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
;
70 struct radv_gs_state
{
71 uint32_t vgt_gs_onchip_cntl
;
72 uint32_t vgt_gs_max_prims_per_subgroup
;
73 uint32_t vgt_esgs_ring_itemsize
;
78 radv_pipeline_destroy(struct radv_device
*device
,
79 struct radv_pipeline
*pipeline
,
80 const VkAllocationCallbacks
* allocator
)
82 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; ++i
)
83 if (pipeline
->shaders
[i
])
84 radv_shader_variant_destroy(device
, pipeline
->shaders
[i
]);
86 if (pipeline
->gs_copy_shader
)
87 radv_shader_variant_destroy(device
, pipeline
->gs_copy_shader
);
90 free(pipeline
->cs
.buf
);
91 vk_free2(&device
->alloc
, allocator
, pipeline
);
94 void radv_DestroyPipeline(
97 const VkAllocationCallbacks
* pAllocator
)
99 RADV_FROM_HANDLE(radv_device
, device
, _device
);
100 RADV_FROM_HANDLE(radv_pipeline
, pipeline
, _pipeline
);
105 radv_pipeline_destroy(device
, pipeline
, pAllocator
);
108 static uint32_t get_hash_flags(struct radv_device
*device
)
110 uint32_t hash_flags
= 0;
112 if (device
->instance
->debug_flags
& RADV_DEBUG_UNSAFE_MATH
)
113 hash_flags
|= RADV_HASH_SHADER_UNSAFE_MATH
;
114 if (device
->instance
->perftest_flags
& RADV_PERFTEST_SISCHED
)
115 hash_flags
|= RADV_HASH_SHADER_SISCHED
;
120 radv_pipeline_scratch_init(struct radv_device
*device
,
121 struct radv_pipeline
*pipeline
)
123 unsigned scratch_bytes_per_wave
= 0;
124 unsigned max_waves
= 0;
125 unsigned min_waves
= 1;
127 for (int i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
128 if (pipeline
->shaders
[i
]) {
129 unsigned max_stage_waves
= device
->scratch_waves
;
131 scratch_bytes_per_wave
= MAX2(scratch_bytes_per_wave
,
132 pipeline
->shaders
[i
]->config
.scratch_bytes_per_wave
);
134 max_stage_waves
= MIN2(max_stage_waves
,
135 4 * device
->physical_device
->rad_info
.num_good_compute_units
*
136 (256 / pipeline
->shaders
[i
]->config
.num_vgprs
));
137 max_waves
= MAX2(max_waves
, max_stage_waves
);
141 if (pipeline
->shaders
[MESA_SHADER_COMPUTE
]) {
142 unsigned group_size
= pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.cs
.block_size
[0] *
143 pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.cs
.block_size
[1] *
144 pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.cs
.block_size
[2];
145 min_waves
= MAX2(min_waves
, round_up_u32(group_size
, 64));
148 if (scratch_bytes_per_wave
)
149 max_waves
= MIN2(max_waves
, 0xffffffffu
/ scratch_bytes_per_wave
);
151 if (scratch_bytes_per_wave
&& max_waves
< min_waves
) {
152 /* Not really true at this moment, but will be true on first
153 * execution. Avoid having hanging shaders. */
154 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
156 pipeline
->scratch_bytes_per_wave
= scratch_bytes_per_wave
;
157 pipeline
->max_waves
= max_waves
;
161 static uint32_t si_translate_blend_function(VkBlendOp op
)
164 case VK_BLEND_OP_ADD
:
165 return V_028780_COMB_DST_PLUS_SRC
;
166 case VK_BLEND_OP_SUBTRACT
:
167 return V_028780_COMB_SRC_MINUS_DST
;
168 case VK_BLEND_OP_REVERSE_SUBTRACT
:
169 return V_028780_COMB_DST_MINUS_SRC
;
170 case VK_BLEND_OP_MIN
:
171 return V_028780_COMB_MIN_DST_SRC
;
172 case VK_BLEND_OP_MAX
:
173 return V_028780_COMB_MAX_DST_SRC
;
179 static uint32_t si_translate_blend_factor(VkBlendFactor factor
)
182 case VK_BLEND_FACTOR_ZERO
:
183 return V_028780_BLEND_ZERO
;
184 case VK_BLEND_FACTOR_ONE
:
185 return V_028780_BLEND_ONE
;
186 case VK_BLEND_FACTOR_SRC_COLOR
:
187 return V_028780_BLEND_SRC_COLOR
;
188 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR
:
189 return V_028780_BLEND_ONE_MINUS_SRC_COLOR
;
190 case VK_BLEND_FACTOR_DST_COLOR
:
191 return V_028780_BLEND_DST_COLOR
;
192 case VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR
:
193 return V_028780_BLEND_ONE_MINUS_DST_COLOR
;
194 case VK_BLEND_FACTOR_SRC_ALPHA
:
195 return V_028780_BLEND_SRC_ALPHA
;
196 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
:
197 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA
;
198 case VK_BLEND_FACTOR_DST_ALPHA
:
199 return V_028780_BLEND_DST_ALPHA
;
200 case VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA
:
201 return V_028780_BLEND_ONE_MINUS_DST_ALPHA
;
202 case VK_BLEND_FACTOR_CONSTANT_COLOR
:
203 return V_028780_BLEND_CONSTANT_COLOR
;
204 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR
:
205 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR
;
206 case VK_BLEND_FACTOR_CONSTANT_ALPHA
:
207 return V_028780_BLEND_CONSTANT_ALPHA
;
208 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA
:
209 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA
;
210 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
:
211 return V_028780_BLEND_SRC_ALPHA_SATURATE
;
212 case VK_BLEND_FACTOR_SRC1_COLOR
:
213 return V_028780_BLEND_SRC1_COLOR
;
214 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
:
215 return V_028780_BLEND_INV_SRC1_COLOR
;
216 case VK_BLEND_FACTOR_SRC1_ALPHA
:
217 return V_028780_BLEND_SRC1_ALPHA
;
218 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
:
219 return V_028780_BLEND_INV_SRC1_ALPHA
;
225 static uint32_t si_translate_blend_opt_function(VkBlendOp op
)
228 case VK_BLEND_OP_ADD
:
229 return V_028760_OPT_COMB_ADD
;
230 case VK_BLEND_OP_SUBTRACT
:
231 return V_028760_OPT_COMB_SUBTRACT
;
232 case VK_BLEND_OP_REVERSE_SUBTRACT
:
233 return V_028760_OPT_COMB_REVSUBTRACT
;
234 case VK_BLEND_OP_MIN
:
235 return V_028760_OPT_COMB_MIN
;
236 case VK_BLEND_OP_MAX
:
237 return V_028760_OPT_COMB_MAX
;
239 return V_028760_OPT_COMB_BLEND_DISABLED
;
243 static uint32_t si_translate_blend_opt_factor(VkBlendFactor factor
, bool is_alpha
)
246 case VK_BLEND_FACTOR_ZERO
:
247 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL
;
248 case VK_BLEND_FACTOR_ONE
:
249 return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
;
250 case VK_BLEND_FACTOR_SRC_COLOR
:
251 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
252 : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0
;
253 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR
:
254 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
255 : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1
;
256 case VK_BLEND_FACTOR_SRC_ALPHA
:
257 return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
;
258 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
:
259 return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
;
260 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
:
261 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
262 : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0
;
264 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
269 * Get rid of DST in the blend factors by commuting the operands:
270 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
272 static void si_blend_remove_dst(unsigned *func
, unsigned *src_factor
,
273 unsigned *dst_factor
, unsigned expected_dst
,
274 unsigned replacement_src
)
276 if (*src_factor
== expected_dst
&&
277 *dst_factor
== VK_BLEND_FACTOR_ZERO
) {
278 *src_factor
= VK_BLEND_FACTOR_ZERO
;
279 *dst_factor
= replacement_src
;
281 /* Commuting the operands requires reversing subtractions. */
282 if (*func
== VK_BLEND_OP_SUBTRACT
)
283 *func
= VK_BLEND_OP_REVERSE_SUBTRACT
;
284 else if (*func
== VK_BLEND_OP_REVERSE_SUBTRACT
)
285 *func
= VK_BLEND_OP_SUBTRACT
;
289 static bool si_blend_factor_uses_dst(unsigned factor
)
291 return factor
== VK_BLEND_FACTOR_DST_COLOR
||
292 factor
== VK_BLEND_FACTOR_DST_ALPHA
||
293 factor
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
294 factor
== VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA
||
295 factor
== VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR
;
298 static bool is_dual_src(VkBlendFactor factor
)
301 case VK_BLEND_FACTOR_SRC1_COLOR
:
302 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
:
303 case VK_BLEND_FACTOR_SRC1_ALPHA
:
304 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
:
311 static unsigned si_choose_spi_color_format(VkFormat vk_format
,
313 bool blend_need_alpha
)
315 const struct vk_format_description
*desc
= vk_format_description(vk_format
);
316 unsigned format
, ntype
, swap
;
318 /* Alpha is needed for alpha-to-coverage.
319 * Blending may be with or without alpha.
321 unsigned normal
= 0; /* most optimal, may not support blending or export alpha */
322 unsigned alpha
= 0; /* exports alpha, but may not support blending */
323 unsigned blend
= 0; /* supports blending, but may not export alpha */
324 unsigned blend_alpha
= 0; /* least optimal, supports blending and exports alpha */
326 format
= radv_translate_colorformat(vk_format
);
327 ntype
= radv_translate_color_numformat(vk_format
, desc
,
328 vk_format_get_first_non_void_channel(vk_format
));
329 swap
= radv_translate_colorswap(vk_format
, false);
331 /* Choose the SPI color formats. These are required values for Stoney/RB+.
332 * Other chips have multiple choices, though they are not necessarily better.
335 case V_028C70_COLOR_5_6_5
:
336 case V_028C70_COLOR_1_5_5_5
:
337 case V_028C70_COLOR_5_5_5_1
:
338 case V_028C70_COLOR_4_4_4_4
:
339 case V_028C70_COLOR_10_11_11
:
340 case V_028C70_COLOR_11_11_10
:
341 case V_028C70_COLOR_8
:
342 case V_028C70_COLOR_8_8
:
343 case V_028C70_COLOR_8_8_8_8
:
344 case V_028C70_COLOR_10_10_10_2
:
345 case V_028C70_COLOR_2_10_10_10
:
346 if (ntype
== V_028C70_NUMBER_UINT
)
347 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
348 else if (ntype
== V_028C70_NUMBER_SINT
)
349 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
351 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
354 case V_028C70_COLOR_16
:
355 case V_028C70_COLOR_16_16
:
356 case V_028C70_COLOR_16_16_16_16
:
357 if (ntype
== V_028C70_NUMBER_UNORM
||
358 ntype
== V_028C70_NUMBER_SNORM
) {
359 /* UNORM16 and SNORM16 don't support blending */
360 if (ntype
== V_028C70_NUMBER_UNORM
)
361 normal
= alpha
= V_028714_SPI_SHADER_UNORM16_ABGR
;
363 normal
= alpha
= V_028714_SPI_SHADER_SNORM16_ABGR
;
365 /* Use 32 bits per channel for blending. */
366 if (format
== V_028C70_COLOR_16
) {
367 if (swap
== V_028C70_SWAP_STD
) { /* R */
368 blend
= V_028714_SPI_SHADER_32_R
;
369 blend_alpha
= V_028714_SPI_SHADER_32_AR
;
370 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
371 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
374 } else if (format
== V_028C70_COLOR_16_16
) {
375 if (swap
== V_028C70_SWAP_STD
) { /* RG */
376 blend
= V_028714_SPI_SHADER_32_GR
;
377 blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
378 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
379 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
382 } else /* 16_16_16_16 */
383 blend
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
384 } else if (ntype
== V_028C70_NUMBER_UINT
)
385 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
386 else if (ntype
== V_028C70_NUMBER_SINT
)
387 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
388 else if (ntype
== V_028C70_NUMBER_FLOAT
)
389 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
394 case V_028C70_COLOR_32
:
395 if (swap
== V_028C70_SWAP_STD
) { /* R */
396 blend
= normal
= V_028714_SPI_SHADER_32_R
;
397 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
398 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
399 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
404 case V_028C70_COLOR_32_32
:
405 if (swap
== V_028C70_SWAP_STD
) { /* RG */
406 blend
= normal
= V_028714_SPI_SHADER_32_GR
;
407 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
408 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
409 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
414 case V_028C70_COLOR_32_32_32_32
:
415 case V_028C70_COLOR_8_24
:
416 case V_028C70_COLOR_24_8
:
417 case V_028C70_COLOR_X24_8_32_FLOAT
:
418 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_ABGR
;
422 unreachable("unhandled blend format");
425 if (blend_enable
&& blend_need_alpha
)
427 else if(blend_need_alpha
)
429 else if(blend_enable
)
436 radv_pipeline_compute_spi_color_formats(struct radv_pipeline
*pipeline
,
437 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
438 uint32_t blend_enable
,
439 uint32_t blend_need_alpha
,
440 bool single_cb_enable
,
441 bool blend_mrt0_is_dual_src
,
442 struct radv_blend_state
*blend
)
444 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
445 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
446 unsigned col_format
= 0;
448 for (unsigned i
= 0; i
< (single_cb_enable
? 1 : subpass
->color_count
); ++i
) {
451 if (subpass
->color_attachments
[i
].attachment
== VK_ATTACHMENT_UNUSED
) {
452 cf
= V_028714_SPI_SHADER_ZERO
;
454 struct radv_render_pass_attachment
*attachment
= pass
->attachments
+ subpass
->color_attachments
[i
].attachment
;
456 cf
= si_choose_spi_color_format(attachment
->format
,
457 blend_enable
& (1 << i
),
458 blend_need_alpha
& (1 << i
));
461 col_format
|= cf
<< (4 * i
);
464 blend
->cb_shader_mask
= ac_get_cb_shader_mask(col_format
);
466 if (blend_mrt0_is_dual_src
)
467 col_format
|= (col_format
& 0xf) << 4;
468 blend
->spi_shader_col_format
= col_format
;
472 format_is_int8(VkFormat format
)
474 const struct vk_format_description
*desc
= vk_format_description(format
);
475 int channel
= vk_format_get_first_non_void_channel(format
);
477 return channel
>= 0 && desc
->channel
[channel
].pure_integer
&&
478 desc
->channel
[channel
].size
== 8;
482 format_is_int10(VkFormat format
)
484 const struct vk_format_description
*desc
= vk_format_description(format
);
486 if (desc
->nr_channels
!= 4)
488 for (unsigned i
= 0; i
< 4; i
++) {
489 if (desc
->channel
[i
].pure_integer
&& desc
->channel
[i
].size
== 10)
495 unsigned radv_format_meta_fs_key(VkFormat format
)
497 unsigned col_format
= si_choose_spi_color_format(format
, false, false) - 1;
498 bool is_int8
= format_is_int8(format
);
499 bool is_int10
= format_is_int10(format
);
501 return col_format
+ (is_int8
? 3 : is_int10
? 5 : 0);
505 radv_pipeline_compute_get_int_clamp(const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
506 unsigned *is_int8
, unsigned *is_int10
)
508 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
509 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
513 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
514 struct radv_render_pass_attachment
*attachment
;
516 if (subpass
->color_attachments
[i
].attachment
== VK_ATTACHMENT_UNUSED
)
519 attachment
= pass
->attachments
+ subpass
->color_attachments
[i
].attachment
;
521 if (format_is_int8(attachment
->format
))
523 if (format_is_int10(attachment
->format
))
528 static struct radv_blend_state
529 radv_pipeline_init_blend_state(struct radv_pipeline
*pipeline
,
530 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
531 const struct radv_graphics_pipeline_create_info
*extra
)
533 const VkPipelineColorBlendStateCreateInfo
*vkblend
= pCreateInfo
->pColorBlendState
;
534 const VkPipelineMultisampleStateCreateInfo
*vkms
= pCreateInfo
->pMultisampleState
;
535 struct radv_blend_state blend
= {0};
536 unsigned mode
= V_028808_CB_NORMAL
;
537 uint32_t blend_enable
= 0, blend_need_alpha
= 0;
538 bool blend_mrt0_is_dual_src
= false;
540 bool single_cb_enable
= false;
545 if (extra
&& extra
->custom_blend_mode
) {
546 single_cb_enable
= true;
547 mode
= extra
->custom_blend_mode
;
549 blend
.cb_color_control
= 0;
550 if (vkblend
->logicOpEnable
)
551 blend
.cb_color_control
|= S_028808_ROP3(vkblend
->logicOp
| (vkblend
->logicOp
<< 4));
553 blend
.cb_color_control
|= S_028808_ROP3(0xcc);
555 blend
.db_alpha_to_mask
= S_028B70_ALPHA_TO_MASK_OFFSET0(2) |
556 S_028B70_ALPHA_TO_MASK_OFFSET1(2) |
557 S_028B70_ALPHA_TO_MASK_OFFSET2(2) |
558 S_028B70_ALPHA_TO_MASK_OFFSET3(2);
560 if (vkms
&& vkms
->alphaToCoverageEnable
) {
561 blend
.db_alpha_to_mask
|= S_028B70_ALPHA_TO_MASK_ENABLE(1);
564 blend
.cb_target_mask
= 0;
565 for (i
= 0; i
< vkblend
->attachmentCount
; i
++) {
566 const VkPipelineColorBlendAttachmentState
*att
= &vkblend
->pAttachments
[i
];
567 unsigned blend_cntl
= 0;
568 unsigned srcRGB_opt
, dstRGB_opt
, srcA_opt
, dstA_opt
;
569 VkBlendOp eqRGB
= att
->colorBlendOp
;
570 VkBlendFactor srcRGB
= att
->srcColorBlendFactor
;
571 VkBlendFactor dstRGB
= att
->dstColorBlendFactor
;
572 VkBlendOp eqA
= att
->alphaBlendOp
;
573 VkBlendFactor srcA
= att
->srcAlphaBlendFactor
;
574 VkBlendFactor dstA
= att
->dstAlphaBlendFactor
;
576 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
);
578 if (!att
->colorWriteMask
)
581 blend
.cb_target_mask
|= (unsigned)att
->colorWriteMask
<< (4 * i
);
582 if (!att
->blendEnable
) {
583 blend
.cb_blend_control
[i
] = blend_cntl
;
587 if (is_dual_src(srcRGB
) || is_dual_src(dstRGB
) || is_dual_src(srcA
) || is_dual_src(dstA
))
589 blend_mrt0_is_dual_src
= true;
591 if (eqRGB
== VK_BLEND_OP_MIN
|| eqRGB
== VK_BLEND_OP_MAX
) {
592 srcRGB
= VK_BLEND_FACTOR_ONE
;
593 dstRGB
= VK_BLEND_FACTOR_ONE
;
595 if (eqA
== VK_BLEND_OP_MIN
|| eqA
== VK_BLEND_OP_MAX
) {
596 srcA
= VK_BLEND_FACTOR_ONE
;
597 dstA
= VK_BLEND_FACTOR_ONE
;
600 /* Blending optimizations for RB+.
601 * These transformations don't change the behavior.
603 * First, get rid of DST in the blend factors:
604 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
606 si_blend_remove_dst(&eqRGB
, &srcRGB
, &dstRGB
,
607 VK_BLEND_FACTOR_DST_COLOR
,
608 VK_BLEND_FACTOR_SRC_COLOR
);
610 si_blend_remove_dst(&eqA
, &srcA
, &dstA
,
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_ALPHA
,
616 VK_BLEND_FACTOR_SRC_ALPHA
);
618 /* Look up the ideal settings from tables. */
619 srcRGB_opt
= si_translate_blend_opt_factor(srcRGB
, false);
620 dstRGB_opt
= si_translate_blend_opt_factor(dstRGB
, false);
621 srcA_opt
= si_translate_blend_opt_factor(srcA
, true);
622 dstA_opt
= si_translate_blend_opt_factor(dstA
, true);
624 /* Handle interdependencies. */
625 if (si_blend_factor_uses_dst(srcRGB
))
626 dstRGB_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
627 if (si_blend_factor_uses_dst(srcA
))
628 dstA_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
630 if (srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
&&
631 (dstRGB
== VK_BLEND_FACTOR_ZERO
||
632 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
633 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
))
634 dstRGB_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0
;
636 /* Set the final value. */
637 blend
.sx_mrt_blend_opt
[i
] =
638 S_028760_COLOR_SRC_OPT(srcRGB_opt
) |
639 S_028760_COLOR_DST_OPT(dstRGB_opt
) |
640 S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB
)) |
641 S_028760_ALPHA_SRC_OPT(srcA_opt
) |
642 S_028760_ALPHA_DST_OPT(dstA_opt
) |
643 S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA
));
644 blend_cntl
|= S_028780_ENABLE(1);
646 blend_cntl
|= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB
));
647 blend_cntl
|= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB
));
648 blend_cntl
|= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB
));
649 if (srcA
!= srcRGB
|| dstA
!= dstRGB
|| eqA
!= eqRGB
) {
650 blend_cntl
|= S_028780_SEPARATE_ALPHA_BLEND(1);
651 blend_cntl
|= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA
));
652 blend_cntl
|= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA
));
653 blend_cntl
|= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA
));
655 blend
.cb_blend_control
[i
] = blend_cntl
;
657 blend_enable
|= 1 << i
;
659 if (srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
660 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
661 srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
662 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
663 srcRGB
== VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
||
664 dstRGB
== VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
)
665 blend_need_alpha
|= 1 << i
;
667 for (i
= vkblend
->attachmentCount
; i
< 8; i
++) {
668 blend
.cb_blend_control
[i
] = 0;
669 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
);
672 /* disable RB+ for now */
673 if (pipeline
->device
->physical_device
->has_rbplus
)
674 blend
.cb_color_control
|= S_028808_DISABLE_DUAL_QUAD(1);
676 if (blend
.cb_target_mask
)
677 blend
.cb_color_control
|= S_028808_MODE(mode
);
679 blend
.cb_color_control
|= S_028808_MODE(V_028808_CB_DISABLE
);
681 radv_pipeline_compute_spi_color_formats(pipeline
, pCreateInfo
,
682 blend_enable
, blend_need_alpha
, single_cb_enable
, blend_mrt0_is_dual_src
,
687 static uint32_t si_translate_stencil_op(enum VkStencilOp op
)
690 case VK_STENCIL_OP_KEEP
:
691 return V_02842C_STENCIL_KEEP
;
692 case VK_STENCIL_OP_ZERO
:
693 return V_02842C_STENCIL_ZERO
;
694 case VK_STENCIL_OP_REPLACE
:
695 return V_02842C_STENCIL_REPLACE_TEST
;
696 case VK_STENCIL_OP_INCREMENT_AND_CLAMP
:
697 return V_02842C_STENCIL_ADD_CLAMP
;
698 case VK_STENCIL_OP_DECREMENT_AND_CLAMP
:
699 return V_02842C_STENCIL_SUB_CLAMP
;
700 case VK_STENCIL_OP_INVERT
:
701 return V_02842C_STENCIL_INVERT
;
702 case VK_STENCIL_OP_INCREMENT_AND_WRAP
:
703 return V_02842C_STENCIL_ADD_WRAP
;
704 case VK_STENCIL_OP_DECREMENT_AND_WRAP
:
705 return V_02842C_STENCIL_SUB_WRAP
;
711 static uint32_t si_translate_fill(VkPolygonMode func
)
714 case VK_POLYGON_MODE_FILL
:
715 return V_028814_X_DRAW_TRIANGLES
;
716 case VK_POLYGON_MODE_LINE
:
717 return V_028814_X_DRAW_LINES
;
718 case VK_POLYGON_MODE_POINT
:
719 return V_028814_X_DRAW_POINTS
;
722 return V_028814_X_DRAW_POINTS
;
726 static uint8_t radv_pipeline_get_ps_iter_samples(const VkPipelineMultisampleStateCreateInfo
*vkms
)
728 uint32_t num_samples
= vkms
->rasterizationSamples
;
729 uint32_t ps_iter_samples
= 1;
731 if (vkms
->sampleShadingEnable
) {
732 ps_iter_samples
= ceil(vkms
->minSampleShading
* num_samples
);
733 ps_iter_samples
= util_next_power_of_two(ps_iter_samples
);
735 return ps_iter_samples
;
739 radv_pipeline_init_multisample_state(struct radv_pipeline
*pipeline
,
740 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
742 const VkPipelineMultisampleStateCreateInfo
*vkms
= pCreateInfo
->pMultisampleState
;
743 struct radv_multisample_state
*ms
= &pipeline
->graphics
.ms
;
744 unsigned num_tile_pipes
= pipeline
->device
->physical_device
->rad_info
.num_tile_pipes
;
745 int ps_iter_samples
= 1;
746 uint32_t mask
= 0xffff;
749 ms
->num_samples
= vkms
->rasterizationSamples
;
754 ps_iter_samples
= radv_pipeline_get_ps_iter_samples(vkms
);
755 if (vkms
&& !vkms
->sampleShadingEnable
&& pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.force_persample
) {
756 ps_iter_samples
= ms
->num_samples
;
759 ms
->pa_sc_line_cntl
= S_028BDC_DX10_DIAMOND_TEST_ENA(1);
760 ms
->pa_sc_aa_config
= 0;
761 ms
->db_eqaa
= S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
762 S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
763 ms
->pa_sc_mode_cntl_1
=
764 S_028A4C_WALK_FENCE_ENABLE(1) | //TODO linear dst fixes
765 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes
== 2 ? 2 : 3) |
767 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
768 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
769 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
770 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
771 S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
772 S_028A4C_FORCE_EOV_REZ_ENABLE(1);
773 ms
->pa_sc_mode_cntl_0
= S_028A48_ALTERNATE_RBS_PER_TILE(pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) |
774 S_028A48_VPORT_SCISSOR_ENABLE(1);
776 if (ms
->num_samples
> 1) {
777 unsigned log_samples
= util_logbase2(ms
->num_samples
);
778 unsigned log_ps_iter_samples
= util_logbase2(ps_iter_samples
);
779 ms
->pa_sc_mode_cntl_0
|= S_028A48_MSAA_ENABLE(1);
780 ms
->pa_sc_line_cntl
|= S_028BDC_EXPAND_LINE_WIDTH(1); /* CM_R_028BDC_PA_SC_LINE_CNTL */
781 ms
->db_eqaa
|= S_028804_MAX_ANCHOR_SAMPLES(log_samples
) |
782 S_028804_PS_ITER_SAMPLES(log_ps_iter_samples
) |
783 S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples
) |
784 S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples
);
785 ms
->pa_sc_aa_config
|= S_028BE0_MSAA_NUM_SAMPLES(log_samples
) |
786 S_028BE0_MAX_SAMPLE_DIST(radv_cayman_get_maxdist(log_samples
)) |
787 S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples
); /* CM_R_028BE0_PA_SC_AA_CONFIG */
788 ms
->pa_sc_mode_cntl_1
|= S_028A4C_PS_ITER_SAMPLE(ps_iter_samples
> 1);
789 if (ps_iter_samples
> 1)
790 pipeline
->graphics
.spi_baryc_cntl
|= S_0286E0_POS_FLOAT_LOCATION(2);
793 const struct VkPipelineRasterizationStateRasterizationOrderAMD
*raster_order
=
794 vk_find_struct_const(pCreateInfo
->pRasterizationState
->pNext
, PIPELINE_RASTERIZATION_STATE_RASTERIZATION_ORDER_AMD
);
795 if (raster_order
&& raster_order
->rasterizationOrder
== VK_RASTERIZATION_ORDER_RELAXED_AMD
) {
796 ms
->pa_sc_mode_cntl_1
|= S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(1) |
797 S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7);
800 if (vkms
&& vkms
->pSampleMask
) {
801 mask
= vkms
->pSampleMask
[0] & 0xffff;
804 ms
->pa_sc_aa_mask
[0] = mask
| (mask
<< 16);
805 ms
->pa_sc_aa_mask
[1] = mask
| (mask
<< 16);
809 radv_prim_can_use_guardband(enum VkPrimitiveTopology topology
)
812 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
813 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
814 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
815 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
816 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
818 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
819 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
820 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
821 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
822 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
823 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
826 unreachable("unhandled primitive type");
831 si_translate_prim(enum VkPrimitiveTopology topology
)
834 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
835 return V_008958_DI_PT_POINTLIST
;
836 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
837 return V_008958_DI_PT_LINELIST
;
838 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
839 return V_008958_DI_PT_LINESTRIP
;
840 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
841 return V_008958_DI_PT_TRILIST
;
842 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
843 return V_008958_DI_PT_TRISTRIP
;
844 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
845 return V_008958_DI_PT_TRIFAN
;
846 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
847 return V_008958_DI_PT_LINELIST_ADJ
;
848 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
849 return V_008958_DI_PT_LINESTRIP_ADJ
;
850 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
851 return V_008958_DI_PT_TRILIST_ADJ
;
852 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
853 return V_008958_DI_PT_TRISTRIP_ADJ
;
854 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
855 return V_008958_DI_PT_PATCH
;
863 si_conv_gl_prim_to_gs_out(unsigned gl_prim
)
866 case 0: /* GL_POINTS */
867 return V_028A6C_OUTPRIM_TYPE_POINTLIST
;
868 case 1: /* GL_LINES */
869 case 3: /* GL_LINE_STRIP */
870 case 0xA: /* GL_LINE_STRIP_ADJACENCY_ARB */
871 case 0x8E7A: /* GL_ISOLINES */
872 return V_028A6C_OUTPRIM_TYPE_LINESTRIP
;
874 case 4: /* GL_TRIANGLES */
875 case 0xc: /* GL_TRIANGLES_ADJACENCY_ARB */
876 case 5: /* GL_TRIANGLE_STRIP */
877 case 7: /* GL_QUADS */
878 return V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
886 si_conv_prim_to_gs_out(enum VkPrimitiveTopology topology
)
889 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
890 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
891 return V_028A6C_OUTPRIM_TYPE_POINTLIST
;
892 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
893 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
894 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
895 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
896 return V_028A6C_OUTPRIM_TYPE_LINESTRIP
;
897 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
898 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
899 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
900 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
901 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
902 return V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
909 static unsigned si_map_swizzle(unsigned swizzle
)
913 return V_008F0C_SQ_SEL_Y
;
915 return V_008F0C_SQ_SEL_Z
;
917 return V_008F0C_SQ_SEL_W
;
919 return V_008F0C_SQ_SEL_0
;
921 return V_008F0C_SQ_SEL_1
;
922 default: /* VK_SWIZZLE_X */
923 return V_008F0C_SQ_SEL_X
;
928 static unsigned radv_dynamic_state_mask(VkDynamicState state
)
931 case VK_DYNAMIC_STATE_VIEWPORT
:
932 return RADV_DYNAMIC_VIEWPORT
;
933 case VK_DYNAMIC_STATE_SCISSOR
:
934 return RADV_DYNAMIC_SCISSOR
;
935 case VK_DYNAMIC_STATE_LINE_WIDTH
:
936 return RADV_DYNAMIC_LINE_WIDTH
;
937 case VK_DYNAMIC_STATE_DEPTH_BIAS
:
938 return RADV_DYNAMIC_DEPTH_BIAS
;
939 case VK_DYNAMIC_STATE_BLEND_CONSTANTS
:
940 return RADV_DYNAMIC_BLEND_CONSTANTS
;
941 case VK_DYNAMIC_STATE_DEPTH_BOUNDS
:
942 return RADV_DYNAMIC_DEPTH_BOUNDS
;
943 case VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK
:
944 return RADV_DYNAMIC_STENCIL_COMPARE_MASK
;
945 case VK_DYNAMIC_STATE_STENCIL_WRITE_MASK
:
946 return RADV_DYNAMIC_STENCIL_WRITE_MASK
;
947 case VK_DYNAMIC_STATE_STENCIL_REFERENCE
:
948 return RADV_DYNAMIC_STENCIL_REFERENCE
;
949 case VK_DYNAMIC_STATE_DISCARD_RECTANGLE_EXT
:
950 return RADV_DYNAMIC_DISCARD_RECTANGLE
;
952 unreachable("Unhandled dynamic state");
956 static uint32_t radv_pipeline_needed_dynamic_state(const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
958 uint32_t states
= RADV_DYNAMIC_ALL
;
960 /* If rasterization is disabled we do not care about any of the dynamic states,
961 * since they are all rasterization related only. */
962 if (pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
)
965 if (!pCreateInfo
->pRasterizationState
->depthBiasEnable
)
966 states
&= ~RADV_DYNAMIC_DEPTH_BIAS
;
968 if (!pCreateInfo
->pDepthStencilState
||
969 !pCreateInfo
->pDepthStencilState
->depthBoundsTestEnable
)
970 states
&= ~RADV_DYNAMIC_DEPTH_BOUNDS
;
972 if (!pCreateInfo
->pDepthStencilState
||
973 !pCreateInfo
->pDepthStencilState
->stencilTestEnable
)
974 states
&= ~(RADV_DYNAMIC_STENCIL_COMPARE_MASK
|
975 RADV_DYNAMIC_STENCIL_WRITE_MASK
|
976 RADV_DYNAMIC_STENCIL_REFERENCE
);
978 if (!vk_find_struct_const(pCreateInfo
->pNext
, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT
))
979 states
&= ~RADV_DYNAMIC_DISCARD_RECTANGLE
;
981 /* TODO: blend constants & line width. */
988 radv_pipeline_init_dynamic_state(struct radv_pipeline
*pipeline
,
989 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
991 uint32_t needed_states
= radv_pipeline_needed_dynamic_state(pCreateInfo
);
992 uint32_t states
= needed_states
;
993 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
994 struct radv_subpass
*subpass
= &pass
->subpasses
[pCreateInfo
->subpass
];
996 pipeline
->dynamic_state
= default_dynamic_state
;
997 pipeline
->graphics
.needed_dynamic_state
= needed_states
;
999 if (pCreateInfo
->pDynamicState
) {
1000 /* Remove all of the states that are marked as dynamic */
1001 uint32_t count
= pCreateInfo
->pDynamicState
->dynamicStateCount
;
1002 for (uint32_t s
= 0; s
< count
; s
++)
1003 states
&= ~radv_dynamic_state_mask(pCreateInfo
->pDynamicState
->pDynamicStates
[s
]);
1006 struct radv_dynamic_state
*dynamic
= &pipeline
->dynamic_state
;
1008 if (needed_states
& RADV_DYNAMIC_VIEWPORT
) {
1009 assert(pCreateInfo
->pViewportState
);
1011 dynamic
->viewport
.count
= pCreateInfo
->pViewportState
->viewportCount
;
1012 if (states
& RADV_DYNAMIC_VIEWPORT
) {
1013 typed_memcpy(dynamic
->viewport
.viewports
,
1014 pCreateInfo
->pViewportState
->pViewports
,
1015 pCreateInfo
->pViewportState
->viewportCount
);
1019 if (needed_states
& RADV_DYNAMIC_SCISSOR
) {
1020 dynamic
->scissor
.count
= pCreateInfo
->pViewportState
->scissorCount
;
1021 if (states
& RADV_DYNAMIC_SCISSOR
) {
1022 typed_memcpy(dynamic
->scissor
.scissors
,
1023 pCreateInfo
->pViewportState
->pScissors
,
1024 pCreateInfo
->pViewportState
->scissorCount
);
1028 if (states
& RADV_DYNAMIC_LINE_WIDTH
) {
1029 assert(pCreateInfo
->pRasterizationState
);
1030 dynamic
->line_width
= pCreateInfo
->pRasterizationState
->lineWidth
;
1033 if (states
& RADV_DYNAMIC_DEPTH_BIAS
) {
1034 assert(pCreateInfo
->pRasterizationState
);
1035 dynamic
->depth_bias
.bias
=
1036 pCreateInfo
->pRasterizationState
->depthBiasConstantFactor
;
1037 dynamic
->depth_bias
.clamp
=
1038 pCreateInfo
->pRasterizationState
->depthBiasClamp
;
1039 dynamic
->depth_bias
.slope
=
1040 pCreateInfo
->pRasterizationState
->depthBiasSlopeFactor
;
1043 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1045 * pColorBlendState is [...] NULL if the pipeline has rasterization
1046 * disabled or if the subpass of the render pass the pipeline is
1047 * created against does not use any color attachments.
1049 bool uses_color_att
= false;
1050 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
1051 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
) {
1052 uses_color_att
= true;
1057 if (uses_color_att
&& states
& RADV_DYNAMIC_BLEND_CONSTANTS
) {
1058 assert(pCreateInfo
->pColorBlendState
);
1059 typed_memcpy(dynamic
->blend_constants
,
1060 pCreateInfo
->pColorBlendState
->blendConstants
, 4);
1063 /* If there is no depthstencil attachment, then don't read
1064 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
1065 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
1066 * no need to override the depthstencil defaults in
1067 * radv_pipeline::dynamic_state when there is no depthstencil attachment.
1069 * Section 9.2 of the Vulkan 1.0.15 spec says:
1071 * pDepthStencilState is [...] NULL if the pipeline has rasterization
1072 * disabled or if the subpass of the render pass the pipeline is created
1073 * against does not use a depth/stencil attachment.
1075 if (needed_states
&&
1076 subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
1077 assert(pCreateInfo
->pDepthStencilState
);
1079 if (states
& RADV_DYNAMIC_DEPTH_BOUNDS
) {
1080 dynamic
->depth_bounds
.min
=
1081 pCreateInfo
->pDepthStencilState
->minDepthBounds
;
1082 dynamic
->depth_bounds
.max
=
1083 pCreateInfo
->pDepthStencilState
->maxDepthBounds
;
1086 if (states
& RADV_DYNAMIC_STENCIL_COMPARE_MASK
) {
1087 dynamic
->stencil_compare_mask
.front
=
1088 pCreateInfo
->pDepthStencilState
->front
.compareMask
;
1089 dynamic
->stencil_compare_mask
.back
=
1090 pCreateInfo
->pDepthStencilState
->back
.compareMask
;
1093 if (states
& RADV_DYNAMIC_STENCIL_WRITE_MASK
) {
1094 dynamic
->stencil_write_mask
.front
=
1095 pCreateInfo
->pDepthStencilState
->front
.writeMask
;
1096 dynamic
->stencil_write_mask
.back
=
1097 pCreateInfo
->pDepthStencilState
->back
.writeMask
;
1100 if (states
& RADV_DYNAMIC_STENCIL_REFERENCE
) {
1101 dynamic
->stencil_reference
.front
=
1102 pCreateInfo
->pDepthStencilState
->front
.reference
;
1103 dynamic
->stencil_reference
.back
=
1104 pCreateInfo
->pDepthStencilState
->back
.reference
;
1108 const VkPipelineDiscardRectangleStateCreateInfoEXT
*discard_rectangle_info
=
1109 vk_find_struct_const(pCreateInfo
->pNext
, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT
);
1110 if (states
& RADV_DYNAMIC_DISCARD_RECTANGLE
) {
1111 dynamic
->discard_rectangle
.count
= discard_rectangle_info
->discardRectangleCount
;
1112 typed_memcpy(dynamic
->discard_rectangle
.rectangles
,
1113 discard_rectangle_info
->pDiscardRectangles
,
1114 discard_rectangle_info
->discardRectangleCount
);
1117 pipeline
->dynamic_state
.mask
= states
;
1120 static struct radv_gs_state
1121 calculate_gs_info(const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1122 const struct radv_pipeline
*pipeline
)
1124 struct radv_gs_state gs
= {0};
1125 struct radv_shader_variant_info
*gs_info
= &pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
;
1126 struct radv_es_output_info
*es_info
;
1127 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1128 es_info
= radv_pipeline_has_tess(pipeline
) ? &gs_info
->tes
.es_info
: &gs_info
->vs
.es_info
;
1130 es_info
= radv_pipeline_has_tess(pipeline
) ?
1131 &pipeline
->shaders
[MESA_SHADER_TESS_EVAL
]->info
.tes
.es_info
:
1132 &pipeline
->shaders
[MESA_SHADER_VERTEX
]->info
.vs
.es_info
;
1134 unsigned gs_num_invocations
= MAX2(gs_info
->gs
.invocations
, 1);
1135 bool uses_adjacency
;
1136 switch(pCreateInfo
->pInputAssemblyState
->topology
) {
1137 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
1138 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
1139 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
1140 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
1141 uses_adjacency
= true;
1144 uses_adjacency
= false;
1148 /* All these are in dwords: */
1149 /* We can't allow using the whole LDS, because GS waves compete with
1150 * other shader stages for LDS space. */
1151 const unsigned max_lds_size
= 8 * 1024;
1152 const unsigned esgs_itemsize
= es_info
->esgs_itemsize
/ 4;
1153 unsigned esgs_lds_size
;
1155 /* All these are per subgroup: */
1156 const unsigned max_out_prims
= 32 * 1024;
1157 const unsigned max_es_verts
= 255;
1158 const unsigned ideal_gs_prims
= 64;
1159 unsigned max_gs_prims
, gs_prims
;
1160 unsigned min_es_verts
, es_verts
, worst_case_es_verts
;
1162 if (uses_adjacency
|| gs_num_invocations
> 1)
1163 max_gs_prims
= 127 / gs_num_invocations
;
1167 /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
1168 * Make sure we don't go over the maximum value.
1170 if (gs_info
->gs
.vertices_out
> 0) {
1171 max_gs_prims
= MIN2(max_gs_prims
,
1173 (gs_info
->gs
.vertices_out
* gs_num_invocations
));
1175 assert(max_gs_prims
> 0);
1177 /* If the primitive has adjacency, halve the number of vertices
1178 * that will be reused in multiple primitives.
1180 min_es_verts
= gs_info
->gs
.vertices_in
/ (uses_adjacency
? 2 : 1);
1182 gs_prims
= MIN2(ideal_gs_prims
, max_gs_prims
);
1183 worst_case_es_verts
= MIN2(min_es_verts
* gs_prims
, max_es_verts
);
1185 /* Compute ESGS LDS size based on the worst case number of ES vertices
1186 * needed to create the target number of GS prims per subgroup.
1188 esgs_lds_size
= esgs_itemsize
* worst_case_es_verts
;
1190 /* If total LDS usage is too big, refactor partitions based on ratio
1191 * of ESGS item sizes.
1193 if (esgs_lds_size
> max_lds_size
) {
1194 /* Our target GS Prims Per Subgroup was too large. Calculate
1195 * the maximum number of GS Prims Per Subgroup that will fit
1196 * into LDS, capped by the maximum that the hardware can support.
1198 gs_prims
= MIN2((max_lds_size
/ (esgs_itemsize
* min_es_verts
)),
1200 assert(gs_prims
> 0);
1201 worst_case_es_verts
= MIN2(min_es_verts
* gs_prims
,
1204 esgs_lds_size
= esgs_itemsize
* worst_case_es_verts
;
1205 assert(esgs_lds_size
<= max_lds_size
);
1208 /* Now calculate remaining ESGS information. */
1210 es_verts
= MIN2(esgs_lds_size
/ esgs_itemsize
, max_es_verts
);
1212 es_verts
= max_es_verts
;
1214 /* Vertices for adjacency primitives are not always reused, so restore
1215 * it for ES_VERTS_PER_SUBGRP.
1217 min_es_verts
= gs_info
->gs
.vertices_in
;
1219 /* For normal primitives, the VGT only checks if they are past the ES
1220 * verts per subgroup after allocating a full GS primitive and if they
1221 * are, kick off a new subgroup. But if those additional ES verts are
1222 * unique (e.g. not reused) we need to make sure there is enough LDS
1223 * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
1225 es_verts
-= min_es_verts
- 1;
1227 uint32_t es_verts_per_subgroup
= es_verts
;
1228 uint32_t gs_prims_per_subgroup
= gs_prims
;
1229 uint32_t gs_inst_prims_in_subgroup
= gs_prims
* gs_num_invocations
;
1230 uint32_t max_prims_per_subgroup
= gs_inst_prims_in_subgroup
* gs_info
->gs
.vertices_out
;
1231 gs
.lds_size
= align(esgs_lds_size
, 128) / 128;
1232 gs
.vgt_gs_onchip_cntl
= S_028A44_ES_VERTS_PER_SUBGRP(es_verts_per_subgroup
) |
1233 S_028A44_GS_PRIMS_PER_SUBGRP(gs_prims_per_subgroup
) |
1234 S_028A44_GS_INST_PRIMS_IN_SUBGRP(gs_inst_prims_in_subgroup
);
1235 gs
.vgt_gs_max_prims_per_subgroup
= S_028A94_MAX_PRIMS_PER_SUBGROUP(max_prims_per_subgroup
);
1236 gs
.vgt_esgs_ring_itemsize
= esgs_itemsize
;
1237 assert(max_prims_per_subgroup
<= max_out_prims
);
1243 calculate_gs_ring_sizes(struct radv_pipeline
*pipeline
, const struct radv_gs_state
*gs
)
1245 struct radv_device
*device
= pipeline
->device
;
1246 unsigned num_se
= device
->physical_device
->rad_info
.max_se
;
1247 unsigned wave_size
= 64;
1248 unsigned max_gs_waves
= 32 * num_se
; /* max 32 per SE on GCN */
1249 unsigned gs_vertex_reuse
= 16 * num_se
; /* GS_VERTEX_REUSE register (per SE) */
1250 unsigned alignment
= 256 * num_se
;
1251 /* The maximum size is 63.999 MB per SE. */
1252 unsigned max_size
= ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se
;
1253 struct radv_shader_variant_info
*gs_info
= &pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
;
1255 /* Calculate the minimum size. */
1256 unsigned min_esgs_ring_size
= align(gs
->vgt_esgs_ring_itemsize
* 4 * gs_vertex_reuse
*
1257 wave_size
, alignment
);
1258 /* These are recommended sizes, not minimum sizes. */
1259 unsigned esgs_ring_size
= max_gs_waves
* 2 * wave_size
*
1260 gs
->vgt_esgs_ring_itemsize
* 4 * gs_info
->gs
.vertices_in
;
1261 unsigned gsvs_ring_size
= max_gs_waves
* 2 * wave_size
*
1262 gs_info
->gs
.max_gsvs_emit_size
* 1; // no streams in VK (gs->max_gs_stream + 1);
1264 min_esgs_ring_size
= align(min_esgs_ring_size
, alignment
);
1265 esgs_ring_size
= align(esgs_ring_size
, alignment
);
1266 gsvs_ring_size
= align(gsvs_ring_size
, alignment
);
1268 if (pipeline
->device
->physical_device
->rad_info
.chip_class
<= VI
)
1269 pipeline
->graphics
.esgs_ring_size
= CLAMP(esgs_ring_size
, min_esgs_ring_size
, max_size
);
1271 pipeline
->graphics
.gsvs_ring_size
= MIN2(gsvs_ring_size
, max_size
);
1274 static void si_multiwave_lds_size_workaround(struct radv_device
*device
,
1277 /* SPI barrier management bug:
1278 * Make sure we have at least 4k of LDS in use to avoid the bug.
1279 * It applies to workgroup sizes of more than one wavefront.
1281 if (device
->physical_device
->rad_info
.family
== CHIP_BONAIRE
||
1282 device
->physical_device
->rad_info
.family
== CHIP_KABINI
||
1283 device
->physical_device
->rad_info
.family
== CHIP_MULLINS
)
1284 *lds_size
= MAX2(*lds_size
, 8);
1287 struct radv_shader_variant
*
1288 radv_get_vertex_shader(struct radv_pipeline
*pipeline
)
1290 if (pipeline
->shaders
[MESA_SHADER_VERTEX
])
1291 return pipeline
->shaders
[MESA_SHADER_VERTEX
];
1292 if (pipeline
->shaders
[MESA_SHADER_TESS_CTRL
])
1293 return pipeline
->shaders
[MESA_SHADER_TESS_CTRL
];
1294 return pipeline
->shaders
[MESA_SHADER_GEOMETRY
];
1297 static struct radv_shader_variant
*
1298 radv_get_tess_eval_shader(struct radv_pipeline
*pipeline
)
1300 if (pipeline
->shaders
[MESA_SHADER_TESS_EVAL
])
1301 return pipeline
->shaders
[MESA_SHADER_TESS_EVAL
];
1302 return pipeline
->shaders
[MESA_SHADER_GEOMETRY
];
1305 static struct radv_tessellation_state
1306 calculate_tess_state(struct radv_pipeline
*pipeline
,
1307 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1309 unsigned num_tcs_input_cp
= pCreateInfo
->pTessellationState
->patchControlPoints
;
1310 unsigned num_tcs_output_cp
, num_tcs_inputs
, num_tcs_outputs
;
1311 unsigned num_tcs_patch_outputs
;
1312 unsigned input_vertex_size
, output_vertex_size
, pervertex_output_patch_size
;
1313 unsigned input_patch_size
, output_patch_size
, output_patch0_offset
;
1314 unsigned lds_size
, hardware_lds_size
;
1315 unsigned num_patches
;
1316 struct radv_tessellation_state tess
= {0};
1318 /* This calculates how shader inputs and outputs among VS, TCS, and TES
1319 * are laid out in LDS. */
1320 num_tcs_inputs
= util_last_bit64(radv_get_vertex_shader(pipeline
)->info
.vs
.outputs_written
);
1322 num_tcs_outputs
= util_last_bit64(pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.outputs_written
); //tcs->outputs_written
1323 num_tcs_output_cp
= pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.tcs_vertices_out
; //TCS VERTICES OUT
1324 num_tcs_patch_outputs
= util_last_bit64(pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.patch_outputs_written
);
1326 /* Ensure that we only need one wave per SIMD so we don't need to check
1327 * resource usage. Also ensures that the number of tcs in and out
1328 * vertices per threadgroup are at most 256.
1330 input_vertex_size
= num_tcs_inputs
* 16;
1331 output_vertex_size
= num_tcs_outputs
* 16;
1333 input_patch_size
= num_tcs_input_cp
* input_vertex_size
;
1335 pervertex_output_patch_size
= num_tcs_output_cp
* output_vertex_size
;
1336 output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
1337 /* Ensure that we only need one wave per SIMD so we don't need to check
1338 * resource usage. Also ensures that the number of tcs in and out
1339 * vertices per threadgroup are at most 256.
1341 num_patches
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
) * 4;
1343 /* Make sure that the data fits in LDS. This assumes the shaders only
1344 * use LDS for the inputs and outputs.
1346 hardware_lds_size
= pipeline
->device
->physical_device
->rad_info
.chip_class
>= CIK
? 65536 : 32768;
1347 num_patches
= MIN2(num_patches
, hardware_lds_size
/ (input_patch_size
+ output_patch_size
));
1349 /* Make sure the output data fits in the offchip buffer */
1350 num_patches
= MIN2(num_patches
,
1351 (pipeline
->device
->tess_offchip_block_dw_size
* 4) /
1354 /* Not necessary for correctness, but improves performance. The
1355 * specific value is taken from the proprietary driver.
1357 num_patches
= MIN2(num_patches
, 40);
1359 /* SI bug workaround - limit LS-HS threadgroups to only one wave. */
1360 if (pipeline
->device
->physical_device
->rad_info
.chip_class
== SI
) {
1361 unsigned one_wave
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
);
1362 num_patches
= MIN2(num_patches
, one_wave
);
1365 output_patch0_offset
= input_patch_size
* num_patches
;
1367 lds_size
= output_patch0_offset
+ output_patch_size
* num_patches
;
1369 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1370 assert(lds_size
<= 65536);
1371 lds_size
= align(lds_size
, 512) / 512;
1373 assert(lds_size
<= 32768);
1374 lds_size
= align(lds_size
, 256) / 256;
1376 si_multiwave_lds_size_workaround(pipeline
->device
, &lds_size
);
1378 tess
.lds_size
= lds_size
;
1380 tess
.ls_hs_config
= S_028B58_NUM_PATCHES(num_patches
) |
1381 S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp
) |
1382 S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp
);
1383 tess
.num_patches
= num_patches
;
1385 struct radv_shader_variant
*tes
= radv_get_tess_eval_shader(pipeline
);
1386 unsigned type
= 0, partitioning
= 0, topology
= 0, distribution_mode
= 0;
1388 switch (tes
->info
.tes
.primitive_mode
) {
1390 type
= V_028B6C_TESS_TRIANGLE
;
1393 type
= V_028B6C_TESS_QUAD
;
1396 type
= V_028B6C_TESS_ISOLINE
;
1400 switch (tes
->info
.tes
.spacing
) {
1401 case TESS_SPACING_EQUAL
:
1402 partitioning
= V_028B6C_PART_INTEGER
;
1404 case TESS_SPACING_FRACTIONAL_ODD
:
1405 partitioning
= V_028B6C_PART_FRAC_ODD
;
1407 case TESS_SPACING_FRACTIONAL_EVEN
:
1408 partitioning
= V_028B6C_PART_FRAC_EVEN
;
1414 bool ccw
= tes
->info
.tes
.ccw
;
1415 const VkPipelineTessellationDomainOriginStateCreateInfoKHR
*domain_origin_state
=
1416 vk_find_struct_const(pCreateInfo
->pTessellationState
,
1417 PIPELINE_TESSELLATION_DOMAIN_ORIGIN_STATE_CREATE_INFO_KHR
);
1419 if (domain_origin_state
&& domain_origin_state
->domainOrigin
!= VK_TESSELLATION_DOMAIN_ORIGIN_UPPER_LEFT_KHR
)
1422 if (tes
->info
.tes
.point_mode
)
1423 topology
= V_028B6C_OUTPUT_POINT
;
1424 else if (tes
->info
.tes
.primitive_mode
== GL_ISOLINES
)
1425 topology
= V_028B6C_OUTPUT_LINE
;
1427 topology
= V_028B6C_OUTPUT_TRIANGLE_CCW
;
1429 topology
= V_028B6C_OUTPUT_TRIANGLE_CW
;
1431 if (pipeline
->device
->has_distributed_tess
) {
1432 if (pipeline
->device
->physical_device
->rad_info
.family
== CHIP_FIJI
||
1433 pipeline
->device
->physical_device
->rad_info
.family
>= CHIP_POLARIS10
)
1434 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS
;
1436 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_DONUTS
;
1438 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_NO_DIST
;
1440 tess
.tf_param
= S_028B6C_TYPE(type
) |
1441 S_028B6C_PARTITIONING(partitioning
) |
1442 S_028B6C_TOPOLOGY(topology
) |
1443 S_028B6C_DISTRIBUTION_MODE(distribution_mode
);
1448 static const struct radv_prim_vertex_count prim_size_table
[] = {
1449 [V_008958_DI_PT_NONE
] = {0, 0},
1450 [V_008958_DI_PT_POINTLIST
] = {1, 1},
1451 [V_008958_DI_PT_LINELIST
] = {2, 2},
1452 [V_008958_DI_PT_LINESTRIP
] = {2, 1},
1453 [V_008958_DI_PT_TRILIST
] = {3, 3},
1454 [V_008958_DI_PT_TRIFAN
] = {3, 1},
1455 [V_008958_DI_PT_TRISTRIP
] = {3, 1},
1456 [V_008958_DI_PT_LINELIST_ADJ
] = {4, 4},
1457 [V_008958_DI_PT_LINESTRIP_ADJ
] = {4, 1},
1458 [V_008958_DI_PT_TRILIST_ADJ
] = {6, 6},
1459 [V_008958_DI_PT_TRISTRIP_ADJ
] = {6, 2},
1460 [V_008958_DI_PT_RECTLIST
] = {3, 3},
1461 [V_008958_DI_PT_LINELOOP
] = {2, 1},
1462 [V_008958_DI_PT_POLYGON
] = {3, 1},
1463 [V_008958_DI_PT_2D_TRI_STRIP
] = {0, 0},
1466 static const struct radv_vs_output_info
*get_vs_output_info(const struct radv_pipeline
*pipeline
)
1468 if (radv_pipeline_has_gs(pipeline
))
1469 return &pipeline
->gs_copy_shader
->info
.vs
.outinfo
;
1470 else if (radv_pipeline_has_tess(pipeline
))
1471 return &pipeline
->shaders
[MESA_SHADER_TESS_EVAL
]->info
.tes
.outinfo
;
1473 return &pipeline
->shaders
[MESA_SHADER_VERTEX
]->info
.vs
.outinfo
;
1477 radv_link_shaders(struct radv_pipeline
*pipeline
, nir_shader
**shaders
)
1479 nir_shader
* ordered_shaders
[MESA_SHADER_STAGES
];
1480 int shader_count
= 0;
1482 if(shaders
[MESA_SHADER_FRAGMENT
]) {
1483 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_FRAGMENT
];
1485 if(shaders
[MESA_SHADER_GEOMETRY
]) {
1486 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_GEOMETRY
];
1488 if(shaders
[MESA_SHADER_TESS_EVAL
]) {
1489 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_TESS_EVAL
];
1491 if(shaders
[MESA_SHADER_TESS_CTRL
]) {
1492 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_TESS_CTRL
];
1494 if(shaders
[MESA_SHADER_VERTEX
]) {
1495 ordered_shaders
[shader_count
++] = shaders
[MESA_SHADER_VERTEX
];
1498 for (int i
= 1; i
< shader_count
; ++i
) {
1499 nir_lower_io_arrays_to_elements(ordered_shaders
[i
],
1500 ordered_shaders
[i
- 1]);
1502 nir_remove_dead_variables(ordered_shaders
[i
],
1503 nir_var_shader_out
);
1504 nir_remove_dead_variables(ordered_shaders
[i
- 1],
1507 bool progress
= nir_remove_unused_varyings(ordered_shaders
[i
],
1508 ordered_shaders
[i
- 1]);
1510 nir_compact_varyings(ordered_shaders
[i
],
1511 ordered_shaders
[i
- 1], true);
1514 if (nir_lower_global_vars_to_local(ordered_shaders
[i
])) {
1515 ac_lower_indirect_derefs(ordered_shaders
[i
],
1516 pipeline
->device
->physical_device
->rad_info
.chip_class
);
1518 radv_optimize_nir(ordered_shaders
[i
]);
1520 if (nir_lower_global_vars_to_local(ordered_shaders
[i
- 1])) {
1521 ac_lower_indirect_derefs(ordered_shaders
[i
- 1],
1522 pipeline
->device
->physical_device
->rad_info
.chip_class
);
1524 radv_optimize_nir(ordered_shaders
[i
- 1]);
1530 static struct radv_pipeline_key
1531 radv_generate_graphics_pipeline_key(struct radv_pipeline
*pipeline
,
1532 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1533 const struct radv_blend_state
*blend
,
1534 bool has_view_index
)
1536 const VkPipelineVertexInputStateCreateInfo
*input_state
=
1537 pCreateInfo
->pVertexInputState
;
1538 struct radv_pipeline_key key
;
1539 memset(&key
, 0, sizeof(key
));
1541 key
.has_multiview_view_index
= has_view_index
;
1543 uint32_t binding_input_rate
= 0;
1544 for (unsigned i
= 0; i
< input_state
->vertexBindingDescriptionCount
; ++i
) {
1545 if (input_state
->pVertexBindingDescriptions
[i
].inputRate
)
1546 binding_input_rate
|= 1u << input_state
->pVertexBindingDescriptions
[i
].binding
;
1549 for (unsigned i
= 0; i
< input_state
->vertexAttributeDescriptionCount
; ++i
) {
1551 binding
= input_state
->pVertexAttributeDescriptions
[i
].binding
;
1552 if (binding_input_rate
& (1u << binding
))
1553 key
.instance_rate_inputs
|= 1u << input_state
->pVertexAttributeDescriptions
[i
].location
;
1556 if (pCreateInfo
->pTessellationState
)
1557 key
.tess_input_vertices
= pCreateInfo
->pTessellationState
->patchControlPoints
;
1560 if (pCreateInfo
->pMultisampleState
&&
1561 pCreateInfo
->pMultisampleState
->rasterizationSamples
> 1) {
1562 uint32_t num_samples
= pCreateInfo
->pMultisampleState
->rasterizationSamples
;
1563 uint32_t ps_iter_samples
= radv_pipeline_get_ps_iter_samples(pCreateInfo
->pMultisampleState
);
1564 key
.multisample
= true;
1565 key
.log2_num_samples
= util_logbase2(num_samples
);
1566 key
.log2_ps_iter_samples
= util_logbase2(ps_iter_samples
);
1569 key
.col_format
= blend
->spi_shader_col_format
;
1570 if (pipeline
->device
->physical_device
->rad_info
.chip_class
< VI
)
1571 radv_pipeline_compute_get_int_clamp(pCreateInfo
, &key
.is_int8
, &key
.is_int10
);
1577 radv_fill_shader_keys(struct radv_shader_variant_key
*keys
,
1578 const struct radv_pipeline_key
*key
,
1581 keys
[MESA_SHADER_VERTEX
].vs
.instance_rate_inputs
= key
->instance_rate_inputs
;
1583 if (nir
[MESA_SHADER_TESS_CTRL
]) {
1584 keys
[MESA_SHADER_VERTEX
].vs
.as_ls
= true;
1585 keys
[MESA_SHADER_TESS_CTRL
].tcs
.num_inputs
= 0;
1586 keys
[MESA_SHADER_TESS_CTRL
].tcs
.input_vertices
= key
->tess_input_vertices
;
1587 keys
[MESA_SHADER_TESS_CTRL
].tcs
.primitive_mode
= nir
[MESA_SHADER_TESS_EVAL
]->info
.tess
.primitive_mode
;
1589 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
));
1592 if (nir
[MESA_SHADER_GEOMETRY
]) {
1593 if (nir
[MESA_SHADER_TESS_CTRL
])
1594 keys
[MESA_SHADER_TESS_EVAL
].tes
.as_es
= true;
1596 keys
[MESA_SHADER_VERTEX
].vs
.as_es
= true;
1599 for(int i
= 0; i
< MESA_SHADER_STAGES
; ++i
)
1600 keys
[i
].has_multiview_view_index
= key
->has_multiview_view_index
;
1602 keys
[MESA_SHADER_FRAGMENT
].fs
.multisample
= key
->multisample
;
1603 keys
[MESA_SHADER_FRAGMENT
].fs
.col_format
= key
->col_format
;
1604 keys
[MESA_SHADER_FRAGMENT
].fs
.is_int8
= key
->is_int8
;
1605 keys
[MESA_SHADER_FRAGMENT
].fs
.is_int10
= key
->is_int10
;
1606 keys
[MESA_SHADER_FRAGMENT
].fs
.log2_ps_iter_samples
= key
->log2_ps_iter_samples
;
1607 keys
[MESA_SHADER_FRAGMENT
].fs
.log2_num_samples
= key
->log2_num_samples
;
1611 merge_tess_info(struct shader_info
*tes_info
,
1612 const struct shader_info
*tcs_info
)
1614 /* The Vulkan 1.0.38 spec, section 21.1 Tessellator says:
1616 * "PointMode. Controls generation of points rather than triangles
1617 * or lines. This functionality defaults to disabled, and is
1618 * enabled if either shader stage includes the execution mode.
1620 * and about Triangles, Quads, IsoLines, VertexOrderCw, VertexOrderCcw,
1621 * PointMode, SpacingEqual, SpacingFractionalEven, SpacingFractionalOdd,
1622 * and OutputVertices, it says:
1624 * "One mode must be set in at least one of the tessellation
1627 * So, the fields can be set in either the TCS or TES, but they must
1628 * agree if set in both. Our backend looks at TES, so bitwise-or in
1629 * the values from the TCS.
1631 assert(tcs_info
->tess
.tcs_vertices_out
== 0 ||
1632 tes_info
->tess
.tcs_vertices_out
== 0 ||
1633 tcs_info
->tess
.tcs_vertices_out
== tes_info
->tess
.tcs_vertices_out
);
1634 tes_info
->tess
.tcs_vertices_out
|= tcs_info
->tess
.tcs_vertices_out
;
1636 assert(tcs_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
1637 tes_info
->tess
.spacing
== TESS_SPACING_UNSPECIFIED
||
1638 tcs_info
->tess
.spacing
== tes_info
->tess
.spacing
);
1639 tes_info
->tess
.spacing
|= tcs_info
->tess
.spacing
;
1641 assert(tcs_info
->tess
.primitive_mode
== 0 ||
1642 tes_info
->tess
.primitive_mode
== 0 ||
1643 tcs_info
->tess
.primitive_mode
== tes_info
->tess
.primitive_mode
);
1644 tes_info
->tess
.primitive_mode
|= tcs_info
->tess
.primitive_mode
;
1645 tes_info
->tess
.ccw
|= tcs_info
->tess
.ccw
;
1646 tes_info
->tess
.point_mode
|= tcs_info
->tess
.point_mode
;
1650 void radv_create_shaders(struct radv_pipeline
*pipeline
,
1651 struct radv_device
*device
,
1652 struct radv_pipeline_cache
*cache
,
1653 struct radv_pipeline_key key
,
1654 const VkPipelineShaderStageCreateInfo
**pStages
)
1656 struct radv_shader_module fs_m
= {0};
1657 struct radv_shader_module
*modules
[MESA_SHADER_STAGES
] = { 0, };
1658 nir_shader
*nir
[MESA_SHADER_STAGES
] = {0};
1659 void *codes
[MESA_SHADER_STAGES
] = {0};
1660 unsigned code_sizes
[MESA_SHADER_STAGES
] = {0};
1661 struct radv_shader_variant_key keys
[MESA_SHADER_STAGES
] = {{{{0}}}};
1662 unsigned char hash
[20], gs_copy_hash
[20];
1664 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1666 modules
[i
] = radv_shader_module_from_handle(pStages
[i
]->module
);
1667 if (modules
[i
]->nir
)
1668 _mesa_sha1_compute(modules
[i
]->nir
->info
.name
,
1669 strlen(modules
[i
]->nir
->info
.name
),
1674 radv_hash_shaders(hash
, pStages
, pipeline
->layout
, &key
, get_hash_flags(device
));
1675 memcpy(gs_copy_hash
, hash
, 20);
1676 gs_copy_hash
[0] ^= 1;
1678 if (modules
[MESA_SHADER_GEOMETRY
]) {
1679 struct radv_shader_variant
*variants
[MESA_SHADER_STAGES
] = {0};
1680 radv_create_shader_variants_from_pipeline_cache(device
, cache
, gs_copy_hash
, variants
);
1681 pipeline
->gs_copy_shader
= variants
[MESA_SHADER_GEOMETRY
];
1684 if (radv_create_shader_variants_from_pipeline_cache(device
, cache
, hash
, pipeline
->shaders
) &&
1685 (!modules
[MESA_SHADER_GEOMETRY
] || pipeline
->gs_copy_shader
)) {
1686 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1687 if (pipeline
->shaders
[i
])
1688 pipeline
->active_stages
|= mesa_to_vk_shader_stage(i
);
1693 if (!modules
[MESA_SHADER_FRAGMENT
] && !modules
[MESA_SHADER_COMPUTE
]) {
1695 nir_builder_init_simple_shader(&fs_b
, NULL
, MESA_SHADER_FRAGMENT
, NULL
);
1696 fs_b
.shader
->info
.name
= ralloc_strdup(fs_b
.shader
, "noop_fs");
1697 fs_m
.nir
= fs_b
.shader
;
1698 modules
[MESA_SHADER_FRAGMENT
] = &fs_m
;
1701 /* Determine first and last stage. */
1702 unsigned first
= MESA_SHADER_STAGES
;
1704 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
1707 if (first
== MESA_SHADER_STAGES
)
1712 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1713 const VkPipelineShaderStageCreateInfo
*stage
= pStages
[i
];
1718 nir
[i
] = radv_shader_compile_to_nir(device
, modules
[i
],
1719 stage
? stage
->pName
: "main", i
,
1720 stage
? stage
->pSpecializationInfo
: NULL
);
1721 pipeline
->active_stages
|= mesa_to_vk_shader_stage(i
);
1723 /* We don't want to alter meta shaders IR directly so clone it
1726 if (nir
[i
]->info
.name
) {
1727 nir
[i
] = nir_shader_clone(NULL
, nir
[i
]);
1730 if (first
!= last
) {
1731 nir_variable_mode mask
= 0;
1734 mask
= mask
| nir_var_shader_in
;
1737 mask
= mask
| nir_var_shader_out
;
1739 nir_lower_io_to_scalar_early(nir
[i
], mask
);
1740 radv_optimize_nir(nir
[i
]);
1744 if (nir
[MESA_SHADER_TESS_CTRL
]) {
1745 nir_lower_tes_patch_vertices(nir
[MESA_SHADER_TESS_EVAL
], nir
[MESA_SHADER_TESS_CTRL
]->info
.tess
.tcs_vertices_out
);
1746 merge_tess_info(&nir
[MESA_SHADER_TESS_EVAL
]->info
, &nir
[MESA_SHADER_TESS_CTRL
]->info
);
1749 radv_link_shaders(pipeline
, nir
);
1751 for (int i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1752 if (modules
[i
] && radv_can_dump_shader(device
, modules
[i
]))
1753 nir_print_shader(nir
[i
], stderr
);
1756 radv_fill_shader_keys(keys
, &key
, nir
);
1758 if (nir
[MESA_SHADER_FRAGMENT
]) {
1759 if (!pipeline
->shaders
[MESA_SHADER_FRAGMENT
]) {
1760 pipeline
->shaders
[MESA_SHADER_FRAGMENT
] =
1761 radv_shader_variant_create(device
, modules
[MESA_SHADER_FRAGMENT
], &nir
[MESA_SHADER_FRAGMENT
], 1,
1762 pipeline
->layout
, keys
+ MESA_SHADER_FRAGMENT
,
1763 &codes
[MESA_SHADER_FRAGMENT
], &code_sizes
[MESA_SHADER_FRAGMENT
]);
1766 /* TODO: These are no longer used as keys we should refactor this */
1767 keys
[MESA_SHADER_VERTEX
].vs
.export_prim_id
=
1768 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.prim_id_input
;
1769 keys
[MESA_SHADER_TESS_EVAL
].tes
.export_prim_id
=
1770 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.prim_id_input
;
1773 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
&& modules
[MESA_SHADER_TESS_CTRL
]) {
1774 if (!pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]) {
1775 struct nir_shader
*combined_nir
[] = {nir
[MESA_SHADER_VERTEX
], nir
[MESA_SHADER_TESS_CTRL
]};
1776 struct radv_shader_variant_key key
= keys
[MESA_SHADER_TESS_CTRL
];
1777 key
.tcs
.vs_key
= keys
[MESA_SHADER_VERTEX
].vs
;
1778 pipeline
->shaders
[MESA_SHADER_TESS_CTRL
] = radv_shader_variant_create(device
, modules
[MESA_SHADER_TESS_CTRL
], combined_nir
, 2,
1780 &key
, &codes
[MESA_SHADER_TESS_CTRL
],
1781 &code_sizes
[MESA_SHADER_TESS_CTRL
]);
1783 modules
[MESA_SHADER_VERTEX
] = NULL
;
1784 keys
[MESA_SHADER_TESS_EVAL
].tes
.num_patches
= pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.num_patches
;
1785 keys
[MESA_SHADER_TESS_EVAL
].tes
.tcs_num_outputs
= util_last_bit64(pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.info
.tcs
.outputs_written
);
1788 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
&& modules
[MESA_SHADER_GEOMETRY
]) {
1789 gl_shader_stage pre_stage
= modules
[MESA_SHADER_TESS_EVAL
] ? MESA_SHADER_TESS_EVAL
: MESA_SHADER_VERTEX
;
1790 if (!pipeline
->shaders
[MESA_SHADER_GEOMETRY
]) {
1791 struct nir_shader
*combined_nir
[] = {nir
[pre_stage
], nir
[MESA_SHADER_GEOMETRY
]};
1792 pipeline
->shaders
[MESA_SHADER_GEOMETRY
] = radv_shader_variant_create(device
, modules
[MESA_SHADER_GEOMETRY
], combined_nir
, 2,
1794 &keys
[pre_stage
] , &codes
[MESA_SHADER_GEOMETRY
],
1795 &code_sizes
[MESA_SHADER_GEOMETRY
]);
1797 modules
[pre_stage
] = NULL
;
1800 for (int i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1801 if(modules
[i
] && !pipeline
->shaders
[i
]) {
1802 if (i
== MESA_SHADER_TESS_CTRL
) {
1803 keys
[MESA_SHADER_TESS_CTRL
].tcs
.num_inputs
= util_last_bit64(pipeline
->shaders
[MESA_SHADER_VERTEX
]->info
.info
.vs
.ls_outputs_written
);
1805 if (i
== MESA_SHADER_TESS_EVAL
) {
1806 keys
[MESA_SHADER_TESS_EVAL
].tes
.num_patches
= pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.num_patches
;
1807 keys
[MESA_SHADER_TESS_EVAL
].tes
.tcs_num_outputs
= util_last_bit64(pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.info
.tcs
.outputs_written
);
1809 pipeline
->shaders
[i
] = radv_shader_variant_create(device
, modules
[i
], &nir
[i
], 1,
1811 keys
+ i
, &codes
[i
],
1816 if(modules
[MESA_SHADER_GEOMETRY
]) {
1817 void *gs_copy_code
= NULL
;
1818 unsigned gs_copy_code_size
= 0;
1819 if (!pipeline
->gs_copy_shader
) {
1820 pipeline
->gs_copy_shader
= radv_create_gs_copy_shader(
1821 device
, nir
[MESA_SHADER_GEOMETRY
], &gs_copy_code
,
1823 keys
[MESA_SHADER_GEOMETRY
].has_multiview_view_index
);
1826 if (pipeline
->gs_copy_shader
) {
1827 void *code
[MESA_SHADER_STAGES
] = {0};
1828 unsigned code_size
[MESA_SHADER_STAGES
] = {0};
1829 struct radv_shader_variant
*variants
[MESA_SHADER_STAGES
] = {0};
1831 code
[MESA_SHADER_GEOMETRY
] = gs_copy_code
;
1832 code_size
[MESA_SHADER_GEOMETRY
] = gs_copy_code_size
;
1833 variants
[MESA_SHADER_GEOMETRY
] = pipeline
->gs_copy_shader
;
1835 radv_pipeline_cache_insert_shaders(device
, cache
,
1844 radv_pipeline_cache_insert_shaders(device
, cache
, hash
, pipeline
->shaders
,
1845 (const void**)codes
, code_sizes
);
1847 for (int i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
1850 if (!pipeline
->device
->keep_shader_info
)
1851 ralloc_free(nir
[i
]);
1853 if (radv_can_dump_shader_stats(device
, modules
[i
]))
1854 radv_shader_dump_stats(device
,
1855 pipeline
->shaders
[i
],
1861 ralloc_free(fs_m
.nir
);
1865 radv_pipeline_stage_to_user_data_0(struct radv_pipeline
*pipeline
,
1866 gl_shader_stage stage
, enum chip_class chip_class
)
1868 bool has_gs
= radv_pipeline_has_gs(pipeline
);
1869 bool has_tess
= radv_pipeline_has_tess(pipeline
);
1871 case MESA_SHADER_FRAGMENT
:
1872 return R_00B030_SPI_SHADER_USER_DATA_PS_0
;
1873 case MESA_SHADER_VERTEX
:
1874 if (chip_class
>= GFX9
) {
1875 return has_tess
? R_00B430_SPI_SHADER_USER_DATA_LS_0
:
1876 has_gs
? R_00B330_SPI_SHADER_USER_DATA_ES_0
:
1877 R_00B130_SPI_SHADER_USER_DATA_VS_0
;
1880 return R_00B530_SPI_SHADER_USER_DATA_LS_0
;
1882 return has_gs
? R_00B330_SPI_SHADER_USER_DATA_ES_0
: R_00B130_SPI_SHADER_USER_DATA_VS_0
;
1883 case MESA_SHADER_GEOMETRY
:
1884 return chip_class
>= GFX9
? R_00B330_SPI_SHADER_USER_DATA_ES_0
:
1885 R_00B230_SPI_SHADER_USER_DATA_GS_0
;
1886 case MESA_SHADER_COMPUTE
:
1887 return R_00B900_COMPUTE_USER_DATA_0
;
1888 case MESA_SHADER_TESS_CTRL
:
1889 return chip_class
>= GFX9
? R_00B430_SPI_SHADER_USER_DATA_LS_0
:
1890 R_00B430_SPI_SHADER_USER_DATA_HS_0
;
1891 case MESA_SHADER_TESS_EVAL
:
1892 if (chip_class
>= GFX9
) {
1893 return has_gs
? R_00B330_SPI_SHADER_USER_DATA_ES_0
:
1894 R_00B130_SPI_SHADER_USER_DATA_VS_0
;
1897 return R_00B330_SPI_SHADER_USER_DATA_ES_0
;
1899 return R_00B130_SPI_SHADER_USER_DATA_VS_0
;
1901 unreachable("unknown shader");
1905 struct radv_bin_size_entry
{
1911 radv_compute_bin_size(struct radv_pipeline
*pipeline
, const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1913 static const struct radv_bin_size_entry color_size_table
[][3][9] = {
1917 /* One shader engine */
1923 { UINT_MAX
, { 0, 0}},
1926 /* Two shader engines */
1932 { UINT_MAX
, { 0, 0}},
1935 /* Four shader engines */
1940 { UINT_MAX
, { 0, 0}},
1946 /* One shader engine */
1952 { UINT_MAX
, { 0, 0}},
1955 /* Two shader engines */
1961 { UINT_MAX
, { 0, 0}},
1964 /* Four shader engines */
1971 { UINT_MAX
, { 0, 0}},
1977 /* One shader engine */
1984 { UINT_MAX
, { 0, 0}},
1987 /* Two shader engines */
1995 { UINT_MAX
, { 0, 0}},
1998 /* Four shader engines */
2006 { UINT_MAX
, { 0, 0}},
2010 static const struct radv_bin_size_entry ds_size_table
[][3][9] = {
2014 // One shader engine
2021 { UINT_MAX
, { 0, 0}},
2024 // Two shader engines
2032 { UINT_MAX
, { 0, 0}},
2035 // Four shader engines
2043 { UINT_MAX
, { 0, 0}},
2049 // One shader engine
2057 { UINT_MAX
, { 0, 0}},
2060 // Two shader engines
2069 { UINT_MAX
, { 0, 0}},
2072 // Four shader engines
2081 { UINT_MAX
, { 0, 0}},
2087 // One shader engine
2095 { UINT_MAX
, { 0, 0}},
2098 // Two shader engines
2107 { UINT_MAX
, { 0, 0}},
2110 // Four shader engines
2118 { UINT_MAX
, { 0, 0}},
2123 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
2124 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
2125 VkExtent2D extent
= {512, 512};
2127 unsigned log_num_rb_per_se
=
2128 util_logbase2_ceil(pipeline
->device
->physical_device
->rad_info
.num_render_backends
/
2129 pipeline
->device
->physical_device
->rad_info
.max_se
);
2130 unsigned log_num_se
= util_logbase2_ceil(pipeline
->device
->physical_device
->rad_info
.max_se
);
2132 unsigned total_samples
= 1u << G_028BE0_MSAA_NUM_SAMPLES(pipeline
->graphics
.ms
.pa_sc_mode_cntl_1
);
2133 unsigned ps_iter_samples
= 1u << G_028804_PS_ITER_SAMPLES(pipeline
->graphics
.ms
.db_eqaa
);
2134 unsigned effective_samples
= total_samples
;
2135 unsigned color_bytes_per_pixel
= 0;
2137 const VkPipelineColorBlendStateCreateInfo
*vkblend
= pCreateInfo
->pColorBlendState
;
2139 for (unsigned i
= 0; i
< subpass
->color_count
; i
++) {
2140 if (!vkblend
->pAttachments
[i
].colorWriteMask
)
2143 if (subpass
->color_attachments
[i
].attachment
== VK_ATTACHMENT_UNUSED
)
2146 VkFormat format
= pass
->attachments
[subpass
->color_attachments
[i
].attachment
].format
;
2147 color_bytes_per_pixel
+= vk_format_get_blocksize(format
);
2150 /* MSAA images typically don't use all samples all the time. */
2151 if (effective_samples
>= 2 && ps_iter_samples
<= 1)
2152 effective_samples
= 2;
2153 color_bytes_per_pixel
*= effective_samples
;
2156 const struct radv_bin_size_entry
*color_entry
= color_size_table
[log_num_rb_per_se
][log_num_se
];
2157 while(color_entry
->bpp
<= color_bytes_per_pixel
)
2160 extent
= color_entry
->extent
;
2162 if (subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
2163 struct radv_render_pass_attachment
*attachment
= pass
->attachments
+ subpass
->depth_stencil_attachment
.attachment
;
2165 /* Coefficients taken from AMDVLK */
2166 unsigned depth_coeff
= vk_format_is_depth(attachment
->format
) ? 5 : 0;
2167 unsigned stencil_coeff
= vk_format_is_stencil(attachment
->format
) ? 1 : 0;
2168 unsigned ds_bytes_per_pixel
= 4 * (depth_coeff
+ stencil_coeff
) * total_samples
;
2170 const struct radv_bin_size_entry
*ds_entry
= ds_size_table
[log_num_rb_per_se
][log_num_se
];
2171 while(ds_entry
->bpp
<= ds_bytes_per_pixel
)
2174 extent
.width
= MIN2(extent
.width
, ds_entry
->extent
.width
);
2175 extent
.height
= MIN2(extent
.height
, ds_entry
->extent
.height
);
2182 radv_pipeline_generate_binning_state(struct radeon_winsys_cs
*cs
,
2183 struct radv_pipeline
*pipeline
,
2184 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
2186 if (pipeline
->device
->physical_device
->rad_info
.chip_class
< GFX9
)
2189 uint32_t pa_sc_binner_cntl_0
=
2190 S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_LEGACY_SC
) |
2191 S_028C44_DISABLE_START_OF_PRIM(1);
2192 uint32_t db_dfsm_control
= S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF
);
2194 VkExtent2D bin_size
= radv_compute_bin_size(pipeline
, pCreateInfo
);
2196 unsigned context_states_per_bin
; /* allowed range: [1, 6] */
2197 unsigned persistent_states_per_bin
; /* allowed range: [1, 32] */
2198 unsigned fpovs_per_batch
; /* allowed range: [0, 255], 0 = unlimited */
2200 switch (pipeline
->device
->physical_device
->rad_info
.family
) {
2202 context_states_per_bin
= 1;
2203 persistent_states_per_bin
= 1;
2204 fpovs_per_batch
= 63;
2207 context_states_per_bin
= 6;
2208 persistent_states_per_bin
= 32;
2209 fpovs_per_batch
= 63;
2212 unreachable("unhandled family while determining binning state.");
2215 if (pipeline
->device
->pbb_allowed
&& bin_size
.width
&& bin_size
.height
) {
2216 pa_sc_binner_cntl_0
=
2217 S_028C44_BINNING_MODE(V_028C44_BINNING_ALLOWED
) |
2218 S_028C44_BIN_SIZE_X(bin_size
.width
== 16) |
2219 S_028C44_BIN_SIZE_Y(bin_size
.height
== 16) |
2220 S_028C44_BIN_SIZE_X_EXTEND(util_logbase2(MAX2(bin_size
.width
, 32)) - 5) |
2221 S_028C44_BIN_SIZE_Y_EXTEND(util_logbase2(MAX2(bin_size
.height
, 32)) - 5) |
2222 S_028C44_CONTEXT_STATES_PER_BIN(context_states_per_bin
- 1) |
2223 S_028C44_PERSISTENT_STATES_PER_BIN(persistent_states_per_bin
- 1) |
2224 S_028C44_DISABLE_START_OF_PRIM(1) |
2225 S_028C44_FPOVS_PER_BATCH(fpovs_per_batch
) |
2226 S_028C44_OPTIMAL_BIN_SELECTION(1);
2229 radeon_set_context_reg(cs
, R_028C44_PA_SC_BINNER_CNTL_0
,
2230 pa_sc_binner_cntl_0
);
2231 radeon_set_context_reg(cs
, R_028060_DB_DFSM_CONTROL
,
2237 radv_pipeline_generate_depth_stencil_state(struct radeon_winsys_cs
*cs
,
2238 struct radv_pipeline
*pipeline
,
2239 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
2240 const struct radv_graphics_pipeline_create_info
*extra
)
2242 const VkPipelineDepthStencilStateCreateInfo
*vkds
= pCreateInfo
->pDepthStencilState
;
2243 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
2244 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
2245 struct radv_render_pass_attachment
*attachment
= NULL
;
2246 uint32_t db_depth_control
= 0, db_stencil_control
= 0;
2247 uint32_t db_render_control
= 0, db_render_override2
= 0;
2249 if (subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
)
2250 attachment
= pass
->attachments
+ subpass
->depth_stencil_attachment
.attachment
;
2252 bool has_depth_attachment
= attachment
&& vk_format_is_depth(attachment
->format
);
2253 bool has_stencil_attachment
= attachment
&& vk_format_is_stencil(attachment
->format
);
2255 if (vkds
&& has_depth_attachment
) {
2256 db_depth_control
= S_028800_Z_ENABLE(vkds
->depthTestEnable
? 1 : 0) |
2257 S_028800_Z_WRITE_ENABLE(vkds
->depthWriteEnable
? 1 : 0) |
2258 S_028800_ZFUNC(vkds
->depthCompareOp
) |
2259 S_028800_DEPTH_BOUNDS_ENABLE(vkds
->depthBoundsTestEnable
? 1 : 0);
2261 /* from amdvlk: For 4xAA and 8xAA need to decompress on flush for better performance */
2262 db_render_override2
|= S_028010_DECOMPRESS_Z_ON_FLUSH(attachment
->samples
> 2);
2265 if (has_stencil_attachment
&& vkds
&& vkds
->stencilTestEnable
) {
2266 db_depth_control
|= S_028800_STENCIL_ENABLE(1) | S_028800_BACKFACE_ENABLE(1);
2267 db_depth_control
|= S_028800_STENCILFUNC(vkds
->front
.compareOp
);
2268 db_stencil_control
|= S_02842C_STENCILFAIL(si_translate_stencil_op(vkds
->front
.failOp
));
2269 db_stencil_control
|= S_02842C_STENCILZPASS(si_translate_stencil_op(vkds
->front
.passOp
));
2270 db_stencil_control
|= S_02842C_STENCILZFAIL(si_translate_stencil_op(vkds
->front
.depthFailOp
));
2272 db_depth_control
|= S_028800_STENCILFUNC_BF(vkds
->back
.compareOp
);
2273 db_stencil_control
|= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(vkds
->back
.failOp
));
2274 db_stencil_control
|= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(vkds
->back
.passOp
));
2275 db_stencil_control
|= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(vkds
->back
.depthFailOp
));
2278 if (attachment
&& extra
) {
2279 db_render_control
|= S_028000_DEPTH_CLEAR_ENABLE(extra
->db_depth_clear
);
2280 db_render_control
|= S_028000_STENCIL_CLEAR_ENABLE(extra
->db_stencil_clear
);
2282 db_render_control
|= S_028000_RESUMMARIZE_ENABLE(extra
->db_resummarize
);
2283 db_render_control
|= S_028000_DEPTH_COMPRESS_DISABLE(extra
->db_flush_depth_inplace
);
2284 db_render_control
|= S_028000_STENCIL_COMPRESS_DISABLE(extra
->db_flush_stencil_inplace
);
2285 db_render_override2
|= S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(extra
->db_depth_disable_expclear
);
2286 db_render_override2
|= S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(extra
->db_stencil_disable_expclear
);
2289 radeon_set_context_reg(cs
, R_028800_DB_DEPTH_CONTROL
, db_depth_control
);
2290 radeon_set_context_reg(cs
, R_02842C_DB_STENCIL_CONTROL
, db_stencil_control
);
2292 radeon_set_context_reg(cs
, R_028000_DB_RENDER_CONTROL
, db_render_control
);
2293 radeon_set_context_reg(cs
, R_028010_DB_RENDER_OVERRIDE2
, db_render_override2
);
2297 radv_pipeline_generate_blend_state(struct radeon_winsys_cs
*cs
,
2298 struct radv_pipeline
*pipeline
,
2299 const struct radv_blend_state
*blend
)
2301 radeon_set_context_reg_seq(cs
, R_028780_CB_BLEND0_CONTROL
, 8);
2302 radeon_emit_array(cs
, blend
->cb_blend_control
,
2304 radeon_set_context_reg(cs
, R_028808_CB_COLOR_CONTROL
, blend
->cb_color_control
);
2305 radeon_set_context_reg(cs
, R_028B70_DB_ALPHA_TO_MASK
, blend
->db_alpha_to_mask
);
2307 if (pipeline
->device
->physical_device
->has_rbplus
) {
2309 radeon_set_context_reg_seq(cs
, R_028760_SX_MRT0_BLEND_OPT
, 8);
2310 radeon_emit_array(cs
, blend
->sx_mrt_blend_opt
, 8);
2312 radeon_set_context_reg_seq(cs
, R_028754_SX_PS_DOWNCONVERT
, 3);
2313 radeon_emit(cs
, 0); /* R_028754_SX_PS_DOWNCONVERT */
2314 radeon_emit(cs
, 0); /* R_028758_SX_BLEND_OPT_EPSILON */
2315 radeon_emit(cs
, 0); /* R_02875C_SX_BLEND_OPT_CONTROL */
2318 radeon_set_context_reg(cs
, R_028714_SPI_SHADER_COL_FORMAT
, blend
->spi_shader_col_format
);
2320 radeon_set_context_reg(cs
, R_028238_CB_TARGET_MASK
, blend
->cb_target_mask
);
2321 radeon_set_context_reg(cs
, R_02823C_CB_SHADER_MASK
, blend
->cb_shader_mask
);
2326 radv_pipeline_generate_raster_state(struct radeon_winsys_cs
*cs
,
2327 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
2329 const VkPipelineRasterizationStateCreateInfo
*vkraster
= pCreateInfo
->pRasterizationState
;
2331 radeon_set_context_reg(cs
, R_028810_PA_CL_CLIP_CNTL
,
2332 S_028810_PS_UCP_MODE(3) |
2333 S_028810_DX_CLIP_SPACE_DEF(1) | // vulkan uses DX conventions.
2334 S_028810_ZCLIP_NEAR_DISABLE(vkraster
->depthClampEnable
? 1 : 0) |
2335 S_028810_ZCLIP_FAR_DISABLE(vkraster
->depthClampEnable
? 1 : 0) |
2336 S_028810_DX_RASTERIZATION_KILL(vkraster
->rasterizerDiscardEnable
? 1 : 0) |
2337 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1));
2339 radeon_set_context_reg(cs
, R_0286D4_SPI_INTERP_CONTROL_0
,
2340 S_0286D4_FLAT_SHADE_ENA(1) |
2341 S_0286D4_PNT_SPRITE_ENA(1) |
2342 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S
) |
2343 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T
) |
2344 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0
) |
2345 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1
) |
2346 S_0286D4_PNT_SPRITE_TOP_1(0)); /* vulkan is top to bottom - 1.0 at bottom */
2348 radeon_set_context_reg(cs
, R_028BE4_PA_SU_VTX_CNTL
,
2349 S_028BE4_PIX_CENTER(1) | // TODO verify
2350 S_028BE4_ROUND_MODE(V_028BE4_X_ROUND_TO_EVEN
) |
2351 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH
));
2353 radeon_set_context_reg(cs
, R_028814_PA_SU_SC_MODE_CNTL
,
2354 S_028814_FACE(vkraster
->frontFace
) |
2355 S_028814_CULL_FRONT(!!(vkraster
->cullMode
& VK_CULL_MODE_FRONT_BIT
)) |
2356 S_028814_CULL_BACK(!!(vkraster
->cullMode
& VK_CULL_MODE_BACK_BIT
)) |
2357 S_028814_POLY_MODE(vkraster
->polygonMode
!= VK_POLYGON_MODE_FILL
) |
2358 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(vkraster
->polygonMode
)) |
2359 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(vkraster
->polygonMode
)) |
2360 S_028814_POLY_OFFSET_FRONT_ENABLE(vkraster
->depthBiasEnable
? 1 : 0) |
2361 S_028814_POLY_OFFSET_BACK_ENABLE(vkraster
->depthBiasEnable
? 1 : 0) |
2362 S_028814_POLY_OFFSET_PARA_ENABLE(vkraster
->depthBiasEnable
? 1 : 0));
2367 radv_pipeline_generate_multisample_state(struct radeon_winsys_cs
*cs
,
2368 struct radv_pipeline
*pipeline
)
2370 struct radv_multisample_state
*ms
= &pipeline
->graphics
.ms
;
2372 radeon_set_context_reg_seq(cs
, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0
, 2);
2373 radeon_emit(cs
, ms
->pa_sc_aa_mask
[0]);
2374 radeon_emit(cs
, ms
->pa_sc_aa_mask
[1]);
2376 radeon_set_context_reg(cs
, R_028804_DB_EQAA
, ms
->db_eqaa
);
2377 radeon_set_context_reg(cs
, R_028A4C_PA_SC_MODE_CNTL_1
, ms
->pa_sc_mode_cntl_1
);
2379 if (pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.needs_sample_positions
) {
2381 struct radv_userdata_info
*loc
= radv_lookup_user_sgpr(pipeline
, MESA_SHADER_FRAGMENT
, AC_UD_PS_SAMPLE_POS_OFFSET
);
2382 uint32_t base_reg
= pipeline
->user_data_0
[MESA_SHADER_FRAGMENT
];
2383 if (loc
->sgpr_idx
== -1)
2385 assert(loc
->num_sgprs
== 1);
2386 assert(!loc
->indirect
);
2387 switch (pipeline
->graphics
.ms
.num_samples
) {
2405 radeon_set_sh_reg(cs
, base_reg
+ loc
->sgpr_idx
* 4, offset
);
2410 radv_pipeline_generate_vgt_gs_mode(struct radeon_winsys_cs
*cs
,
2411 const struct radv_pipeline
*pipeline
)
2413 const struct radv_vs_output_info
*outinfo
= get_vs_output_info(pipeline
);
2415 uint32_t vgt_primitiveid_en
= false;
2416 uint32_t vgt_gs_mode
= 0;
2418 if (radv_pipeline_has_gs(pipeline
)) {
2419 const struct radv_shader_variant
*gs
=
2420 pipeline
->shaders
[MESA_SHADER_GEOMETRY
];
2422 vgt_gs_mode
= ac_vgt_gs_mode(gs
->info
.gs
.vertices_out
,
2423 pipeline
->device
->physical_device
->rad_info
.chip_class
);
2424 } else if (outinfo
->export_prim_id
) {
2425 vgt_gs_mode
= S_028A40_MODE(V_028A40_GS_SCENARIO_A
);
2426 vgt_primitiveid_en
= true;
2429 radeon_set_context_reg(cs
, R_028A84_VGT_PRIMITIVEID_EN
, vgt_primitiveid_en
);
2430 radeon_set_context_reg(cs
, R_028A40_VGT_GS_MODE
, vgt_gs_mode
);
2434 radv_pipeline_generate_hw_vs(struct radeon_winsys_cs
*cs
,
2435 struct radv_pipeline
*pipeline
,
2436 struct radv_shader_variant
*shader
)
2438 uint64_t va
= radv_buffer_get_va(shader
->bo
) + shader
->bo_offset
;
2440 radeon_set_sh_reg_seq(cs
, R_00B120_SPI_SHADER_PGM_LO_VS
, 4);
2441 radeon_emit(cs
, va
>> 8);
2442 radeon_emit(cs
, va
>> 40);
2443 radeon_emit(cs
, shader
->rsrc1
);
2444 radeon_emit(cs
, shader
->rsrc2
);
2446 const struct radv_vs_output_info
*outinfo
= get_vs_output_info(pipeline
);
2447 unsigned clip_dist_mask
, cull_dist_mask
, total_mask
;
2448 clip_dist_mask
= outinfo
->clip_dist_mask
;
2449 cull_dist_mask
= outinfo
->cull_dist_mask
;
2450 total_mask
= clip_dist_mask
| cull_dist_mask
;
2451 bool misc_vec_ena
= outinfo
->writes_pointsize
||
2452 outinfo
->writes_layer
||
2453 outinfo
->writes_viewport_index
;
2455 radeon_set_context_reg(cs
, R_0286C4_SPI_VS_OUT_CONFIG
,
2456 S_0286C4_VS_EXPORT_COUNT(MAX2(1, outinfo
->param_exports
) - 1));
2458 radeon_set_context_reg(cs
, R_02870C_SPI_SHADER_POS_FORMAT
,
2459 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP
) |
2460 S_02870C_POS1_EXPORT_FORMAT(outinfo
->pos_exports
> 1 ?
2461 V_02870C_SPI_SHADER_4COMP
:
2462 V_02870C_SPI_SHADER_NONE
) |
2463 S_02870C_POS2_EXPORT_FORMAT(outinfo
->pos_exports
> 2 ?
2464 V_02870C_SPI_SHADER_4COMP
:
2465 V_02870C_SPI_SHADER_NONE
) |
2466 S_02870C_POS3_EXPORT_FORMAT(outinfo
->pos_exports
> 3 ?
2467 V_02870C_SPI_SHADER_4COMP
:
2468 V_02870C_SPI_SHADER_NONE
));
2470 radeon_set_context_reg(cs
, R_028818_PA_CL_VTE_CNTL
,
2471 S_028818_VTX_W0_FMT(1) |
2472 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
2473 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
2474 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
2476 radeon_set_context_reg(cs
, R_02881C_PA_CL_VS_OUT_CNTL
,
2477 S_02881C_USE_VTX_POINT_SIZE(outinfo
->writes_pointsize
) |
2478 S_02881C_USE_VTX_RENDER_TARGET_INDX(outinfo
->writes_layer
) |
2479 S_02881C_USE_VTX_VIEWPORT_INDX(outinfo
->writes_viewport_index
) |
2480 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena
) |
2481 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena
) |
2482 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask
& 0x0f) != 0) |
2483 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask
& 0xf0) != 0) |
2484 cull_dist_mask
<< 8 |
2487 if (pipeline
->device
->physical_device
->rad_info
.chip_class
<= VI
)
2488 radeon_set_context_reg(cs
, R_028AB4_VGT_REUSE_OFF
,
2489 outinfo
->writes_viewport_index
);
2493 radv_pipeline_generate_hw_es(struct radeon_winsys_cs
*cs
,
2494 struct radv_pipeline
*pipeline
,
2495 struct radv_shader_variant
*shader
)
2497 uint64_t va
= radv_buffer_get_va(shader
->bo
) + shader
->bo_offset
;
2499 radeon_set_sh_reg_seq(cs
, R_00B320_SPI_SHADER_PGM_LO_ES
, 4);
2500 radeon_emit(cs
, va
>> 8);
2501 radeon_emit(cs
, va
>> 40);
2502 radeon_emit(cs
, shader
->rsrc1
);
2503 radeon_emit(cs
, shader
->rsrc2
);
2507 radv_pipeline_generate_hw_ls(struct radeon_winsys_cs
*cs
,
2508 struct radv_pipeline
*pipeline
,
2509 struct radv_shader_variant
*shader
,
2510 const struct radv_tessellation_state
*tess
)
2512 uint64_t va
= radv_buffer_get_va(shader
->bo
) + shader
->bo_offset
;
2513 uint32_t rsrc2
= shader
->rsrc2
;
2515 radeon_set_sh_reg_seq(cs
, R_00B520_SPI_SHADER_PGM_LO_LS
, 2);
2516 radeon_emit(cs
, va
>> 8);
2517 radeon_emit(cs
, va
>> 40);
2519 rsrc2
|= S_00B52C_LDS_SIZE(tess
->lds_size
);
2520 if (pipeline
->device
->physical_device
->rad_info
.chip_class
== CIK
&&
2521 pipeline
->device
->physical_device
->rad_info
.family
!= CHIP_HAWAII
)
2522 radeon_set_sh_reg(cs
, R_00B52C_SPI_SHADER_PGM_RSRC2_LS
, rsrc2
);
2524 radeon_set_sh_reg_seq(cs
, R_00B528_SPI_SHADER_PGM_RSRC1_LS
, 2);
2525 radeon_emit(cs
, shader
->rsrc1
);
2526 radeon_emit(cs
, rsrc2
);
2530 radv_pipeline_generate_hw_hs(struct radeon_winsys_cs
*cs
,
2531 struct radv_pipeline
*pipeline
,
2532 struct radv_shader_variant
*shader
,
2533 const struct radv_tessellation_state
*tess
)
2535 uint64_t va
= radv_buffer_get_va(shader
->bo
) + shader
->bo_offset
;
2537 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2538 radeon_set_sh_reg_seq(cs
, R_00B410_SPI_SHADER_PGM_LO_LS
, 2);
2539 radeon_emit(cs
, va
>> 8);
2540 radeon_emit(cs
, va
>> 40);
2542 radeon_set_sh_reg_seq(cs
, R_00B428_SPI_SHADER_PGM_RSRC1_HS
, 2);
2543 radeon_emit(cs
, shader
->rsrc1
);
2544 radeon_emit(cs
, shader
->rsrc2
|
2545 S_00B42C_LDS_SIZE(tess
->lds_size
));
2547 radeon_set_sh_reg_seq(cs
, R_00B420_SPI_SHADER_PGM_LO_HS
, 4);
2548 radeon_emit(cs
, va
>> 8);
2549 radeon_emit(cs
, va
>> 40);
2550 radeon_emit(cs
, shader
->rsrc1
);
2551 radeon_emit(cs
, shader
->rsrc2
);
2556 radv_pipeline_generate_vertex_shader(struct radeon_winsys_cs
*cs
,
2557 struct radv_pipeline
*pipeline
,
2558 const struct radv_tessellation_state
*tess
)
2560 struct radv_shader_variant
*vs
;
2562 /* Skip shaders merged into HS/GS */
2563 vs
= pipeline
->shaders
[MESA_SHADER_VERTEX
];
2567 if (vs
->info
.vs
.as_ls
)
2568 radv_pipeline_generate_hw_ls(cs
, pipeline
, vs
, tess
);
2569 else if (vs
->info
.vs
.as_es
)
2570 radv_pipeline_generate_hw_es(cs
, pipeline
, vs
);
2572 radv_pipeline_generate_hw_vs(cs
, pipeline
, vs
);
2576 radv_pipeline_generate_tess_shaders(struct radeon_winsys_cs
*cs
,
2577 struct radv_pipeline
*pipeline
,
2578 const struct radv_tessellation_state
*tess
)
2580 if (!radv_pipeline_has_tess(pipeline
))
2583 struct radv_shader_variant
*tes
, *tcs
;
2585 tcs
= pipeline
->shaders
[MESA_SHADER_TESS_CTRL
];
2586 tes
= pipeline
->shaders
[MESA_SHADER_TESS_EVAL
];
2589 if (tes
->info
.tes
.as_es
)
2590 radv_pipeline_generate_hw_es(cs
, pipeline
, tes
);
2592 radv_pipeline_generate_hw_vs(cs
, pipeline
, tes
);
2595 radv_pipeline_generate_hw_hs(cs
, pipeline
, tcs
, tess
);
2597 radeon_set_context_reg(cs
, R_028B6C_VGT_TF_PARAM
,
2600 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= CIK
)
2601 radeon_set_context_reg_idx(cs
, R_028B58_VGT_LS_HS_CONFIG
, 2,
2602 tess
->ls_hs_config
);
2604 radeon_set_context_reg(cs
, R_028B58_VGT_LS_HS_CONFIG
,
2605 tess
->ls_hs_config
);
2609 radv_pipeline_generate_geometry_shader(struct radeon_winsys_cs
*cs
,
2610 struct radv_pipeline
*pipeline
,
2611 const struct radv_gs_state
*gs_state
)
2613 struct radv_shader_variant
*gs
;
2616 gs
= pipeline
->shaders
[MESA_SHADER_GEOMETRY
];
2620 uint32_t gsvs_itemsize
= gs
->info
.gs
.max_gsvs_emit_size
>> 2;
2622 radeon_set_context_reg_seq(cs
, R_028A60_VGT_GSVS_RING_OFFSET_1
, 3);
2623 radeon_emit(cs
, gsvs_itemsize
);
2624 radeon_emit(cs
, gsvs_itemsize
);
2625 radeon_emit(cs
, gsvs_itemsize
);
2627 radeon_set_context_reg(cs
, R_028AB0_VGT_GSVS_RING_ITEMSIZE
, gsvs_itemsize
);
2629 radeon_set_context_reg(cs
, R_028B38_VGT_GS_MAX_VERT_OUT
, gs
->info
.gs
.vertices_out
);
2631 uint32_t gs_vert_itemsize
= gs
->info
.gs
.gsvs_vertex_size
;
2632 radeon_set_context_reg_seq(cs
, R_028B5C_VGT_GS_VERT_ITEMSIZE
, 4);
2633 radeon_emit(cs
, gs_vert_itemsize
>> 2);
2638 uint32_t gs_num_invocations
= gs
->info
.gs
.invocations
;
2639 radeon_set_context_reg(cs
, R_028B90_VGT_GS_INSTANCE_CNT
,
2640 S_028B90_CNT(MIN2(gs_num_invocations
, 127)) |
2641 S_028B90_ENABLE(gs_num_invocations
> 0));
2643 radeon_set_context_reg(cs
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
2644 gs_state
->vgt_esgs_ring_itemsize
);
2646 va
= radv_buffer_get_va(gs
->bo
) + gs
->bo_offset
;
2648 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2649 radeon_set_sh_reg_seq(cs
, R_00B210_SPI_SHADER_PGM_LO_ES
, 2);
2650 radeon_emit(cs
, va
>> 8);
2651 radeon_emit(cs
, va
>> 40);
2653 radeon_set_sh_reg_seq(cs
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
, 2);
2654 radeon_emit(cs
, gs
->rsrc1
);
2655 radeon_emit(cs
, gs
->rsrc2
| S_00B22C_LDS_SIZE(gs_state
->lds_size
));
2657 radeon_set_context_reg(cs
, R_028A44_VGT_GS_ONCHIP_CNTL
, gs_state
->vgt_gs_onchip_cntl
);
2658 radeon_set_context_reg(cs
, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP
, gs_state
->vgt_gs_max_prims_per_subgroup
);
2660 radeon_set_sh_reg_seq(cs
, R_00B220_SPI_SHADER_PGM_LO_GS
, 4);
2661 radeon_emit(cs
, va
>> 8);
2662 radeon_emit(cs
, va
>> 40);
2663 radeon_emit(cs
, gs
->rsrc1
);
2664 radeon_emit(cs
, gs
->rsrc2
);
2667 radv_pipeline_generate_hw_vs(cs
, pipeline
, pipeline
->gs_copy_shader
);
2669 struct radv_userdata_info
*loc
= radv_lookup_user_sgpr(pipeline
, MESA_SHADER_GEOMETRY
,
2670 AC_UD_GS_VS_RING_STRIDE_ENTRIES
);
2671 if (loc
->sgpr_idx
!= -1) {
2672 uint32_t stride
= gs
->info
.gs
.max_gsvs_emit_size
;
2673 uint32_t num_entries
= 64;
2674 bool is_vi
= pipeline
->device
->physical_device
->rad_info
.chip_class
>= VI
;
2677 num_entries
*= stride
;
2679 stride
= S_008F04_STRIDE(stride
);
2680 radeon_set_sh_reg_seq(cs
, R_00B230_SPI_SHADER_USER_DATA_GS_0
+ loc
->sgpr_idx
* 4, 2);
2681 radeon_emit(cs
, stride
);
2682 radeon_emit(cs
, num_entries
);
2686 static uint32_t offset_to_ps_input(uint32_t offset
, bool flat_shade
)
2688 uint32_t ps_input_cntl
;
2689 if (offset
<= AC_EXP_PARAM_OFFSET_31
) {
2690 ps_input_cntl
= S_028644_OFFSET(offset
);
2692 ps_input_cntl
|= S_028644_FLAT_SHADE(1);
2694 /* The input is a DEFAULT_VAL constant. */
2695 assert(offset
>= AC_EXP_PARAM_DEFAULT_VAL_0000
&&
2696 offset
<= AC_EXP_PARAM_DEFAULT_VAL_1111
);
2697 offset
-= AC_EXP_PARAM_DEFAULT_VAL_0000
;
2698 ps_input_cntl
= S_028644_OFFSET(0x20) |
2699 S_028644_DEFAULT_VAL(offset
);
2701 return ps_input_cntl
;
2705 radv_pipeline_generate_ps_inputs(struct radeon_winsys_cs
*cs
,
2706 struct radv_pipeline
*pipeline
)
2708 struct radv_shader_variant
*ps
= pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
2709 const struct radv_vs_output_info
*outinfo
= get_vs_output_info(pipeline
);
2710 uint32_t ps_input_cntl
[32];
2712 unsigned ps_offset
= 0;
2714 if (ps
->info
.info
.ps
.prim_id_input
) {
2715 unsigned vs_offset
= outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
];
2716 if (vs_offset
!= AC_EXP_PARAM_UNDEFINED
) {
2717 ps_input_cntl
[ps_offset
] = offset_to_ps_input(vs_offset
, true);
2722 if (ps
->info
.info
.ps
.layer_input
||
2723 ps
->info
.info
.ps
.uses_input_attachments
||
2724 ps
->info
.info
.needs_multiview_view_index
) {
2725 unsigned vs_offset
= outinfo
->vs_output_param_offset
[VARYING_SLOT_LAYER
];
2726 if (vs_offset
!= AC_EXP_PARAM_UNDEFINED
)
2727 ps_input_cntl
[ps_offset
] = offset_to_ps_input(vs_offset
, true);
2729 ps_input_cntl
[ps_offset
] = offset_to_ps_input(AC_EXP_PARAM_DEFAULT_VAL_0000
, true);
2733 if (ps
->info
.info
.ps
.has_pcoord
) {
2735 val
= S_028644_PT_SPRITE_TEX(1) | S_028644_OFFSET(0x20);
2736 ps_input_cntl
[ps_offset
] = val
;
2740 for (unsigned i
= 0; i
< 32 && (1u << i
) <= ps
->info
.fs
.input_mask
; ++i
) {
2743 if (!(ps
->info
.fs
.input_mask
& (1u << i
)))
2746 vs_offset
= outinfo
->vs_output_param_offset
[VARYING_SLOT_VAR0
+ i
];
2747 if (vs_offset
== AC_EXP_PARAM_UNDEFINED
) {
2748 ps_input_cntl
[ps_offset
] = S_028644_OFFSET(0x20);
2753 flat_shade
= !!(ps
->info
.fs
.flat_shaded_mask
& (1u << ps_offset
));
2755 ps_input_cntl
[ps_offset
] = offset_to_ps_input(vs_offset
, flat_shade
);
2760 radeon_set_context_reg_seq(cs
, R_028644_SPI_PS_INPUT_CNTL_0
, ps_offset
);
2761 for (unsigned i
= 0; i
< ps_offset
; i
++) {
2762 radeon_emit(cs
, ps_input_cntl
[i
]);
2768 radv_compute_db_shader_control(const struct radv_device
*device
,
2769 const struct radv_shader_variant
*ps
)
2772 if (ps
->info
.fs
.early_fragment_test
|| !ps
->info
.info
.ps
.writes_memory
)
2773 z_order
= V_02880C_EARLY_Z_THEN_LATE_Z
;
2775 z_order
= V_02880C_LATE_Z
;
2777 return S_02880C_Z_EXPORT_ENABLE(ps
->info
.info
.ps
.writes_z
) |
2778 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(ps
->info
.info
.ps
.writes_stencil
) |
2779 S_02880C_KILL_ENABLE(!!ps
->info
.fs
.can_discard
) |
2780 S_02880C_MASK_EXPORT_ENABLE(ps
->info
.info
.ps
.writes_sample_mask
) |
2781 S_02880C_Z_ORDER(z_order
) |
2782 S_02880C_DEPTH_BEFORE_SHADER(ps
->info
.fs
.early_fragment_test
) |
2783 S_02880C_EXEC_ON_HIER_FAIL(ps
->info
.info
.ps
.writes_memory
) |
2784 S_02880C_EXEC_ON_NOOP(ps
->info
.info
.ps
.writes_memory
) |
2785 S_02880C_DUAL_QUAD_DISABLE(!!device
->physical_device
->has_rbplus
);
2789 radv_pipeline_generate_fragment_shader(struct radeon_winsys_cs
*cs
,
2790 struct radv_pipeline
*pipeline
)
2792 struct radv_shader_variant
*ps
;
2794 assert (pipeline
->shaders
[MESA_SHADER_FRAGMENT
]);
2796 ps
= pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
2797 va
= radv_buffer_get_va(ps
->bo
) + ps
->bo_offset
;
2799 radeon_set_sh_reg_seq(cs
, R_00B020_SPI_SHADER_PGM_LO_PS
, 4);
2800 radeon_emit(cs
, va
>> 8);
2801 radeon_emit(cs
, va
>> 40);
2802 radeon_emit(cs
, ps
->rsrc1
);
2803 radeon_emit(cs
, ps
->rsrc2
);
2805 radeon_set_context_reg(cs
, R_02880C_DB_SHADER_CONTROL
,
2806 radv_compute_db_shader_control(pipeline
->device
, ps
));
2808 radeon_set_context_reg(cs
, R_0286CC_SPI_PS_INPUT_ENA
,
2809 ps
->config
.spi_ps_input_ena
);
2811 radeon_set_context_reg(cs
, R_0286D0_SPI_PS_INPUT_ADDR
,
2812 ps
->config
.spi_ps_input_addr
);
2814 radeon_set_context_reg(cs
, R_0286D8_SPI_PS_IN_CONTROL
,
2815 S_0286D8_NUM_INTERP(ps
->info
.fs
.num_interp
));
2817 radeon_set_context_reg(cs
, R_0286E0_SPI_BARYC_CNTL
, pipeline
->graphics
.spi_baryc_cntl
);
2819 radeon_set_context_reg(cs
, R_028710_SPI_SHADER_Z_FORMAT
,
2820 ac_get_spi_shader_z_format(ps
->info
.info
.ps
.writes_z
,
2821 ps
->info
.info
.ps
.writes_stencil
,
2822 ps
->info
.info
.ps
.writes_sample_mask
));
2824 if (pipeline
->device
->dfsm_allowed
) {
2825 /* optimise this? */
2826 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2827 radeon_emit(cs
, EVENT_TYPE(V_028A90_FLUSH_DFSM
) | EVENT_INDEX(0));
2832 radv_pipeline_generate_vgt_vertex_reuse(struct radeon_winsys_cs
*cs
,
2833 struct radv_pipeline
*pipeline
)
2835 if (pipeline
->device
->physical_device
->rad_info
.family
< CHIP_POLARIS10
)
2838 unsigned vtx_reuse_depth
= 30;
2839 if (radv_pipeline_has_tess(pipeline
) &&
2840 radv_get_tess_eval_shader(pipeline
)->info
.tes
.spacing
== TESS_SPACING_FRACTIONAL_ODD
) {
2841 vtx_reuse_depth
= 14;
2843 radeon_set_context_reg(cs
, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL
,
2844 S_028C58_VTX_REUSE_DEPTH(vtx_reuse_depth
));
2848 radv_compute_vgt_shader_stages_en(const struct radv_pipeline
*pipeline
)
2850 uint32_t stages
= 0;
2851 if (radv_pipeline_has_tess(pipeline
)) {
2852 stages
|= S_028B54_LS_EN(V_028B54_LS_STAGE_ON
) |
2853 S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
2855 if (radv_pipeline_has_gs(pipeline
))
2856 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_DS
) |
2858 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
2860 stages
|= S_028B54_VS_EN(V_028B54_VS_STAGE_DS
);
2862 } else if (radv_pipeline_has_gs(pipeline
))
2863 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL
) |
2865 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
2867 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
2868 stages
|= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
2874 radv_compute_cliprect_rule(const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
2876 const VkPipelineDiscardRectangleStateCreateInfoEXT
*discard_rectangle_info
=
2877 vk_find_struct_const(pCreateInfo
->pNext
, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT
);
2879 if (!discard_rectangle_info
)
2884 for (unsigned i
= 0; i
< (1u << MAX_DISCARD_RECTANGLES
); ++i
) {
2885 /* Interpret i as a bitmask, and then set the bit in the mask if
2886 * that combination of rectangles in which the pixel is contained
2887 * should pass the cliprect test. */
2888 unsigned relevant_subset
= i
& ((1u << discard_rectangle_info
->discardRectangleCount
) - 1);
2890 if (discard_rectangle_info
->discardRectangleMode
== VK_DISCARD_RECTANGLE_MODE_INCLUSIVE_EXT
&&
2894 if (discard_rectangle_info
->discardRectangleMode
== VK_DISCARD_RECTANGLE_MODE_EXCLUSIVE_EXT
&&
2905 radv_pipeline_generate_pm4(struct radv_pipeline
*pipeline
,
2906 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
2907 const struct radv_graphics_pipeline_create_info
*extra
,
2908 const struct radv_blend_state
*blend
,
2909 const struct radv_tessellation_state
*tess
,
2910 const struct radv_gs_state
*gs
,
2911 unsigned prim
, unsigned gs_out
)
2913 pipeline
->cs
.buf
= malloc(4 * 256);
2914 pipeline
->cs
.max_dw
= 256;
2916 radv_pipeline_generate_depth_stencil_state(&pipeline
->cs
, pipeline
, pCreateInfo
, extra
);
2917 radv_pipeline_generate_blend_state(&pipeline
->cs
, pipeline
, blend
);
2918 radv_pipeline_generate_raster_state(&pipeline
->cs
, pCreateInfo
);
2919 radv_pipeline_generate_multisample_state(&pipeline
->cs
, pipeline
);
2920 radv_pipeline_generate_vgt_gs_mode(&pipeline
->cs
, pipeline
);
2921 radv_pipeline_generate_vertex_shader(&pipeline
->cs
, pipeline
, tess
);
2922 radv_pipeline_generate_tess_shaders(&pipeline
->cs
, pipeline
, tess
);
2923 radv_pipeline_generate_geometry_shader(&pipeline
->cs
, pipeline
, gs
);
2924 radv_pipeline_generate_fragment_shader(&pipeline
->cs
, pipeline
);
2925 radv_pipeline_generate_ps_inputs(&pipeline
->cs
, pipeline
);
2926 radv_pipeline_generate_vgt_vertex_reuse(&pipeline
->cs
, pipeline
);
2927 radv_pipeline_generate_binning_state(&pipeline
->cs
, pipeline
, pCreateInfo
);
2929 radeon_set_context_reg(&pipeline
->cs
, R_0286E8_SPI_TMPRING_SIZE
,
2930 S_0286E8_WAVES(pipeline
->max_waves
) |
2931 S_0286E8_WAVESIZE(pipeline
->scratch_bytes_per_wave
>> 10));
2933 radeon_set_context_reg(&pipeline
->cs
, R_028B54_VGT_SHADER_STAGES_EN
, radv_compute_vgt_shader_stages_en(pipeline
));
2935 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2936 radeon_set_uconfig_reg_idx(&pipeline
->cs
, R_030908_VGT_PRIMITIVE_TYPE
, 1, prim
);
2938 radeon_set_config_reg(&pipeline
->cs
, R_008958_VGT_PRIMITIVE_TYPE
, prim
);
2940 radeon_set_context_reg(&pipeline
->cs
, R_028A6C_VGT_GS_OUT_PRIM_TYPE
, gs_out
);
2942 radeon_set_context_reg(&pipeline
->cs
, R_02820C_PA_SC_CLIPRECT_RULE
, radv_compute_cliprect_rule(pCreateInfo
));
2944 assert(pipeline
->cs
.cdw
<= pipeline
->cs
.max_dw
);
2947 static struct radv_ia_multi_vgt_param_helpers
2948 radv_compute_ia_multi_vgt_param_helpers(struct radv_pipeline
*pipeline
,
2949 const struct radv_tessellation_state
*tess
,
2952 struct radv_ia_multi_vgt_param_helpers ia_multi_vgt_param
= {0};
2953 const struct radv_device
*device
= pipeline
->device
;
2955 if (radv_pipeline_has_tess(pipeline
))
2956 ia_multi_vgt_param
.primgroup_size
= tess
->num_patches
;
2957 else if (radv_pipeline_has_gs(pipeline
))
2958 ia_multi_vgt_param
.primgroup_size
= 64;
2960 ia_multi_vgt_param
.primgroup_size
= 128; /* recommended without a GS */
2962 ia_multi_vgt_param
.partial_es_wave
= false;
2963 if (pipeline
->device
->has_distributed_tess
) {
2964 if (radv_pipeline_has_gs(pipeline
)) {
2965 if (device
->physical_device
->rad_info
.chip_class
<= VI
)
2966 ia_multi_vgt_param
.partial_es_wave
= true;
2969 /* GS requirement. */
2970 if (SI_GS_PER_ES
/ ia_multi_vgt_param
.primgroup_size
>= pipeline
->device
->gs_table_depth
- 3)
2971 ia_multi_vgt_param
.partial_es_wave
= true;
2973 ia_multi_vgt_param
.wd_switch_on_eop
= false;
2974 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2975 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
2976 * 4 shader engines. Set 1 to pass the assertion below.
2977 * The other cases are hardware requirements. */
2978 if (device
->physical_device
->rad_info
.max_se
< 4 ||
2979 prim
== V_008958_DI_PT_POLYGON
||
2980 prim
== V_008958_DI_PT_LINELOOP
||
2981 prim
== V_008958_DI_PT_TRIFAN
||
2982 prim
== V_008958_DI_PT_TRISTRIP_ADJ
||
2983 (pipeline
->graphics
.prim_restart_enable
&&
2984 (device
->physical_device
->rad_info
.family
< CHIP_POLARIS10
||
2985 (prim
!= V_008958_DI_PT_POINTLIST
&&
2986 prim
!= V_008958_DI_PT_LINESTRIP
&&
2987 prim
!= V_008958_DI_PT_TRISTRIP
))))
2988 ia_multi_vgt_param
.wd_switch_on_eop
= true;
2991 ia_multi_vgt_param
.ia_switch_on_eoi
= false;
2992 if (pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.prim_id_input
)
2993 ia_multi_vgt_param
.ia_switch_on_eoi
= true;
2994 if (radv_pipeline_has_gs(pipeline
) &&
2995 pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
.info
.uses_prim_id
)
2996 ia_multi_vgt_param
.ia_switch_on_eoi
= true;
2997 if (radv_pipeline_has_tess(pipeline
)) {
2998 /* SWITCH_ON_EOI must be set if PrimID is used. */
2999 if (pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.info
.uses_prim_id
||
3000 radv_get_tess_eval_shader(pipeline
)->info
.info
.uses_prim_id
)
3001 ia_multi_vgt_param
.ia_switch_on_eoi
= true;
3004 ia_multi_vgt_param
.partial_vs_wave
= false;
3005 if (radv_pipeline_has_tess(pipeline
)) {
3006 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
3007 if ((device
->physical_device
->rad_info
.family
== CHIP_TAHITI
||
3008 device
->physical_device
->rad_info
.family
== CHIP_PITCAIRN
||
3009 device
->physical_device
->rad_info
.family
== CHIP_BONAIRE
) &&
3010 radv_pipeline_has_gs(pipeline
))
3011 ia_multi_vgt_param
.partial_vs_wave
= true;
3012 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
3013 if (device
->has_distributed_tess
) {
3014 if (radv_pipeline_has_gs(pipeline
)) {
3015 if (device
->physical_device
->rad_info
.family
== CHIP_TONGA
||
3016 device
->physical_device
->rad_info
.family
== CHIP_FIJI
||
3017 device
->physical_device
->rad_info
.family
== CHIP_POLARIS10
||
3018 device
->physical_device
->rad_info
.family
== CHIP_POLARIS11
||
3019 device
->physical_device
->rad_info
.family
== CHIP_POLARIS12
)
3020 ia_multi_vgt_param
.partial_vs_wave
= true;
3022 ia_multi_vgt_param
.partial_vs_wave
= true;
3027 ia_multi_vgt_param
.base
=
3028 S_028AA8_PRIMGROUP_SIZE(ia_multi_vgt_param
.primgroup_size
- 1) |
3029 /* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
3030 S_028AA8_MAX_PRIMGRP_IN_WAVE(device
->physical_device
->rad_info
.chip_class
== VI
? 2 : 0) |
3031 S_030960_EN_INST_OPT_BASIC(device
->physical_device
->rad_info
.chip_class
>= GFX9
) |
3032 S_030960_EN_INST_OPT_ADV(device
->physical_device
->rad_info
.chip_class
>= GFX9
);
3034 return ia_multi_vgt_param
;
3039 radv_compute_vertex_input_state(struct radv_pipeline
*pipeline
,
3040 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
3042 const VkPipelineVertexInputStateCreateInfo
*vi_info
=
3043 pCreateInfo
->pVertexInputState
;
3044 struct radv_vertex_elements_info
*velems
= &pipeline
->vertex_elements
;
3046 for (uint32_t i
= 0; i
< vi_info
->vertexAttributeDescriptionCount
; i
++) {
3047 const VkVertexInputAttributeDescription
*desc
=
3048 &vi_info
->pVertexAttributeDescriptions
[i
];
3049 unsigned loc
= desc
->location
;
3050 const struct vk_format_description
*format_desc
;
3052 uint32_t num_format
, data_format
;
3053 format_desc
= vk_format_description(desc
->format
);
3054 first_non_void
= vk_format_get_first_non_void_channel(desc
->format
);
3056 num_format
= radv_translate_buffer_numformat(format_desc
, first_non_void
);
3057 data_format
= radv_translate_buffer_dataformat(format_desc
, first_non_void
);
3059 velems
->rsrc_word3
[loc
] = S_008F0C_DST_SEL_X(si_map_swizzle(format_desc
->swizzle
[0])) |
3060 S_008F0C_DST_SEL_Y(si_map_swizzle(format_desc
->swizzle
[1])) |
3061 S_008F0C_DST_SEL_Z(si_map_swizzle(format_desc
->swizzle
[2])) |
3062 S_008F0C_DST_SEL_W(si_map_swizzle(format_desc
->swizzle
[3])) |
3063 S_008F0C_NUM_FORMAT(num_format
) |
3064 S_008F0C_DATA_FORMAT(data_format
);
3065 velems
->format_size
[loc
] = format_desc
->block
.bits
/ 8;
3066 velems
->offset
[loc
] = desc
->offset
;
3067 velems
->binding
[loc
] = desc
->binding
;
3068 velems
->count
= MAX2(velems
->count
, loc
+ 1);
3071 for (uint32_t i
= 0; i
< vi_info
->vertexBindingDescriptionCount
; i
++) {
3072 const VkVertexInputBindingDescription
*desc
=
3073 &vi_info
->pVertexBindingDescriptions
[i
];
3075 pipeline
->binding_stride
[desc
->binding
] = desc
->stride
;
3080 radv_pipeline_init(struct radv_pipeline
*pipeline
,
3081 struct radv_device
*device
,
3082 struct radv_pipeline_cache
*cache
,
3083 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
3084 const struct radv_graphics_pipeline_create_info
*extra
,
3085 const VkAllocationCallbacks
*alloc
)
3088 bool has_view_index
= false;
3090 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
3091 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
3092 if (subpass
->view_mask
)
3093 has_view_index
= true;
3095 alloc
= &device
->alloc
;
3097 pipeline
->device
= device
;
3098 pipeline
->layout
= radv_pipeline_layout_from_handle(pCreateInfo
->layout
);
3099 assert(pipeline
->layout
);
3101 struct radv_blend_state blend
= radv_pipeline_init_blend_state(pipeline
, pCreateInfo
, extra
);
3103 const VkPipelineShaderStageCreateInfo
*pStages
[MESA_SHADER_STAGES
] = { 0, };
3104 for (uint32_t i
= 0; i
< pCreateInfo
->stageCount
; i
++) {
3105 gl_shader_stage stage
= ffs(pCreateInfo
->pStages
[i
].stage
) - 1;
3106 pStages
[stage
] = &pCreateInfo
->pStages
[i
];
3109 radv_create_shaders(pipeline
, device
, cache
,
3110 radv_generate_graphics_pipeline_key(pipeline
, pCreateInfo
, &blend
, has_view_index
),
3113 pipeline
->graphics
.spi_baryc_cntl
= S_0286E0_FRONT_FACE_ALL_BITS(1);
3114 radv_pipeline_init_multisample_state(pipeline
, pCreateInfo
);
3116 uint32_t prim
= si_translate_prim(pCreateInfo
->pInputAssemblyState
->topology
);
3118 pipeline
->graphics
.can_use_guardband
= radv_prim_can_use_guardband(pCreateInfo
->pInputAssemblyState
->topology
);
3120 if (radv_pipeline_has_gs(pipeline
)) {
3121 gs_out
= si_conv_gl_prim_to_gs_out(pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
.gs
.output_prim
);
3122 pipeline
->graphics
.can_use_guardband
= gs_out
== V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
3124 gs_out
= si_conv_prim_to_gs_out(pCreateInfo
->pInputAssemblyState
->topology
);
3126 if (extra
&& extra
->use_rectlist
) {
3127 prim
= V_008958_DI_PT_RECTLIST
;
3128 gs_out
= V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
3129 pipeline
->graphics
.can_use_guardband
= true;
3131 pipeline
->graphics
.prim_restart_enable
= !!pCreateInfo
->pInputAssemblyState
->primitiveRestartEnable
;
3132 /* prim vertex count will need TESS changes */
3133 pipeline
->graphics
.prim_vertex_count
= prim_size_table
[prim
];
3135 radv_pipeline_init_dynamic_state(pipeline
, pCreateInfo
);
3137 /* Ensure that some export memory is always allocated, for two reasons:
3139 * 1) Correctness: The hardware ignores the EXEC mask if no export
3140 * memory is allocated, so KILL and alpha test do not work correctly
3142 * 2) Performance: Every shader needs at least a NULL export, even when
3143 * it writes no color/depth output. The NULL export instruction
3144 * stalls without this setting.
3146 * Don't add this to CB_SHADER_MASK.
3148 struct radv_shader_variant
*ps
= pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
3149 if (!blend
.spi_shader_col_format
) {
3150 if (!ps
->info
.info
.ps
.writes_z
&&
3151 !ps
->info
.info
.ps
.writes_stencil
&&
3152 !ps
->info
.info
.ps
.writes_sample_mask
)
3153 blend
.spi_shader_col_format
= V_028714_SPI_SHADER_32_R
;
3156 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
3157 if (pipeline
->shaders
[i
]) {
3158 pipeline
->need_indirect_descriptor_sets
|= pipeline
->shaders
[i
]->info
.need_indirect_descriptor_sets
;
3162 struct radv_gs_state gs
= {0};
3163 if (radv_pipeline_has_gs(pipeline
)) {
3164 gs
= calculate_gs_info(pCreateInfo
, pipeline
);
3165 calculate_gs_ring_sizes(pipeline
, &gs
);
3168 struct radv_tessellation_state tess
= {0};
3169 if (radv_pipeline_has_tess(pipeline
)) {
3170 if (prim
== V_008958_DI_PT_PATCH
) {
3171 pipeline
->graphics
.prim_vertex_count
.min
= pCreateInfo
->pTessellationState
->patchControlPoints
;
3172 pipeline
->graphics
.prim_vertex_count
.incr
= 1;
3174 tess
= calculate_tess_state(pipeline
, pCreateInfo
);
3177 pipeline
->graphics
.ia_multi_vgt_param
= radv_compute_ia_multi_vgt_param_helpers(pipeline
, &tess
, prim
);
3179 radv_compute_vertex_input_state(pipeline
, pCreateInfo
);
3181 for (uint32_t i
= 0; i
< MESA_SHADER_STAGES
; i
++)
3182 pipeline
->user_data_0
[i
] = radv_pipeline_stage_to_user_data_0(pipeline
, i
, device
->physical_device
->rad_info
.chip_class
);
3184 struct radv_userdata_info
*loc
= radv_lookup_user_sgpr(pipeline
, MESA_SHADER_VERTEX
,
3185 AC_UD_VS_BASE_VERTEX_START_INSTANCE
);
3186 if (loc
->sgpr_idx
!= -1) {
3187 pipeline
->graphics
.vtx_base_sgpr
= pipeline
->user_data_0
[MESA_SHADER_VERTEX
];
3188 pipeline
->graphics
.vtx_base_sgpr
+= loc
->sgpr_idx
* 4;
3189 if (radv_get_vertex_shader(pipeline
)->info
.info
.vs
.needs_draw_id
)
3190 pipeline
->graphics
.vtx_emit_num
= 3;
3192 pipeline
->graphics
.vtx_emit_num
= 2;
3195 result
= radv_pipeline_scratch_init(device
, pipeline
);
3196 radv_pipeline_generate_pm4(pipeline
, pCreateInfo
, extra
, &blend
, &tess
, &gs
, prim
, gs_out
);
3202 radv_graphics_pipeline_create(
3204 VkPipelineCache _cache
,
3205 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
3206 const struct radv_graphics_pipeline_create_info
*extra
,
3207 const VkAllocationCallbacks
*pAllocator
,
3208 VkPipeline
*pPipeline
)
3210 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3211 RADV_FROM_HANDLE(radv_pipeline_cache
, cache
, _cache
);
3212 struct radv_pipeline
*pipeline
;
3215 pipeline
= vk_zalloc2(&device
->alloc
, pAllocator
, sizeof(*pipeline
), 8,
3216 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3217 if (pipeline
== NULL
)
3218 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3220 result
= radv_pipeline_init(pipeline
, device
, cache
,
3221 pCreateInfo
, extra
, pAllocator
);
3222 if (result
!= VK_SUCCESS
) {
3223 radv_pipeline_destroy(device
, pipeline
, pAllocator
);
3227 *pPipeline
= radv_pipeline_to_handle(pipeline
);
3232 VkResult
radv_CreateGraphicsPipelines(
3234 VkPipelineCache pipelineCache
,
3236 const VkGraphicsPipelineCreateInfo
* pCreateInfos
,
3237 const VkAllocationCallbacks
* pAllocator
,
3238 VkPipeline
* pPipelines
)
3240 VkResult result
= VK_SUCCESS
;
3243 for (; i
< count
; i
++) {
3245 r
= radv_graphics_pipeline_create(_device
,
3248 NULL
, pAllocator
, &pPipelines
[i
]);
3249 if (r
!= VK_SUCCESS
) {
3251 pPipelines
[i
] = VK_NULL_HANDLE
;
3260 radv_compute_generate_pm4(struct radv_pipeline
*pipeline
)
3262 struct radv_shader_variant
*compute_shader
;
3263 struct radv_device
*device
= pipeline
->device
;
3264 unsigned compute_resource_limits
;
3265 unsigned waves_per_threadgroup
;
3268 pipeline
->cs
.buf
= malloc(20 * 4);
3269 pipeline
->cs
.max_dw
= 20;
3271 compute_shader
= pipeline
->shaders
[MESA_SHADER_COMPUTE
];
3272 va
= radv_buffer_get_va(compute_shader
->bo
) + compute_shader
->bo_offset
;
3274 radeon_set_sh_reg_seq(&pipeline
->cs
, R_00B830_COMPUTE_PGM_LO
, 2);
3275 radeon_emit(&pipeline
->cs
, va
>> 8);
3276 radeon_emit(&pipeline
->cs
, va
>> 40);
3278 radeon_set_sh_reg_seq(&pipeline
->cs
, R_00B848_COMPUTE_PGM_RSRC1
, 2);
3279 radeon_emit(&pipeline
->cs
, compute_shader
->rsrc1
);
3280 radeon_emit(&pipeline
->cs
, compute_shader
->rsrc2
);
3282 radeon_set_sh_reg(&pipeline
->cs
, R_00B860_COMPUTE_TMPRING_SIZE
,
3283 S_00B860_WAVES(pipeline
->max_waves
) |
3284 S_00B860_WAVESIZE(pipeline
->scratch_bytes_per_wave
>> 10));
3286 /* Calculate best compute resource limits. */
3287 waves_per_threadgroup
=
3288 DIV_ROUND_UP(compute_shader
->info
.cs
.block_size
[0] *
3289 compute_shader
->info
.cs
.block_size
[1] *
3290 compute_shader
->info
.cs
.block_size
[2], 64);
3291 compute_resource_limits
=
3292 S_00B854_SIMD_DEST_CNTL(waves_per_threadgroup
% 4 == 0);
3294 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
3295 unsigned num_cu_per_se
=
3296 device
->physical_device
->rad_info
.num_good_compute_units
/
3297 device
->physical_device
->rad_info
.max_se
;
3299 /* Force even distribution on all SIMDs in CU if the workgroup
3300 * size is 64. This has shown some good improvements if # of
3301 * CUs per SE is not a multiple of 4.
3303 if (num_cu_per_se
% 4 && waves_per_threadgroup
== 1)
3304 compute_resource_limits
|= S_00B854_FORCE_SIMD_DIST(1);
3307 radeon_set_sh_reg(&pipeline
->cs
, R_00B854_COMPUTE_RESOURCE_LIMITS
,
3308 compute_resource_limits
);
3310 radeon_set_sh_reg_seq(&pipeline
->cs
, R_00B81C_COMPUTE_NUM_THREAD_X
, 3);
3311 radeon_emit(&pipeline
->cs
,
3312 S_00B81C_NUM_THREAD_FULL(compute_shader
->info
.cs
.block_size
[0]));
3313 radeon_emit(&pipeline
->cs
,
3314 S_00B81C_NUM_THREAD_FULL(compute_shader
->info
.cs
.block_size
[1]));
3315 radeon_emit(&pipeline
->cs
,
3316 S_00B81C_NUM_THREAD_FULL(compute_shader
->info
.cs
.block_size
[2]));
3318 assert(pipeline
->cs
.cdw
<= pipeline
->cs
.max_dw
);
3321 static VkResult
radv_compute_pipeline_create(
3323 VkPipelineCache _cache
,
3324 const VkComputePipelineCreateInfo
* pCreateInfo
,
3325 const VkAllocationCallbacks
* pAllocator
,
3326 VkPipeline
* pPipeline
)
3328 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3329 RADV_FROM_HANDLE(radv_pipeline_cache
, cache
, _cache
);
3330 const VkPipelineShaderStageCreateInfo
*pStages
[MESA_SHADER_STAGES
] = { 0, };
3331 struct radv_pipeline
*pipeline
;
3334 pipeline
= vk_zalloc2(&device
->alloc
, pAllocator
, sizeof(*pipeline
), 8,
3335 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3336 if (pipeline
== NULL
)
3337 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3339 pipeline
->device
= device
;
3340 pipeline
->layout
= radv_pipeline_layout_from_handle(pCreateInfo
->layout
);
3341 assert(pipeline
->layout
);
3343 pStages
[MESA_SHADER_COMPUTE
] = &pCreateInfo
->stage
;
3344 radv_create_shaders(pipeline
, device
, cache
, (struct radv_pipeline_key
) {0}, pStages
);
3346 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
);
3347 pipeline
->need_indirect_descriptor_sets
|= pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.need_indirect_descriptor_sets
;
3348 result
= radv_pipeline_scratch_init(device
, pipeline
);
3349 if (result
!= VK_SUCCESS
) {
3350 radv_pipeline_destroy(device
, pipeline
, pAllocator
);
3354 radv_compute_generate_pm4(pipeline
);
3356 *pPipeline
= radv_pipeline_to_handle(pipeline
);
3361 VkResult
radv_CreateComputePipelines(
3363 VkPipelineCache pipelineCache
,
3365 const VkComputePipelineCreateInfo
* pCreateInfos
,
3366 const VkAllocationCallbacks
* pAllocator
,
3367 VkPipeline
* pPipelines
)
3369 VkResult result
= VK_SUCCESS
;
3372 for (; i
< count
; i
++) {
3374 r
= radv_compute_pipeline_create(_device
, pipelineCache
,
3376 pAllocator
, &pPipelines
[i
]);
3377 if (r
!= VK_SUCCESS
) {
3379 pPipelines
[i
] = VK_NULL_HANDLE
;