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"
52 struct radv_blend_state
{
53 uint32_t cb_color_control
;
54 uint32_t cb_target_mask
;
55 uint32_t sx_mrt_blend_opt
[8];
56 uint32_t cb_blend_control
[8];
58 uint32_t spi_shader_col_format
;
59 uint32_t cb_shader_mask
;
60 uint32_t db_alpha_to_mask
;
63 struct radv_tessellation_state
{
64 uint32_t ls_hs_config
;
65 uint32_t tcs_in_layout
;
66 uint32_t tcs_out_layout
;
67 uint32_t tcs_out_offsets
;
68 uint32_t offchip_layout
;
74 struct radv_gs_state
{
75 uint32_t vgt_gs_onchip_cntl
;
76 uint32_t vgt_gs_max_prims_per_subgroup
;
77 uint32_t vgt_esgs_ring_itemsize
;
82 radv_pipeline_destroy(struct radv_device
*device
,
83 struct radv_pipeline
*pipeline
,
84 const VkAllocationCallbacks
* allocator
)
86 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; ++i
)
87 if (pipeline
->shaders
[i
])
88 radv_shader_variant_destroy(device
, pipeline
->shaders
[i
]);
90 if (pipeline
->gs_copy_shader
)
91 radv_shader_variant_destroy(device
, pipeline
->gs_copy_shader
);
94 free(pipeline
->cs
.buf
);
95 vk_free2(&device
->alloc
, allocator
, pipeline
);
98 void radv_DestroyPipeline(
100 VkPipeline _pipeline
,
101 const VkAllocationCallbacks
* pAllocator
)
103 RADV_FROM_HANDLE(radv_device
, device
, _device
);
104 RADV_FROM_HANDLE(radv_pipeline
, pipeline
, _pipeline
);
109 radv_pipeline_destroy(device
, pipeline
, pAllocator
);
112 static uint32_t get_hash_flags(struct radv_device
*device
)
114 uint32_t hash_flags
= 0;
116 if (device
->instance
->debug_flags
& RADV_DEBUG_UNSAFE_MATH
)
117 hash_flags
|= RADV_HASH_SHADER_UNSAFE_MATH
;
118 if (device
->instance
->perftest_flags
& RADV_PERFTEST_SISCHED
)
119 hash_flags
|= RADV_HASH_SHADER_SISCHED
;
124 radv_pipeline_scratch_init(struct radv_device
*device
,
125 struct radv_pipeline
*pipeline
)
127 unsigned scratch_bytes_per_wave
= 0;
128 unsigned max_waves
= 0;
129 unsigned min_waves
= 1;
131 for (int i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
132 if (pipeline
->shaders
[i
]) {
133 unsigned max_stage_waves
= device
->scratch_waves
;
135 scratch_bytes_per_wave
= MAX2(scratch_bytes_per_wave
,
136 pipeline
->shaders
[i
]->config
.scratch_bytes_per_wave
);
138 max_stage_waves
= MIN2(max_stage_waves
,
139 4 * device
->physical_device
->rad_info
.num_good_compute_units
*
140 (256 / pipeline
->shaders
[i
]->config
.num_vgprs
));
141 max_waves
= MAX2(max_waves
, max_stage_waves
);
145 if (pipeline
->shaders
[MESA_SHADER_COMPUTE
]) {
146 unsigned group_size
= pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.cs
.block_size
[0] *
147 pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.cs
.block_size
[1] *
148 pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.cs
.block_size
[2];
149 min_waves
= MAX2(min_waves
, round_up_u32(group_size
, 64));
152 if (scratch_bytes_per_wave
)
153 max_waves
= MIN2(max_waves
, 0xffffffffu
/ scratch_bytes_per_wave
);
155 if (scratch_bytes_per_wave
&& max_waves
< min_waves
) {
156 /* Not really true at this moment, but will be true on first
157 * execution. Avoid having hanging shaders. */
158 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
160 pipeline
->scratch_bytes_per_wave
= scratch_bytes_per_wave
;
161 pipeline
->max_waves
= max_waves
;
165 static uint32_t si_translate_blend_function(VkBlendOp op
)
168 case VK_BLEND_OP_ADD
:
169 return V_028780_COMB_DST_PLUS_SRC
;
170 case VK_BLEND_OP_SUBTRACT
:
171 return V_028780_COMB_SRC_MINUS_DST
;
172 case VK_BLEND_OP_REVERSE_SUBTRACT
:
173 return V_028780_COMB_DST_MINUS_SRC
;
174 case VK_BLEND_OP_MIN
:
175 return V_028780_COMB_MIN_DST_SRC
;
176 case VK_BLEND_OP_MAX
:
177 return V_028780_COMB_MAX_DST_SRC
;
183 static uint32_t si_translate_blend_factor(VkBlendFactor factor
)
186 case VK_BLEND_FACTOR_ZERO
:
187 return V_028780_BLEND_ZERO
;
188 case VK_BLEND_FACTOR_ONE
:
189 return V_028780_BLEND_ONE
;
190 case VK_BLEND_FACTOR_SRC_COLOR
:
191 return V_028780_BLEND_SRC_COLOR
;
192 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR
:
193 return V_028780_BLEND_ONE_MINUS_SRC_COLOR
;
194 case VK_BLEND_FACTOR_DST_COLOR
:
195 return V_028780_BLEND_DST_COLOR
;
196 case VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR
:
197 return V_028780_BLEND_ONE_MINUS_DST_COLOR
;
198 case VK_BLEND_FACTOR_SRC_ALPHA
:
199 return V_028780_BLEND_SRC_ALPHA
;
200 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
:
201 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA
;
202 case VK_BLEND_FACTOR_DST_ALPHA
:
203 return V_028780_BLEND_DST_ALPHA
;
204 case VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA
:
205 return V_028780_BLEND_ONE_MINUS_DST_ALPHA
;
206 case VK_BLEND_FACTOR_CONSTANT_COLOR
:
207 return V_028780_BLEND_CONSTANT_COLOR
;
208 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR
:
209 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR
;
210 case VK_BLEND_FACTOR_CONSTANT_ALPHA
:
211 return V_028780_BLEND_CONSTANT_ALPHA
;
212 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA
:
213 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA
;
214 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
:
215 return V_028780_BLEND_SRC_ALPHA_SATURATE
;
216 case VK_BLEND_FACTOR_SRC1_COLOR
:
217 return V_028780_BLEND_SRC1_COLOR
;
218 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
:
219 return V_028780_BLEND_INV_SRC1_COLOR
;
220 case VK_BLEND_FACTOR_SRC1_ALPHA
:
221 return V_028780_BLEND_SRC1_ALPHA
;
222 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
:
223 return V_028780_BLEND_INV_SRC1_ALPHA
;
229 static uint32_t si_translate_blend_opt_function(VkBlendOp op
)
232 case VK_BLEND_OP_ADD
:
233 return V_028760_OPT_COMB_ADD
;
234 case VK_BLEND_OP_SUBTRACT
:
235 return V_028760_OPT_COMB_SUBTRACT
;
236 case VK_BLEND_OP_REVERSE_SUBTRACT
:
237 return V_028760_OPT_COMB_REVSUBTRACT
;
238 case VK_BLEND_OP_MIN
:
239 return V_028760_OPT_COMB_MIN
;
240 case VK_BLEND_OP_MAX
:
241 return V_028760_OPT_COMB_MAX
;
243 return V_028760_OPT_COMB_BLEND_DISABLED
;
247 static uint32_t si_translate_blend_opt_factor(VkBlendFactor factor
, bool is_alpha
)
250 case VK_BLEND_FACTOR_ZERO
:
251 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL
;
252 case VK_BLEND_FACTOR_ONE
:
253 return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
;
254 case VK_BLEND_FACTOR_SRC_COLOR
:
255 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
256 : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0
;
257 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR
:
258 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
259 : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1
;
260 case VK_BLEND_FACTOR_SRC_ALPHA
:
261 return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
;
262 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
:
263 return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
;
264 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
:
265 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
266 : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0
;
268 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
273 * Get rid of DST in the blend factors by commuting the operands:
274 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
276 static void si_blend_remove_dst(unsigned *func
, unsigned *src_factor
,
277 unsigned *dst_factor
, unsigned expected_dst
,
278 unsigned replacement_src
)
280 if (*src_factor
== expected_dst
&&
281 *dst_factor
== VK_BLEND_FACTOR_ZERO
) {
282 *src_factor
= VK_BLEND_FACTOR_ZERO
;
283 *dst_factor
= replacement_src
;
285 /* Commuting the operands requires reversing subtractions. */
286 if (*func
== VK_BLEND_OP_SUBTRACT
)
287 *func
= VK_BLEND_OP_REVERSE_SUBTRACT
;
288 else if (*func
== VK_BLEND_OP_REVERSE_SUBTRACT
)
289 *func
= VK_BLEND_OP_SUBTRACT
;
293 static bool si_blend_factor_uses_dst(unsigned factor
)
295 return factor
== VK_BLEND_FACTOR_DST_COLOR
||
296 factor
== VK_BLEND_FACTOR_DST_ALPHA
||
297 factor
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
298 factor
== VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA
||
299 factor
== VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR
;
302 static bool is_dual_src(VkBlendFactor factor
)
305 case VK_BLEND_FACTOR_SRC1_COLOR
:
306 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
:
307 case VK_BLEND_FACTOR_SRC1_ALPHA
:
308 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
:
315 static unsigned si_choose_spi_color_format(VkFormat vk_format
,
317 bool blend_need_alpha
)
319 const struct vk_format_description
*desc
= vk_format_description(vk_format
);
320 unsigned format
, ntype
, swap
;
322 /* Alpha is needed for alpha-to-coverage.
323 * Blending may be with or without alpha.
325 unsigned normal
= 0; /* most optimal, may not support blending or export alpha */
326 unsigned alpha
= 0; /* exports alpha, but may not support blending */
327 unsigned blend
= 0; /* supports blending, but may not export alpha */
328 unsigned blend_alpha
= 0; /* least optimal, supports blending and exports alpha */
330 format
= radv_translate_colorformat(vk_format
);
331 ntype
= radv_translate_color_numformat(vk_format
, desc
,
332 vk_format_get_first_non_void_channel(vk_format
));
333 swap
= radv_translate_colorswap(vk_format
, false);
335 /* Choose the SPI color formats. These are required values for Stoney/RB+.
336 * Other chips have multiple choices, though they are not necessarily better.
339 case V_028C70_COLOR_5_6_5
:
340 case V_028C70_COLOR_1_5_5_5
:
341 case V_028C70_COLOR_5_5_5_1
:
342 case V_028C70_COLOR_4_4_4_4
:
343 case V_028C70_COLOR_10_11_11
:
344 case V_028C70_COLOR_11_11_10
:
345 case V_028C70_COLOR_8
:
346 case V_028C70_COLOR_8_8
:
347 case V_028C70_COLOR_8_8_8_8
:
348 case V_028C70_COLOR_10_10_10_2
:
349 case V_028C70_COLOR_2_10_10_10
:
350 if (ntype
== V_028C70_NUMBER_UINT
)
351 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
352 else if (ntype
== V_028C70_NUMBER_SINT
)
353 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
355 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
358 case V_028C70_COLOR_16
:
359 case V_028C70_COLOR_16_16
:
360 case V_028C70_COLOR_16_16_16_16
:
361 if (ntype
== V_028C70_NUMBER_UNORM
||
362 ntype
== V_028C70_NUMBER_SNORM
) {
363 /* UNORM16 and SNORM16 don't support blending */
364 if (ntype
== V_028C70_NUMBER_UNORM
)
365 normal
= alpha
= V_028714_SPI_SHADER_UNORM16_ABGR
;
367 normal
= alpha
= V_028714_SPI_SHADER_SNORM16_ABGR
;
369 /* Use 32 bits per channel for blending. */
370 if (format
== V_028C70_COLOR_16
) {
371 if (swap
== V_028C70_SWAP_STD
) { /* R */
372 blend
= V_028714_SPI_SHADER_32_R
;
373 blend_alpha
= V_028714_SPI_SHADER_32_AR
;
374 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
375 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
378 } else if (format
== V_028C70_COLOR_16_16
) {
379 if (swap
== V_028C70_SWAP_STD
) { /* RG */
380 blend
= V_028714_SPI_SHADER_32_GR
;
381 blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
382 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
383 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
386 } else /* 16_16_16_16 */
387 blend
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
388 } else if (ntype
== V_028C70_NUMBER_UINT
)
389 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
390 else if (ntype
== V_028C70_NUMBER_SINT
)
391 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
392 else if (ntype
== V_028C70_NUMBER_FLOAT
)
393 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
398 case V_028C70_COLOR_32
:
399 if (swap
== V_028C70_SWAP_STD
) { /* R */
400 blend
= normal
= V_028714_SPI_SHADER_32_R
;
401 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
402 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
403 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
408 case V_028C70_COLOR_32_32
:
409 if (swap
== V_028C70_SWAP_STD
) { /* RG */
410 blend
= normal
= V_028714_SPI_SHADER_32_GR
;
411 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
412 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
413 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
418 case V_028C70_COLOR_32_32_32_32
:
419 case V_028C70_COLOR_8_24
:
420 case V_028C70_COLOR_24_8
:
421 case V_028C70_COLOR_X24_8_32_FLOAT
:
422 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_ABGR
;
426 unreachable("unhandled blend format");
429 if (blend_enable
&& blend_need_alpha
)
431 else if(blend_need_alpha
)
433 else if(blend_enable
)
440 radv_pipeline_compute_spi_color_formats(struct radv_pipeline
*pipeline
,
441 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
442 uint32_t blend_enable
,
443 uint32_t blend_need_alpha
,
444 bool single_cb_enable
,
445 bool blend_mrt0_is_dual_src
,
446 struct radv_blend_state
*blend
)
448 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
449 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
450 unsigned col_format
= 0;
452 for (unsigned i
= 0; i
< (single_cb_enable
? 1 : subpass
->color_count
); ++i
) {
455 if (subpass
->color_attachments
[i
].attachment
== VK_ATTACHMENT_UNUSED
) {
456 cf
= V_028714_SPI_SHADER_ZERO
;
458 struct radv_render_pass_attachment
*attachment
= pass
->attachments
+ subpass
->color_attachments
[i
].attachment
;
460 cf
= si_choose_spi_color_format(attachment
->format
,
461 blend_enable
& (1 << i
),
462 blend_need_alpha
& (1 << i
));
465 col_format
|= cf
<< (4 * i
);
468 blend
->cb_shader_mask
= ac_get_cb_shader_mask(col_format
);
470 if (blend_mrt0_is_dual_src
)
471 col_format
|= (col_format
& 0xf) << 4;
472 blend
->spi_shader_col_format
= col_format
;
476 format_is_int8(VkFormat format
)
478 const struct vk_format_description
*desc
= vk_format_description(format
);
479 int channel
= vk_format_get_first_non_void_channel(format
);
481 return channel
>= 0 && desc
->channel
[channel
].pure_integer
&&
482 desc
->channel
[channel
].size
== 8;
486 format_is_int10(VkFormat format
)
488 const struct vk_format_description
*desc
= vk_format_description(format
);
490 if (desc
->nr_channels
!= 4)
492 for (unsigned i
= 0; i
< 4; i
++) {
493 if (desc
->channel
[i
].pure_integer
&& desc
->channel
[i
].size
== 10)
499 unsigned radv_format_meta_fs_key(VkFormat format
)
501 unsigned col_format
= si_choose_spi_color_format(format
, false, false) - 1;
502 bool is_int8
= format_is_int8(format
);
503 bool is_int10
= format_is_int10(format
);
505 return col_format
+ (is_int8
? 3 : is_int10
? 5 : 0);
509 radv_pipeline_compute_get_int_clamp(const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
510 unsigned *is_int8
, unsigned *is_int10
)
512 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
513 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
517 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
518 struct radv_render_pass_attachment
*attachment
;
520 if (subpass
->color_attachments
[i
].attachment
== VK_ATTACHMENT_UNUSED
)
523 attachment
= pass
->attachments
+ subpass
->color_attachments
[i
].attachment
;
525 if (format_is_int8(attachment
->format
))
527 if (format_is_int10(attachment
->format
))
532 static struct radv_blend_state
533 radv_pipeline_init_blend_state(struct radv_pipeline
*pipeline
,
534 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
535 const struct radv_graphics_pipeline_create_info
*extra
)
537 const VkPipelineColorBlendStateCreateInfo
*vkblend
= pCreateInfo
->pColorBlendState
;
538 const VkPipelineMultisampleStateCreateInfo
*vkms
= pCreateInfo
->pMultisampleState
;
539 struct radv_blend_state blend
= {0};
540 unsigned mode
= V_028808_CB_NORMAL
;
541 uint32_t blend_enable
= 0, blend_need_alpha
= 0;
542 bool blend_mrt0_is_dual_src
= false;
544 bool single_cb_enable
= false;
549 if (extra
&& extra
->custom_blend_mode
) {
550 single_cb_enable
= true;
551 mode
= extra
->custom_blend_mode
;
553 blend
.cb_color_control
= 0;
554 if (vkblend
->logicOpEnable
)
555 blend
.cb_color_control
|= S_028808_ROP3(vkblend
->logicOp
| (vkblend
->logicOp
<< 4));
557 blend
.cb_color_control
|= S_028808_ROP3(0xcc);
559 blend
.db_alpha_to_mask
= S_028B70_ALPHA_TO_MASK_OFFSET0(2) |
560 S_028B70_ALPHA_TO_MASK_OFFSET1(2) |
561 S_028B70_ALPHA_TO_MASK_OFFSET2(2) |
562 S_028B70_ALPHA_TO_MASK_OFFSET3(2);
564 if (vkms
&& vkms
->alphaToCoverageEnable
) {
565 blend
.db_alpha_to_mask
|= S_028B70_ALPHA_TO_MASK_ENABLE(1);
568 blend
.cb_target_mask
= 0;
569 for (i
= 0; i
< vkblend
->attachmentCount
; i
++) {
570 const VkPipelineColorBlendAttachmentState
*att
= &vkblend
->pAttachments
[i
];
571 unsigned blend_cntl
= 0;
572 unsigned srcRGB_opt
, dstRGB_opt
, srcA_opt
, dstA_opt
;
573 VkBlendOp eqRGB
= att
->colorBlendOp
;
574 VkBlendFactor srcRGB
= att
->srcColorBlendFactor
;
575 VkBlendFactor dstRGB
= att
->dstColorBlendFactor
;
576 VkBlendOp eqA
= att
->alphaBlendOp
;
577 VkBlendFactor srcA
= att
->srcAlphaBlendFactor
;
578 VkBlendFactor dstA
= att
->dstAlphaBlendFactor
;
580 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
);
582 if (!att
->colorWriteMask
)
585 blend
.cb_target_mask
|= (unsigned)att
->colorWriteMask
<< (4 * i
);
586 if (!att
->blendEnable
) {
587 blend
.cb_blend_control
[i
] = blend_cntl
;
591 if (is_dual_src(srcRGB
) || is_dual_src(dstRGB
) || is_dual_src(srcA
) || is_dual_src(dstA
))
593 blend_mrt0_is_dual_src
= true;
595 if (eqRGB
== VK_BLEND_OP_MIN
|| eqRGB
== VK_BLEND_OP_MAX
) {
596 srcRGB
= VK_BLEND_FACTOR_ONE
;
597 dstRGB
= VK_BLEND_FACTOR_ONE
;
599 if (eqA
== VK_BLEND_OP_MIN
|| eqA
== VK_BLEND_OP_MAX
) {
600 srcA
= VK_BLEND_FACTOR_ONE
;
601 dstA
= VK_BLEND_FACTOR_ONE
;
604 /* Blending optimizations for RB+.
605 * These transformations don't change the behavior.
607 * First, get rid of DST in the blend factors:
608 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
610 si_blend_remove_dst(&eqRGB
, &srcRGB
, &dstRGB
,
611 VK_BLEND_FACTOR_DST_COLOR
,
612 VK_BLEND_FACTOR_SRC_COLOR
);
614 si_blend_remove_dst(&eqA
, &srcA
, &dstA
,
615 VK_BLEND_FACTOR_DST_COLOR
,
616 VK_BLEND_FACTOR_SRC_COLOR
);
618 si_blend_remove_dst(&eqA
, &srcA
, &dstA
,
619 VK_BLEND_FACTOR_DST_ALPHA
,
620 VK_BLEND_FACTOR_SRC_ALPHA
);
622 /* Look up the ideal settings from tables. */
623 srcRGB_opt
= si_translate_blend_opt_factor(srcRGB
, false);
624 dstRGB_opt
= si_translate_blend_opt_factor(dstRGB
, false);
625 srcA_opt
= si_translate_blend_opt_factor(srcA
, true);
626 dstA_opt
= si_translate_blend_opt_factor(dstA
, true);
628 /* Handle interdependencies. */
629 if (si_blend_factor_uses_dst(srcRGB
))
630 dstRGB_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
631 if (si_blend_factor_uses_dst(srcA
))
632 dstA_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
634 if (srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
&&
635 (dstRGB
== VK_BLEND_FACTOR_ZERO
||
636 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
637 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
))
638 dstRGB_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0
;
640 /* Set the final value. */
641 blend
.sx_mrt_blend_opt
[i
] =
642 S_028760_COLOR_SRC_OPT(srcRGB_opt
) |
643 S_028760_COLOR_DST_OPT(dstRGB_opt
) |
644 S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB
)) |
645 S_028760_ALPHA_SRC_OPT(srcA_opt
) |
646 S_028760_ALPHA_DST_OPT(dstA_opt
) |
647 S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA
));
648 blend_cntl
|= S_028780_ENABLE(1);
650 blend_cntl
|= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB
));
651 blend_cntl
|= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB
));
652 blend_cntl
|= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB
));
653 if (srcA
!= srcRGB
|| dstA
!= dstRGB
|| eqA
!= eqRGB
) {
654 blend_cntl
|= S_028780_SEPARATE_ALPHA_BLEND(1);
655 blend_cntl
|= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA
));
656 blend_cntl
|= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA
));
657 blend_cntl
|= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA
));
659 blend
.cb_blend_control
[i
] = blend_cntl
;
661 blend_enable
|= 1 << i
;
663 if (srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
664 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
665 srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
666 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
667 srcRGB
== VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
||
668 dstRGB
== VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
)
669 blend_need_alpha
|= 1 << i
;
671 for (i
= vkblend
->attachmentCount
; i
< 8; i
++) {
672 blend
.cb_blend_control
[i
] = 0;
673 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
);
676 /* disable RB+ for now */
677 if (pipeline
->device
->physical_device
->has_rbplus
)
678 blend
.cb_color_control
|= S_028808_DISABLE_DUAL_QUAD(1);
680 if (blend
.cb_target_mask
)
681 blend
.cb_color_control
|= S_028808_MODE(mode
);
683 blend
.cb_color_control
|= S_028808_MODE(V_028808_CB_DISABLE
);
685 radv_pipeline_compute_spi_color_formats(pipeline
, pCreateInfo
,
686 blend_enable
, blend_need_alpha
, single_cb_enable
, blend_mrt0_is_dual_src
,
691 static uint32_t si_translate_stencil_op(enum VkStencilOp op
)
694 case VK_STENCIL_OP_KEEP
:
695 return V_02842C_STENCIL_KEEP
;
696 case VK_STENCIL_OP_ZERO
:
697 return V_02842C_STENCIL_ZERO
;
698 case VK_STENCIL_OP_REPLACE
:
699 return V_02842C_STENCIL_REPLACE_TEST
;
700 case VK_STENCIL_OP_INCREMENT_AND_CLAMP
:
701 return V_02842C_STENCIL_ADD_CLAMP
;
702 case VK_STENCIL_OP_DECREMENT_AND_CLAMP
:
703 return V_02842C_STENCIL_SUB_CLAMP
;
704 case VK_STENCIL_OP_INVERT
:
705 return V_02842C_STENCIL_INVERT
;
706 case VK_STENCIL_OP_INCREMENT_AND_WRAP
:
707 return V_02842C_STENCIL_ADD_WRAP
;
708 case VK_STENCIL_OP_DECREMENT_AND_WRAP
:
709 return V_02842C_STENCIL_SUB_WRAP
;
715 static uint32_t si_translate_fill(VkPolygonMode func
)
718 case VK_POLYGON_MODE_FILL
:
719 return V_028814_X_DRAW_TRIANGLES
;
720 case VK_POLYGON_MODE_LINE
:
721 return V_028814_X_DRAW_LINES
;
722 case VK_POLYGON_MODE_POINT
:
723 return V_028814_X_DRAW_POINTS
;
726 return V_028814_X_DRAW_POINTS
;
730 static uint8_t radv_pipeline_get_ps_iter_samples(const VkPipelineMultisampleStateCreateInfo
*vkms
)
732 uint32_t num_samples
= vkms
->rasterizationSamples
;
733 uint32_t ps_iter_samples
= 1;
735 if (vkms
->sampleShadingEnable
) {
736 ps_iter_samples
= ceil(vkms
->minSampleShading
* num_samples
);
737 ps_iter_samples
= util_next_power_of_two(ps_iter_samples
);
739 return ps_iter_samples
;
743 radv_pipeline_init_multisample_state(struct radv_pipeline
*pipeline
,
744 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
746 const VkPipelineMultisampleStateCreateInfo
*vkms
= pCreateInfo
->pMultisampleState
;
747 struct radv_multisample_state
*ms
= &pipeline
->graphics
.ms
;
748 unsigned num_tile_pipes
= pipeline
->device
->physical_device
->rad_info
.num_tile_pipes
;
749 int ps_iter_samples
= 1;
750 uint32_t mask
= 0xffff;
753 ms
->num_samples
= vkms
->rasterizationSamples
;
758 ps_iter_samples
= radv_pipeline_get_ps_iter_samples(vkms
);
759 if (vkms
&& !vkms
->sampleShadingEnable
&& pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.force_persample
) {
760 ps_iter_samples
= ms
->num_samples
;
763 ms
->pa_sc_line_cntl
= S_028BDC_DX10_DIAMOND_TEST_ENA(1);
764 ms
->pa_sc_aa_config
= 0;
765 ms
->db_eqaa
= S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
766 S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
767 ms
->pa_sc_mode_cntl_1
=
768 S_028A4C_WALK_FENCE_ENABLE(1) | //TODO linear dst fixes
769 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes
== 2 ? 2 : 3) |
771 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
772 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
773 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
774 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
775 S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
776 S_028A4C_FORCE_EOV_REZ_ENABLE(1);
777 ms
->pa_sc_mode_cntl_0
= S_028A48_ALTERNATE_RBS_PER_TILE(pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) |
778 S_028A48_VPORT_SCISSOR_ENABLE(1);
780 if (ms
->num_samples
> 1) {
781 unsigned log_samples
= util_logbase2(ms
->num_samples
);
782 unsigned log_ps_iter_samples
= util_logbase2(ps_iter_samples
);
783 ms
->pa_sc_mode_cntl_0
|= S_028A48_MSAA_ENABLE(1);
784 ms
->pa_sc_line_cntl
|= S_028BDC_EXPAND_LINE_WIDTH(1); /* CM_R_028BDC_PA_SC_LINE_CNTL */
785 ms
->db_eqaa
|= S_028804_MAX_ANCHOR_SAMPLES(log_samples
) |
786 S_028804_PS_ITER_SAMPLES(log_ps_iter_samples
) |
787 S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples
) |
788 S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples
);
789 ms
->pa_sc_aa_config
|= S_028BE0_MSAA_NUM_SAMPLES(log_samples
) |
790 S_028BE0_MAX_SAMPLE_DIST(radv_cayman_get_maxdist(log_samples
)) |
791 S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples
); /* CM_R_028BE0_PA_SC_AA_CONFIG */
792 ms
->pa_sc_mode_cntl_1
|= S_028A4C_PS_ITER_SAMPLE(ps_iter_samples
> 1);
793 if (ps_iter_samples
> 1)
794 pipeline
->graphics
.spi_baryc_cntl
|= S_0286E0_POS_FLOAT_LOCATION(2);
797 const struct VkPipelineRasterizationStateRasterizationOrderAMD
*raster_order
=
798 vk_find_struct_const(pCreateInfo
->pRasterizationState
->pNext
, PIPELINE_RASTERIZATION_STATE_RASTERIZATION_ORDER_AMD
);
799 if (raster_order
&& raster_order
->rasterizationOrder
== VK_RASTERIZATION_ORDER_RELAXED_AMD
) {
800 ms
->pa_sc_mode_cntl_1
|= S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(1) |
801 S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7);
804 if (vkms
&& vkms
->pSampleMask
) {
805 mask
= vkms
->pSampleMask
[0] & 0xffff;
808 ms
->pa_sc_aa_mask
[0] = mask
| (mask
<< 16);
809 ms
->pa_sc_aa_mask
[1] = mask
| (mask
<< 16);
813 radv_prim_can_use_guardband(enum VkPrimitiveTopology topology
)
816 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
817 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
818 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
819 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
820 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
822 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
823 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
824 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
825 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
826 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
827 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
830 unreachable("unhandled primitive type");
835 si_translate_prim(enum VkPrimitiveTopology topology
)
838 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
839 return V_008958_DI_PT_POINTLIST
;
840 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
841 return V_008958_DI_PT_LINELIST
;
842 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
843 return V_008958_DI_PT_LINESTRIP
;
844 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
845 return V_008958_DI_PT_TRILIST
;
846 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
847 return V_008958_DI_PT_TRISTRIP
;
848 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
849 return V_008958_DI_PT_TRIFAN
;
850 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
851 return V_008958_DI_PT_LINELIST_ADJ
;
852 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
853 return V_008958_DI_PT_LINESTRIP_ADJ
;
854 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
855 return V_008958_DI_PT_TRILIST_ADJ
;
856 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
857 return V_008958_DI_PT_TRISTRIP_ADJ
;
858 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
859 return V_008958_DI_PT_PATCH
;
867 si_conv_gl_prim_to_gs_out(unsigned gl_prim
)
870 case 0: /* GL_POINTS */
871 return V_028A6C_OUTPRIM_TYPE_POINTLIST
;
872 case 1: /* GL_LINES */
873 case 3: /* GL_LINE_STRIP */
874 case 0xA: /* GL_LINE_STRIP_ADJACENCY_ARB */
875 case 0x8E7A: /* GL_ISOLINES */
876 return V_028A6C_OUTPRIM_TYPE_LINESTRIP
;
878 case 4: /* GL_TRIANGLES */
879 case 0xc: /* GL_TRIANGLES_ADJACENCY_ARB */
880 case 5: /* GL_TRIANGLE_STRIP */
881 case 7: /* GL_QUADS */
882 return V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
890 si_conv_prim_to_gs_out(enum VkPrimitiveTopology topology
)
893 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
894 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
895 return V_028A6C_OUTPRIM_TYPE_POINTLIST
;
896 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
897 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
898 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
899 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
900 return V_028A6C_OUTPRIM_TYPE_LINESTRIP
;
901 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
902 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
903 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
904 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
905 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
906 return V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
913 static unsigned si_map_swizzle(unsigned swizzle
)
917 return V_008F0C_SQ_SEL_Y
;
919 return V_008F0C_SQ_SEL_Z
;
921 return V_008F0C_SQ_SEL_W
;
923 return V_008F0C_SQ_SEL_0
;
925 return V_008F0C_SQ_SEL_1
;
926 default: /* VK_SWIZZLE_X */
927 return V_008F0C_SQ_SEL_X
;
932 static unsigned radv_dynamic_state_mask(VkDynamicState state
)
935 case VK_DYNAMIC_STATE_VIEWPORT
:
936 return RADV_DYNAMIC_VIEWPORT
;
937 case VK_DYNAMIC_STATE_SCISSOR
:
938 return RADV_DYNAMIC_SCISSOR
;
939 case VK_DYNAMIC_STATE_LINE_WIDTH
:
940 return RADV_DYNAMIC_LINE_WIDTH
;
941 case VK_DYNAMIC_STATE_DEPTH_BIAS
:
942 return RADV_DYNAMIC_DEPTH_BIAS
;
943 case VK_DYNAMIC_STATE_BLEND_CONSTANTS
:
944 return RADV_DYNAMIC_BLEND_CONSTANTS
;
945 case VK_DYNAMIC_STATE_DEPTH_BOUNDS
:
946 return RADV_DYNAMIC_DEPTH_BOUNDS
;
947 case VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK
:
948 return RADV_DYNAMIC_STENCIL_COMPARE_MASK
;
949 case VK_DYNAMIC_STATE_STENCIL_WRITE_MASK
:
950 return RADV_DYNAMIC_STENCIL_WRITE_MASK
;
951 case VK_DYNAMIC_STATE_STENCIL_REFERENCE
:
952 return RADV_DYNAMIC_STENCIL_REFERENCE
;
953 case VK_DYNAMIC_STATE_DISCARD_RECTANGLE_EXT
:
954 return RADV_DYNAMIC_DISCARD_RECTANGLE
;
956 unreachable("Unhandled dynamic state");
960 static uint32_t radv_pipeline_needed_dynamic_state(const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
962 uint32_t states
= RADV_DYNAMIC_ALL
;
964 /* If rasterization is disabled we do not care about any of the dynamic states,
965 * since they are all rasterization related only. */
966 if (pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
)
969 if (!pCreateInfo
->pRasterizationState
->depthBiasEnable
)
970 states
&= ~RADV_DYNAMIC_DEPTH_BIAS
;
972 if (!pCreateInfo
->pDepthStencilState
||
973 !pCreateInfo
->pDepthStencilState
->depthBoundsTestEnable
)
974 states
&= ~RADV_DYNAMIC_DEPTH_BOUNDS
;
976 if (!pCreateInfo
->pDepthStencilState
||
977 !pCreateInfo
->pDepthStencilState
->stencilTestEnable
)
978 states
&= ~(RADV_DYNAMIC_STENCIL_COMPARE_MASK
|
979 RADV_DYNAMIC_STENCIL_WRITE_MASK
|
980 RADV_DYNAMIC_STENCIL_REFERENCE
);
982 if (!vk_find_struct_const(pCreateInfo
->pNext
, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT
))
983 states
&= ~RADV_DYNAMIC_DISCARD_RECTANGLE
;
985 /* TODO: blend constants & line width. */
992 radv_pipeline_init_dynamic_state(struct radv_pipeline
*pipeline
,
993 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
995 uint32_t needed_states
= radv_pipeline_needed_dynamic_state(pCreateInfo
);
996 uint32_t states
= needed_states
;
997 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
998 struct radv_subpass
*subpass
= &pass
->subpasses
[pCreateInfo
->subpass
];
1000 pipeline
->dynamic_state
= default_dynamic_state
;
1001 pipeline
->graphics
.needed_dynamic_state
= needed_states
;
1003 if (pCreateInfo
->pDynamicState
) {
1004 /* Remove all of the states that are marked as dynamic */
1005 uint32_t count
= pCreateInfo
->pDynamicState
->dynamicStateCount
;
1006 for (uint32_t s
= 0; s
< count
; s
++)
1007 states
&= ~radv_dynamic_state_mask(pCreateInfo
->pDynamicState
->pDynamicStates
[s
]);
1010 struct radv_dynamic_state
*dynamic
= &pipeline
->dynamic_state
;
1012 if (needed_states
& RADV_DYNAMIC_VIEWPORT
) {
1013 assert(pCreateInfo
->pViewportState
);
1015 dynamic
->viewport
.count
= pCreateInfo
->pViewportState
->viewportCount
;
1016 if (states
& RADV_DYNAMIC_VIEWPORT
) {
1017 typed_memcpy(dynamic
->viewport
.viewports
,
1018 pCreateInfo
->pViewportState
->pViewports
,
1019 pCreateInfo
->pViewportState
->viewportCount
);
1023 if (needed_states
& RADV_DYNAMIC_SCISSOR
) {
1024 dynamic
->scissor
.count
= pCreateInfo
->pViewportState
->scissorCount
;
1025 if (states
& RADV_DYNAMIC_SCISSOR
) {
1026 typed_memcpy(dynamic
->scissor
.scissors
,
1027 pCreateInfo
->pViewportState
->pScissors
,
1028 pCreateInfo
->pViewportState
->scissorCount
);
1032 if (states
& RADV_DYNAMIC_LINE_WIDTH
) {
1033 assert(pCreateInfo
->pRasterizationState
);
1034 dynamic
->line_width
= pCreateInfo
->pRasterizationState
->lineWidth
;
1037 if (states
& RADV_DYNAMIC_DEPTH_BIAS
) {
1038 assert(pCreateInfo
->pRasterizationState
);
1039 dynamic
->depth_bias
.bias
=
1040 pCreateInfo
->pRasterizationState
->depthBiasConstantFactor
;
1041 dynamic
->depth_bias
.clamp
=
1042 pCreateInfo
->pRasterizationState
->depthBiasClamp
;
1043 dynamic
->depth_bias
.slope
=
1044 pCreateInfo
->pRasterizationState
->depthBiasSlopeFactor
;
1047 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1049 * pColorBlendState is [...] NULL if the pipeline has rasterization
1050 * disabled or if the subpass of the render pass the pipeline is
1051 * created against does not use any color attachments.
1053 bool uses_color_att
= false;
1054 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
1055 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
) {
1056 uses_color_att
= true;
1061 if (uses_color_att
&& states
& RADV_DYNAMIC_BLEND_CONSTANTS
) {
1062 assert(pCreateInfo
->pColorBlendState
);
1063 typed_memcpy(dynamic
->blend_constants
,
1064 pCreateInfo
->pColorBlendState
->blendConstants
, 4);
1067 /* If there is no depthstencil attachment, then don't read
1068 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
1069 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
1070 * no need to override the depthstencil defaults in
1071 * radv_pipeline::dynamic_state when there is no depthstencil attachment.
1073 * Section 9.2 of the Vulkan 1.0.15 spec says:
1075 * pDepthStencilState is [...] NULL if the pipeline has rasterization
1076 * disabled or if the subpass of the render pass the pipeline is created
1077 * against does not use a depth/stencil attachment.
1079 if (needed_states
&&
1080 subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
1081 assert(pCreateInfo
->pDepthStencilState
);
1083 if (states
& RADV_DYNAMIC_DEPTH_BOUNDS
) {
1084 dynamic
->depth_bounds
.min
=
1085 pCreateInfo
->pDepthStencilState
->minDepthBounds
;
1086 dynamic
->depth_bounds
.max
=
1087 pCreateInfo
->pDepthStencilState
->maxDepthBounds
;
1090 if (states
& RADV_DYNAMIC_STENCIL_COMPARE_MASK
) {
1091 dynamic
->stencil_compare_mask
.front
=
1092 pCreateInfo
->pDepthStencilState
->front
.compareMask
;
1093 dynamic
->stencil_compare_mask
.back
=
1094 pCreateInfo
->pDepthStencilState
->back
.compareMask
;
1097 if (states
& RADV_DYNAMIC_STENCIL_WRITE_MASK
) {
1098 dynamic
->stencil_write_mask
.front
=
1099 pCreateInfo
->pDepthStencilState
->front
.writeMask
;
1100 dynamic
->stencil_write_mask
.back
=
1101 pCreateInfo
->pDepthStencilState
->back
.writeMask
;
1104 if (states
& RADV_DYNAMIC_STENCIL_REFERENCE
) {
1105 dynamic
->stencil_reference
.front
=
1106 pCreateInfo
->pDepthStencilState
->front
.reference
;
1107 dynamic
->stencil_reference
.back
=
1108 pCreateInfo
->pDepthStencilState
->back
.reference
;
1112 const VkPipelineDiscardRectangleStateCreateInfoEXT
*discard_rectangle_info
=
1113 vk_find_struct_const(pCreateInfo
->pNext
, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT
);
1114 if (states
& RADV_DYNAMIC_DISCARD_RECTANGLE
) {
1115 dynamic
->discard_rectangle
.count
= discard_rectangle_info
->discardRectangleCount
;
1116 typed_memcpy(dynamic
->discard_rectangle
.rectangles
,
1117 discard_rectangle_info
->pDiscardRectangles
,
1118 discard_rectangle_info
->discardRectangleCount
);
1121 pipeline
->dynamic_state
.mask
= states
;
1124 static struct radv_gs_state
1125 calculate_gs_info(const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1126 const struct radv_pipeline
*pipeline
)
1128 struct radv_gs_state gs
= {0};
1129 struct radv_shader_variant_info
*gs_info
= &pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
;
1130 struct radv_es_output_info
*es_info
;
1131 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1132 es_info
= radv_pipeline_has_tess(pipeline
) ? &gs_info
->tes
.es_info
: &gs_info
->vs
.es_info
;
1134 es_info
= radv_pipeline_has_tess(pipeline
) ?
1135 &pipeline
->shaders
[MESA_SHADER_TESS_EVAL
]->info
.tes
.es_info
:
1136 &pipeline
->shaders
[MESA_SHADER_VERTEX
]->info
.vs
.es_info
;
1138 unsigned gs_num_invocations
= MAX2(gs_info
->gs
.invocations
, 1);
1139 bool uses_adjacency
;
1140 switch(pCreateInfo
->pInputAssemblyState
->topology
) {
1141 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
1142 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
1143 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
1144 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
1145 uses_adjacency
= true;
1148 uses_adjacency
= false;
1152 /* All these are in dwords: */
1153 /* We can't allow using the whole LDS, because GS waves compete with
1154 * other shader stages for LDS space. */
1155 const unsigned max_lds_size
= 8 * 1024;
1156 const unsigned esgs_itemsize
= es_info
->esgs_itemsize
/ 4;
1157 unsigned esgs_lds_size
;
1159 /* All these are per subgroup: */
1160 const unsigned max_out_prims
= 32 * 1024;
1161 const unsigned max_es_verts
= 255;
1162 const unsigned ideal_gs_prims
= 64;
1163 unsigned max_gs_prims
, gs_prims
;
1164 unsigned min_es_verts
, es_verts
, worst_case_es_verts
;
1166 if (uses_adjacency
|| gs_num_invocations
> 1)
1167 max_gs_prims
= 127 / gs_num_invocations
;
1171 /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
1172 * Make sure we don't go over the maximum value.
1174 if (gs_info
->gs
.vertices_out
> 0) {
1175 max_gs_prims
= MIN2(max_gs_prims
,
1177 (gs_info
->gs
.vertices_out
* gs_num_invocations
));
1179 assert(max_gs_prims
> 0);
1181 /* If the primitive has adjacency, halve the number of vertices
1182 * that will be reused in multiple primitives.
1184 min_es_verts
= gs_info
->gs
.vertices_in
/ (uses_adjacency
? 2 : 1);
1186 gs_prims
= MIN2(ideal_gs_prims
, max_gs_prims
);
1187 worst_case_es_verts
= MIN2(min_es_verts
* gs_prims
, max_es_verts
);
1189 /* Compute ESGS LDS size based on the worst case number of ES vertices
1190 * needed to create the target number of GS prims per subgroup.
1192 esgs_lds_size
= esgs_itemsize
* worst_case_es_verts
;
1194 /* If total LDS usage is too big, refactor partitions based on ratio
1195 * of ESGS item sizes.
1197 if (esgs_lds_size
> max_lds_size
) {
1198 /* Our target GS Prims Per Subgroup was too large. Calculate
1199 * the maximum number of GS Prims Per Subgroup that will fit
1200 * into LDS, capped by the maximum that the hardware can support.
1202 gs_prims
= MIN2((max_lds_size
/ (esgs_itemsize
* min_es_verts
)),
1204 assert(gs_prims
> 0);
1205 worst_case_es_verts
= MIN2(min_es_verts
* gs_prims
,
1208 esgs_lds_size
= esgs_itemsize
* worst_case_es_verts
;
1209 assert(esgs_lds_size
<= max_lds_size
);
1212 /* Now calculate remaining ESGS information. */
1214 es_verts
= MIN2(esgs_lds_size
/ esgs_itemsize
, max_es_verts
);
1216 es_verts
= max_es_verts
;
1218 /* Vertices for adjacency primitives are not always reused, so restore
1219 * it for ES_VERTS_PER_SUBGRP.
1221 min_es_verts
= gs_info
->gs
.vertices_in
;
1223 /* For normal primitives, the VGT only checks if they are past the ES
1224 * verts per subgroup after allocating a full GS primitive and if they
1225 * are, kick off a new subgroup. But if those additional ES verts are
1226 * unique (e.g. not reused) we need to make sure there is enough LDS
1227 * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
1229 es_verts
-= min_es_verts
- 1;
1231 uint32_t es_verts_per_subgroup
= es_verts
;
1232 uint32_t gs_prims_per_subgroup
= gs_prims
;
1233 uint32_t gs_inst_prims_in_subgroup
= gs_prims
* gs_num_invocations
;
1234 uint32_t max_prims_per_subgroup
= gs_inst_prims_in_subgroup
* gs_info
->gs
.vertices_out
;
1235 gs
.lds_size
= align(esgs_lds_size
, 128) / 128;
1236 gs
.vgt_gs_onchip_cntl
= S_028A44_ES_VERTS_PER_SUBGRP(es_verts_per_subgroup
) |
1237 S_028A44_GS_PRIMS_PER_SUBGRP(gs_prims_per_subgroup
) |
1238 S_028A44_GS_INST_PRIMS_IN_SUBGRP(gs_inst_prims_in_subgroup
);
1239 gs
.vgt_gs_max_prims_per_subgroup
= S_028A94_MAX_PRIMS_PER_SUBGROUP(max_prims_per_subgroup
);
1240 gs
.vgt_esgs_ring_itemsize
= esgs_itemsize
;
1241 assert(max_prims_per_subgroup
<= max_out_prims
);
1247 calculate_gs_ring_sizes(struct radv_pipeline
*pipeline
, const struct radv_gs_state
*gs
)
1249 struct radv_device
*device
= pipeline
->device
;
1250 unsigned num_se
= device
->physical_device
->rad_info
.max_se
;
1251 unsigned wave_size
= 64;
1252 unsigned max_gs_waves
= 32 * num_se
; /* max 32 per SE on GCN */
1253 unsigned gs_vertex_reuse
= 16 * num_se
; /* GS_VERTEX_REUSE register (per SE) */
1254 unsigned alignment
= 256 * num_se
;
1255 /* The maximum size is 63.999 MB per SE. */
1256 unsigned max_size
= ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se
;
1257 struct radv_shader_variant_info
*gs_info
= &pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
;
1259 /* Calculate the minimum size. */
1260 unsigned min_esgs_ring_size
= align(gs
->vgt_esgs_ring_itemsize
* 4 * gs_vertex_reuse
*
1261 wave_size
, alignment
);
1262 /* These are recommended sizes, not minimum sizes. */
1263 unsigned esgs_ring_size
= max_gs_waves
* 2 * wave_size
*
1264 gs
->vgt_esgs_ring_itemsize
* 4 * gs_info
->gs
.vertices_in
;
1265 unsigned gsvs_ring_size
= max_gs_waves
* 2 * wave_size
*
1266 gs_info
->gs
.max_gsvs_emit_size
* 1; // no streams in VK (gs->max_gs_stream + 1);
1268 min_esgs_ring_size
= align(min_esgs_ring_size
, alignment
);
1269 esgs_ring_size
= align(esgs_ring_size
, alignment
);
1270 gsvs_ring_size
= align(gsvs_ring_size
, alignment
);
1272 if (pipeline
->device
->physical_device
->rad_info
.chip_class
<= VI
)
1273 pipeline
->graphics
.esgs_ring_size
= CLAMP(esgs_ring_size
, min_esgs_ring_size
, max_size
);
1275 pipeline
->graphics
.gsvs_ring_size
= MIN2(gsvs_ring_size
, max_size
);
1278 static void si_multiwave_lds_size_workaround(struct radv_device
*device
,
1281 /* SPI barrier management bug:
1282 * Make sure we have at least 4k of LDS in use to avoid the bug.
1283 * It applies to workgroup sizes of more than one wavefront.
1285 if (device
->physical_device
->rad_info
.family
== CHIP_BONAIRE
||
1286 device
->physical_device
->rad_info
.family
== CHIP_KABINI
||
1287 device
->physical_device
->rad_info
.family
== CHIP_MULLINS
)
1288 *lds_size
= MAX2(*lds_size
, 8);
1291 struct radv_shader_variant
*
1292 radv_get_vertex_shader(struct radv_pipeline
*pipeline
)
1294 if (pipeline
->shaders
[MESA_SHADER_VERTEX
])
1295 return pipeline
->shaders
[MESA_SHADER_VERTEX
];
1296 if (pipeline
->shaders
[MESA_SHADER_TESS_CTRL
])
1297 return pipeline
->shaders
[MESA_SHADER_TESS_CTRL
];
1298 return pipeline
->shaders
[MESA_SHADER_GEOMETRY
];
1301 static struct radv_shader_variant
*
1302 radv_get_tess_eval_shader(struct radv_pipeline
*pipeline
)
1304 if (pipeline
->shaders
[MESA_SHADER_TESS_EVAL
])
1305 return pipeline
->shaders
[MESA_SHADER_TESS_EVAL
];
1306 return pipeline
->shaders
[MESA_SHADER_GEOMETRY
];
1309 static struct radv_tessellation_state
1310 calculate_tess_state(struct radv_pipeline
*pipeline
,
1311 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1313 unsigned num_tcs_input_cp
= pCreateInfo
->pTessellationState
->patchControlPoints
;
1314 unsigned num_tcs_output_cp
, num_tcs_inputs
, num_tcs_outputs
;
1315 unsigned num_tcs_patch_outputs
;
1316 unsigned input_vertex_size
, output_vertex_size
, pervertex_output_patch_size
;
1317 unsigned input_patch_size
, output_patch_size
, output_patch0_offset
;
1318 unsigned lds_size
, hardware_lds_size
;
1319 unsigned perpatch_output_offset
;
1320 unsigned num_patches
;
1321 struct radv_tessellation_state tess
= {0};
1323 /* This calculates how shader inputs and outputs among VS, TCS, and TES
1324 * are laid out in LDS. */
1325 num_tcs_inputs
= util_last_bit64(radv_get_vertex_shader(pipeline
)->info
.vs
.outputs_written
);
1327 num_tcs_outputs
= util_last_bit64(pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.outputs_written
); //tcs->outputs_written
1328 num_tcs_output_cp
= pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.tcs_vertices_out
; //TCS VERTICES OUT
1329 num_tcs_patch_outputs
= util_last_bit64(pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.patch_outputs_written
);
1331 /* Ensure that we only need one wave per SIMD so we don't need to check
1332 * resource usage. Also ensures that the number of tcs in and out
1333 * vertices per threadgroup are at most 256.
1335 input_vertex_size
= num_tcs_inputs
* 16;
1336 output_vertex_size
= num_tcs_outputs
* 16;
1338 input_patch_size
= num_tcs_input_cp
* input_vertex_size
;
1340 pervertex_output_patch_size
= num_tcs_output_cp
* output_vertex_size
;
1341 output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
1342 /* Ensure that we only need one wave per SIMD so we don't need to check
1343 * resource usage. Also ensures that the number of tcs in and out
1344 * vertices per threadgroup are at most 256.
1346 num_patches
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
) * 4;
1348 /* Make sure that the data fits in LDS. This assumes the shaders only
1349 * use LDS for the inputs and outputs.
1351 hardware_lds_size
= pipeline
->device
->physical_device
->rad_info
.chip_class
>= CIK
? 65536 : 32768;
1352 num_patches
= MIN2(num_patches
, hardware_lds_size
/ (input_patch_size
+ output_patch_size
));
1354 /* Make sure the output data fits in the offchip buffer */
1355 num_patches
= MIN2(num_patches
,
1356 (pipeline
->device
->tess_offchip_block_dw_size
* 4) /
1359 /* Not necessary for correctness, but improves performance. The
1360 * specific value is taken from the proprietary driver.
1362 num_patches
= MIN2(num_patches
, 40);
1364 /* SI bug workaround - limit LS-HS threadgroups to only one wave. */
1365 if (pipeline
->device
->physical_device
->rad_info
.chip_class
== SI
) {
1366 unsigned one_wave
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
);
1367 num_patches
= MIN2(num_patches
, one_wave
);
1370 output_patch0_offset
= input_patch_size
* num_patches
;
1371 perpatch_output_offset
= output_patch0_offset
+ pervertex_output_patch_size
;
1373 lds_size
= output_patch0_offset
+ output_patch_size
* num_patches
;
1375 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1376 assert(lds_size
<= 65536);
1377 lds_size
= align(lds_size
, 512) / 512;
1379 assert(lds_size
<= 32768);
1380 lds_size
= align(lds_size
, 256) / 256;
1382 si_multiwave_lds_size_workaround(pipeline
->device
, &lds_size
);
1384 tess
.lds_size
= lds_size
;
1386 tess
.tcs_in_layout
= (input_patch_size
/ 4) |
1387 ((input_vertex_size
/ 4) << 13);
1388 tess
.tcs_out_layout
= (output_patch_size
/ 4) |
1389 ((output_vertex_size
/ 4) << 13);
1390 tess
.tcs_out_offsets
= (output_patch0_offset
/ 16) |
1391 ((perpatch_output_offset
/ 16) << 16);
1392 tess
.offchip_layout
= (pervertex_output_patch_size
* num_patches
<< 16) |
1395 tess
.ls_hs_config
= S_028B58_NUM_PATCHES(num_patches
) |
1396 S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp
) |
1397 S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp
);
1398 tess
.num_patches
= num_patches
;
1400 struct radv_shader_variant
*tes
= radv_get_tess_eval_shader(pipeline
);
1401 unsigned type
= 0, partitioning
= 0, topology
= 0, distribution_mode
= 0;
1403 switch (tes
->info
.tes
.primitive_mode
) {
1405 type
= V_028B6C_TESS_TRIANGLE
;
1408 type
= V_028B6C_TESS_QUAD
;
1411 type
= V_028B6C_TESS_ISOLINE
;
1415 switch (tes
->info
.tes
.spacing
) {
1416 case TESS_SPACING_EQUAL
:
1417 partitioning
= V_028B6C_PART_INTEGER
;
1419 case TESS_SPACING_FRACTIONAL_ODD
:
1420 partitioning
= V_028B6C_PART_FRAC_ODD
;
1422 case TESS_SPACING_FRACTIONAL_EVEN
:
1423 partitioning
= V_028B6C_PART_FRAC_EVEN
;
1429 bool ccw
= tes
->info
.tes
.ccw
;
1430 const VkPipelineTessellationDomainOriginStateCreateInfoKHR
*domain_origin_state
=
1431 vk_find_struct_const(pCreateInfo
->pTessellationState
,
1432 PIPELINE_TESSELLATION_DOMAIN_ORIGIN_STATE_CREATE_INFO_KHR
);
1434 if (domain_origin_state
&& domain_origin_state
->domainOrigin
!= VK_TESSELLATION_DOMAIN_ORIGIN_UPPER_LEFT_KHR
)
1437 if (tes
->info
.tes
.point_mode
)
1438 topology
= V_028B6C_OUTPUT_POINT
;
1439 else if (tes
->info
.tes
.primitive_mode
== GL_ISOLINES
)
1440 topology
= V_028B6C_OUTPUT_LINE
;
1442 topology
= V_028B6C_OUTPUT_TRIANGLE_CCW
;
1444 topology
= V_028B6C_OUTPUT_TRIANGLE_CW
;
1446 if (pipeline
->device
->has_distributed_tess
) {
1447 if (pipeline
->device
->physical_device
->rad_info
.family
== CHIP_FIJI
||
1448 pipeline
->device
->physical_device
->rad_info
.family
>= CHIP_POLARIS10
)
1449 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS
;
1451 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_DONUTS
;
1453 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_NO_DIST
;
1455 tess
.tf_param
= S_028B6C_TYPE(type
) |
1456 S_028B6C_PARTITIONING(partitioning
) |
1457 S_028B6C_TOPOLOGY(topology
) |
1458 S_028B6C_DISTRIBUTION_MODE(distribution_mode
);
1463 static const struct radv_prim_vertex_count prim_size_table
[] = {
1464 [V_008958_DI_PT_NONE
] = {0, 0},
1465 [V_008958_DI_PT_POINTLIST
] = {1, 1},
1466 [V_008958_DI_PT_LINELIST
] = {2, 2},
1467 [V_008958_DI_PT_LINESTRIP
] = {2, 1},
1468 [V_008958_DI_PT_TRILIST
] = {3, 3},
1469 [V_008958_DI_PT_TRIFAN
] = {3, 1},
1470 [V_008958_DI_PT_TRISTRIP
] = {3, 1},
1471 [V_008958_DI_PT_LINELIST_ADJ
] = {4, 4},
1472 [V_008958_DI_PT_LINESTRIP_ADJ
] = {4, 1},
1473 [V_008958_DI_PT_TRILIST_ADJ
] = {6, 6},
1474 [V_008958_DI_PT_TRISTRIP_ADJ
] = {6, 2},
1475 [V_008958_DI_PT_RECTLIST
] = {3, 3},
1476 [V_008958_DI_PT_LINELOOP
] = {2, 1},
1477 [V_008958_DI_PT_POLYGON
] = {3, 1},
1478 [V_008958_DI_PT_2D_TRI_STRIP
] = {0, 0},
1481 static const struct radv_vs_output_info
*get_vs_output_info(const struct radv_pipeline
*pipeline
)
1483 if (radv_pipeline_has_gs(pipeline
))
1484 return &pipeline
->gs_copy_shader
->info
.vs
.outinfo
;
1485 else if (radv_pipeline_has_tess(pipeline
))
1486 return &pipeline
->shaders
[MESA_SHADER_TESS_EVAL
]->info
.tes
.outinfo
;
1488 return &pipeline
->shaders
[MESA_SHADER_VERTEX
]->info
.vs
.outinfo
;
1492 radv_link_shaders(struct radv_pipeline
*pipeline
, nir_shader
**shaders
)
1494 nir_shader
* ordered_shaders
[MESA_SHADER_STAGES
];
1495 int shader_count
= 0;
1497 if(shaders
[MESA_SHADER_FRAGMENT
]) {
1498 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_FRAGMENT
];
1500 if(shaders
[MESA_SHADER_GEOMETRY
]) {
1501 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_GEOMETRY
];
1503 if(shaders
[MESA_SHADER_TESS_EVAL
]) {
1504 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_TESS_EVAL
];
1506 if(shaders
[MESA_SHADER_TESS_CTRL
]) {
1507 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_TESS_CTRL
];
1509 if(shaders
[MESA_SHADER_VERTEX
]) {
1510 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_VERTEX
];
1513 for (int i
= 1; i
< shader_count
; ++i
) {
1514 nir_lower_io_arrays_to_elements(ordered_shaders
[i
],
1515 ordered_shaders
[i
- 1]);
1517 nir_remove_dead_variables(ordered_shaders
[i
],
1518 nir_var_shader_out
);
1519 nir_remove_dead_variables(ordered_shaders
[i
- 1],
1522 bool progress
= nir_remove_unused_varyings(ordered_shaders
[i
],
1523 ordered_shaders
[i
- 1]);
1525 nir_compact_varyings(ordered_shaders
[i
],
1526 ordered_shaders
[i
- 1], true);
1529 if (nir_lower_global_vars_to_local(ordered_shaders
[i
])) {
1530 ac_lower_indirect_derefs(ordered_shaders
[i
],
1531 pipeline
->device
->physical_device
->rad_info
.chip_class
);
1533 radv_optimize_nir(ordered_shaders
[i
]);
1535 if (nir_lower_global_vars_to_local(ordered_shaders
[i
- 1])) {
1536 ac_lower_indirect_derefs(ordered_shaders
[i
- 1],
1537 pipeline
->device
->physical_device
->rad_info
.chip_class
);
1539 radv_optimize_nir(ordered_shaders
[i
- 1]);
1545 static struct radv_pipeline_key
1546 radv_generate_graphics_pipeline_key(struct radv_pipeline
*pipeline
,
1547 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1548 const struct radv_blend_state
*blend
,
1549 bool has_view_index
)
1551 const VkPipelineVertexInputStateCreateInfo
*input_state
=
1552 pCreateInfo
->pVertexInputState
;
1553 struct radv_pipeline_key key
;
1554 memset(&key
, 0, sizeof(key
));
1556 key
.has_multiview_view_index
= has_view_index
;
1558 uint32_t binding_input_rate
= 0;
1559 for (unsigned i
= 0; i
< input_state
->vertexBindingDescriptionCount
; ++i
) {
1560 if (input_state
->pVertexBindingDescriptions
[i
].inputRate
)
1561 binding_input_rate
|= 1u << input_state
->pVertexBindingDescriptions
[i
].binding
;
1564 for (unsigned i
= 0; i
< input_state
->vertexAttributeDescriptionCount
; ++i
) {
1566 binding
= input_state
->pVertexAttributeDescriptions
[i
].binding
;
1567 if (binding_input_rate
& (1u << binding
))
1568 key
.instance_rate_inputs
|= 1u << input_state
->pVertexAttributeDescriptions
[i
].location
;
1571 if (pCreateInfo
->pTessellationState
)
1572 key
.tess_input_vertices
= pCreateInfo
->pTessellationState
->patchControlPoints
;
1575 if (pCreateInfo
->pMultisampleState
&&
1576 pCreateInfo
->pMultisampleState
->rasterizationSamples
> 1) {
1577 uint32_t num_samples
= pCreateInfo
->pMultisampleState
->rasterizationSamples
;
1578 uint32_t ps_iter_samples
= radv_pipeline_get_ps_iter_samples(pCreateInfo
->pMultisampleState
);
1579 key
.multisample
= true;
1580 key
.log2_num_samples
= util_logbase2(num_samples
);
1581 key
.log2_ps_iter_samples
= util_logbase2(ps_iter_samples
);
1584 key
.col_format
= blend
->spi_shader_col_format
;
1585 if (pipeline
->device
->physical_device
->rad_info
.chip_class
< VI
)
1586 radv_pipeline_compute_get_int_clamp(pCreateInfo
, &key
.is_int8
, &key
.is_int10
);
1592 radv_fill_shader_keys(struct radv_shader_variant_key
*keys
,
1593 const struct radv_pipeline_key
*key
,
1596 keys
[MESA_SHADER_VERTEX
].vs
.instance_rate_inputs
= key
->instance_rate_inputs
;
1598 if (nir
[MESA_SHADER_TESS_CTRL
]) {
1599 keys
[MESA_SHADER_VERTEX
].vs
.as_ls
= true;
1600 keys
[MESA_SHADER_TESS_CTRL
].tcs
.input_vertices
= key
->tess_input_vertices
;
1601 keys
[MESA_SHADER_TESS_CTRL
].tcs
.primitive_mode
= nir
[MESA_SHADER_TESS_EVAL
]->info
.tess
.primitive_mode
;
1603 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
));
1606 if (nir
[MESA_SHADER_GEOMETRY
]) {
1607 if (nir
[MESA_SHADER_TESS_CTRL
])
1608 keys
[MESA_SHADER_TESS_EVAL
].tes
.as_es
= true;
1610 keys
[MESA_SHADER_VERTEX
].vs
.as_es
= true;
1613 for(int i
= 0; i
< MESA_SHADER_STAGES
; ++i
)
1614 keys
[i
].has_multiview_view_index
= key
->has_multiview_view_index
;
1616 keys
[MESA_SHADER_FRAGMENT
].fs
.multisample
= key
->multisample
;
1617 keys
[MESA_SHADER_FRAGMENT
].fs
.col_format
= key
->col_format
;
1618 keys
[MESA_SHADER_FRAGMENT
].fs
.is_int8
= key
->is_int8
;
1619 keys
[MESA_SHADER_FRAGMENT
].fs
.is_int10
= key
->is_int10
;
1620 keys
[MESA_SHADER_FRAGMENT
].fs
.log2_ps_iter_samples
= key
->log2_ps_iter_samples
;
1621 keys
[MESA_SHADER_FRAGMENT
].fs
.log2_num_samples
= key
->log2_num_samples
;
1625 merge_tess_info(struct shader_info
*tes_info
,
1626 const struct shader_info
*tcs_info
)
1628 /* The Vulkan 1.0.38 spec, section 21.1 Tessellator says:
1630 * "PointMode. Controls generation of points rather than triangles
1631 * or lines. This functionality defaults to disabled, and is
1632 * enabled if either shader stage includes the execution mode.
1634 * and about Triangles, Quads, IsoLines, VertexOrderCw, VertexOrderCcw,
1635 * PointMode, SpacingEqual, SpacingFractionalEven, SpacingFractionalOdd,
1636 * and OutputVertices, it says:
1638 * "One mode must be set in at least one of the tessellation
1641 * So, the fields can be set in either the TCS or TES, but they must
1642 * agree if set in both. Our backend looks at TES, so bitwise-or in
1643 * the values from the TCS.
1645 assert(tcs_info
->tess
.tcs_vertices_out
== 0 ||
1646 tes_info
->tess
.tcs_vertices_out
== 0 ||
1647 tcs_info
->tess
.tcs_vertices_out
== tes_info
->tess
.tcs_vertices_out
);
1648 tes_info
->tess
.tcs_vertices_out
|= tcs_info
->tess
.tcs_vertices_out
;
1650 assert(tcs_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
1651 tes_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
1652 tcs_info
->tess
.spacing
== tes_info
->tess
.spacing
);
1653 tes_info
->tess
.spacing
|= tcs_info
->tess
.spacing
;
1655 assert(tcs_info
->tess
.primitive_mode
== 0 ||
1656 tes_info
->tess
.primitive_mode
== 0 ||
1657 tcs_info
->tess
.primitive_mode
== tes_info
->tess
.primitive_mode
);
1658 tes_info
->tess
.primitive_mode
|= tcs_info
->tess
.primitive_mode
;
1659 tes_info
->tess
.ccw
|= tcs_info
->tess
.ccw
;
1660 tes_info
->tess
.point_mode
|= tcs_info
->tess
.point_mode
;
1664 void radv_create_shaders(struct radv_pipeline
*pipeline
,
1665 struct radv_device
*device
,
1666 struct radv_pipeline_cache
*cache
,
1667 struct radv_pipeline_key key
,
1668 const VkPipelineShaderStageCreateInfo
**pStages
)
1670 struct radv_shader_module fs_m
= {0};
1671 struct radv_shader_module
*modules
[MESA_SHADER_STAGES
] = { 0, };
1672 nir_shader
*nir
[MESA_SHADER_STAGES
] = {0};
1673 void *codes
[MESA_SHADER_STAGES
] = {0};
1674 unsigned code_sizes
[MESA_SHADER_STAGES
] = {0};
1675 struct radv_shader_variant_key keys
[MESA_SHADER_STAGES
] = {{{{0}}}};
1676 unsigned char hash
[20], gs_copy_hash
[20];
1678 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1680 modules
[i
] = radv_shader_module_from_handle(pStages
[i
]->module
);
1681 if (modules
[i
]->nir
)
1682 _mesa_sha1_compute(modules
[i
]->nir
->info
.name
,
1683 strlen(modules
[i
]->nir
->info
.name
),
1688 radv_hash_shaders(hash
, pStages
, pipeline
->layout
, &key
, get_hash_flags(device
));
1689 memcpy(gs_copy_hash
, hash
, 20);
1690 gs_copy_hash
[0] ^= 1;
1692 if (modules
[MESA_SHADER_GEOMETRY
]) {
1693 struct radv_shader_variant
*variants
[MESA_SHADER_STAGES
] = {0};
1694 radv_create_shader_variants_from_pipeline_cache(device
, cache
, gs_copy_hash
, variants
);
1695 pipeline
->gs_copy_shader
= variants
[MESA_SHADER_GEOMETRY
];
1698 if (radv_create_shader_variants_from_pipeline_cache(device
, cache
, hash
, pipeline
->shaders
) &&
1699 (!modules
[MESA_SHADER_GEOMETRY
] || pipeline
->gs_copy_shader
)) {
1700 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1701 if (pipeline
->shaders
[i
])
1702 pipeline
->active_stages
|= mesa_to_vk_shader_stage(i
);
1707 if (!modules
[MESA_SHADER_FRAGMENT
] && !modules
[MESA_SHADER_COMPUTE
]) {
1709 nir_builder_init_simple_shader(&fs_b
, NULL
, MESA_SHADER_FRAGMENT
, NULL
);
1710 fs_b
.shader
->info
.name
= ralloc_strdup(fs_b
.shader
, "noop_fs");
1711 fs_m
.nir
= fs_b
.shader
;
1712 modules
[MESA_SHADER_FRAGMENT
] = &fs_m
;
1715 /* Determine first and last stage. */
1716 unsigned first
= MESA_SHADER_STAGES
;
1718 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
1721 if (first
== MESA_SHADER_STAGES
)
1726 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1727 const VkPipelineShaderStageCreateInfo
*stage
= pStages
[i
];
1732 nir
[i
] = radv_shader_compile_to_nir(device
, modules
[i
],
1733 stage
? stage
->pName
: "main", i
,
1734 stage
? stage
->pSpecializationInfo
: NULL
);
1735 pipeline
->active_stages
|= mesa_to_vk_shader_stage(i
);
1737 /* We don't want to alter meta shaders IR directly so clone it
1740 if (nir
[i
]->info
.name
) {
1741 nir
[i
] = nir_shader_clone(NULL
, nir
[i
]);
1744 if (first
!= last
) {
1745 nir_variable_mode mask
= 0;
1748 mask
= mask
| nir_var_shader_in
;
1751 mask
= mask
| nir_var_shader_out
;
1753 nir_lower_io_to_scalar_early(nir
[i
], mask
);
1754 radv_optimize_nir(nir
[i
]);
1758 if (nir
[MESA_SHADER_TESS_CTRL
]) {
1759 nir_lower_tes_patch_vertices(nir
[MESA_SHADER_TESS_EVAL
], nir
[MESA_SHADER_TESS_CTRL
]->info
.tess
.tcs_vertices_out
);
1760 merge_tess_info(&nir
[MESA_SHADER_TESS_EVAL
]->info
, &nir
[MESA_SHADER_TESS_CTRL
]->info
);
1763 radv_link_shaders(pipeline
, nir
);
1765 for (int i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1766 if (modules
[i
] && radv_can_dump_shader(device
, modules
[i
]))
1767 nir_print_shader(nir
[i
], stderr
);
1770 radv_fill_shader_keys(keys
, &key
, nir
);
1772 if (nir
[MESA_SHADER_FRAGMENT
]) {
1773 if (!pipeline
->shaders
[MESA_SHADER_FRAGMENT
]) {
1774 pipeline
->shaders
[MESA_SHADER_FRAGMENT
] =
1775 radv_shader_variant_create(device
, modules
[MESA_SHADER_FRAGMENT
], &nir
[MESA_SHADER_FRAGMENT
], 1,
1776 pipeline
->layout
, keys
+ MESA_SHADER_FRAGMENT
,
1777 &codes
[MESA_SHADER_FRAGMENT
], &code_sizes
[MESA_SHADER_FRAGMENT
]);
1780 /* TODO: These are no longer used as keys we should refactor this */
1781 keys
[MESA_SHADER_VERTEX
].vs
.export_prim_id
=
1782 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.prim_id_input
;
1783 keys
[MESA_SHADER_TESS_EVAL
].tes
.export_prim_id
=
1784 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.prim_id_input
;
1787 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
&& modules
[MESA_SHADER_TESS_CTRL
]) {
1788 if (!pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]) {
1789 struct nir_shader
*combined_nir
[] = {nir
[MESA_SHADER_VERTEX
], nir
[MESA_SHADER_TESS_CTRL
]};
1790 struct radv_shader_variant_key key
= keys
[MESA_SHADER_TESS_CTRL
];
1791 key
.tcs
.vs_key
= keys
[MESA_SHADER_VERTEX
].vs
;
1792 pipeline
->shaders
[MESA_SHADER_TESS_CTRL
] = radv_shader_variant_create(device
, modules
[MESA_SHADER_TESS_CTRL
], combined_nir
, 2,
1794 &key
, &codes
[MESA_SHADER_TESS_CTRL
],
1795 &code_sizes
[MESA_SHADER_TESS_CTRL
]);
1797 modules
[MESA_SHADER_VERTEX
] = NULL
;
1800 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
&& modules
[MESA_SHADER_GEOMETRY
]) {
1801 gl_shader_stage pre_stage
= modules
[MESA_SHADER_TESS_EVAL
] ? MESA_SHADER_TESS_EVAL
: MESA_SHADER_VERTEX
;
1802 if (!pipeline
->shaders
[MESA_SHADER_GEOMETRY
]) {
1803 struct nir_shader
*combined_nir
[] = {nir
[pre_stage
], nir
[MESA_SHADER_GEOMETRY
]};
1804 pipeline
->shaders
[MESA_SHADER_GEOMETRY
] = radv_shader_variant_create(device
, modules
[MESA_SHADER_GEOMETRY
], combined_nir
, 2,
1806 &keys
[pre_stage
] , &codes
[MESA_SHADER_GEOMETRY
],
1807 &code_sizes
[MESA_SHADER_GEOMETRY
]);
1809 modules
[pre_stage
] = NULL
;
1812 for (int i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1813 if(modules
[i
] && !pipeline
->shaders
[i
]) {
1814 pipeline
->shaders
[i
] = radv_shader_variant_create(device
, modules
[i
], &nir
[i
], 1,
1816 keys
+ i
, &codes
[i
],
1821 if(modules
[MESA_SHADER_GEOMETRY
]) {
1822 void *gs_copy_code
= NULL
;
1823 unsigned gs_copy_code_size
= 0;
1824 if (!pipeline
->gs_copy_shader
) {
1825 pipeline
->gs_copy_shader
= radv_create_gs_copy_shader(
1826 device
, nir
[MESA_SHADER_GEOMETRY
], &gs_copy_code
,
1828 keys
[MESA_SHADER_GEOMETRY
].has_multiview_view_index
);
1831 if (pipeline
->gs_copy_shader
) {
1832 void *code
[MESA_SHADER_STAGES
] = {0};
1833 unsigned code_size
[MESA_SHADER_STAGES
] = {0};
1834 struct radv_shader_variant
*variants
[MESA_SHADER_STAGES
] = {0};
1836 code
[MESA_SHADER_GEOMETRY
] = gs_copy_code
;
1837 code_size
[MESA_SHADER_GEOMETRY
] = gs_copy_code_size
;
1838 variants
[MESA_SHADER_GEOMETRY
] = pipeline
->gs_copy_shader
;
1840 radv_pipeline_cache_insert_shaders(device
, cache
,
1849 radv_pipeline_cache_insert_shaders(device
, cache
, hash
, pipeline
->shaders
,
1850 (const void**)codes
, code_sizes
);
1852 for (int i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1855 if (!pipeline
->device
->keep_shader_info
)
1856 ralloc_free(nir
[i
]);
1858 if (radv_can_dump_shader_stats(device
, modules
[i
]))
1859 radv_shader_dump_stats(device
,
1860 pipeline
->shaders
[i
],
1866 ralloc_free(fs_m
.nir
);
1870 radv_pipeline_stage_to_user_data_0(struct radv_pipeline
*pipeline
,
1871 gl_shader_stage stage
, enum chip_class chip_class
)
1873 bool has_gs
= radv_pipeline_has_gs(pipeline
);
1874 bool has_tess
= radv_pipeline_has_tess(pipeline
);
1876 case MESA_SHADER_FRAGMENT
:
1877 return R_00B030_SPI_SHADER_USER_DATA_PS_0
;
1878 case MESA_SHADER_VERTEX
:
1879 if (chip_class
>= GFX9
) {
1880 return has_tess
? R_00B430_SPI_SHADER_USER_DATA_LS_0
:
1881 has_gs
? R_00B330_SPI_SHADER_USER_DATA_ES_0
:
1882 R_00B130_SPI_SHADER_USER_DATA_VS_0
;
1885 return R_00B530_SPI_SHADER_USER_DATA_LS_0
;
1887 return has_gs
? R_00B330_SPI_SHADER_USER_DATA_ES_0
: R_00B130_SPI_SHADER_USER_DATA_VS_0
;
1888 case MESA_SHADER_GEOMETRY
:
1889 return chip_class
>= GFX9
? R_00B330_SPI_SHADER_USER_DATA_ES_0
:
1890 R_00B230_SPI_SHADER_USER_DATA_GS_0
;
1891 case MESA_SHADER_COMPUTE
:
1892 return R_00B900_COMPUTE_USER_DATA_0
;
1893 case MESA_SHADER_TESS_CTRL
:
1894 return chip_class
>= GFX9
? R_00B430_SPI_SHADER_USER_DATA_LS_0
:
1895 R_00B430_SPI_SHADER_USER_DATA_HS_0
;
1896 case MESA_SHADER_TESS_EVAL
:
1897 if (chip_class
>= GFX9
) {
1898 return has_gs
? R_00B330_SPI_SHADER_USER_DATA_ES_0
:
1899 R_00B130_SPI_SHADER_USER_DATA_VS_0
;
1902 return R_00B330_SPI_SHADER_USER_DATA_ES_0
;
1904 return R_00B130_SPI_SHADER_USER_DATA_VS_0
;
1906 unreachable("unknown shader");
1910 struct radv_bin_size_entry
{
1916 radv_compute_bin_size(struct radv_pipeline
*pipeline
, const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1918 static const struct radv_bin_size_entry color_size_table
[][3][9] = {
1922 /* One shader engine */
1928 { UINT_MAX
, { 0, 0}},
1931 /* Two shader engines */
1937 { UINT_MAX
, { 0, 0}},
1940 /* Four shader engines */
1945 { UINT_MAX
, { 0, 0}},
1951 /* One shader engine */
1957 { UINT_MAX
, { 0, 0}},
1960 /* Two shader engines */
1966 { UINT_MAX
, { 0, 0}},
1969 /* Four shader engines */
1976 { UINT_MAX
, { 0, 0}},
1982 /* One shader engine */
1989 { UINT_MAX
, { 0, 0}},
1992 /* Two shader engines */
2000 { UINT_MAX
, { 0, 0}},
2003 /* Four shader engines */
2011 { UINT_MAX
, { 0, 0}},
2015 static const struct radv_bin_size_entry ds_size_table
[][3][9] = {
2019 // One shader engine
2026 { UINT_MAX
, { 0, 0}},
2029 // Two shader engines
2037 { UINT_MAX
, { 0, 0}},
2040 // Four shader engines
2048 { UINT_MAX
, { 0, 0}},
2054 // One shader engine
2062 { UINT_MAX
, { 0, 0}},
2065 // Two shader engines
2074 { UINT_MAX
, { 0, 0}},
2077 // Four shader engines
2086 { UINT_MAX
, { 0, 0}},
2092 // One shader engine
2100 { UINT_MAX
, { 0, 0}},
2103 // Two shader engines
2112 { UINT_MAX
, { 0, 0}},
2115 // Four shader engines
2123 { UINT_MAX
, { 0, 0}},
2128 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
2129 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
2130 VkExtent2D extent
= {512, 512};
2132 unsigned log_num_rb_per_se
=
2133 util_logbase2_ceil(pipeline
->device
->physical_device
->rad_info
.num_render_backends
/
2134 pipeline
->device
->physical_device
->rad_info
.max_se
);
2135 unsigned log_num_se
= util_logbase2_ceil(pipeline
->device
->physical_device
->rad_info
.max_se
);
2137 unsigned total_samples
= 1u << G_028BE0_MSAA_NUM_SAMPLES(pipeline
->graphics
.ms
.pa_sc_mode_cntl_1
);
2138 unsigned ps_iter_samples
= 1u << G_028804_PS_ITER_SAMPLES(pipeline
->graphics
.ms
.db_eqaa
);
2139 unsigned effective_samples
= total_samples
;
2140 unsigned color_bytes_per_pixel
= 0;
2142 const VkPipelineColorBlendStateCreateInfo
*vkblend
= pCreateInfo
->pColorBlendState
;
2144 for (unsigned i
= 0; i
< subpass
->color_count
; i
++) {
2145 if (!vkblend
->pAttachments
[i
].colorWriteMask
)
2148 if (subpass
->color_attachments
[i
].attachment
== VK_ATTACHMENT_UNUSED
)
2151 VkFormat format
= pass
->attachments
[subpass
->color_attachments
[i
].attachment
].format
;
2152 color_bytes_per_pixel
+= vk_format_get_blocksize(format
);
2155 /* MSAA images typically don't use all samples all the time. */
2156 if (effective_samples
>= 2 && ps_iter_samples
<= 1)
2157 effective_samples
= 2;
2158 color_bytes_per_pixel
*= effective_samples
;
2161 const struct radv_bin_size_entry
*color_entry
= color_size_table
[log_num_rb_per_se
][log_num_se
];
2162 while(color_entry
->bpp
<= color_bytes_per_pixel
)
2165 extent
= color_entry
->extent
;
2167 if (subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
2168 struct radv_render_pass_attachment
*attachment
= pass
->attachments
+ subpass
->depth_stencil_attachment
.attachment
;
2170 /* Coefficients taken from AMDVLK */
2171 unsigned depth_coeff
= vk_format_is_depth(attachment
->format
) ? 5 : 0;
2172 unsigned stencil_coeff
= vk_format_is_stencil(attachment
->format
) ? 1 : 0;
2173 unsigned ds_bytes_per_pixel
= 4 * (depth_coeff
+ stencil_coeff
) * total_samples
;
2175 const struct radv_bin_size_entry
*ds_entry
= ds_size_table
[log_num_rb_per_se
][log_num_se
];
2176 while(ds_entry
->bpp
<= ds_bytes_per_pixel
)
2179 extent
.width
= MIN2(extent
.width
, ds_entry
->extent
.width
);
2180 extent
.height
= MIN2(extent
.height
, ds_entry
->extent
.height
);
2187 radv_pipeline_generate_binning_state(struct radeon_winsys_cs
*cs
,
2188 struct radv_pipeline
*pipeline
,
2189 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
2191 if (pipeline
->device
->physical_device
->rad_info
.chip_class
< GFX9
)
2194 uint32_t pa_sc_binner_cntl_0
=
2195 S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_LEGACY_SC
) |
2196 S_028C44_DISABLE_START_OF_PRIM(1);
2197 uint32_t db_dfsm_control
= S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF
);
2199 VkExtent2D bin_size
= radv_compute_bin_size(pipeline
, pCreateInfo
);
2201 unsigned context_states_per_bin
; /* allowed range: [1, 6] */
2202 unsigned persistent_states_per_bin
; /* allowed range: [1, 32] */
2203 unsigned fpovs_per_batch
; /* allowed range: [0, 255], 0 = unlimited */
2205 switch (pipeline
->device
->physical_device
->rad_info
.family
) {
2207 context_states_per_bin
= 1;
2208 persistent_states_per_bin
= 1;
2209 fpovs_per_batch
= 63;
2212 context_states_per_bin
= 6;
2213 persistent_states_per_bin
= 32;
2214 fpovs_per_batch
= 63;
2217 unreachable("unhandled family while determining binning state.");
2220 if (pipeline
->device
->pbb_allowed
&& bin_size
.width
&& bin_size
.height
) {
2221 pa_sc_binner_cntl_0
=
2222 S_028C44_BINNING_MODE(V_028C44_BINNING_ALLOWED
) |
2223 S_028C44_BIN_SIZE_X(bin_size
.width
== 16) |
2224 S_028C44_BIN_SIZE_Y(bin_size
.height
== 16) |
2225 S_028C44_BIN_SIZE_X_EXTEND(util_logbase2(MAX2(bin_size
.width
, 32)) - 5) |
2226 S_028C44_BIN_SIZE_Y_EXTEND(util_logbase2(MAX2(bin_size
.height
, 32)) - 5) |
2227 S_028C44_CONTEXT_STATES_PER_BIN(context_states_per_bin
- 1) |
2228 S_028C44_PERSISTENT_STATES_PER_BIN(persistent_states_per_bin
- 1) |
2229 S_028C44_DISABLE_START_OF_PRIM(1) |
2230 S_028C44_FPOVS_PER_BATCH(fpovs_per_batch
) |
2231 S_028C44_OPTIMAL_BIN_SELECTION(1);
2234 radeon_set_context_reg(cs
, R_028C44_PA_SC_BINNER_CNTL_0
,
2235 pa_sc_binner_cntl_0
);
2236 radeon_set_context_reg(cs
, R_028060_DB_DFSM_CONTROL
,
2242 radv_pipeline_generate_depth_stencil_state(struct radeon_winsys_cs
*cs
,
2243 struct radv_pipeline
*pipeline
,
2244 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
2245 const struct radv_graphics_pipeline_create_info
*extra
)
2247 const VkPipelineDepthStencilStateCreateInfo
*vkds
= pCreateInfo
->pDepthStencilState
;
2248 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
2249 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
2250 struct radv_render_pass_attachment
*attachment
= NULL
;
2251 uint32_t db_depth_control
= 0, db_stencil_control
= 0;
2252 uint32_t db_render_control
= 0, db_render_override2
= 0;
2254 if (subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
)
2255 attachment
= pass
->attachments
+ subpass
->depth_stencil_attachment
.attachment
;
2257 bool has_depth_attachment
= attachment
&& vk_format_is_depth(attachment
->format
);
2258 bool has_stencil_attachment
= attachment
&& vk_format_is_stencil(attachment
->format
);
2260 if (vkds
&& has_depth_attachment
) {
2261 db_depth_control
= S_028800_Z_ENABLE(vkds
->depthTestEnable
? 1 : 0) |
2262 S_028800_Z_WRITE_ENABLE(vkds
->depthWriteEnable
? 1 : 0) |
2263 S_028800_ZFUNC(vkds
->depthCompareOp
) |
2264 S_028800_DEPTH_BOUNDS_ENABLE(vkds
->depthBoundsTestEnable
? 1 : 0);
2266 /* from amdvlk: For 4xAA and 8xAA need to decompress on flush for better performance */
2267 db_render_override2
|= S_028010_DECOMPRESS_Z_ON_FLUSH(attachment
->samples
> 2);
2270 if (has_stencil_attachment
&& vkds
&& vkds
->stencilTestEnable
) {
2271 db_depth_control
|= S_028800_STENCIL_ENABLE(1) | S_028800_BACKFACE_ENABLE(1);
2272 db_depth_control
|= S_028800_STENCILFUNC(vkds
->front
.compareOp
);
2273 db_stencil_control
|= S_02842C_STENCILFAIL(si_translate_stencil_op(vkds
->front
.failOp
));
2274 db_stencil_control
|= S_02842C_STENCILZPASS(si_translate_stencil_op(vkds
->front
.passOp
));
2275 db_stencil_control
|= S_02842C_STENCILZFAIL(si_translate_stencil_op(vkds
->front
.depthFailOp
));
2277 db_depth_control
|= S_028800_STENCILFUNC_BF(vkds
->back
.compareOp
);
2278 db_stencil_control
|= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(vkds
->back
.failOp
));
2279 db_stencil_control
|= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(vkds
->back
.passOp
));
2280 db_stencil_control
|= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(vkds
->back
.depthFailOp
));
2283 if (attachment
&& extra
) {
2284 db_render_control
|= S_028000_DEPTH_CLEAR_ENABLE(extra
->db_depth_clear
);
2285 db_render_control
|= S_028000_STENCIL_CLEAR_ENABLE(extra
->db_stencil_clear
);
2287 db_render_control
|= S_028000_RESUMMARIZE_ENABLE(extra
->db_resummarize
);
2288 db_render_control
|= S_028000_DEPTH_COMPRESS_DISABLE(extra
->db_flush_depth_inplace
);
2289 db_render_control
|= S_028000_STENCIL_COMPRESS_DISABLE(extra
->db_flush_stencil_inplace
);
2290 db_render_override2
|= S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(extra
->db_depth_disable_expclear
);
2291 db_render_override2
|= S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(extra
->db_stencil_disable_expclear
);
2294 radeon_set_context_reg(cs
, R_028800_DB_DEPTH_CONTROL
, db_depth_control
);
2295 radeon_set_context_reg(cs
, R_02842C_DB_STENCIL_CONTROL
, db_stencil_control
);
2297 radeon_set_context_reg(cs
, R_028000_DB_RENDER_CONTROL
, db_render_control
);
2298 radeon_set_context_reg(cs
, R_028010_DB_RENDER_OVERRIDE2
, db_render_override2
);
2302 radv_pipeline_generate_blend_state(struct radeon_winsys_cs
*cs
,
2303 struct radv_pipeline
*pipeline
,
2304 const struct radv_blend_state
*blend
)
2306 radeon_set_context_reg_seq(cs
, R_028780_CB_BLEND0_CONTROL
, 8);
2307 radeon_emit_array(cs
, blend
->cb_blend_control
,
2309 radeon_set_context_reg(cs
, R_028808_CB_COLOR_CONTROL
, blend
->cb_color_control
);
2310 radeon_set_context_reg(cs
, R_028B70_DB_ALPHA_TO_MASK
, blend
->db_alpha_to_mask
);
2312 if (pipeline
->device
->physical_device
->has_rbplus
) {
2314 radeon_set_context_reg_seq(cs
, R_028760_SX_MRT0_BLEND_OPT
, 8);
2315 radeon_emit_array(cs
, blend
->sx_mrt_blend_opt
, 8);
2317 radeon_set_context_reg_seq(cs
, R_028754_SX_PS_DOWNCONVERT
, 3);
2318 radeon_emit(cs
, 0); /* R_028754_SX_PS_DOWNCONVERT */
2319 radeon_emit(cs
, 0); /* R_028758_SX_BLEND_OPT_EPSILON */
2320 radeon_emit(cs
, 0); /* R_02875C_SX_BLEND_OPT_CONTROL */
2323 radeon_set_context_reg(cs
, R_028714_SPI_SHADER_COL_FORMAT
, blend
->spi_shader_col_format
);
2325 radeon_set_context_reg(cs
, R_028238_CB_TARGET_MASK
, blend
->cb_target_mask
);
2326 radeon_set_context_reg(cs
, R_02823C_CB_SHADER_MASK
, blend
->cb_shader_mask
);
2331 radv_pipeline_generate_raster_state(struct radeon_winsys_cs
*cs
,
2332 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
2334 const VkPipelineRasterizationStateCreateInfo
*vkraster
= pCreateInfo
->pRasterizationState
;
2336 radeon_set_context_reg(cs
, R_028810_PA_CL_CLIP_CNTL
,
2337 S_028810_PS_UCP_MODE(3) |
2338 S_028810_DX_CLIP_SPACE_DEF(1) | // vulkan uses DX conventions.
2339 S_028810_ZCLIP_NEAR_DISABLE(vkraster
->depthClampEnable
? 1 : 0) |
2340 S_028810_ZCLIP_FAR_DISABLE(vkraster
->depthClampEnable
? 1 : 0) |
2341 S_028810_DX_RASTERIZATION_KILL(vkraster
->rasterizerDiscardEnable
? 1 : 0) |
2342 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1));
2344 radeon_set_context_reg(cs
, R_0286D4_SPI_INTERP_CONTROL_0
,
2345 S_0286D4_FLAT_SHADE_ENA(1) |
2346 S_0286D4_PNT_SPRITE_ENA(1) |
2347 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S
) |
2348 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T
) |
2349 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0
) |
2350 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1
) |
2351 S_0286D4_PNT_SPRITE_TOP_1(0)); /* vulkan is top to bottom - 1.0 at bottom */
2353 radeon_set_context_reg(cs
, R_028BE4_PA_SU_VTX_CNTL
,
2354 S_028BE4_PIX_CENTER(1) | // TODO verify
2355 S_028BE4_ROUND_MODE(V_028BE4_X_ROUND_TO_EVEN
) |
2356 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH
));
2358 radeon_set_context_reg(cs
, R_028814_PA_SU_SC_MODE_CNTL
,
2359 S_028814_FACE(vkraster
->frontFace
) |
2360 S_028814_CULL_FRONT(!!(vkraster
->cullMode
& VK_CULL_MODE_FRONT_BIT
)) |
2361 S_028814_CULL_BACK(!!(vkraster
->cullMode
& VK_CULL_MODE_BACK_BIT
)) |
2362 S_028814_POLY_MODE(vkraster
->polygonMode
!= VK_POLYGON_MODE_FILL
) |
2363 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(vkraster
->polygonMode
)) |
2364 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(vkraster
->polygonMode
)) |
2365 S_028814_POLY_OFFSET_FRONT_ENABLE(vkraster
->depthBiasEnable
? 1 : 0) |
2366 S_028814_POLY_OFFSET_BACK_ENABLE(vkraster
->depthBiasEnable
? 1 : 0) |
2367 S_028814_POLY_OFFSET_PARA_ENABLE(vkraster
->depthBiasEnable
? 1 : 0));
2372 radv_pipeline_generate_multisample_state(struct radeon_winsys_cs
*cs
,
2373 struct radv_pipeline
*pipeline
)
2375 struct radv_multisample_state
*ms
= &pipeline
->graphics
.ms
;
2377 radeon_set_context_reg_seq(cs
, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0
, 2);
2378 radeon_emit(cs
, ms
->pa_sc_aa_mask
[0]);
2379 radeon_emit(cs
, ms
->pa_sc_aa_mask
[1]);
2381 radeon_set_context_reg(cs
, R_028804_DB_EQAA
, ms
->db_eqaa
);
2382 radeon_set_context_reg(cs
, R_028A4C_PA_SC_MODE_CNTL_1
, ms
->pa_sc_mode_cntl_1
);
2384 if (pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.needs_sample_positions
) {
2386 struct radv_userdata_info
*loc
= radv_lookup_user_sgpr(pipeline
, MESA_SHADER_FRAGMENT
, AC_UD_PS_SAMPLE_POS_OFFSET
);
2387 uint32_t base_reg
= pipeline
->user_data_0
[MESA_SHADER_FRAGMENT
];
2388 if (loc
->sgpr_idx
== -1)
2390 assert(loc
->num_sgprs
== 1);
2391 assert(!loc
->indirect
);
2392 switch (pipeline
->graphics
.ms
.num_samples
) {
2410 radeon_set_sh_reg(cs
, base_reg
+ loc
->sgpr_idx
* 4, offset
);
2415 radv_pipeline_generate_vgt_gs_mode(struct radeon_winsys_cs
*cs
,
2416 const struct radv_pipeline
*pipeline
)
2418 const struct radv_vs_output_info
*outinfo
= get_vs_output_info(pipeline
);
2420 uint32_t vgt_primitiveid_en
= false;
2421 uint32_t vgt_gs_mode
= 0;
2423 if (radv_pipeline_has_gs(pipeline
)) {
2424 const struct radv_shader_variant
*gs
=
2425 pipeline
->shaders
[MESA_SHADER_GEOMETRY
];
2427 vgt_gs_mode
= ac_vgt_gs_mode(gs
->info
.gs
.vertices_out
,
2428 pipeline
->device
->physical_device
->rad_info
.chip_class
);
2429 } else if (outinfo
->export_prim_id
) {
2430 vgt_gs_mode
= S_028A40_MODE(V_028A40_GS_SCENARIO_A
);
2431 vgt_primitiveid_en
= true;
2434 radeon_set_context_reg(cs
, R_028A84_VGT_PRIMITIVEID_EN
, vgt_primitiveid_en
);
2435 radeon_set_context_reg(cs
, R_028A40_VGT_GS_MODE
, vgt_gs_mode
);
2439 radv_pipeline_generate_hw_vs(struct radeon_winsys_cs
*cs
,
2440 struct radv_pipeline
*pipeline
,
2441 struct radv_shader_variant
*shader
)
2443 uint64_t va
= radv_buffer_get_va(shader
->bo
) + shader
->bo_offset
;
2445 radeon_set_sh_reg_seq(cs
, R_00B120_SPI_SHADER_PGM_LO_VS
, 4);
2446 radeon_emit(cs
, va
>> 8);
2447 radeon_emit(cs
, va
>> 40);
2448 radeon_emit(cs
, shader
->rsrc1
);
2449 radeon_emit(cs
, shader
->rsrc2
);
2451 const struct radv_vs_output_info
*outinfo
= get_vs_output_info(pipeline
);
2452 unsigned clip_dist_mask
, cull_dist_mask
, total_mask
;
2453 clip_dist_mask
= outinfo
->clip_dist_mask
;
2454 cull_dist_mask
= outinfo
->cull_dist_mask
;
2455 total_mask
= clip_dist_mask
| cull_dist_mask
;
2456 bool misc_vec_ena
= outinfo
->writes_pointsize
||
2457 outinfo
->writes_layer
||
2458 outinfo
->writes_viewport_index
;
2460 radeon_set_context_reg(cs
, R_0286C4_SPI_VS_OUT_CONFIG
,
2461 S_0286C4_VS_EXPORT_COUNT(MAX2(1, outinfo
->param_exports
) - 1));
2463 radeon_set_context_reg(cs
, R_02870C_SPI_SHADER_POS_FORMAT
,
2464 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP
) |
2465 S_02870C_POS1_EXPORT_FORMAT(outinfo
->pos_exports
> 1 ?
2466 V_02870C_SPI_SHADER_4COMP
:
2467 V_02870C_SPI_SHADER_NONE
) |
2468 S_02870C_POS2_EXPORT_FORMAT(outinfo
->pos_exports
> 2 ?
2469 V_02870C_SPI_SHADER_4COMP
:
2470 V_02870C_SPI_SHADER_NONE
) |
2471 S_02870C_POS3_EXPORT_FORMAT(outinfo
->pos_exports
> 3 ?
2472 V_02870C_SPI_SHADER_4COMP
:
2473 V_02870C_SPI_SHADER_NONE
));
2475 radeon_set_context_reg(cs
, R_028818_PA_CL_VTE_CNTL
,
2476 S_028818_VTX_W0_FMT(1) |
2477 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
2478 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
2479 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
2481 radeon_set_context_reg(cs
, R_02881C_PA_CL_VS_OUT_CNTL
,
2482 S_02881C_USE_VTX_POINT_SIZE(outinfo
->writes_pointsize
) |
2483 S_02881C_USE_VTX_RENDER_TARGET_INDX(outinfo
->writes_layer
) |
2484 S_02881C_USE_VTX_VIEWPORT_INDX(outinfo
->writes_viewport_index
) |
2485 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena
) |
2486 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena
) |
2487 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask
& 0x0f) != 0) |
2488 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask
& 0xf0) != 0) |
2489 cull_dist_mask
<< 8 |
2492 if (pipeline
->device
->physical_device
->rad_info
.chip_class
<= VI
)
2493 radeon_set_context_reg(cs
, R_028AB4_VGT_REUSE_OFF
,
2494 outinfo
->writes_viewport_index
);
2498 radv_pipeline_generate_hw_es(struct radeon_winsys_cs
*cs
,
2499 struct radv_pipeline
*pipeline
,
2500 struct radv_shader_variant
*shader
)
2502 uint64_t va
= radv_buffer_get_va(shader
->bo
) + shader
->bo_offset
;
2504 radeon_set_sh_reg_seq(cs
, R_00B320_SPI_SHADER_PGM_LO_ES
, 4);
2505 radeon_emit(cs
, va
>> 8);
2506 radeon_emit(cs
, va
>> 40);
2507 radeon_emit(cs
, shader
->rsrc1
);
2508 radeon_emit(cs
, shader
->rsrc2
);
2512 radv_pipeline_generate_hw_ls(struct radeon_winsys_cs
*cs
,
2513 struct radv_pipeline
*pipeline
,
2514 struct radv_shader_variant
*shader
,
2515 const struct radv_tessellation_state
*tess
)
2517 uint64_t va
= radv_buffer_get_va(shader
->bo
) + shader
->bo_offset
;
2518 uint32_t rsrc2
= shader
->rsrc2
;
2520 radeon_set_sh_reg_seq(cs
, R_00B520_SPI_SHADER_PGM_LO_LS
, 2);
2521 radeon_emit(cs
, va
>> 8);
2522 radeon_emit(cs
, va
>> 40);
2524 rsrc2
|= S_00B52C_LDS_SIZE(tess
->lds_size
);
2525 if (pipeline
->device
->physical_device
->rad_info
.chip_class
== CIK
&&
2526 pipeline
->device
->physical_device
->rad_info
.family
!= CHIP_HAWAII
)
2527 radeon_set_sh_reg(cs
, R_00B52C_SPI_SHADER_PGM_RSRC2_LS
, rsrc2
);
2529 radeon_set_sh_reg_seq(cs
, R_00B528_SPI_SHADER_PGM_RSRC1_LS
, 2);
2530 radeon_emit(cs
, shader
->rsrc1
);
2531 radeon_emit(cs
, rsrc2
);
2535 radv_pipeline_generate_hw_hs(struct radeon_winsys_cs
*cs
,
2536 struct radv_pipeline
*pipeline
,
2537 struct radv_shader_variant
*shader
,
2538 const struct radv_tessellation_state
*tess
)
2540 uint64_t va
= radv_buffer_get_va(shader
->bo
) + shader
->bo_offset
;
2542 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2543 radeon_set_sh_reg_seq(cs
, R_00B410_SPI_SHADER_PGM_LO_LS
, 2);
2544 radeon_emit(cs
, va
>> 8);
2545 radeon_emit(cs
, va
>> 40);
2547 radeon_set_sh_reg_seq(cs
, R_00B428_SPI_SHADER_PGM_RSRC1_HS
, 2);
2548 radeon_emit(cs
, shader
->rsrc1
);
2549 radeon_emit(cs
, shader
->rsrc2
|
2550 S_00B42C_LDS_SIZE(tess
->lds_size
));
2552 radeon_set_sh_reg_seq(cs
, R_00B420_SPI_SHADER_PGM_LO_HS
, 4);
2553 radeon_emit(cs
, va
>> 8);
2554 radeon_emit(cs
, va
>> 40);
2555 radeon_emit(cs
, shader
->rsrc1
);
2556 radeon_emit(cs
, shader
->rsrc2
);
2561 radv_pipeline_generate_vertex_shader(struct radeon_winsys_cs
*cs
,
2562 struct radv_pipeline
*pipeline
,
2563 const struct radv_tessellation_state
*tess
)
2565 struct radv_shader_variant
*vs
;
2567 /* Skip shaders merged into HS/GS */
2568 vs
= pipeline
->shaders
[MESA_SHADER_VERTEX
];
2572 if (vs
->info
.vs
.as_ls
)
2573 radv_pipeline_generate_hw_ls(cs
, pipeline
, vs
, tess
);
2574 else if (vs
->info
.vs
.as_es
)
2575 radv_pipeline_generate_hw_es(cs
, pipeline
, vs
);
2577 radv_pipeline_generate_hw_vs(cs
, pipeline
, vs
);
2581 radv_pipeline_generate_tess_shaders(struct radeon_winsys_cs
*cs
,
2582 struct radv_pipeline
*pipeline
,
2583 const struct radv_tessellation_state
*tess
)
2585 if (!radv_pipeline_has_tess(pipeline
))
2588 struct radv_shader_variant
*tes
, *tcs
;
2590 tcs
= pipeline
->shaders
[MESA_SHADER_TESS_CTRL
];
2591 tes
= pipeline
->shaders
[MESA_SHADER_TESS_EVAL
];
2594 if (tes
->info
.tes
.as_es
)
2595 radv_pipeline_generate_hw_es(cs
, pipeline
, tes
);
2597 radv_pipeline_generate_hw_vs(cs
, pipeline
, tes
);
2600 radv_pipeline_generate_hw_hs(cs
, pipeline
, tcs
, tess
);
2602 radeon_set_context_reg(cs
, R_028B6C_VGT_TF_PARAM
,
2605 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= CIK
)
2606 radeon_set_context_reg_idx(cs
, R_028B58_VGT_LS_HS_CONFIG
, 2,
2607 tess
->ls_hs_config
);
2609 radeon_set_context_reg(cs
, R_028B58_VGT_LS_HS_CONFIG
,
2610 tess
->ls_hs_config
);
2612 struct radv_userdata_info
*loc
;
2614 loc
= radv_lookup_user_sgpr(pipeline
, MESA_SHADER_TESS_CTRL
, AC_UD_TCS_OFFCHIP_LAYOUT
);
2615 if (loc
->sgpr_idx
!= -1) {
2616 uint32_t base_reg
= pipeline
->user_data_0
[MESA_SHADER_TESS_CTRL
];
2617 assert(loc
->num_sgprs
== 4);
2618 assert(!loc
->indirect
);
2619 radeon_set_sh_reg_seq(cs
, base_reg
+ loc
->sgpr_idx
* 4, 4);
2620 radeon_emit(cs
, tess
->offchip_layout
);
2621 radeon_emit(cs
, tess
->tcs_out_offsets
);
2622 radeon_emit(cs
, tess
->tcs_out_layout
);
2623 radeon_emit(cs
, tess
->tcs_in_layout
);
2626 loc
= radv_lookup_user_sgpr(pipeline
, MESA_SHADER_TESS_EVAL
, AC_UD_TES_OFFCHIP_LAYOUT
);
2627 if (loc
->sgpr_idx
!= -1) {
2628 uint32_t base_reg
= pipeline
->user_data_0
[MESA_SHADER_TESS_EVAL
];
2629 assert(loc
->num_sgprs
== 1);
2630 assert(!loc
->indirect
);
2632 radeon_set_sh_reg(cs
, base_reg
+ loc
->sgpr_idx
* 4,
2633 tess
->offchip_layout
);
2638 radv_pipeline_generate_geometry_shader(struct radeon_winsys_cs
*cs
,
2639 struct radv_pipeline
*pipeline
,
2640 const struct radv_gs_state
*gs_state
)
2642 struct radv_shader_variant
*gs
;
2645 gs
= pipeline
->shaders
[MESA_SHADER_GEOMETRY
];
2649 uint32_t gsvs_itemsize
= gs
->info
.gs
.max_gsvs_emit_size
>> 2;
2651 radeon_set_context_reg_seq(cs
, R_028A60_VGT_GSVS_RING_OFFSET_1
, 3);
2652 radeon_emit(cs
, gsvs_itemsize
);
2653 radeon_emit(cs
, gsvs_itemsize
);
2654 radeon_emit(cs
, gsvs_itemsize
);
2656 radeon_set_context_reg(cs
, R_028AB0_VGT_GSVS_RING_ITEMSIZE
, gsvs_itemsize
);
2658 radeon_set_context_reg(cs
, R_028B38_VGT_GS_MAX_VERT_OUT
, gs
->info
.gs
.vertices_out
);
2660 uint32_t gs_vert_itemsize
= gs
->info
.gs
.gsvs_vertex_size
;
2661 radeon_set_context_reg_seq(cs
, R_028B5C_VGT_GS_VERT_ITEMSIZE
, 4);
2662 radeon_emit(cs
, gs_vert_itemsize
>> 2);
2667 uint32_t gs_num_invocations
= gs
->info
.gs
.invocations
;
2668 radeon_set_context_reg(cs
, R_028B90_VGT_GS_INSTANCE_CNT
,
2669 S_028B90_CNT(MIN2(gs_num_invocations
, 127)) |
2670 S_028B90_ENABLE(gs_num_invocations
> 0));
2672 radeon_set_context_reg(cs
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
2673 gs_state
->vgt_esgs_ring_itemsize
);
2675 va
= radv_buffer_get_va(gs
->bo
) + gs
->bo_offset
;
2677 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2678 radeon_set_sh_reg_seq(cs
, R_00B210_SPI_SHADER_PGM_LO_ES
, 2);
2679 radeon_emit(cs
, va
>> 8);
2680 radeon_emit(cs
, va
>> 40);
2682 radeon_set_sh_reg_seq(cs
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
, 2);
2683 radeon_emit(cs
, gs
->rsrc1
);
2684 radeon_emit(cs
, gs
->rsrc2
| S_00B22C_LDS_SIZE(gs_state
->lds_size
));
2686 radeon_set_context_reg(cs
, R_028A44_VGT_GS_ONCHIP_CNTL
, gs_state
->vgt_gs_onchip_cntl
);
2687 radeon_set_context_reg(cs
, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP
, gs_state
->vgt_gs_max_prims_per_subgroup
);
2689 radeon_set_sh_reg_seq(cs
, R_00B220_SPI_SHADER_PGM_LO_GS
, 4);
2690 radeon_emit(cs
, va
>> 8);
2691 radeon_emit(cs
, va
>> 40);
2692 radeon_emit(cs
, gs
->rsrc1
);
2693 radeon_emit(cs
, gs
->rsrc2
);
2696 radv_pipeline_generate_hw_vs(cs
, pipeline
, pipeline
->gs_copy_shader
);
2698 struct radv_userdata_info
*loc
= radv_lookup_user_sgpr(pipeline
, MESA_SHADER_GEOMETRY
,
2699 AC_UD_GS_VS_RING_STRIDE_ENTRIES
);
2700 if (loc
->sgpr_idx
!= -1) {
2701 uint32_t stride
= gs
->info
.gs
.max_gsvs_emit_size
;
2702 uint32_t num_entries
= 64;
2703 bool is_vi
= pipeline
->device
->physical_device
->rad_info
.chip_class
>= VI
;
2706 num_entries
*= stride
;
2708 stride
= S_008F04_STRIDE(stride
);
2709 radeon_set_sh_reg_seq(cs
, R_00B230_SPI_SHADER_USER_DATA_GS_0
+ loc
->sgpr_idx
* 4, 2);
2710 radeon_emit(cs
, stride
);
2711 radeon_emit(cs
, num_entries
);
2715 static uint32_t offset_to_ps_input(uint32_t offset
, bool flat_shade
)
2717 uint32_t ps_input_cntl
;
2718 if (offset
<= AC_EXP_PARAM_OFFSET_31
) {
2719 ps_input_cntl
= S_028644_OFFSET(offset
);
2721 ps_input_cntl
|= S_028644_FLAT_SHADE(1);
2723 /* The input is a DEFAULT_VAL constant. */
2724 assert(offset
>= AC_EXP_PARAM_DEFAULT_VAL_0000
&&
2725 offset
<= AC_EXP_PARAM_DEFAULT_VAL_1111
);
2726 offset
-= AC_EXP_PARAM_DEFAULT_VAL_0000
;
2727 ps_input_cntl
= S_028644_OFFSET(0x20) |
2728 S_028644_DEFAULT_VAL(offset
);
2730 return ps_input_cntl
;
2734 radv_pipeline_generate_ps_inputs(struct radeon_winsys_cs
*cs
,
2735 struct radv_pipeline
*pipeline
)
2737 struct radv_shader_variant
*ps
= pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
2738 const struct radv_vs_output_info
*outinfo
= get_vs_output_info(pipeline
);
2739 uint32_t ps_input_cntl
[32];
2741 unsigned ps_offset
= 0;
2743 if (ps
->info
.info
.ps
.prim_id_input
) {
2744 unsigned vs_offset
= outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
];
2745 if (vs_offset
!= AC_EXP_PARAM_UNDEFINED
) {
2746 ps_input_cntl
[ps_offset
] = offset_to_ps_input(vs_offset
, true);
2751 if (ps
->info
.info
.ps
.layer_input
||
2752 ps
->info
.info
.ps
.uses_input_attachments
||
2753 ps
->info
.info
.needs_multiview_view_index
) {
2754 unsigned vs_offset
= outinfo
->vs_output_param_offset
[VARYING_SLOT_LAYER
];
2755 if (vs_offset
!= AC_EXP_PARAM_UNDEFINED
)
2756 ps_input_cntl
[ps_offset
] = offset_to_ps_input(vs_offset
, true);
2758 ps_input_cntl
[ps_offset
] = offset_to_ps_input(AC_EXP_PARAM_DEFAULT_VAL_0000
, true);
2762 if (ps
->info
.info
.ps
.has_pcoord
) {
2764 val
= S_028644_PT_SPRITE_TEX(1) | S_028644_OFFSET(0x20);
2765 ps_input_cntl
[ps_offset
] = val
;
2769 for (unsigned i
= 0; i
< 32 && (1u << i
) <= ps
->info
.fs
.input_mask
; ++i
) {
2772 if (!(ps
->info
.fs
.input_mask
& (1u << i
)))
2775 vs_offset
= outinfo
->vs_output_param_offset
[VARYING_SLOT_VAR0
+ i
];
2776 if (vs_offset
== AC_EXP_PARAM_UNDEFINED
) {
2777 ps_input_cntl
[ps_offset
] = S_028644_OFFSET(0x20);
2782 flat_shade
= !!(ps
->info
.fs
.flat_shaded_mask
& (1u << ps_offset
));
2784 ps_input_cntl
[ps_offset
] = offset_to_ps_input(vs_offset
, flat_shade
);
2789 radeon_set_context_reg_seq(cs
, R_028644_SPI_PS_INPUT_CNTL_0
, ps_offset
);
2790 for (unsigned i
= 0; i
< ps_offset
; i
++) {
2791 radeon_emit(cs
, ps_input_cntl
[i
]);
2797 radv_compute_db_shader_control(const struct radv_device
*device
,
2798 const struct radv_shader_variant
*ps
)
2801 if (ps
->info
.fs
.early_fragment_test
|| !ps
->info
.info
.ps
.writes_memory
)
2802 z_order
= V_02880C_EARLY_Z_THEN_LATE_Z
;
2804 z_order
= V_02880C_LATE_Z
;
2806 return S_02880C_Z_EXPORT_ENABLE(ps
->info
.info
.ps
.writes_z
) |
2807 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(ps
->info
.info
.ps
.writes_stencil
) |
2808 S_02880C_KILL_ENABLE(!!ps
->info
.fs
.can_discard
) |
2809 S_02880C_MASK_EXPORT_ENABLE(ps
->info
.info
.ps
.writes_sample_mask
) |
2810 S_02880C_Z_ORDER(z_order
) |
2811 S_02880C_DEPTH_BEFORE_SHADER(ps
->info
.fs
.early_fragment_test
) |
2812 S_02880C_EXEC_ON_HIER_FAIL(ps
->info
.info
.ps
.writes_memory
) |
2813 S_02880C_EXEC_ON_NOOP(ps
->info
.info
.ps
.writes_memory
) |
2814 S_02880C_DUAL_QUAD_DISABLE(!!device
->physical_device
->has_rbplus
);
2818 radv_pipeline_generate_fragment_shader(struct radeon_winsys_cs
*cs
,
2819 struct radv_pipeline
*pipeline
)
2821 struct radv_shader_variant
*ps
;
2823 assert (pipeline
->shaders
[MESA_SHADER_FRAGMENT
]);
2825 ps
= pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
2826 va
= radv_buffer_get_va(ps
->bo
) + ps
->bo_offset
;
2828 radeon_set_sh_reg_seq(cs
, R_00B020_SPI_SHADER_PGM_LO_PS
, 4);
2829 radeon_emit(cs
, va
>> 8);
2830 radeon_emit(cs
, va
>> 40);
2831 radeon_emit(cs
, ps
->rsrc1
);
2832 radeon_emit(cs
, ps
->rsrc2
);
2834 radeon_set_context_reg(cs
, R_02880C_DB_SHADER_CONTROL
,
2835 radv_compute_db_shader_control(pipeline
->device
, ps
));
2837 radeon_set_context_reg(cs
, R_0286CC_SPI_PS_INPUT_ENA
,
2838 ps
->config
.spi_ps_input_ena
);
2840 radeon_set_context_reg(cs
, R_0286D0_SPI_PS_INPUT_ADDR
,
2841 ps
->config
.spi_ps_input_addr
);
2843 radeon_set_context_reg(cs
, R_0286D8_SPI_PS_IN_CONTROL
,
2844 S_0286D8_NUM_INTERP(ps
->info
.fs
.num_interp
));
2846 radeon_set_context_reg(cs
, R_0286E0_SPI_BARYC_CNTL
, pipeline
->graphics
.spi_baryc_cntl
);
2848 radeon_set_context_reg(cs
, R_028710_SPI_SHADER_Z_FORMAT
,
2849 ac_get_spi_shader_z_format(ps
->info
.info
.ps
.writes_z
,
2850 ps
->info
.info
.ps
.writes_stencil
,
2851 ps
->info
.info
.ps
.writes_sample_mask
));
2853 if (pipeline
->device
->dfsm_allowed
) {
2854 /* optimise this? */
2855 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2856 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_DFSM
) | EVENT_INDEX(0));
2861 radv_pipeline_generate_vgt_vertex_reuse(struct radeon_winsys_cs
*cs
,
2862 struct radv_pipeline
*pipeline
)
2864 if (pipeline
->device
->physical_device
->rad_info
.family
< CHIP_POLARIS10
)
2867 unsigned vtx_reuse_depth
= 30;
2868 if (radv_pipeline_has_tess(pipeline
) &&
2869 radv_get_tess_eval_shader(pipeline
)->info
.tes
.spacing
== TESS_SPACING_FRACTIONAL_ODD
) {
2870 vtx_reuse_depth
= 14;
2872 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
2873 S_028C58_VTX_REUSE_DEPTH(vtx_reuse_depth
));
2877 radv_compute_vgt_shader_stages_en(const struct radv_pipeline
*pipeline
)
2879 uint32_t stages
= 0;
2880 if (radv_pipeline_has_tess(pipeline
)) {
2881 stages
|= S_028B54_LS_EN(V_028B54_LS_STAGE_ON
) |
2882 S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
2884 if (radv_pipeline_has_gs(pipeline
))
2885 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_DS
) |
2887 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
2889 stages
|= S_028B54_VS_EN(V_028B54_VS_STAGE_DS
);
2891 } else if (radv_pipeline_has_gs(pipeline
))
2892 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL
) |
2894 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
2896 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
2897 stages
|= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
2903 radv_compute_cliprect_rule(const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
2905 const VkPipelineDiscardRectangleStateCreateInfoEXT
*discard_rectangle_info
=
2906 vk_find_struct_const(pCreateInfo
->pNext
, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT
);
2908 if (!discard_rectangle_info
)
2913 for (unsigned i
= 0; i
< (1u << MAX_DISCARD_RECTANGLES
); ++i
) {
2914 /* Interpret i as a bitmask, and then set the bit in the mask if
2915 * that combination of rectangles in which the pixel is contained
2916 * should pass the cliprect test. */
2917 unsigned relevant_subset
= i
& ((1u << discard_rectangle_info
->discardRectangleCount
) - 1);
2919 if (discard_rectangle_info
->discardRectangleMode
== VK_DISCARD_RECTANGLE_MODE_INCLUSIVE_EXT
&&
2923 if (discard_rectangle_info
->discardRectangleMode
== VK_DISCARD_RECTANGLE_MODE_EXCLUSIVE_EXT
&&
2934 radv_pipeline_generate_pm4(struct radv_pipeline
*pipeline
,
2935 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
2936 const struct radv_graphics_pipeline_create_info
*extra
,
2937 const struct radv_blend_state
*blend
,
2938 const struct radv_tessellation_state
*tess
,
2939 const struct radv_gs_state
*gs
,
2940 unsigned prim
, unsigned gs_out
)
2942 pipeline
->cs
.buf
= malloc(4 * 256);
2943 pipeline
->cs
.max_dw
= 256;
2945 radv_pipeline_generate_depth_stencil_state(&pipeline
->cs
, pipeline
, pCreateInfo
, extra
);
2946 radv_pipeline_generate_blend_state(&pipeline
->cs
, pipeline
, blend
);
2947 radv_pipeline_generate_raster_state(&pipeline
->cs
, pCreateInfo
);
2948 radv_pipeline_generate_multisample_state(&pipeline
->cs
, pipeline
);
2949 radv_pipeline_generate_vgt_gs_mode(&pipeline
->cs
, pipeline
);
2950 radv_pipeline_generate_vertex_shader(&pipeline
->cs
, pipeline
, tess
);
2951 radv_pipeline_generate_tess_shaders(&pipeline
->cs
, pipeline
, tess
);
2952 radv_pipeline_generate_geometry_shader(&pipeline
->cs
, pipeline
, gs
);
2953 radv_pipeline_generate_fragment_shader(&pipeline
->cs
, pipeline
);
2954 radv_pipeline_generate_ps_inputs(&pipeline
->cs
, pipeline
);
2955 radv_pipeline_generate_vgt_vertex_reuse(&pipeline
->cs
, pipeline
);
2956 radv_pipeline_generate_binning_state(&pipeline
->cs
, pipeline
, pCreateInfo
);
2958 radeon_set_context_reg(&pipeline
->cs
, R_0286E8_SPI_TMPRING_SIZE
,
2959 S_0286E8_WAVES(pipeline
->max_waves
) |
2960 S_0286E8_WAVESIZE(pipeline
->scratch_bytes_per_wave
>> 10));
2962 radeon_set_context_reg(&pipeline
->cs
, R_028B54_VGT_SHADER_STAGES_EN
, radv_compute_vgt_shader_stages_en(pipeline
));
2964 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2965 radeon_set_uconfig_reg_idx(&pipeline
->cs
, R_030908_VGT_PRIMITIVE_TYPE
, 1, prim
);
2967 radeon_set_config_reg(&pipeline
->cs
, R_008958_VGT_PRIMITIVE_TYPE
, prim
);
2969 radeon_set_context_reg(&pipeline
->cs
, R_028A6C_VGT_GS_OUT_PRIM_TYPE
, gs_out
);
2971 radeon_set_context_reg(&pipeline
->cs
, R_02820C_PA_SC_CLIPRECT_RULE
, radv_compute_cliprect_rule(pCreateInfo
));
2973 assert(pipeline
->cs
.cdw
<= pipeline
->cs
.max_dw
);
2976 static struct radv_ia_multi_vgt_param_helpers
2977 radv_compute_ia_multi_vgt_param_helpers(struct radv_pipeline
*pipeline
,
2978 const struct radv_tessellation_state
*tess
,
2981 struct radv_ia_multi_vgt_param_helpers ia_multi_vgt_param
= {0};
2982 const struct radv_device
*device
= pipeline
->device
;
2984 if (radv_pipeline_has_tess(pipeline
))
2985 ia_multi_vgt_param
.primgroup_size
= tess
->num_patches
;
2986 else if (radv_pipeline_has_gs(pipeline
))
2987 ia_multi_vgt_param
.primgroup_size
= 64;
2989 ia_multi_vgt_param
.primgroup_size
= 128; /* recommended without a GS */
2991 ia_multi_vgt_param
.partial_es_wave
= false;
2992 if (pipeline
->device
->has_distributed_tess
) {
2993 if (radv_pipeline_has_gs(pipeline
)) {
2994 if (device
->physical_device
->rad_info
.chip_class
<= VI
)
2995 ia_multi_vgt_param
.partial_es_wave
= true;
2998 /* GS requirement. */
2999 if (SI_GS_PER_ES
/ ia_multi_vgt_param
.primgroup_size
>= pipeline
->device
->gs_table_depth
- 3)
3000 ia_multi_vgt_param
.partial_es_wave
= true;
3002 ia_multi_vgt_param
.wd_switch_on_eop
= false;
3003 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
3004 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
3005 * 4 shader engines. Set 1 to pass the assertion below.
3006 * The other cases are hardware requirements. */
3007 if (device
->physical_device
->rad_info
.max_se
< 4 ||
3008 prim
== V_008958_DI_PT_POLYGON
||
3009 prim
== V_008958_DI_PT_LINELOOP
||
3010 prim
== V_008958_DI_PT_TRIFAN
||
3011 prim
== V_008958_DI_PT_TRISTRIP_ADJ
||
3012 (pipeline
->graphics
.prim_restart_enable
&&
3013 (device
->physical_device
->rad_info
.family
< CHIP_POLARIS10
||
3014 (prim
!= V_008958_DI_PT_POINTLIST
&&
3015 prim
!= V_008958_DI_PT_LINESTRIP
&&
3016 prim
!= V_008958_DI_PT_TRISTRIP
))))
3017 ia_multi_vgt_param
.wd_switch_on_eop
= true;
3020 ia_multi_vgt_param
.ia_switch_on_eoi
= false;
3021 if (pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.prim_id_input
)
3022 ia_multi_vgt_param
.ia_switch_on_eoi
= true;
3023 if (radv_pipeline_has_gs(pipeline
) &&
3024 pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
.info
.uses_prim_id
)
3025 ia_multi_vgt_param
.ia_switch_on_eoi
= true;
3026 if (radv_pipeline_has_tess(pipeline
)) {
3027 /* SWITCH_ON_EOI must be set if PrimID is used. */
3028 if (pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.info
.uses_prim_id
||
3029 radv_get_tess_eval_shader(pipeline
)->info
.info
.uses_prim_id
)
3030 ia_multi_vgt_param
.ia_switch_on_eoi
= true;
3033 ia_multi_vgt_param
.partial_vs_wave
= false;
3034 if (radv_pipeline_has_tess(pipeline
)) {
3035 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
3036 if ((device
->physical_device
->rad_info
.family
== CHIP_TAHITI
||
3037 device
->physical_device
->rad_info
.family
== CHIP_PITCAIRN
||
3038 device
->physical_device
->rad_info
.family
== CHIP_BONAIRE
) &&
3039 radv_pipeline_has_gs(pipeline
))
3040 ia_multi_vgt_param
.partial_vs_wave
= true;
3041 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
3042 if (device
->has_distributed_tess
) {
3043 if (radv_pipeline_has_gs(pipeline
)) {
3044 if (device
->physical_device
->rad_info
.family
== CHIP_TONGA
||
3045 device
->physical_device
->rad_info
.family
== CHIP_FIJI
||
3046 device
->physical_device
->rad_info
.family
== CHIP_POLARIS10
||
3047 device
->physical_device
->rad_info
.family
== CHIP_POLARIS11
||
3048 device
->physical_device
->rad_info
.family
== CHIP_POLARIS12
)
3049 ia_multi_vgt_param
.partial_vs_wave
= true;
3051 ia_multi_vgt_param
.partial_vs_wave
= true;
3056 ia_multi_vgt_param
.base
=
3057 S_028AA8_PRIMGROUP_SIZE(ia_multi_vgt_param
.primgroup_size
- 1) |
3058 /* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
3059 S_028AA8_MAX_PRIMGRP_IN_WAVE(device
->physical_device
->rad_info
.chip_class
== VI
? 2 : 0) |
3060 S_030960_EN_INST_OPT_BASIC(device
->physical_device
->rad_info
.chip_class
>= GFX9
) |
3061 S_030960_EN_INST_OPT_ADV(device
->physical_device
->rad_info
.chip_class
>= GFX9
);
3063 return ia_multi_vgt_param
;
3068 radv_compute_vertex_input_state(struct radv_pipeline
*pipeline
,
3069 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
3071 const VkPipelineVertexInputStateCreateInfo
*vi_info
=
3072 pCreateInfo
->pVertexInputState
;
3073 struct radv_vertex_elements_info
*velems
= &pipeline
->vertex_elements
;
3075 for (uint32_t i
= 0; i
< vi_info
->vertexAttributeDescriptionCount
; i
++) {
3076 const VkVertexInputAttributeDescription
*desc
=
3077 &vi_info
->pVertexAttributeDescriptions
[i
];
3078 unsigned loc
= desc
->location
;
3079 const struct vk_format_description
*format_desc
;
3081 uint32_t num_format
, data_format
;
3082 format_desc
= vk_format_description(desc
->format
);
3083 first_non_void
= vk_format_get_first_non_void_channel(desc
->format
);
3085 num_format
= radv_translate_buffer_numformat(format_desc
, first_non_void
);
3086 data_format
= radv_translate_buffer_dataformat(format_desc
, first_non_void
);
3088 velems
->rsrc_word3
[loc
] = S_008F0C_DST_SEL_X(si_map_swizzle(format_desc
->swizzle
[0])) |
3089 S_008F0C_DST_SEL_Y(si_map_swizzle(format_desc
->swizzle
[1])) |
3090 S_008F0C_DST_SEL_Z(si_map_swizzle(format_desc
->swizzle
[2])) |
3091 S_008F0C_DST_SEL_W(si_map_swizzle(format_desc
->swizzle
[3])) |
3092 S_008F0C_NUM_FORMAT(num_format
) |
3093 S_008F0C_DATA_FORMAT(data_format
);
3094 velems
->format_size
[loc
] = format_desc
->block
.bits
/ 8;
3095 velems
->offset
[loc
] = desc
->offset
;
3096 velems
->binding
[loc
] = desc
->binding
;
3097 velems
->count
= MAX2(velems
->count
, loc
+ 1);
3100 for (uint32_t i
= 0; i
< vi_info
->vertexBindingDescriptionCount
; i
++) {
3101 const VkVertexInputBindingDescription
*desc
=
3102 &vi_info
->pVertexBindingDescriptions
[i
];
3104 pipeline
->binding_stride
[desc
->binding
] = desc
->stride
;
3109 radv_pipeline_init(struct radv_pipeline
*pipeline
,
3110 struct radv_device
*device
,
3111 struct radv_pipeline_cache
*cache
,
3112 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
3113 const struct radv_graphics_pipeline_create_info
*extra
,
3114 const VkAllocationCallbacks
*alloc
)
3117 bool has_view_index
= false;
3119 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
3120 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
3121 if (subpass
->view_mask
)
3122 has_view_index
= true;
3124 alloc
= &device
->alloc
;
3126 pipeline
->device
= device
;
3127 pipeline
->layout
= radv_pipeline_layout_from_handle(pCreateInfo
->layout
);
3128 assert(pipeline
->layout
);
3130 struct radv_blend_state blend
= radv_pipeline_init_blend_state(pipeline
, pCreateInfo
, extra
);
3132 const VkPipelineShaderStageCreateInfo
*pStages
[MESA_SHADER_STAGES
] = { 0, };
3133 for (uint32_t i
= 0; i
< pCreateInfo
->stageCount
; i
++) {
3134 gl_shader_stage stage
= ffs(pCreateInfo
->pStages
[i
].stage
) - 1;
3135 pStages
[stage
] = &pCreateInfo
->pStages
[i
];
3138 radv_create_shaders(pipeline
, device
, cache
,
3139 radv_generate_graphics_pipeline_key(pipeline
, pCreateInfo
, &blend
, has_view_index
),
3142 pipeline
->graphics
.spi_baryc_cntl
= S_0286E0_FRONT_FACE_ALL_BITS(1);
3143 radv_pipeline_init_multisample_state(pipeline
, pCreateInfo
);
3145 uint32_t prim
= si_translate_prim(pCreateInfo
->pInputAssemblyState
->topology
);
3147 pipeline
->graphics
.can_use_guardband
= radv_prim_can_use_guardband(pCreateInfo
->pInputAssemblyState
->topology
);
3149 if (radv_pipeline_has_gs(pipeline
)) {
3150 gs_out
= si_conv_gl_prim_to_gs_out(pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
.gs
.output_prim
);
3151 pipeline
->graphics
.can_use_guardband
= gs_out
== V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
3153 gs_out
= si_conv_prim_to_gs_out(pCreateInfo
->pInputAssemblyState
->topology
);
3155 if (extra
&& extra
->use_rectlist
) {
3156 prim
= V_008958_DI_PT_RECTLIST
;
3157 gs_out
= V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
3158 pipeline
->graphics
.can_use_guardband
= true;
3160 pipeline
->graphics
.prim_restart_enable
= !!pCreateInfo
->pInputAssemblyState
->primitiveRestartEnable
;
3161 /* prim vertex count will need TESS changes */
3162 pipeline
->graphics
.prim_vertex_count
= prim_size_table
[prim
];
3164 radv_pipeline_init_dynamic_state(pipeline
, pCreateInfo
);
3166 /* Ensure that some export memory is always allocated, for two reasons:
3168 * 1) Correctness: The hardware ignores the EXEC mask if no export
3169 * memory is allocated, so KILL and alpha test do not work correctly
3171 * 2) Performance: Every shader needs at least a NULL export, even when
3172 * it writes no color/depth output. The NULL export instruction
3173 * stalls without this setting.
3175 * Don't add this to CB_SHADER_MASK.
3177 struct radv_shader_variant
*ps
= pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
3178 if (!blend
.spi_shader_col_format
) {
3179 if (!ps
->info
.info
.ps
.writes_z
&&
3180 !ps
->info
.info
.ps
.writes_stencil
&&
3181 !ps
->info
.info
.ps
.writes_sample_mask
)
3182 blend
.spi_shader_col_format
= V_028714_SPI_SHADER_32_R
;
3185 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
3186 if (pipeline
->shaders
[i
]) {
3187 pipeline
->need_indirect_descriptor_sets
|= pipeline
->shaders
[i
]->info
.need_indirect_descriptor_sets
;
3191 struct radv_gs_state gs
= {0};
3192 if (radv_pipeline_has_gs(pipeline
)) {
3193 gs
= calculate_gs_info(pCreateInfo
, pipeline
);
3194 calculate_gs_ring_sizes(pipeline
, &gs
);
3197 struct radv_tessellation_state tess
= {0};
3198 if (radv_pipeline_has_tess(pipeline
)) {
3199 if (prim
== V_008958_DI_PT_PATCH
) {
3200 pipeline
->graphics
.prim_vertex_count
.min
= pCreateInfo
->pTessellationState
->patchControlPoints
;
3201 pipeline
->graphics
.prim_vertex_count
.incr
= 1;
3203 tess
= calculate_tess_state(pipeline
, pCreateInfo
);
3206 pipeline
->graphics
.ia_multi_vgt_param
= radv_compute_ia_multi_vgt_param_helpers(pipeline
, &tess
, prim
);
3208 radv_compute_vertex_input_state(pipeline
, pCreateInfo
);
3210 for (uint32_t i
= 0; i
< MESA_SHADER_STAGES
; i
++)
3211 pipeline
->user_data_0
[i
] = radv_pipeline_stage_to_user_data_0(pipeline
, i
, device
->physical_device
->rad_info
.chip_class
);
3213 struct radv_userdata_info
*loc
= radv_lookup_user_sgpr(pipeline
, MESA_SHADER_VERTEX
,
3214 AC_UD_VS_BASE_VERTEX_START_INSTANCE
);
3215 if (loc
->sgpr_idx
!= -1) {
3216 pipeline
->graphics
.vtx_base_sgpr
= pipeline
->user_data_0
[MESA_SHADER_VERTEX
];
3217 pipeline
->graphics
.vtx_base_sgpr
+= loc
->sgpr_idx
* 4;
3218 if (radv_get_vertex_shader(pipeline
)->info
.info
.vs
.needs_draw_id
)
3219 pipeline
->graphics
.vtx_emit_num
= 3;
3221 pipeline
->graphics
.vtx_emit_num
= 2;
3224 result
= radv_pipeline_scratch_init(device
, pipeline
);
3225 radv_pipeline_generate_pm4(pipeline
, pCreateInfo
, extra
, &blend
, &tess
, &gs
, prim
, gs_out
);
3231 radv_graphics_pipeline_create(
3233 VkPipelineCache _cache
,
3234 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
3235 const struct radv_graphics_pipeline_create_info
*extra
,
3236 const VkAllocationCallbacks
*pAllocator
,
3237 VkPipeline
*pPipeline
)
3239 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3240 RADV_FROM_HANDLE(radv_pipeline_cache
, cache
, _cache
);
3241 struct radv_pipeline
*pipeline
;
3244 pipeline
= vk_zalloc2(&device
->alloc
, pAllocator
, sizeof(*pipeline
), 8,
3245 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3246 if (pipeline
== NULL
)
3247 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3249 result
= radv_pipeline_init(pipeline
, device
, cache
,
3250 pCreateInfo
, extra
, pAllocator
);
3251 if (result
!= VK_SUCCESS
) {
3252 radv_pipeline_destroy(device
, pipeline
, pAllocator
);
3256 *pPipeline
= radv_pipeline_to_handle(pipeline
);
3261 VkResult
radv_CreateGraphicsPipelines(
3263 VkPipelineCache pipelineCache
,
3265 const VkGraphicsPipelineCreateInfo
* pCreateInfos
,
3266 const VkAllocationCallbacks
* pAllocator
,
3267 VkPipeline
* pPipelines
)
3269 VkResult result
= VK_SUCCESS
;
3272 for (; i
< count
; i
++) {
3274 r
= radv_graphics_pipeline_create(_device
,
3277 NULL
, pAllocator
, &pPipelines
[i
]);
3278 if (r
!= VK_SUCCESS
) {
3280 pPipelines
[i
] = VK_NULL_HANDLE
;
3289 radv_compute_generate_pm4(struct radv_pipeline
*pipeline
)
3291 struct radv_shader_variant
*compute_shader
;
3292 struct radv_device
*device
= pipeline
->device
;
3293 unsigned compute_resource_limits
;
3294 unsigned waves_per_threadgroup
;
3297 pipeline
->cs
.buf
= malloc(20 * 4);
3298 pipeline
->cs
.max_dw
= 20;
3300 compute_shader
= pipeline
->shaders
[MESA_SHADER_COMPUTE
];
3301 va
= radv_buffer_get_va(compute_shader
->bo
) + compute_shader
->bo_offset
;
3303 radeon_set_sh_reg_seq(&pipeline
->cs
, R_00B830_COMPUTE_PGM_LO
, 2);
3304 radeon_emit(&pipeline
->cs
, va
>> 8);
3305 radeon_emit(&pipeline
->cs
, va
>> 40);
3307 radeon_set_sh_reg_seq(&pipeline
->cs
, R_00B848_COMPUTE_PGM_RSRC1
, 2);
3308 radeon_emit(&pipeline
->cs
, compute_shader
->rsrc1
);
3309 radeon_emit(&pipeline
->cs
, compute_shader
->rsrc2
);
3311 radeon_set_sh_reg(&pipeline
->cs
, R_00B860_COMPUTE_TMPRING_SIZE
,
3312 S_00B860_WAVES(pipeline
->max_waves
) |
3313 S_00B860_WAVESIZE(pipeline
->scratch_bytes_per_wave
>> 10));
3315 /* Calculate best compute resource limits. */
3316 waves_per_threadgroup
=
3317 DIV_ROUND_UP(compute_shader
->info
.cs
.block_size
[0] *
3318 compute_shader
->info
.cs
.block_size
[1] *
3319 compute_shader
->info
.cs
.block_size
[2], 64);
3320 compute_resource_limits
=
3321 S_00B854_SIMD_DEST_CNTL(waves_per_threadgroup
% 4 == 0);
3323 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
3324 unsigned num_cu_per_se
=
3325 device
->physical_device
->rad_info
.num_good_compute_units
/
3326 device
->physical_device
->rad_info
.max_se
;
3328 /* Force even distribution on all SIMDs in CU if the workgroup
3329 * size is 64. This has shown some good improvements if # of
3330 * CUs per SE is not a multiple of 4.
3332 if (num_cu_per_se
% 4 && waves_per_threadgroup
== 1)
3333 compute_resource_limits
|= S_00B854_FORCE_SIMD_DIST(1);
3336 radeon_set_sh_reg(&pipeline
->cs
, R_00B854_COMPUTE_RESOURCE_LIMITS
,
3337 compute_resource_limits
);
3339 radeon_set_sh_reg_seq(&pipeline
->cs
, R_00B81C_COMPUTE_NUM_THREAD_X
, 3);
3340 radeon_emit(&pipeline
->cs
,
3341 S_00B81C_NUM_THREAD_FULL(compute_shader
->info
.cs
.block_size
[0]));
3342 radeon_emit(&pipeline
->cs
,
3343 S_00B81C_NUM_THREAD_FULL(compute_shader
->info
.cs
.block_size
[1]));
3344 radeon_emit(&pipeline
->cs
,
3345 S_00B81C_NUM_THREAD_FULL(compute_shader
->info
.cs
.block_size
[2]));
3347 assert(pipeline
->cs
.cdw
<= pipeline
->cs
.max_dw
);
3350 static VkResult
radv_compute_pipeline_create(
3352 VkPipelineCache _cache
,
3353 const VkComputePipelineCreateInfo
* pCreateInfo
,
3354 const VkAllocationCallbacks
* pAllocator
,
3355 VkPipeline
* pPipeline
)
3357 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3358 RADV_FROM_HANDLE(radv_pipeline_cache
, cache
, _cache
);
3359 const VkPipelineShaderStageCreateInfo
*pStages
[MESA_SHADER_STAGES
] = { 0, };
3360 struct radv_pipeline
*pipeline
;
3363 pipeline
= vk_zalloc2(&device
->alloc
, pAllocator
, sizeof(*pipeline
), 8,
3364 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3365 if (pipeline
== NULL
)
3366 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3368 pipeline
->device
= device
;
3369 pipeline
->layout
= radv_pipeline_layout_from_handle(pCreateInfo
->layout
);
3370 assert(pipeline
->layout
);
3372 pStages
[MESA_SHADER_COMPUTE
] = &pCreateInfo
->stage
;
3373 radv_create_shaders(pipeline
, device
, cache
, (struct radv_pipeline_key
) {0}, pStages
);
3375 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
);
3376 pipeline
->need_indirect_descriptor_sets
|= pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.need_indirect_descriptor_sets
;
3377 result
= radv_pipeline_scratch_init(device
, pipeline
);
3378 if (result
!= VK_SUCCESS
) {
3379 radv_pipeline_destroy(device
, pipeline
, pAllocator
);
3383 radv_compute_generate_pm4(pipeline
);
3385 *pPipeline
= radv_pipeline_to_handle(pipeline
);
3390 VkResult
radv_CreateComputePipelines(
3392 VkPipelineCache pipelineCache
,
3394 const VkComputePipelineCreateInfo
* pCreateInfos
,
3395 const VkAllocationCallbacks
* pAllocator
,
3396 VkPipeline
* pPipelines
)
3398 VkResult result
= VK_SUCCESS
;
3401 for (; i
< count
; i
++) {
3403 r
= radv_compute_pipeline_create(_device
, pipelineCache
,
3405 pAllocator
, &pPipelines
[i
]);
3406 if (r
!= VK_SUCCESS
) {
3408 pPipelines
[i
] = VK_NULL_HANDLE
;