeline
, tes
);
2844 radv_pipeline_generate_hw_vs(cs
, pipeline
, tes
);
2847 radv_pipeline_generate_hw_hs(cs
, pipeline
, tcs
, tess
);
2849 radeon_set_context_reg(cs
, R_028B6C_VGT_TF_PARAM
,
2852 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= CIK
)
2853 radeon_set_context_reg_idx(cs
, R_028B58_VGT_LS_HS_CONFIG
, 2,
2854 tess
->ls_hs_config
);
2856 radeon_set_context_reg(cs
, R_028B58_VGT_LS_HS_CONFIG
,
2857 tess
->ls_hs_config
);
2861 radv_pipeline_generate_geometry_shader(struct radeon_winsys_cs
*cs
,
2862 struct radv_pipeline
*pipeline
,
2863 const struct radv_gs_state
*gs_state
)
2865 struct radv_shader_variant
*gs
;
2868 gs
= pipeline
->shaders
[MESA_SHADER_GEOMETRY
];
2872 uint32_t gsvs_itemsize
= gs
->info
.gs
.max_gsvs_emit_size
>> 2;
2874 radeon_set_context_reg_seq(cs
, R_028A60_VGT_GSVS_RING_OFFSET_1
, 3);
2875 radeon_emit(cs
, gsvs_itemsize
);
2876 radeon_emit(cs
, gsvs_itemsize
);
2877 radeon_emit(cs
, gsvs_itemsize
);
2879 radeon_set_context_reg(cs
, R_028AB0_VGT_GSVS_RING_ITEMSIZE
, gsvs_itemsize
);
2881 radeon_set_context_reg(cs
, R_028B38_VGT_GS_MAX_VERT_OUT
, gs
->info
.gs
.vertices_out
);
2883 uint32_t gs_vert_itemsize
= gs
->info
.gs
.gsvs_vertex_size
;
2884 radeon_set_context_reg_seq(cs
, R_028B5C_VGT_GS_VERT_ITEMSIZE
, 4);
2885 radeon_emit(cs
, gs_vert_itemsize
>> 2);
2890 uint32_t gs_num_invocations
= gs
->info
.gs
.invocations
;
2891 radeon_set_context_reg(cs
, R_028B90_VGT_GS_INSTANCE_CNT
,
2892 S_028B90_CNT(MIN2(gs_num_invocations
, 127)) |
2893 S_028B90_ENABLE(gs_num_invocations
> 0));
2895 radeon_set_context_reg(cs
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
2896 gs_state
->vgt_esgs_ring_itemsize
);
2898 va
= radv_buffer_get_va(gs
->bo
) + gs
->bo_offset
;
2900 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2901 radeon_set_sh_reg_seq(cs
, R_00B210_SPI_SHADER_PGM_LO_ES
, 2);
2902 radeon_emit(cs
, va
>> 8);
2903 radeon_emit(cs
, S_00B214_MEM_BASE(va
>> 40));
2905 radeon_set_sh_reg_seq(cs
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
, 2);
2906 radeon_emit(cs
, gs
->rsrc1
);
2907 radeon_emit(cs
, gs
->rsrc2
| S_00B22C_LDS_SIZE(gs_state
->lds_size
));
2909 radeon_set_context_reg(cs
, R_028A44_VGT_GS_ONCHIP_CNTL
, gs_state
->vgt_gs_onchip_cntl
);
2910 radeon_set_context_reg(cs
, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP
, gs_state
->vgt_gs_max_prims_per_subgroup
);
2912 radeon_set_sh_reg_seq(cs
, R_00B220_SPI_SHADER_PGM_LO_GS
, 4);
2913 radeon_emit(cs
, va
>> 8);
2914 radeon_emit(cs
, S_00B224_MEM_BASE(va
>> 40));
2915 radeon_emit(cs
, gs
->rsrc1
);
2916 radeon_emit(cs
, gs
->rsrc2
);
2919 radv_pipeline_generate_hw_vs(cs
, pipeline
, pipeline
->gs_copy_shader
);
2922 static uint32_t offset_to_ps_input(uint32_t offset
, bool flat_shade
)
2924 uint32_t ps_input_cntl
;
2925 if (offset
<= AC_EXP_PARAM_OFFSET_31
) {
2926 ps_input_cntl
= S_028644_OFFSET(offset
);
2928 ps_input_cntl
|= S_028644_FLAT_SHADE(1);
2930 /* The input is a DEFAULT_VAL constant. */
2931 assert(offset
>= AC_EXP_PARAM_DEFAULT_VAL_0000
&&
2932 offset
<= AC_EXP_PARAM_DEFAULT_VAL_1111
);
2933 offset
-= AC_EXP_PARAM_DEFAULT_VAL_0000
;
2934 ps_input_cntl
= S_028644_OFFSET(0x20) |
2935 S_028644_DEFAULT_VAL(offset
);
2937 return ps_input_cntl
;
2941 radv_pipeline_generate_ps_inputs(struct radeon_winsys_cs
*cs
,
2942 struct radv_pipeline
*pipeline
)
2944 struct radv_shader_variant
*ps
= pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
2945 const struct radv_vs_output_info
*outinfo
= get_vs_output_info(pipeline
);
2946 uint32_t ps_input_cntl
[32];
2948 unsigned ps_offset
= 0;
2950 if (ps
->info
.info
.ps
.prim_id_input
) {
2951 unsigned vs_offset
= outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
];
2952 if (vs_offset
!= AC_EXP_PARAM_UNDEFINED
) {
2953 ps_input_cntl
[ps_offset
] = offset_to_ps_input(vs_offset
, true);
2958 if (ps
->info
.info
.ps
.layer_input
||
2959 ps
->info
.info
.ps
.uses_input_attachments
||
2960 ps
->info
.info
.needs_multiview_view_index
) {
2961 unsigned vs_offset
= outinfo
->vs_output_param_offset
[VARYING_SLOT_LAYER
];
2962 if (vs_offset
!= AC_EXP_PARAM_UNDEFINED
)
2963 ps_input_cntl
[ps_offset
] = offset_to_ps_input(vs_offset
, true);
2965 ps_input_cntl
[ps_offset
] = offset_to_ps_input(AC_EXP_PARAM_DEFAULT_VAL_0000
, true);
2969 if (ps
->info
.info
.ps
.has_pcoord
) {
2971 val
= S_028644_PT_SPRITE_TEX(1) | S_028644_OFFSET(0x20);
2972 ps_input_cntl
[ps_offset
] = val
;
2976 for (unsigned i
= 0; i
< 32 && (1u << i
) <= ps
->info
.fs
.input_mask
; ++i
) {
2979 if (!(ps
->info
.fs
.input_mask
& (1u << i
)))
2982 vs_offset
= outinfo
->vs_output_param_offset
[VARYING_SLOT_VAR0
+ i
];
2983 if (vs_offset
== AC_EXP_PARAM_UNDEFINED
) {
2984 ps_input_cntl
[ps_offset
] = S_028644_OFFSET(0x20);
2989 flat_shade
= !!(ps
->info
.fs
.flat_shaded_mask
& (1u << ps_offset
));
2991 ps_input_cntl
[ps_offset
] = offset_to_ps_input(vs_offset
, flat_shade
);
2996 radeon_set_context_reg_seq(cs
, R_028644_SPI_PS_INPUT_CNTL_0
, ps_offset
);
2997 for (unsigned i
= 0; i
< ps_offset
; i
++) {
2998 radeon_emit(cs
, ps_input_cntl
[i
]);
3004 radv_compute_db_shader_control(const struct radv_device
*device
,
3005 const struct radv_shader_variant
*ps
)
3008 if (ps
->info
.fs
.early_fragment_test
|| !ps
->info
.info
.ps
.writes_memory
)
3009 z_order
= V_02880C_EARLY_Z_THEN_LATE_Z
;
3011 z_order
= V_02880C_LATE_Z
;
3013 bool disable_rbplus
= device
->physical_device
->has_rbplus
&&
3014 !device
->physical_device
->rbplus_allowed
;
3016 return S_02880C_Z_EXPORT_ENABLE(ps
->info
.info
.ps
.writes_z
) |
3017 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(ps
->info
.info
.ps
.writes_stencil
) |
3018 S_02880C_KILL_ENABLE(!!ps
->info
.fs
.can_discard
) |
3019 S_02880C_MASK_EXPORT_ENABLE(ps
->info
.info
.ps
.writes_sample_mask
) |
3020 S_02880C_Z_ORDER(z_order
) |
3021 S_02880C_DEPTH_BEFORE_SHADER(ps
->info
.fs
.early_fragment_test
) |
3022 S_02880C_EXEC_ON_HIER_FAIL(ps
->info
.info
.ps
.writes_memory
) |
3023 S_02880C_EXEC_ON_NOOP(ps
->info
.info
.ps
.writes_memory
) |
3024 S_02880C_DUAL_QUAD_DISABLE(disable_rbplus
);
3028 radv_pipeline_generate_fragment_shader(struct radeon_winsys_cs
*cs
,
3029 struct radv_pipeline
*pipeline
)
3031 struct radv_shader_variant
*ps
;
3033 assert (pipeline
->shaders
[MESA_SHADER_FRAGMENT
]);
3035 ps
= pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
3036 va
= radv_buffer_get_va(ps
->bo
) + ps
->bo_offset
;
3038 radeon_set_sh_reg_seq(cs
, R_00B020_SPI_SHADER_PGM_LO_PS
, 4);
3039 radeon_emit(cs
, va
>> 8);
3040 radeon_emit(cs
, S_00B024_MEM_BASE(va
>> 40));
3041 radeon_emit(cs
, ps
->rsrc1
);
3042 radeon_emit(cs
, ps
->rsrc2
);
3044 radeon_set_context_reg(cs
, R_02880C_DB_SHADER_CONTROL
,
3045 radv_compute_db_shader_control(pipeline
->device
, ps
));
3047 radeon_set_context_reg(cs
, R_0286CC_SPI_PS_INPUT_ENA
,
3048 ps
->config
.spi_ps_input_ena
);
3050 radeon_set_context_reg(cs
, R_0286D0_SPI_PS_INPUT_ADDR
,
3051 ps
->config
.spi_ps_input_addr
);
3053 radeon_set_context_reg(cs
, R_0286D8_SPI_PS_IN_CONTROL
,
3054 S_0286D8_NUM_INTERP(ps
->info
.fs
.num_interp
));
3056 radeon_set_context_reg(cs
, R_0286E0_SPI_BARYC_CNTL
, pipeline
->graphics
.spi_baryc_cntl
);
3058 radeon_set_context_reg(cs
, R_028710_SPI_SHADER_Z_FORMAT
,
3059 ac_get_spi_shader_z_format(ps
->info
.info
.ps
.writes_z
,
3060 ps
->info
.info
.ps
.writes_stencil
,
3061 ps
->info
.info
.ps
.writes_sample_mask
));
3063 if (pipeline
->device
->dfsm_allowed
) {
3064 /* optimise this? */
3065 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
3066 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_DFSM
) | EVENT_INDEX(0));
3071 radv_pipeline_generate_vgt_vertex_reuse(struct radeon_winsys_cs
*cs
,
3072 struct radv_pipeline
*pipeline
)
3074 if (pipeline
->device
->physical_device
->rad_info
.family
< CHIP_POLARIS10
)
3077 unsigned vtx_reuse_depth
= 30;
3078 if (radv_pipeline_has_tess(pipeline
) &&
3079 radv_get_tess_eval_shader(pipeline
)->info
.tes
.spacing
== TESS_SPACING_FRACTIONAL_ODD
) {
3080 vtx_reuse_depth
= 14;
3082 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
3083 S_028C58_VTX_REUSE_DEPTH(vtx_reuse_depth
));
3087 radv_compute_vgt_shader_stages_en(const struct radv_pipeline
*pipeline
)
3089 uint32_t stages
= 0;
3090 if (radv_pipeline_has_tess(pipeline
)) {
3091 stages
|= S_028B54_LS_EN(V_028B54_LS_STAGE_ON
) |
3092 S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
3094 if (radv_pipeline_has_gs(pipeline
))
3095 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_DS
) |
3097 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
3099 stages
|= S_028B54_VS_EN(V_028B54_VS_STAGE_DS
);
3101 } else if (radv_pipeline_has_gs(pipeline
))
3102 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL
) |
3104 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
3106 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
3107 stages
|= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
3113 radv_compute_cliprect_rule(const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
3115 const VkPipelineDiscardRectangleStateCreateInfoEXT
*discard_rectangle_info
=
3116 vk_find_struct_const(pCreateInfo
->pNext
, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT
);
3118 if (!discard_rectangle_info
)
3123 for (unsigned i
= 0; i
< (1u << MAX_DISCARD_RECTANGLES
); ++i
) {
3124 /* Interpret i as a bitmask, and then set the bit in the mask if
3125 * that combination of rectangles in which the pixel is contained
3126 * should pass the cliprect test. */
3127 unsigned relevant_subset
= i
& ((1u << discard_rectangle_info
->discardRectangleCount
) - 1);
3129 if (discard_rectangle_info
->discardRectangleMode
== VK_DISCARD_RECTANGLE_MODE_INCLUSIVE_EXT
&&
3133 if (discard_rectangle_info
->discardRectangleMode
== VK_DISCARD_RECTANGLE_MODE_EXCLUSIVE_EXT
&&
3144 radv_pipeline_generate_pm4(struct radv_pipeline
*pipeline
,
3145 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
3146 const struct radv_graphics_pipeline_create_info
*extra
,
3147 const struct radv_blend_state
*blend
,
3148 const struct radv_tessellation_state
*tess
,
3149 const struct radv_gs_state
*gs
,
3150 unsigned prim
, unsigned gs_out
)
3152 pipeline
->cs
.buf
= malloc(4 * 256);
3153 pipeline
->cs
.max_dw
= 256;
3155 radv_pipeline_generate_depth_stencil_state(&pipeline
->cs
, pipeline
, pCreateInfo
, extra
);
3156 radv_pipeline_generate_blend_state(&pipeline
->cs
, pipeline
, blend
);
3157 radv_pipeline_generate_raster_state(&pipeline
->cs
, pCreateInfo
);
3158 radv_pipeline_generate_multisample_state(&pipeline
->cs
, pipeline
);
3159 radv_pipeline_generate_vgt_gs_mode(&pipeline
->cs
, pipeline
);
3160 radv_pipeline_generate_vertex_shader(&pipeline
->cs
, pipeline
, tess
);
3161 radv_pipeline_generate_tess_shaders(&pipeline
->cs
, pipeline
, tess
);
3162 radv_pipeline_generate_geometry_shader(&pipeline
->cs
, pipeline
, gs
);
3163 radv_pipeline_generate_fragment_shader(&pipeline
->cs
, pipeline
);
3164 radv_pipeline_generate_ps_inputs(&pipeline
->cs
, pipeline
);
3165 radv_pipeline_generate_vgt_vertex_reuse(&pipeline
->cs
, pipeline
);
3166 radv_pipeline_generate_binning_state(&pipeline
->cs
, pipeline
, pCreateInfo
);
3168 radeon_set_context_reg(&pipeline
->cs
, R_0286E8_SPI_TMPRING_SIZE
,
3169 S_0286E8_WAVES(pipeline
->max_waves
) |
3170 S_0286E8_WAVESIZE(pipeline
->scratch_bytes_per_wave
>> 10));
3172 radeon_set_context_reg(&pipeline
->cs
, R_028B54_VGT_SHADER_STAGES_EN
, radv_compute_vgt_shader_stages_en(pipeline
));
3174 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
3175 radeon_set_uconfig_reg_idx(&pipeline
->cs
, R_030908_VGT_PRIMITIVE_TYPE
, 1, prim
);
3177 radeon_set_config_reg(&pipeline
->cs
, R_008958_VGT_PRIMITIVE_TYPE
, prim
);
3179 radeon_set_context_reg(&pipeline
->cs
, R_028A6C_VGT_GS_OUT_PRIM_TYPE
, gs_out
);
3181 radeon_set_context_reg(&pipeline
->cs
, R_02820C_PA_SC_CLIPRECT_RULE
, radv_compute_cliprect_rule(pCreateInfo
));
3183 assert(pipeline
->cs
.cdw
<= pipeline
->cs
.max_dw
);
3186 static struct radv_ia_multi_vgt_param_helpers
3187 radv_compute_ia_multi_vgt_param_helpers(struct radv_pipeline
*pipeline
,
3188 const struct radv_tessellation_state
*tess
,
3191 struct radv_ia_multi_vgt_param_helpers ia_multi_vgt_param
= {0};
3192 const struct radv_device
*device
= pipeline
->device
;
3194 if (radv_pipeline_has_tess(pipeline
))
3195 ia_multi_vgt_param
.primgroup_size
= tess
->num_patches
;
3196 else if (radv_pipeline_has_gs(pipeline
))
3197 ia_multi_vgt_param
.primgroup_size
= 64;
3199 ia_multi_vgt_param
.primgroup_size
= 128; /* recommended without a GS */
3201 ia_multi_vgt_param
.partial_es_wave
= false;
3202 if (pipeline
->device
->has_distributed_tess
) {
3203 if (radv_pipeline_has_gs(pipeline
)) {
3204 if (device
->physical_device
->rad_info
.chip_class
<= VI
)
3205 ia_multi_vgt_param
.partial_es_wave
= true;
3208 /* GS requirement. */
3209 if (SI_GS_PER_ES
/ ia_multi_vgt_param
.primgroup_size
>= pipeline
->device
->gs_table_depth
- 3)
3210 ia_multi_vgt_param
.partial_es_wave
= true;
3212 ia_multi_vgt_param
.wd_switch_on_eop
= false;
3213 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
3214 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
3215 * 4 shader engines. Set 1 to pass the assertion below.
3216 * The other cases are hardware requirements. */
3217 if (device
->physical_device
->rad_info
.max_se
< 4 ||
3218 prim
== V_008958_DI_PT_POLYGON
||
3219 prim
== V_008958_DI_PT_LINELOOP
||
3220 prim
== V_008958_DI_PT_TRIFAN
||
3221 prim
== V_008958_DI_PT_TRISTRIP_ADJ
||
3222 (pipeline
->graphics
.prim_restart_enable
&&
3223 (device
->physical_device
->rad_info
.family
< CHIP_POLARIS10
||
3224 (prim
!= V_008958_DI_PT_POINTLIST
&&
3225 prim
!= V_008958_DI_PT_LINESTRIP
&&
3226 prim
!= V_008958_DI_PT_TRISTRIP
))))
3227 ia_multi_vgt_param
.wd_switch_on_eop
= true;
3230 ia_multi_vgt_param
.ia_switch_on_eoi
= false;
3231 if (pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.prim_id_input
)
3232 ia_multi_vgt_param
.ia_switch_on_eoi
= true;
3233 if (radv_pipeline_has_gs(pipeline
) &&
3234 pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
.info
.uses_prim_id
)
3235 ia_multi_vgt_param
.ia_switch_on_eoi
= true;
3236 if (radv_pipeline_has_tess(pipeline
)) {
3237 /* SWITCH_ON_EOI must be set if PrimID is used. */
3238 if (pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.info
.uses_prim_id
||
3239 radv_get_tess_eval_shader(pipeline
)->info
.info
.uses_prim_id
)
3240 ia_multi_vgt_param
.ia_switch_on_eoi
= true;
3243 ia_multi_vgt_param
.partial_vs_wave
= false;
3244 if (radv_pipeline_has_tess(pipeline
)) {
3245 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
3246 if ((device
->physical_device
->rad_info
.family
== CHIP_TAHITI
||
3247 device
->physical_device
->rad_info
.family
== CHIP_PITCAIRN
||
3248 device
->physical_device
->rad_info
.family
== CHIP_BONAIRE
) &&
3249 radv_pipeline_has_gs(pipeline
))
3250 ia_multi_vgt_param
.partial_vs_wave
= true;
3251 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
3252 if (device
->has_distributed_tess
) {
3253 if (radv_pipeline_has_gs(pipeline
)) {
3254 if (device
->physical_device
->rad_info
.family
== CHIP_TONGA
||
3255 device
->physical_device
->rad_info
.family
== CHIP_FIJI
||
3256 device
->physical_device
->rad_info
.family
== CHIP_POLARIS10
||
3257 device
->physical_device
->rad_info
.family
== CHIP_POLARIS11
||
3258 device
->physical_device
->rad_info
.family
== CHIP_POLARIS12
)
3259 ia_multi_vgt_param
.partial_vs_wave
= true;
3261 ia_multi_vgt_param
.partial_vs_wave
= true;
3266 ia_multi_vgt_param
.base
=
3267 S_028AA8_PRIMGROUP_SIZE(ia_multi_vgt_param
.primgroup_size
- 1) |
3268 /* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
3269 S_028AA8_MAX_PRIMGRP_IN_WAVE(device
->physical_device
->rad_info
.chip_class
== VI
? 2 : 0) |
3270 S_030960_EN_INST_OPT_BASIC(device
->physical_device
->rad_info
.chip_class
>= GFX9
) |
3271 S_030960_EN_INST_OPT_ADV(device
->physical_device
->rad_info
.chip_class
>= GFX9
);
3273 return ia_multi_vgt_param
;
3278 radv_compute_vertex_input_state(struct radv_pipeline
*pipeline
,
3279 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
3281 const VkPipelineVertexInputStateCreateInfo
*vi_info
=
3282 pCreateInfo
->pVertexInputState
;
3283 struct radv_vertex_elements_info
*velems
= &pipeline
->vertex_elements
;
3285 for (uint32_t i
= 0; i
< vi_info
->vertexAttributeDescriptionCount
; i
++) {
3286 const VkVertexInputAttributeDescription
*desc
=
3287 &vi_info
->pVertexAttributeDescriptions
[i
];
3288 unsigned loc
= desc
->location
;
3289 const struct vk_format_description
*format_desc
;
3291 uint32_t num_format
, data_format
;
3292 format_desc
= vk_format_description(desc
->format
);
3293 first_non_void
= vk_format_get_first_non_void_channel(desc
->format
);
3295 num_format
= radv_translate_buffer_numformat(format_desc
, first_non_void
);
3296 data_format
= radv_translate_buffer_dataformat(format_desc
, first_non_void
);
3298 velems
->rsrc_word3
[loc
] = S_008F0C_DST_SEL_X(si_map_swizzle(format_desc
->swizzle
[0])) |
3299 S_008F0C_DST_SEL_Y(si_map_swizzle(format_desc
->swizzle
[1])) |
3300 S_008F0C_DST_SEL_Z(si_map_swizzle(format_desc
->swizzle
[2])) |
3301 S_008F0C_DST_SEL_W(si_map_swizzle(format_desc
->swizzle
[3])) |
3302 S_008F0C_NUM_FORMAT(num_format
) |
3303 S_008F0C_DATA_FORMAT(data_format
);
3304 velems
->format_size
[loc
] = format_desc
->block
.bits
/ 8;
3305 velems
->offset
[loc
] = desc
->offset
;
3306 velems
->binding
[loc
] = desc
->binding
;
3307 velems
->count
= MAX2(velems
->count
, loc
+ 1);
3310 for (uint32_t i
= 0; i
< vi_info
->vertexBindingDescriptionCount
; i
++) {
3311 const VkVertexInputBindingDescription
*desc
=
3312 &vi_info
->pVertexBindingDescriptions
[i
];
3314 pipeline
->binding_stride
[desc
->binding
] = desc
->stride
;
3319 radv_pipeline_init(struct radv_pipeline
*pipeline
,
3320 struct radv_device
*device
,
3321 struct radv_pipeline_cache
*cache
,
3322 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
3323 const struct radv_graphics_pipeline_create_info
*extra
,
3324 const VkAllocationCallbacks
*alloc
)
3327 bool has_view_index
= false;
3329 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
3330 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
3331 if (subpass
->view_mask
)
3332 has_view_index
= true;
3334 alloc
= &device
->alloc
;
3336 pipeline
->device
= device
;
3337 pipeline
->layout
= radv_pipeline_layout_from_handle(pCreateInfo
->layout
);
3338 assert(pipeline
->layout
);
3340 struct radv_blend_state blend
= radv_pipeline_init_blend_state(pipeline
, pCreateInfo
, extra
);
3342 const VkPipelineShaderStageCreateInfo
*pStages
[MESA_SHADER_STAGES
] = { 0, };
3343 for (uint32_t i
= 0; i
< pCreateInfo
->stageCount
; i
++) {
3344 gl_shader_stage stage
= ffs(pCreateInfo
->pStages
[i
].stage
) - 1;
3345 pStages
[stage
] = &pCreateInfo
->pStages
[i
];
3348 radv_create_shaders(pipeline
, device
, cache
,
3349 radv_generate_graphics_pipeline_key(pipeline
, pCreateInfo
, &blend
, has_view_index
),
3352 pipeline
->graphics
.spi_baryc_cntl
= S_0286E0_FRONT_FACE_ALL_BITS(1);
3353 radv_pipeline_init_multisample_state(pipeline
, &blend
, pCreateInfo
);
3355 uint32_t prim
= si_translate_prim(pCreateInfo
->pInputAssemblyState
->topology
);
3357 pipeline
->graphics
.can_use_guardband
= radv_prim_can_use_guardband(pCreateInfo
->pInputAssemblyState
->topology
);
3359 if (radv_pipeline_has_gs(pipeline
)) {
3360 gs_out
= si_conv_gl_prim_to_gs_out(pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
.gs
.output_prim
);
3361 pipeline
->graphics
.can_use_guardband
= gs_out
== V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
3363 gs_out
= si_conv_prim_to_gs_out(pCreateInfo
->pInputAssemblyState
->topology
);
3365 if (extra
&& extra
->use_rectlist
) {
3366 prim
= V_008958_DI_PT_RECTLIST
;
3367 gs_out
= V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
3368 pipeline
->graphics
.can_use_guardband
= true;
3370 pipeline
->graphics
.prim_restart_enable
= !!pCreateInfo
->pInputAssemblyState
->primitiveRestartEnable
;
3371 /* prim vertex count will need TESS changes */
3372 pipeline
->graphics
.prim_vertex_count
= prim_size_table
[prim
];
3374 radv_pipeline_init_dynamic_state(pipeline
, pCreateInfo
);
3376 /* Ensure that some export memory is always allocated, for two reasons:
3378 * 1) Correctness: The hardware ignores the EXEC mask if no export
3379 * memory is allocated, so KILL and alpha test do not work correctly
3381 * 2) Performance: Every shader needs at least a NULL export, even when
3382 * it writes no color/depth output. The NULL export instruction
3383 * stalls without this setting.
3385 * Don't add this to CB_SHADER_MASK.
3387 struct radv_shader_variant
*ps
= pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
3388 if (!blend
.spi_shader_col_format
) {
3389 if (!ps
->info
.info
.ps
.writes_z
&&
3390 !ps
->info
.info
.ps
.writes_stencil
&&
3391 !ps
->info
.info
.ps
.writes_sample_mask
)
3392 blend
.spi_shader_col_format
= V_028714_SPI_SHADER_32_R
;
3395 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
3396 if (pipeline
->shaders
[i
]) {
3397 pipeline
->need_indirect_descriptor_sets
|= pipeline
->shaders
[i
]->info
.need_indirect_descriptor_sets
;
3401 struct radv_gs_state gs
= {0};
3402 if (radv_pipeline_has_gs(pipeline
)) {
3403 gs
= calculate_gs_info(pCreateInfo
, pipeline
);
3404 calculate_gs_ring_sizes(pipeline
, &gs
);
3407 struct radv_tessellation_state tess
= {0};
3408 if (radv_pipeline_has_tess(pipeline
)) {
3409 if (prim
== V_008958_DI_PT_PATCH
) {
3410 pipeline
->graphics
.prim_vertex_count
.min
= pCreateInfo
->pTessellationState
->patchControlPoints
;
3411 pipeline
->graphics
.prim_vertex_count
.incr
= 1;
3413 tess
= calculate_tess_state(pipeline
, pCreateInfo
);
3416 pipeline
->graphics
.ia_multi_vgt_param
= radv_compute_ia_multi_vgt_param_helpers(pipeline
, &tess
, prim
);
3418 radv_compute_vertex_input_state(pipeline
, pCreateInfo
);
3420 for (uint32_t i
= 0; i
< MESA_SHADER_STAGES
; i
++)
3421 pipeline
->user_data_0
[i
] = radv_pipeline_stage_to_user_data_0(pipeline
, i
, device
->physical_device
->rad_info
.chip_class
);
3423 struct radv_userdata_info
*loc
= radv_lookup_user_sgpr(pipeline
, MESA_SHADER_VERTEX
,
3424 AC_UD_VS_BASE_VERTEX_START_INSTANCE
);
3425 if (loc
->sgpr_idx
!= -1) {
3426 pipeline
->graphics
.vtx_base_sgpr
= pipeline
->user_data_0
[MESA_SHADER_VERTEX
];
3427 pipeline
->graphics
.vtx_base_sgpr
+= loc
->sgpr_idx
* 4;
3428 if (radv_get_vertex_shader(pipeline
)->info
.info
.vs
.needs_draw_id
)
3429 pipeline
->graphics
.vtx_emit_num
= 3;
3431 pipeline
->graphics
.vtx_emit_num
= 2;
3434 result
= radv_pipeline_scratch_init(device
, pipeline
);
3435 radv_pipeline_generate_pm4(pipeline
, pCreateInfo
, extra
, &blend
, &tess
, &gs
, prim
, gs_out
);
3441 radv_graphics_pipeline_create(
3443 VkPipelineCache _cache
,
3444 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
3445 const struct radv_graphics_pipeline_create_info
*extra
,
3446 const VkAllocationCallbacks
*pAllocator
,
3447 VkPipeline
*pPipeline
)
3449 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3450 RADV_FROM_HANDLE(radv_pipeline_cache
, cache
, _cache
);
3451 struct radv_pipeline
*pipeline
;
3454 pipeline
= vk_zalloc2(&device
->alloc
, pAllocator
, sizeof(*pipeline
), 8,
3455 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3456 if (pipeline
== NULL
)
3457 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3459 result
= radv_pipeline_init(pipeline
, device
, cache
,
3460 pCreateInfo
, extra
, pAllocator
);
3461 if (result
!= VK_SUCCESS
) {
3462 radv_pipeline_destroy(device
, pipeline
, pAllocator
);
3466 *pPipeline
= radv_pipeline_to_handle(pipeline
);
3471 VkResult
radv_CreateGraphicsPipelines(
3473 VkPipelineCache pipelineCache
,
3475 const VkGraphicsPipelineCreateInfo
* pCreateInfos
,
3476 const VkAllocationCallbacks
* pAllocator
,
3477 VkPipeline
* pPipelines
)
3479 VkResult result
= VK_SUCCESS
;
3482 for (; i
< count
; i
++) {
3484 r
= radv_graphics_pipeline_create(_device
,
3487 NULL
, pAllocator
, &pPipelines
[i
]);
3488 if (r
!= VK_SUCCESS
) {
3490 pPipelines
[i
] = VK_NULL_HANDLE
;
3499 radv_compute_generate_pm4(struct radv_pipeline
*pipeline
)
3501 struct radv_shader_variant
*compute_shader
;
3502 struct radv_device
*device
= pipeline
->device
;
3503 unsigned compute_resource_limits
;
3504 unsigned waves_per_threadgroup
;
3507 pipeline
->cs
.buf
= malloc(20 * 4);
3508 pipeline
->cs
.max_dw
= 20;
3510 compute_shader
= pipeline
->shaders
[MESA_SHADER_COMPUTE
];
3511 va
= radv_buffer_get_va(compute_shader
->bo
) + compute_shader
->bo_offset
;
3513 radeon_set_sh_reg_seq(&pipeline
->cs
, R_00B830_COMPUTE_PGM_LO
, 2);
3514 radeon_emit(&pipeline
->cs
, va
>> 8);
3515 radeon_emit(&pipeline
->cs
, S_00B834_DATA(va
>> 40));
3517 radeon_set_sh_reg_seq(&pipeline
->cs
, R_00B848_COMPUTE_PGM_RSRC1
, 2);
3518 radeon_emit(&pipeline
->cs
, compute_shader
->rsrc1
);
3519 radeon_emit(&pipeline
->cs
, compute_shader
->rsrc2
);
3521 radeon_set_sh_reg(&pipeline
->cs
, R_00B860_COMPUTE_TMPRING_SIZE
,
3522 S_00B860_WAVES(pipeline
->max_waves
) |
3523 S_00B860_WAVESIZE(pipeline
->scratch_bytes_per_wave
>> 10));
3525 /* Calculate best compute resource limits. */
3526 waves_per_threadgroup
=
3527 DIV_ROUND_UP(compute_shader
->info
.cs
.block_size
[0] *
3528 compute_shader
->info
.cs
.block_size
[1] *
3529 compute_shader
->info
.cs
.block_size
[2], 64);
3530 compute_resource_limits
=
3531 S_00B854_SIMD_DEST_CNTL(waves_per_threadgroup
% 4 == 0);
3533 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
3534 unsigned num_cu_per_se
=
3535 device
->physical_device
->rad_info
.num_good_compute_units
/
3536 device
->physical_device
->rad_info
.max_se
;
3538 /* Force even distribution on all SIMDs in CU if the workgroup
3539 * size is 64. This has shown some good improvements if # of
3540 * CUs per SE is not a multiple of 4.
3542 if (num_cu_per_se
% 4 && waves_per_threadgroup
== 1)
3543 compute_resource_limits
|= S_00B854_FORCE_SIMD_DIST(1);
3546 radeon_set_sh_reg(&pipeline
->cs
, R_00B854_COMPUTE_RESOURCE_LIMITS
,
3547 compute_resource_limits
);
3549 radeon_set_sh_reg_seq(&pipeline
->cs
, R_00B81C_COMPUTE_NUM_THREAD_X
, 3);
3550 radeon_emit(&pipeline
->cs
,
3551 S_00B81C_NUM_THREAD_FULL(compute_shader
->info
.cs
.block_size
[0]));
3552 radeon_emit(&pipeline
->cs
,
3553 S_00B81C_NUM_THREAD_FULL(compute_shader
->info
.cs
.block_size
[1]));
3554 radeon_emit(&pipeline
->cs
,
3555 S_00B81C_NUM_THREAD_FULL(compute_shader
->info
.cs
.block_size
[2]));
3557 assert(pipeline
->cs
.cdw
<= pipeline
->cs
.max_dw
);
3560 static VkResult
radv_compute_pipeline_create(
3562 VkPipelineCache _cache
,
3563 const VkComputePipelineCreateInfo
* pCreateInfo
,
3564 const VkAllocationCallbacks
* pAllocator
,
3565 VkPipeline
* pPipeline
)
3567 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3568 RADV_FROM_HANDLE(radv_pipeline_cache
, cache
, _cache
);
3569 const VkPipelineShaderStageCreateInfo
*pStages
[MESA_SHADER_STAGES
] = { 0, };
3570 struct radv_pipeline
*pipeline
;
3573 pipeline
= vk_zalloc2(&device
->alloc
, pAllocator
, sizeof(*pipeline
), 8,
3574 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3575 if (pipeline
== NULL
)
3576 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3578 pipeline
->device
= device
;
3579 pipeline
->layout
= radv_pipeline_layout_from_handle(pCreateInfo
->layout
);
3580 assert(pipeline
->layout
);
3582 pStages
[MESA_SHADER_COMPUTE
] = &pCreateInfo
->stage
;
3583 radv_create_shaders(pipeline
, device
, cache
, (struct radv_pipeline_key
) {0}, pStages
);
3585 pipeline
->user_data_0
[MESA_SHADER_COMPUTE
] = radv_pipeline_stage_to_user_data_0(pipeline
, MESA_SHADER_COMPUTE
, device
->physical_device
->rad_info
.chip_class
);
3586 pipeline
->need_indirect_descriptor_sets
|= pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.need_indirect_descriptor_sets
;
3587 result
= radv_pipeline_scratch_init(device
, pipeline
);
3588 if (result
!= VK_SUCCESS
) {
3589 radv_pipeline_destroy(device
, pipeline
, pAllocator
);
3593 radv_compute_generate_pm4(pipeline
);
3595 *pPipeline
= radv_pipeline_to_handle(pipeline
);
3600 VkResult
radv_CreateComputePipelines(
3602 VkPipelineCache pipelineCache
,
3604 const VkComputePipelineCreateInfo
* pCreateInfos
,
3605 const VkAllocationCallbacks
* pAllocator
,
3606 VkPipeline
* pPipelines
)
3608 VkResult result
= VK_SUCCESS
;
3611 for (; i
< count
; i
++) {
3613 r
= radv_compute_pipeline_create(_device
, pipelineCache
,
3615 pAllocator
, &pPipelines
[i
]);
3616 if (r
!= VK_SUCCESS
) {
3618 pPipelines
[i
] = VK_NULL_HANDLE
;