2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 #include "anv_private.h"
26 #include "genxml/gen_macros.h"
27 #include "genxml/genX_pack.h"
29 #include "common/gen_l3_config.h"
30 #include "common/gen_sample_positions.h"
32 #include "vk_format_info.h"
35 vertex_element_comp_control(enum isl_format format
, unsigned comp
)
39 case 0: bits
= isl_format_layouts
[format
].channels
.r
.bits
; break;
40 case 1: bits
= isl_format_layouts
[format
].channels
.g
.bits
; break;
41 case 2: bits
= isl_format_layouts
[format
].channels
.b
.bits
; break;
42 case 3: bits
= isl_format_layouts
[format
].channels
.a
.bits
; break;
43 default: unreachable("Invalid component");
47 * Take in account hardware restrictions when dealing with 64-bit floats.
49 * From Broadwell spec, command reference structures, page 586:
50 * "When SourceElementFormat is set to one of the *64*_PASSTHRU formats,
51 * 64-bit components are stored * in the URB without any conversion. In
52 * this case, vertex elements must be written as 128 or 256 bits, with
53 * VFCOMP_STORE_0 being used to pad the output as required. E.g., if
54 * R64_PASSTHRU is used to copy a 64-bit Red component into the URB,
55 * Component 1 must be specified as VFCOMP_STORE_0 (with Components 2,3
56 * set to VFCOMP_NOSTORE) in order to output a 128-bit vertex element, or
57 * Components 1-3 must be specified as VFCOMP_STORE_0 in order to output
58 * a 256-bit vertex element. Likewise, use of R64G64B64_PASSTHRU requires
59 * Component 3 to be specified as VFCOMP_STORE_0 in order to output a
60 * 256-bit vertex element."
63 return VFCOMP_STORE_SRC
;
64 } else if (comp
>= 2 &&
65 !isl_format_layouts
[format
].channels
.b
.bits
&&
66 isl_format_layouts
[format
].channels
.r
.type
== ISL_RAW
) {
67 /* When emitting 64-bit attributes, we need to write either 128 or 256
68 * bit chunks, using VFCOMP_NOSTORE when not writing the chunk, and
69 * VFCOMP_STORE_0 to pad the written chunk */
70 return VFCOMP_NOSTORE
;
71 } else if (comp
< 3 ||
72 isl_format_layouts
[format
].channels
.r
.type
== ISL_RAW
) {
73 /* Note we need to pad with value 0, not 1, due hardware restrictions
74 * (see comment above) */
75 return VFCOMP_STORE_0
;
76 } else if (isl_format_layouts
[format
].channels
.r
.type
== ISL_UINT
||
77 isl_format_layouts
[format
].channels
.r
.type
== ISL_SINT
) {
79 return VFCOMP_STORE_1_INT
;
82 return VFCOMP_STORE_1_FP
;
87 emit_vertex_input(struct anv_pipeline
*pipeline
,
88 const VkPipelineVertexInputStateCreateInfo
*info
)
90 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
92 /* Pull inputs_read out of the VS prog data */
93 const uint64_t inputs_read
= vs_prog_data
->inputs_read
;
94 const uint64_t double_inputs_read
= vs_prog_data
->double_inputs_read
;
95 assert((inputs_read
& ((1 << VERT_ATTRIB_GENERIC0
) - 1)) == 0);
96 const uint32_t elements
= inputs_read
>> VERT_ATTRIB_GENERIC0
;
97 const uint32_t elements_double
= double_inputs_read
>> VERT_ATTRIB_GENERIC0
;
98 const bool needs_svgs_elem
= vs_prog_data
->uses_vertexid
||
99 vs_prog_data
->uses_instanceid
||
100 vs_prog_data
->uses_basevertex
||
101 vs_prog_data
->uses_baseinstance
;
103 uint32_t elem_count
= __builtin_popcount(elements
) -
104 __builtin_popcount(elements_double
) / 2;
106 const uint32_t total_elems
=
107 elem_count
+ needs_svgs_elem
+ vs_prog_data
->uses_drawid
;
108 if (total_elems
== 0)
113 const uint32_t num_dwords
= 1 + total_elems
* 2;
114 p
= anv_batch_emitn(&pipeline
->batch
, num_dwords
,
115 GENX(3DSTATE_VERTEX_ELEMENTS
));
118 memset(p
+ 1, 0, (num_dwords
- 1) * 4);
120 for (uint32_t i
= 0; i
< info
->vertexAttributeDescriptionCount
; i
++) {
121 const VkVertexInputAttributeDescription
*desc
=
122 &info
->pVertexAttributeDescriptions
[i
];
123 enum isl_format format
= anv_get_isl_format(&pipeline
->device
->info
,
125 VK_IMAGE_ASPECT_COLOR_BIT
,
126 VK_IMAGE_TILING_LINEAR
);
128 assert(desc
->binding
< MAX_VBS
);
130 if ((elements
& (1 << desc
->location
)) == 0)
131 continue; /* Binding unused */
134 __builtin_popcount(elements
& ((1 << desc
->location
) - 1)) -
135 DIV_ROUND_UP(__builtin_popcount(elements_double
&
136 ((1 << desc
->location
) -1)), 2);
138 struct GENX(VERTEX_ELEMENT_STATE
) element
= {
139 .VertexBufferIndex
= desc
->binding
,
141 .SourceElementFormat
= format
,
142 .EdgeFlagEnable
= false,
143 .SourceElementOffset
= desc
->offset
,
144 .Component0Control
= vertex_element_comp_control(format
, 0),
145 .Component1Control
= vertex_element_comp_control(format
, 1),
146 .Component2Control
= vertex_element_comp_control(format
, 2),
147 .Component3Control
= vertex_element_comp_control(format
, 3),
149 GENX(VERTEX_ELEMENT_STATE_pack
)(NULL
, &p
[1 + slot
* 2], &element
);
152 /* On Broadwell and later, we have a separate VF_INSTANCING packet
153 * that controls instancing. On Haswell and prior, that's part of
154 * VERTEX_BUFFER_STATE which we emit later.
156 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_VF_INSTANCING
), vfi
) {
157 vfi
.InstancingEnable
= pipeline
->instancing_enable
[desc
->binding
];
158 vfi
.VertexElementIndex
= slot
;
159 /* Our implementation of VK_KHR_multiview uses instancing to draw
160 * the different views. If the client asks for instancing, we
161 * need to use the Instance Data Step Rate to ensure that we
162 * repeat the client's per-instance data once for each view.
164 vfi
.InstanceDataStepRate
= anv_subpass_view_count(pipeline
->subpass
);
169 const uint32_t id_slot
= elem_count
;
170 if (needs_svgs_elem
) {
171 /* From the Broadwell PRM for the 3D_Vertex_Component_Control enum:
172 * "Within a VERTEX_ELEMENT_STATE structure, if a Component
173 * Control field is set to something other than VFCOMP_STORE_SRC,
174 * no higher-numbered Component Control fields may be set to
177 * This means, that if we have BaseInstance, we need BaseVertex as
178 * well. Just do all or nothing.
180 uint32_t base_ctrl
= (vs_prog_data
->uses_basevertex
||
181 vs_prog_data
->uses_baseinstance
) ?
182 VFCOMP_STORE_SRC
: VFCOMP_STORE_0
;
184 struct GENX(VERTEX_ELEMENT_STATE
) element
= {
185 .VertexBufferIndex
= ANV_SVGS_VB_INDEX
,
187 .SourceElementFormat
= ISL_FORMAT_R32G32_UINT
,
188 .Component0Control
= base_ctrl
,
189 .Component1Control
= base_ctrl
,
191 .Component2Control
= VFCOMP_STORE_0
,
192 .Component3Control
= VFCOMP_STORE_0
,
194 .Component2Control
= VFCOMP_STORE_VID
,
195 .Component3Control
= VFCOMP_STORE_IID
,
198 GENX(VERTEX_ELEMENT_STATE_pack
)(NULL
, &p
[1 + id_slot
* 2], &element
);
202 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_VF_SGVS
), sgvs
) {
203 sgvs
.VertexIDEnable
= vs_prog_data
->uses_vertexid
;
204 sgvs
.VertexIDComponentNumber
= 2;
205 sgvs
.VertexIDElementOffset
= id_slot
;
206 sgvs
.InstanceIDEnable
= vs_prog_data
->uses_instanceid
;
207 sgvs
.InstanceIDComponentNumber
= 3;
208 sgvs
.InstanceIDElementOffset
= id_slot
;
212 const uint32_t drawid_slot
= elem_count
+ needs_svgs_elem
;
213 if (vs_prog_data
->uses_drawid
) {
214 struct GENX(VERTEX_ELEMENT_STATE
) element
= {
215 .VertexBufferIndex
= ANV_DRAWID_VB_INDEX
,
217 .SourceElementFormat
= ISL_FORMAT_R32_UINT
,
218 .Component0Control
= VFCOMP_STORE_SRC
,
219 .Component1Control
= VFCOMP_STORE_0
,
220 .Component2Control
= VFCOMP_STORE_0
,
221 .Component3Control
= VFCOMP_STORE_0
,
223 GENX(VERTEX_ELEMENT_STATE_pack
)(NULL
,
224 &p
[1 + drawid_slot
* 2],
228 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_VF_INSTANCING
), vfi
) {
229 vfi
.VertexElementIndex
= drawid_slot
;
236 genX(emit_urb_setup
)(struct anv_device
*device
, struct anv_batch
*batch
,
237 const struct gen_l3_config
*l3_config
,
238 VkShaderStageFlags active_stages
,
239 const unsigned entry_size
[4])
241 const struct gen_device_info
*devinfo
= &device
->info
;
243 const unsigned push_constant_kb
= devinfo
->gt
== 3 ? 32 : 16;
245 const unsigned push_constant_kb
= GEN_GEN
>= 8 ? 32 : 16;
248 const unsigned urb_size_kb
= gen_get_l3_config_urb_size(devinfo
, l3_config
);
252 gen_get_urb_config(devinfo
,
253 1024 * push_constant_kb
, 1024 * urb_size_kb
,
255 VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT
,
256 active_stages
& VK_SHADER_STAGE_GEOMETRY_BIT
,
257 entry_size
, entries
, start
);
259 #if GEN_GEN == 7 && !GEN_IS_HASWELL
260 /* From the IVB PRM Vol. 2, Part 1, Section 3.2.1:
262 * "A PIPE_CONTROL with Post-Sync Operation set to 1h and a depth stall
263 * needs to be sent just prior to any 3DSTATE_VS, 3DSTATE_URB_VS,
264 * 3DSTATE_CONSTANT_VS, 3DSTATE_BINDING_TABLE_POINTER_VS,
265 * 3DSTATE_SAMPLER_STATE_POINTER_VS command. Only one PIPE_CONTROL
266 * needs to be sent before any combination of VS associated 3DSTATE."
268 anv_batch_emit(batch
, GEN7_PIPE_CONTROL
, pc
) {
269 pc
.DepthStallEnable
= true;
270 pc
.PostSyncOperation
= WriteImmediateData
;
271 pc
.Address
= (struct anv_address
) { &device
->workaround_bo
, 0 };
275 for (int i
= 0; i
<= MESA_SHADER_GEOMETRY
; i
++) {
276 anv_batch_emit(batch
, GENX(3DSTATE_URB_VS
), urb
) {
277 urb
._3DCommandSubOpcode
+= i
;
278 urb
.VSURBStartingAddress
= start
[i
];
279 urb
.VSURBEntryAllocationSize
= entry_size
[i
] - 1;
280 urb
.VSNumberofURBEntries
= entries
[i
];
286 emit_urb_setup(struct anv_pipeline
*pipeline
)
288 unsigned entry_size
[4];
289 for (int i
= MESA_SHADER_VERTEX
; i
<= MESA_SHADER_GEOMETRY
; i
++) {
290 const struct brw_vue_prog_data
*prog_data
=
291 !anv_pipeline_has_stage(pipeline
, i
) ? NULL
:
292 (const struct brw_vue_prog_data
*) pipeline
->shaders
[i
]->prog_data
;
294 entry_size
[i
] = prog_data
? prog_data
->urb_entry_size
: 1;
297 genX(emit_urb_setup
)(pipeline
->device
, &pipeline
->batch
,
298 pipeline
->urb
.l3_config
,
299 pipeline
->active_stages
, entry_size
);
303 emit_3dstate_sbe(struct anv_pipeline
*pipeline
)
305 const struct brw_wm_prog_data
*wm_prog_data
= get_wm_prog_data(pipeline
);
307 if (!anv_pipeline_has_stage(pipeline
, MESA_SHADER_FRAGMENT
)) {
308 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_SBE
), sbe
);
310 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_SBE_SWIZ
), sbe
);
315 const struct brw_vue_map
*fs_input_map
=
316 &anv_pipeline_get_last_vue_prog_data(pipeline
)->vue_map
;
318 struct GENX(3DSTATE_SBE
) sbe
= {
319 GENX(3DSTATE_SBE_header
),
320 .AttributeSwizzleEnable
= true,
321 .PointSpriteTextureCoordinateOrigin
= UPPERLEFT
,
322 .NumberofSFOutputAttributes
= wm_prog_data
->num_varying_inputs
,
323 .ConstantInterpolationEnable
= wm_prog_data
->flat_inputs
,
327 for (unsigned i
= 0; i
< 32; i
++)
328 sbe
.AttributeActiveComponentFormat
[i
] = ACF_XYZW
;
332 /* On Broadwell, they broke 3DSTATE_SBE into two packets */
333 struct GENX(3DSTATE_SBE_SWIZ
) swiz
= {
334 GENX(3DSTATE_SBE_SWIZ_header
),
340 /* Skip the VUE header and position slots by default */
341 unsigned urb_entry_read_offset
= 1;
342 int max_source_attr
= 0;
343 for (int attr
= 0; attr
< VARYING_SLOT_MAX
; attr
++) {
344 int input_index
= wm_prog_data
->urb_setup
[attr
];
349 /* gl_Layer is stored in the VUE header */
350 if (attr
== VARYING_SLOT_LAYER
) {
351 urb_entry_read_offset
= 0;
355 if (attr
== VARYING_SLOT_PNTC
) {
356 sbe
.PointSpriteTextureCoordinateEnable
= 1 << input_index
;
360 const int slot
= fs_input_map
->varying_to_slot
[attr
];
362 if (input_index
>= 16)
366 /* This attribute does not exist in the VUE--that means that the
367 * vertex shader did not write to it. It could be that it's a
368 * regular varying read by the fragment shader but not written by
369 * the vertex shader or it's gl_PrimitiveID. In the first case the
370 * value is undefined, in the second it needs to be
373 swiz
.Attribute
[input_index
].ConstantSource
= PRIM_ID
;
374 swiz
.Attribute
[input_index
].ComponentOverrideX
= true;
375 swiz
.Attribute
[input_index
].ComponentOverrideY
= true;
376 swiz
.Attribute
[input_index
].ComponentOverrideZ
= true;
377 swiz
.Attribute
[input_index
].ComponentOverrideW
= true;
379 /* We have to subtract two slots to accout for the URB entry output
380 * read offset in the VS and GS stages.
382 const int source_attr
= slot
- 2 * urb_entry_read_offset
;
383 assert(source_attr
>= 0 && source_attr
< 32);
384 max_source_attr
= MAX2(max_source_attr
, source_attr
);
385 swiz
.Attribute
[input_index
].SourceAttribute
= source_attr
;
389 sbe
.VertexURBEntryReadOffset
= urb_entry_read_offset
;
390 sbe
.VertexURBEntryReadLength
= DIV_ROUND_UP(max_source_attr
+ 1, 2);
392 sbe
.ForceVertexURBEntryReadOffset
= true;
393 sbe
.ForceVertexURBEntryReadLength
= true;
396 uint32_t *dw
= anv_batch_emit_dwords(&pipeline
->batch
,
397 GENX(3DSTATE_SBE_length
));
400 GENX(3DSTATE_SBE_pack
)(&pipeline
->batch
, dw
, &sbe
);
403 dw
= anv_batch_emit_dwords(&pipeline
->batch
, GENX(3DSTATE_SBE_SWIZ_length
));
406 GENX(3DSTATE_SBE_SWIZ_pack
)(&pipeline
->batch
, dw
, &swiz
);
410 static const uint32_t vk_to_gen_cullmode
[] = {
411 [VK_CULL_MODE_NONE
] = CULLMODE_NONE
,
412 [VK_CULL_MODE_FRONT_BIT
] = CULLMODE_FRONT
,
413 [VK_CULL_MODE_BACK_BIT
] = CULLMODE_BACK
,
414 [VK_CULL_MODE_FRONT_AND_BACK
] = CULLMODE_BOTH
417 static const uint32_t vk_to_gen_fillmode
[] = {
418 [VK_POLYGON_MODE_FILL
] = FILL_MODE_SOLID
,
419 [VK_POLYGON_MODE_LINE
] = FILL_MODE_WIREFRAME
,
420 [VK_POLYGON_MODE_POINT
] = FILL_MODE_POINT
,
423 static const uint32_t vk_to_gen_front_face
[] = {
424 [VK_FRONT_FACE_COUNTER_CLOCKWISE
] = 1,
425 [VK_FRONT_FACE_CLOCKWISE
] = 0
429 emit_rs_state(struct anv_pipeline
*pipeline
,
430 const VkPipelineRasterizationStateCreateInfo
*rs_info
,
431 const VkPipelineMultisampleStateCreateInfo
*ms_info
,
432 const struct anv_render_pass
*pass
,
433 const struct anv_subpass
*subpass
)
435 struct GENX(3DSTATE_SF
) sf
= {
436 GENX(3DSTATE_SF_header
),
439 sf
.ViewportTransformEnable
= true;
440 sf
.StatisticsEnable
= true;
441 sf
.TriangleStripListProvokingVertexSelect
= 0;
442 sf
.LineStripListProvokingVertexSelect
= 0;
443 sf
.TriangleFanProvokingVertexSelect
= 1;
445 const struct brw_vue_prog_data
*last_vue_prog_data
=
446 anv_pipeline_get_last_vue_prog_data(pipeline
);
448 if (last_vue_prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) {
449 sf
.PointWidthSource
= Vertex
;
451 sf
.PointWidthSource
= State
;
456 struct GENX(3DSTATE_RASTER
) raster
= {
457 GENX(3DSTATE_RASTER_header
),
463 /* For details on 3DSTATE_RASTER multisample state, see the BSpec table
464 * "Multisample Modes State".
467 raster
.DXMultisampleRasterizationEnable
= true;
468 /* NOTE: 3DSTATE_RASTER::ForcedSampleCount affects the BDW and SKL PMA fix
469 * computations. If we ever set this bit to a different value, they will
470 * need to be updated accordingly.
472 raster
.ForcedSampleCount
= FSC_NUMRASTSAMPLES_0
;
473 raster
.ForceMultisampling
= false;
475 raster
.MultisampleRasterizationMode
=
476 (ms_info
&& ms_info
->rasterizationSamples
> 1) ?
477 MSRASTMODE_ON_PATTERN
: MSRASTMODE_OFF_PIXEL
;
480 raster
.FrontWinding
= vk_to_gen_front_face
[rs_info
->frontFace
];
481 raster
.CullMode
= vk_to_gen_cullmode
[rs_info
->cullMode
];
482 raster
.FrontFaceFillMode
= vk_to_gen_fillmode
[rs_info
->polygonMode
];
483 raster
.BackFaceFillMode
= vk_to_gen_fillmode
[rs_info
->polygonMode
];
484 raster
.ScissorRectangleEnable
= true;
487 /* GEN9+ splits ViewportZClipTestEnable into near and far enable bits */
488 raster
.ViewportZFarClipTestEnable
= !pipeline
->depth_clamp_enable
;
489 raster
.ViewportZNearClipTestEnable
= !pipeline
->depth_clamp_enable
;
491 raster
.ViewportZClipTestEnable
= !pipeline
->depth_clamp_enable
;
494 raster
.GlobalDepthOffsetEnableSolid
= rs_info
->depthBiasEnable
;
495 raster
.GlobalDepthOffsetEnableWireframe
= rs_info
->depthBiasEnable
;
496 raster
.GlobalDepthOffsetEnablePoint
= rs_info
->depthBiasEnable
;
499 /* Gen7 requires that we provide the depth format in 3DSTATE_SF so that it
500 * can get the depth offsets correct.
502 if (subpass
->depth_stencil_attachment
.attachment
< pass
->attachment_count
) {
504 pass
->attachments
[subpass
->depth_stencil_attachment
.attachment
].format
;
505 assert(vk_format_is_depth_or_stencil(vk_format
));
506 if (vk_format_aspects(vk_format
) & VK_IMAGE_ASPECT_DEPTH_BIT
) {
507 enum isl_format isl_format
=
508 anv_get_isl_format(&pipeline
->device
->info
, vk_format
,
509 VK_IMAGE_ASPECT_DEPTH_BIT
,
510 VK_IMAGE_TILING_OPTIMAL
);
511 sf
.DepthBufferSurfaceFormat
=
512 isl_format_get_depth_format(isl_format
, false);
518 GENX(3DSTATE_SF_pack
)(NULL
, pipeline
->gen8
.sf
, &sf
);
519 GENX(3DSTATE_RASTER_pack
)(NULL
, pipeline
->gen8
.raster
, &raster
);
522 GENX(3DSTATE_SF_pack
)(NULL
, &pipeline
->gen7
.sf
, &sf
);
527 emit_ms_state(struct anv_pipeline
*pipeline
,
528 const VkPipelineMultisampleStateCreateInfo
*info
)
530 uint32_t samples
= 1;
531 uint32_t log2_samples
= 0;
533 /* From the Vulkan 1.0 spec:
534 * If pSampleMask is NULL, it is treated as if the mask has all bits
535 * enabled, i.e. no coverage is removed from fragments.
537 * 3DSTATE_SAMPLE_MASK.SampleMask is 16 bits.
540 uint32_t sample_mask
= 0xffff;
542 uint32_t sample_mask
= 0xff;
546 samples
= info
->rasterizationSamples
;
547 log2_samples
= __builtin_ffs(samples
) - 1;
550 if (info
&& info
->pSampleMask
)
551 sample_mask
&= info
->pSampleMask
[0];
553 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_MULTISAMPLE
), ms
) {
554 ms
.NumberofMultisamples
= log2_samples
;
556 ms
.PixelLocation
= CENTER
;
558 /* The PRM says that this bit is valid only for DX9:
560 * SW can choose to set this bit only for DX9 API. DX10/OGL API's
561 * should not have any effect by setting or not setting this bit.
563 ms
.PixelPositionOffsetEnable
= false;
568 GEN_SAMPLE_POS_1X(ms
.Sample
);
571 GEN_SAMPLE_POS_2X(ms
.Sample
);
574 GEN_SAMPLE_POS_4X(ms
.Sample
);
577 GEN_SAMPLE_POS_8X(ms
.Sample
);
585 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_SAMPLE_MASK
), sm
) {
586 sm
.SampleMask
= sample_mask
;
590 static const uint32_t vk_to_gen_logic_op
[] = {
591 [VK_LOGIC_OP_COPY
] = LOGICOP_COPY
,
592 [VK_LOGIC_OP_CLEAR
] = LOGICOP_CLEAR
,
593 [VK_LOGIC_OP_AND
] = LOGICOP_AND
,
594 [VK_LOGIC_OP_AND_REVERSE
] = LOGICOP_AND_REVERSE
,
595 [VK_LOGIC_OP_AND_INVERTED
] = LOGICOP_AND_INVERTED
,
596 [VK_LOGIC_OP_NO_OP
] = LOGICOP_NOOP
,
597 [VK_LOGIC_OP_XOR
] = LOGICOP_XOR
,
598 [VK_LOGIC_OP_OR
] = LOGICOP_OR
,
599 [VK_LOGIC_OP_NOR
] = LOGICOP_NOR
,
600 [VK_LOGIC_OP_EQUIVALENT
] = LOGICOP_EQUIV
,
601 [VK_LOGIC_OP_INVERT
] = LOGICOP_INVERT
,
602 [VK_LOGIC_OP_OR_REVERSE
] = LOGICOP_OR_REVERSE
,
603 [VK_LOGIC_OP_COPY_INVERTED
] = LOGICOP_COPY_INVERTED
,
604 [VK_LOGIC_OP_OR_INVERTED
] = LOGICOP_OR_INVERTED
,
605 [VK_LOGIC_OP_NAND
] = LOGICOP_NAND
,
606 [VK_LOGIC_OP_SET
] = LOGICOP_SET
,
609 static const uint32_t vk_to_gen_blend
[] = {
610 [VK_BLEND_FACTOR_ZERO
] = BLENDFACTOR_ZERO
,
611 [VK_BLEND_FACTOR_ONE
] = BLENDFACTOR_ONE
,
612 [VK_BLEND_FACTOR_SRC_COLOR
] = BLENDFACTOR_SRC_COLOR
,
613 [VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR
] = BLENDFACTOR_INV_SRC_COLOR
,
614 [VK_BLEND_FACTOR_DST_COLOR
] = BLENDFACTOR_DST_COLOR
,
615 [VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR
] = BLENDFACTOR_INV_DST_COLOR
,
616 [VK_BLEND_FACTOR_SRC_ALPHA
] = BLENDFACTOR_SRC_ALPHA
,
617 [VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
] = BLENDFACTOR_INV_SRC_ALPHA
,
618 [VK_BLEND_FACTOR_DST_ALPHA
] = BLENDFACTOR_DST_ALPHA
,
619 [VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA
] = BLENDFACTOR_INV_DST_ALPHA
,
620 [VK_BLEND_FACTOR_CONSTANT_COLOR
] = BLENDFACTOR_CONST_COLOR
,
621 [VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR
]= BLENDFACTOR_INV_CONST_COLOR
,
622 [VK_BLEND_FACTOR_CONSTANT_ALPHA
] = BLENDFACTOR_CONST_ALPHA
,
623 [VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA
]= BLENDFACTOR_INV_CONST_ALPHA
,
624 [VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
] = BLENDFACTOR_SRC_ALPHA_SATURATE
,
625 [VK_BLEND_FACTOR_SRC1_COLOR
] = BLENDFACTOR_SRC1_COLOR
,
626 [VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
] = BLENDFACTOR_INV_SRC1_COLOR
,
627 [VK_BLEND_FACTOR_SRC1_ALPHA
] = BLENDFACTOR_SRC1_ALPHA
,
628 [VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
] = BLENDFACTOR_INV_SRC1_ALPHA
,
631 static const uint32_t vk_to_gen_blend_op
[] = {
632 [VK_BLEND_OP_ADD
] = BLENDFUNCTION_ADD
,
633 [VK_BLEND_OP_SUBTRACT
] = BLENDFUNCTION_SUBTRACT
,
634 [VK_BLEND_OP_REVERSE_SUBTRACT
] = BLENDFUNCTION_REVERSE_SUBTRACT
,
635 [VK_BLEND_OP_MIN
] = BLENDFUNCTION_MIN
,
636 [VK_BLEND_OP_MAX
] = BLENDFUNCTION_MAX
,
639 static const uint32_t vk_to_gen_compare_op
[] = {
640 [VK_COMPARE_OP_NEVER
] = PREFILTEROPNEVER
,
641 [VK_COMPARE_OP_LESS
] = PREFILTEROPLESS
,
642 [VK_COMPARE_OP_EQUAL
] = PREFILTEROPEQUAL
,
643 [VK_COMPARE_OP_LESS_OR_EQUAL
] = PREFILTEROPLEQUAL
,
644 [VK_COMPARE_OP_GREATER
] = PREFILTEROPGREATER
,
645 [VK_COMPARE_OP_NOT_EQUAL
] = PREFILTEROPNOTEQUAL
,
646 [VK_COMPARE_OP_GREATER_OR_EQUAL
] = PREFILTEROPGEQUAL
,
647 [VK_COMPARE_OP_ALWAYS
] = PREFILTEROPALWAYS
,
650 static const uint32_t vk_to_gen_stencil_op
[] = {
651 [VK_STENCIL_OP_KEEP
] = STENCILOP_KEEP
,
652 [VK_STENCIL_OP_ZERO
] = STENCILOP_ZERO
,
653 [VK_STENCIL_OP_REPLACE
] = STENCILOP_REPLACE
,
654 [VK_STENCIL_OP_INCREMENT_AND_CLAMP
] = STENCILOP_INCRSAT
,
655 [VK_STENCIL_OP_DECREMENT_AND_CLAMP
] = STENCILOP_DECRSAT
,
656 [VK_STENCIL_OP_INVERT
] = STENCILOP_INVERT
,
657 [VK_STENCIL_OP_INCREMENT_AND_WRAP
] = STENCILOP_INCR
,
658 [VK_STENCIL_OP_DECREMENT_AND_WRAP
] = STENCILOP_DECR
,
661 /* This function sanitizes the VkStencilOpState by looking at the compare ops
662 * and trying to determine whether or not a given stencil op can ever actually
663 * occur. Stencil ops which can never occur are set to VK_STENCIL_OP_KEEP.
664 * This function returns true if, after sanitation, any of the stencil ops are
665 * set to something other than VK_STENCIL_OP_KEEP.
668 sanitize_stencil_face(VkStencilOpState
*face
,
669 VkCompareOp depthCompareOp
)
671 /* If compareOp is ALWAYS then the stencil test will never fail and failOp
672 * will never happen. Set failOp to KEEP in this case.
674 if (face
->compareOp
== VK_COMPARE_OP_ALWAYS
)
675 face
->failOp
= VK_STENCIL_OP_KEEP
;
677 /* If compareOp is NEVER or depthCompareOp is NEVER then one of the depth
678 * or stencil tests will fail and passOp will never happen.
680 if (face
->compareOp
== VK_COMPARE_OP_NEVER
||
681 depthCompareOp
== VK_COMPARE_OP_NEVER
)
682 face
->passOp
= VK_STENCIL_OP_KEEP
;
684 /* If compareOp is NEVER or depthCompareOp is ALWAYS then either the
685 * stencil test will fail or the depth test will pass. In either case,
686 * depthFailOp will never happen.
688 if (face
->compareOp
== VK_COMPARE_OP_NEVER
||
689 depthCompareOp
== VK_COMPARE_OP_ALWAYS
)
690 face
->depthFailOp
= VK_STENCIL_OP_KEEP
;
692 return face
->failOp
!= VK_STENCIL_OP_KEEP
||
693 face
->depthFailOp
!= VK_STENCIL_OP_KEEP
||
694 face
->passOp
!= VK_STENCIL_OP_KEEP
;
697 /* Intel hardware is fairly sensitive to whether or not depth/stencil writes
698 * are enabled. In the presence of discards, it's fairly easy to get into the
699 * non-promoted case which means a fairly big performance hit. From the Iron
700 * Lake PRM, Vol 2, pt. 1, section 8.4.3.2, "Early Depth Test Cases":
702 * "Non-promoted depth (N) is active whenever the depth test can be done
703 * early but it cannot determine whether or not to write source depth to
704 * the depth buffer, therefore the depth write must be performed post pixel
705 * shader. This includes cases where the pixel shader can kill pixels,
706 * including via sampler chroma key, as well as cases where the alpha test
707 * function is enabled, which kills pixels based on a programmable alpha
708 * test. In this case, even if the depth test fails, the pixel cannot be
709 * killed if a stencil write is indicated. Whether or not the stencil write
710 * happens depends on whether or not the pixel is killed later. In these
711 * cases if stencil test fails and stencil writes are off, the pixels can
712 * also be killed early. If stencil writes are enabled, the pixels must be
713 * treated as Computed depth (described above)."
715 * The same thing as mentioned in the stencil case can happen in the depth
716 * case as well if it thinks it writes depth but, thanks to the depth test
717 * being GL_EQUAL, the write doesn't actually matter. A little extra work
718 * up-front to try and disable depth and stencil writes can make a big
721 * Unfortunately, the way depth and stencil testing is specified, there are
722 * many case where, regardless of depth/stencil writes being enabled, nothing
723 * actually gets written due to some other bit of state being set. This
724 * function attempts to "sanitize" the depth stencil state and disable writes
725 * and sometimes even testing whenever possible.
728 sanitize_ds_state(VkPipelineDepthStencilStateCreateInfo
*state
,
729 bool *stencilWriteEnable
,
730 VkImageAspectFlags ds_aspects
)
732 *stencilWriteEnable
= state
->stencilTestEnable
;
734 /* If the depth test is disabled, we won't be writing anything. */
735 if (!state
->depthTestEnable
)
736 state
->depthWriteEnable
= false;
738 /* The Vulkan spec requires that if either depth or stencil is not present,
739 * the pipeline is to act as if the test silently passes.
741 if (!(ds_aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
)) {
742 state
->depthWriteEnable
= false;
743 state
->depthCompareOp
= VK_COMPARE_OP_ALWAYS
;
746 if (!(ds_aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
)) {
747 *stencilWriteEnable
= false;
748 state
->front
.compareOp
= VK_COMPARE_OP_ALWAYS
;
749 state
->back
.compareOp
= VK_COMPARE_OP_ALWAYS
;
752 /* If the stencil test is enabled and always fails, then we will never get
753 * to the depth test so we can just disable the depth test entirely.
755 if (state
->stencilTestEnable
&&
756 state
->front
.compareOp
== VK_COMPARE_OP_NEVER
&&
757 state
->back
.compareOp
== VK_COMPARE_OP_NEVER
) {
758 state
->depthTestEnable
= false;
759 state
->depthWriteEnable
= false;
762 /* If depthCompareOp is EQUAL then the value we would be writing to the
763 * depth buffer is the same as the value that's already there so there's no
764 * point in writing it.
766 if (state
->depthCompareOp
== VK_COMPARE_OP_EQUAL
)
767 state
->depthWriteEnable
= false;
769 /* If the stencil ops are such that we don't actually ever modify the
770 * stencil buffer, we should disable writes.
772 if (!sanitize_stencil_face(&state
->front
, state
->depthCompareOp
) &&
773 !sanitize_stencil_face(&state
->back
, state
->depthCompareOp
))
774 *stencilWriteEnable
= false;
776 /* If the depth test always passes and we never write out depth, that's the
777 * same as if the depth test is disabled entirely.
779 if (state
->depthCompareOp
== VK_COMPARE_OP_ALWAYS
&&
780 !state
->depthWriteEnable
)
781 state
->depthTestEnable
= false;
783 /* If the stencil test always passes and we never write out stencil, that's
784 * the same as if the stencil test is disabled entirely.
786 if (state
->front
.compareOp
== VK_COMPARE_OP_ALWAYS
&&
787 state
->back
.compareOp
== VK_COMPARE_OP_ALWAYS
&&
788 !*stencilWriteEnable
)
789 state
->stencilTestEnable
= false;
793 emit_ds_state(struct anv_pipeline
*pipeline
,
794 const VkPipelineDepthStencilStateCreateInfo
*pCreateInfo
,
795 const struct anv_render_pass
*pass
,
796 const struct anv_subpass
*subpass
)
799 # define depth_stencil_dw pipeline->gen7.depth_stencil_state
801 # define depth_stencil_dw pipeline->gen8.wm_depth_stencil
803 # define depth_stencil_dw pipeline->gen9.wm_depth_stencil
806 if (pCreateInfo
== NULL
) {
807 /* We're going to OR this together with the dynamic state. We need
808 * to make sure it's initialized to something useful.
810 pipeline
->writes_stencil
= false;
811 pipeline
->stencil_test_enable
= false;
812 pipeline
->writes_depth
= false;
813 pipeline
->depth_test_enable
= false;
814 memset(depth_stencil_dw
, 0, sizeof(depth_stencil_dw
));
818 VkImageAspectFlags ds_aspects
= 0;
819 if (subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
820 VkFormat depth_stencil_format
=
821 pass
->attachments
[subpass
->depth_stencil_attachment
.attachment
].format
;
822 ds_aspects
= vk_format_aspects(depth_stencil_format
);
825 VkPipelineDepthStencilStateCreateInfo info
= *pCreateInfo
;
826 sanitize_ds_state(&info
, &pipeline
->writes_stencil
, ds_aspects
);
827 pipeline
->stencil_test_enable
= info
.stencilTestEnable
;
828 pipeline
->writes_depth
= info
.depthWriteEnable
;
829 pipeline
->depth_test_enable
= info
.depthTestEnable
;
831 /* VkBool32 depthBoundsTestEnable; // optional (depth_bounds_test) */
834 struct GENX(DEPTH_STENCIL_STATE
) depth_stencil
= {
836 struct GENX(3DSTATE_WM_DEPTH_STENCIL
) depth_stencil
= {
838 .DepthTestEnable
= info
.depthTestEnable
,
839 .DepthBufferWriteEnable
= info
.depthWriteEnable
,
840 .DepthTestFunction
= vk_to_gen_compare_op
[info
.depthCompareOp
],
841 .DoubleSidedStencilEnable
= true,
843 .StencilTestEnable
= info
.stencilTestEnable
,
844 .StencilFailOp
= vk_to_gen_stencil_op
[info
.front
.failOp
],
845 .StencilPassDepthPassOp
= vk_to_gen_stencil_op
[info
.front
.passOp
],
846 .StencilPassDepthFailOp
= vk_to_gen_stencil_op
[info
.front
.depthFailOp
],
847 .StencilTestFunction
= vk_to_gen_compare_op
[info
.front
.compareOp
],
848 .BackfaceStencilFailOp
= vk_to_gen_stencil_op
[info
.back
.failOp
],
849 .BackfaceStencilPassDepthPassOp
= vk_to_gen_stencil_op
[info
.back
.passOp
],
850 .BackfaceStencilPassDepthFailOp
=vk_to_gen_stencil_op
[info
.back
.depthFailOp
],
851 .BackfaceStencilTestFunction
= vk_to_gen_compare_op
[info
.back
.compareOp
],
855 GENX(DEPTH_STENCIL_STATE_pack
)(NULL
, depth_stencil_dw
, &depth_stencil
);
857 GENX(3DSTATE_WM_DEPTH_STENCIL_pack
)(NULL
, depth_stencil_dw
, &depth_stencil
);
862 emit_cb_state(struct anv_pipeline
*pipeline
,
863 const VkPipelineColorBlendStateCreateInfo
*info
,
864 const VkPipelineMultisampleStateCreateInfo
*ms_info
)
866 struct anv_device
*device
= pipeline
->device
;
869 struct GENX(BLEND_STATE
) blend_state
= {
871 .AlphaToCoverageEnable
= ms_info
&& ms_info
->alphaToCoverageEnable
,
872 .AlphaToOneEnable
= ms_info
&& ms_info
->alphaToOneEnable
,
876 uint32_t surface_count
= 0;
877 struct anv_pipeline_bind_map
*map
;
878 if (anv_pipeline_has_stage(pipeline
, MESA_SHADER_FRAGMENT
)) {
879 map
= &pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->bind_map
;
880 surface_count
= map
->surface_count
;
883 const uint32_t num_dwords
= GENX(BLEND_STATE_length
) +
884 GENX(BLEND_STATE_ENTRY_length
) * surface_count
;
885 pipeline
->blend_state
=
886 anv_state_pool_alloc(&device
->dynamic_state_pool
, num_dwords
* 4, 64);
888 bool has_writeable_rt
= false;
889 uint32_t *state_pos
= pipeline
->blend_state
.map
;
890 state_pos
+= GENX(BLEND_STATE_length
);
892 struct GENX(BLEND_STATE_ENTRY
) bs0
= { 0 };
894 for (unsigned i
= 0; i
< surface_count
; i
++) {
895 struct anv_pipeline_binding
*binding
= &map
->surface_to_descriptor
[i
];
897 /* All color attachments are at the beginning of the binding table */
898 if (binding
->set
!= ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
)
901 /* We can have at most 8 attachments */
904 if (info
== NULL
|| binding
->index
>= info
->attachmentCount
) {
905 /* Default everything to disabled */
906 struct GENX(BLEND_STATE_ENTRY
) entry
= {
907 .WriteDisableAlpha
= true,
908 .WriteDisableRed
= true,
909 .WriteDisableGreen
= true,
910 .WriteDisableBlue
= true,
912 GENX(BLEND_STATE_ENTRY_pack
)(NULL
, state_pos
, &entry
);
913 state_pos
+= GENX(BLEND_STATE_ENTRY_length
);
917 assert(binding
->binding
== 0);
918 const VkPipelineColorBlendAttachmentState
*a
=
919 &info
->pAttachments
[binding
->index
];
921 struct GENX(BLEND_STATE_ENTRY
) entry
= {
923 .AlphaToCoverageEnable
= ms_info
&& ms_info
->alphaToCoverageEnable
,
924 .AlphaToOneEnable
= ms_info
&& ms_info
->alphaToOneEnable
,
926 .LogicOpEnable
= info
->logicOpEnable
,
927 .LogicOpFunction
= vk_to_gen_logic_op
[info
->logicOp
],
928 .ColorBufferBlendEnable
= a
->blendEnable
,
929 .ColorClampRange
= COLORCLAMP_RTFORMAT
,
930 .PreBlendColorClampEnable
= true,
931 .PostBlendColorClampEnable
= true,
932 .SourceBlendFactor
= vk_to_gen_blend
[a
->srcColorBlendFactor
],
933 .DestinationBlendFactor
= vk_to_gen_blend
[a
->dstColorBlendFactor
],
934 .ColorBlendFunction
= vk_to_gen_blend_op
[a
->colorBlendOp
],
935 .SourceAlphaBlendFactor
= vk_to_gen_blend
[a
->srcAlphaBlendFactor
],
936 .DestinationAlphaBlendFactor
= vk_to_gen_blend
[a
->dstAlphaBlendFactor
],
937 .AlphaBlendFunction
= vk_to_gen_blend_op
[a
->alphaBlendOp
],
938 .WriteDisableAlpha
= !(a
->colorWriteMask
& VK_COLOR_COMPONENT_A_BIT
),
939 .WriteDisableRed
= !(a
->colorWriteMask
& VK_COLOR_COMPONENT_R_BIT
),
940 .WriteDisableGreen
= !(a
->colorWriteMask
& VK_COLOR_COMPONENT_G_BIT
),
941 .WriteDisableBlue
= !(a
->colorWriteMask
& VK_COLOR_COMPONENT_B_BIT
),
944 if (a
->srcColorBlendFactor
!= a
->srcAlphaBlendFactor
||
945 a
->dstColorBlendFactor
!= a
->dstAlphaBlendFactor
||
946 a
->colorBlendOp
!= a
->alphaBlendOp
) {
948 blend_state
.IndependentAlphaBlendEnable
= true;
950 entry
.IndependentAlphaBlendEnable
= true;
954 if (a
->colorWriteMask
!= 0)
955 has_writeable_rt
= true;
957 /* Our hardware applies the blend factor prior to the blend function
958 * regardless of what function is used. Technically, this means the
959 * hardware can do MORE than GL or Vulkan specify. However, it also
960 * means that, for MIN and MAX, we have to stomp the blend factor to
961 * ONE to make it a no-op.
963 if (a
->colorBlendOp
== VK_BLEND_OP_MIN
||
964 a
->colorBlendOp
== VK_BLEND_OP_MAX
) {
965 entry
.SourceBlendFactor
= BLENDFACTOR_ONE
;
966 entry
.DestinationBlendFactor
= BLENDFACTOR_ONE
;
968 if (a
->alphaBlendOp
== VK_BLEND_OP_MIN
||
969 a
->alphaBlendOp
== VK_BLEND_OP_MAX
) {
970 entry
.SourceAlphaBlendFactor
= BLENDFACTOR_ONE
;
971 entry
.DestinationAlphaBlendFactor
= BLENDFACTOR_ONE
;
973 GENX(BLEND_STATE_ENTRY_pack
)(NULL
, state_pos
, &entry
);
974 state_pos
+= GENX(BLEND_STATE_ENTRY_length
);
982 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_PS_BLEND
), blend
) {
983 blend
.AlphaToCoverageEnable
= blend_state
.AlphaToCoverageEnable
;
984 blend
.HasWriteableRT
= has_writeable_rt
;
985 blend
.ColorBufferBlendEnable
= bs0
.ColorBufferBlendEnable
;
986 blend
.SourceAlphaBlendFactor
= bs0
.SourceAlphaBlendFactor
;
987 blend
.DestinationAlphaBlendFactor
= bs0
.DestinationAlphaBlendFactor
;
988 blend
.SourceBlendFactor
= bs0
.SourceBlendFactor
;
989 blend
.DestinationBlendFactor
= bs0
.DestinationBlendFactor
;
990 blend
.AlphaTestEnable
= false;
991 blend
.IndependentAlphaBlendEnable
=
992 blend_state
.IndependentAlphaBlendEnable
;
995 (void)has_writeable_rt
;
998 GENX(BLEND_STATE_pack
)(NULL
, pipeline
->blend_state
.map
, &blend_state
);
999 anv_state_flush(device
, pipeline
->blend_state
);
1001 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_BLEND_STATE_POINTERS
), bsp
) {
1002 bsp
.BlendStatePointer
= pipeline
->blend_state
.offset
;
1004 bsp
.BlendStatePointerValid
= true;
1010 emit_3dstate_clip(struct anv_pipeline
*pipeline
,
1011 const VkPipelineViewportStateCreateInfo
*vp_info
,
1012 const VkPipelineRasterizationStateCreateInfo
*rs_info
)
1014 const struct brw_wm_prog_data
*wm_prog_data
= get_wm_prog_data(pipeline
);
1015 (void) wm_prog_data
;
1016 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_CLIP
), clip
) {
1017 clip
.ClipEnable
= true;
1018 clip
.StatisticsEnable
= true;
1019 clip
.EarlyCullEnable
= true;
1020 clip
.APIMode
= APIMODE_D3D
,
1021 clip
.ViewportXYClipTestEnable
= true;
1023 clip
.ClipMode
= CLIPMODE_NORMAL
;
1025 clip
.TriangleStripListProvokingVertexSelect
= 0;
1026 clip
.LineStripListProvokingVertexSelect
= 0;
1027 clip
.TriangleFanProvokingVertexSelect
= 1;
1029 clip
.MinimumPointWidth
= 0.125;
1030 clip
.MaximumPointWidth
= 255.875;
1032 const struct brw_vue_prog_data
*last
=
1033 anv_pipeline_get_last_vue_prog_data(pipeline
);
1035 /* From the Vulkan 1.0.45 spec:
1037 * "If the last active vertex processing stage shader entry point's
1038 * interface does not include a variable decorated with
1039 * ViewportIndex, then the first viewport is used."
1041 if (vp_info
&& (last
->vue_map
.slots_valid
& VARYING_BIT_VIEWPORT
)) {
1042 clip
.MaximumVPIndex
= vp_info
->viewportCount
- 1;
1044 clip
.MaximumVPIndex
= 0;
1047 /* From the Vulkan 1.0.45 spec:
1049 * "If the last active vertex processing stage shader entry point's
1050 * interface does not include a variable decorated with Layer, then
1051 * the first layer is used."
1053 clip
.ForceZeroRTAIndexEnable
=
1054 !(last
->vue_map
.slots_valid
& VARYING_BIT_LAYER
);
1057 clip
.FrontWinding
= vk_to_gen_front_face
[rs_info
->frontFace
];
1058 clip
.CullMode
= vk_to_gen_cullmode
[rs_info
->cullMode
];
1059 clip
.ViewportZClipTestEnable
= !pipeline
->depth_clamp_enable
;
1061 clip
.UserClipDistanceClipTestEnableBitmask
= last
->clip_distance_mask
;
1062 clip
.UserClipDistanceCullTestEnableBitmask
= last
->cull_distance_mask
;
1065 clip
.NonPerspectiveBarycentricEnable
= wm_prog_data
?
1066 (wm_prog_data
->barycentric_interp_modes
&
1067 BRW_BARYCENTRIC_NONPERSPECTIVE_BITS
) != 0 : 0;
1073 emit_3dstate_streamout(struct anv_pipeline
*pipeline
,
1074 const VkPipelineRasterizationStateCreateInfo
*rs_info
)
1076 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_STREAMOUT
), so
) {
1077 so
.RenderingDisable
= rs_info
->rasterizerDiscardEnable
;
1082 get_sampler_count(const struct anv_shader_bin
*bin
)
1084 uint32_t count_by_4
= DIV_ROUND_UP(bin
->bind_map
.sampler_count
, 4);
1086 /* We can potentially have way more than 32 samplers and that's ok.
1087 * However, the 3DSTATE_XS packets only have 3 bits to specify how
1088 * many to pre-fetch and all values above 4 are marked reserved.
1090 return MIN2(count_by_4
, 4);
1094 get_binding_table_entry_count(const struct anv_shader_bin
*bin
)
1096 return DIV_ROUND_UP(bin
->bind_map
.surface_count
, 32);
1099 static struct anv_address
1100 get_scratch_address(struct anv_pipeline
*pipeline
,
1101 gl_shader_stage stage
,
1102 const struct anv_shader_bin
*bin
)
1104 return (struct anv_address
) {
1105 .bo
= anv_scratch_pool_alloc(pipeline
->device
,
1106 &pipeline
->device
->scratch_pool
,
1107 stage
, bin
->prog_data
->total_scratch
),
1113 get_scratch_space(const struct anv_shader_bin
*bin
)
1115 return ffs(bin
->prog_data
->total_scratch
/ 2048);
1119 emit_3dstate_vs(struct anv_pipeline
*pipeline
)
1121 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
1122 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
1123 const struct anv_shader_bin
*vs_bin
=
1124 pipeline
->shaders
[MESA_SHADER_VERTEX
];
1126 assert(anv_pipeline_has_stage(pipeline
, MESA_SHADER_VERTEX
));
1128 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_VS
), vs
) {
1130 vs
.StatisticsEnable
= true;
1131 vs
.KernelStartPointer
= vs_bin
->kernel
.offset
;
1133 vs
.SIMD8DispatchEnable
=
1134 vs_prog_data
->base
.dispatch_mode
== DISPATCH_MODE_SIMD8
;
1137 assert(!vs_prog_data
->base
.base
.use_alt_mode
);
1139 vs
.SingleVertexDispatch
= false;
1141 vs
.VectorMaskEnable
= false;
1142 vs
.SamplerCount
= get_sampler_count(vs_bin
);
1143 vs
.BindingTableEntryCount
= get_binding_table_entry_count(vs_bin
);
1144 vs
.FloatingPointMode
= IEEE754
;
1145 vs
.IllegalOpcodeExceptionEnable
= false;
1146 vs
.SoftwareExceptionEnable
= false;
1147 vs
.MaximumNumberofThreads
= devinfo
->max_vs_threads
- 1;
1148 vs
.VertexCacheDisable
= false;
1150 vs
.VertexURBEntryReadLength
= vs_prog_data
->base
.urb_read_length
;
1151 vs
.VertexURBEntryReadOffset
= 0;
1152 vs
.DispatchGRFStartRegisterForURBData
=
1153 vs_prog_data
->base
.base
.dispatch_grf_start_reg
;
1156 vs
.UserClipDistanceClipTestEnableBitmask
=
1157 vs_prog_data
->base
.clip_distance_mask
;
1158 vs
.UserClipDistanceCullTestEnableBitmask
=
1159 vs_prog_data
->base
.cull_distance_mask
;
1162 vs
.PerThreadScratchSpace
= get_scratch_space(vs_bin
);
1163 vs
.ScratchSpaceBasePointer
=
1164 get_scratch_address(pipeline
, MESA_SHADER_VERTEX
, vs_bin
);
1169 emit_3dstate_hs_te_ds(struct anv_pipeline
*pipeline
,
1170 const VkPipelineTessellationStateCreateInfo
*tess_info
)
1172 if (!anv_pipeline_has_stage(pipeline
, MESA_SHADER_TESS_EVAL
)) {
1173 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_HS
), hs
);
1174 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_TE
), te
);
1175 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_DS
), ds
);
1179 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
1180 const struct anv_shader_bin
*tcs_bin
=
1181 pipeline
->shaders
[MESA_SHADER_TESS_CTRL
];
1182 const struct anv_shader_bin
*tes_bin
=
1183 pipeline
->shaders
[MESA_SHADER_TESS_EVAL
];
1185 const struct brw_tcs_prog_data
*tcs_prog_data
= get_tcs_prog_data(pipeline
);
1186 const struct brw_tes_prog_data
*tes_prog_data
= get_tes_prog_data(pipeline
);
1188 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_HS
), hs
) {
1190 hs
.StatisticsEnable
= true;
1191 hs
.KernelStartPointer
= tcs_bin
->kernel
.offset
;
1193 hs
.SamplerCount
= get_sampler_count(tcs_bin
);
1194 hs
.BindingTableEntryCount
= get_binding_table_entry_count(tcs_bin
);
1195 hs
.MaximumNumberofThreads
= devinfo
->max_tcs_threads
- 1;
1196 hs
.IncludeVertexHandles
= true;
1197 hs
.InstanceCount
= tcs_prog_data
->instances
- 1;
1199 hs
.VertexURBEntryReadLength
= 0;
1200 hs
.VertexURBEntryReadOffset
= 0;
1201 hs
.DispatchGRFStartRegisterForURBData
=
1202 tcs_prog_data
->base
.base
.dispatch_grf_start_reg
;
1204 hs
.PerThreadScratchSpace
= get_scratch_space(tcs_bin
);
1205 hs
.ScratchSpaceBasePointer
=
1206 get_scratch_address(pipeline
, MESA_SHADER_TESS_CTRL
, tcs_bin
);
1209 const VkPipelineTessellationDomainOriginStateCreateInfoKHR
*domain_origin_state
=
1210 tess_info
? vk_find_struct_const(tess_info
, PIPELINE_TESSELLATION_DOMAIN_ORIGIN_STATE_CREATE_INFO_KHR
) : NULL
;
1212 VkTessellationDomainOriginKHR uv_origin
=
1213 domain_origin_state
? domain_origin_state
->domainOrigin
:
1214 VK_TESSELLATION_DOMAIN_ORIGIN_UPPER_LEFT_KHR
;
1216 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_TE
), te
) {
1217 te
.Partitioning
= tes_prog_data
->partitioning
;
1219 if (uv_origin
== VK_TESSELLATION_DOMAIN_ORIGIN_LOWER_LEFT_KHR
) {
1220 te
.OutputTopology
= tes_prog_data
->output_topology
;
1222 /* When the origin is upper-left, we have to flip the winding order */
1223 if (tes_prog_data
->output_topology
== OUTPUT_TRI_CCW
) {
1224 te
.OutputTopology
= OUTPUT_TRI_CW
;
1225 } else if (tes_prog_data
->output_topology
== OUTPUT_TRI_CW
) {
1226 te
.OutputTopology
= OUTPUT_TRI_CCW
;
1228 te
.OutputTopology
= tes_prog_data
->output_topology
;
1232 te
.TEDomain
= tes_prog_data
->domain
;
1234 te
.MaximumTessellationFactorOdd
= 63.0;
1235 te
.MaximumTessellationFactorNotOdd
= 64.0;
1238 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_DS
), ds
) {
1240 ds
.StatisticsEnable
= true;
1241 ds
.KernelStartPointer
= tes_bin
->kernel
.offset
;
1243 ds
.SamplerCount
= get_sampler_count(tes_bin
);
1244 ds
.BindingTableEntryCount
= get_binding_table_entry_count(tes_bin
);
1245 ds
.MaximumNumberofThreads
= devinfo
->max_tes_threads
- 1;
1247 ds
.ComputeWCoordinateEnable
=
1248 tes_prog_data
->domain
== BRW_TESS_DOMAIN_TRI
;
1250 ds
.PatchURBEntryReadLength
= tes_prog_data
->base
.urb_read_length
;
1251 ds
.PatchURBEntryReadOffset
= 0;
1252 ds
.DispatchGRFStartRegisterForURBData
=
1253 tes_prog_data
->base
.base
.dispatch_grf_start_reg
;
1258 tes_prog_data
->base
.dispatch_mode
== DISPATCH_MODE_SIMD8
?
1259 DISPATCH_MODE_SIMD8_SINGLE_PATCH
:
1260 DISPATCH_MODE_SIMD4X2
;
1262 assert(tes_prog_data
->base
.dispatch_mode
== DISPATCH_MODE_SIMD8
);
1263 ds
.DispatchMode
= DISPATCH_MODE_SIMD8_SINGLE_PATCH
;
1266 ds
.UserClipDistanceClipTestEnableBitmask
=
1267 tes_prog_data
->base
.clip_distance_mask
;
1268 ds
.UserClipDistanceCullTestEnableBitmask
=
1269 tes_prog_data
->base
.cull_distance_mask
;
1272 ds
.PerThreadScratchSpace
= get_scratch_space(tes_bin
);
1273 ds
.ScratchSpaceBasePointer
=
1274 get_scratch_address(pipeline
, MESA_SHADER_TESS_EVAL
, tes_bin
);
1279 emit_3dstate_gs(struct anv_pipeline
*pipeline
)
1281 const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
1282 const struct anv_shader_bin
*gs_bin
=
1283 pipeline
->shaders
[MESA_SHADER_GEOMETRY
];
1285 if (!anv_pipeline_has_stage(pipeline
, MESA_SHADER_GEOMETRY
)) {
1286 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_GS
), gs
);
1290 const struct brw_gs_prog_data
*gs_prog_data
= get_gs_prog_data(pipeline
);
1292 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_GS
), gs
) {
1294 gs
.StatisticsEnable
= true;
1295 gs
.KernelStartPointer
= gs_bin
->kernel
.offset
;
1296 gs
.DispatchMode
= gs_prog_data
->base
.dispatch_mode
;
1298 gs
.SingleProgramFlow
= false;
1299 gs
.VectorMaskEnable
= false;
1300 gs
.SamplerCount
= get_sampler_count(gs_bin
);
1301 gs
.BindingTableEntryCount
= get_binding_table_entry_count(gs_bin
);
1302 gs
.IncludeVertexHandles
= gs_prog_data
->base
.include_vue_handles
;
1303 gs
.IncludePrimitiveID
= gs_prog_data
->include_primitive_id
;
1306 /* Broadwell is weird. It needs us to divide by 2. */
1307 gs
.MaximumNumberofThreads
= devinfo
->max_gs_threads
/ 2 - 1;
1309 gs
.MaximumNumberofThreads
= devinfo
->max_gs_threads
- 1;
1312 gs
.OutputVertexSize
= gs_prog_data
->output_vertex_size_hwords
* 2 - 1;
1313 gs
.OutputTopology
= gs_prog_data
->output_topology
;
1314 gs
.VertexURBEntryReadLength
= gs_prog_data
->base
.urb_read_length
;
1315 gs
.ControlDataFormat
= gs_prog_data
->control_data_format
;
1316 gs
.ControlDataHeaderSize
= gs_prog_data
->control_data_header_size_hwords
;
1317 gs
.InstanceControl
= MAX2(gs_prog_data
->invocations
, 1) - 1;
1318 gs
.ReorderMode
= TRAILING
;
1321 gs
.ExpectedVertexCount
= gs_prog_data
->vertices_in
;
1322 gs
.StaticOutput
= gs_prog_data
->static_vertex_count
>= 0;
1323 gs
.StaticOutputVertexCount
= gs_prog_data
->static_vertex_count
>= 0 ?
1324 gs_prog_data
->static_vertex_count
: 0;
1327 gs
.VertexURBEntryReadOffset
= 0;
1328 gs
.VertexURBEntryReadLength
= gs_prog_data
->base
.urb_read_length
;
1329 gs
.DispatchGRFStartRegisterForURBData
=
1330 gs_prog_data
->base
.base
.dispatch_grf_start_reg
;
1333 gs
.UserClipDistanceClipTestEnableBitmask
=
1334 gs_prog_data
->base
.clip_distance_mask
;
1335 gs
.UserClipDistanceCullTestEnableBitmask
=
1336 gs_prog_data
->base
.cull_distance_mask
;
1339 gs
.PerThreadScratchSpace
= get_scratch_space(gs_bin
);
1340 gs
.ScratchSpaceBasePointer
=
1341 get_scratch_address(pipeline
, MESA_SHADER_GEOMETRY
, gs_bin
);
1346 has_color_buffer_write_enabled(const struct anv_pipeline
*pipeline
,
1347 const VkPipelineColorBlendStateCreateInfo
*blend
)
1349 const struct anv_shader_bin
*shader_bin
=
1350 pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
1354 const struct anv_pipeline_bind_map
*bind_map
= &shader_bin
->bind_map
;
1355 for (int i
= 0; i
< bind_map
->surface_count
; i
++) {
1356 struct anv_pipeline_binding
*binding
= &bind_map
->surface_to_descriptor
[i
];
1358 if (binding
->set
!= ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
)
1361 if (binding
->index
== UINT32_MAX
)
1364 if (blend
->pAttachments
[binding
->index
].colorWriteMask
!= 0)
1372 emit_3dstate_wm(struct anv_pipeline
*pipeline
, struct anv_subpass
*subpass
,
1373 const VkPipelineColorBlendStateCreateInfo
*blend
,
1374 const VkPipelineMultisampleStateCreateInfo
*multisample
)
1376 const struct brw_wm_prog_data
*wm_prog_data
= get_wm_prog_data(pipeline
);
1378 MAYBE_UNUSED
uint32_t samples
=
1379 multisample
? multisample
->rasterizationSamples
: 1;
1381 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_WM
), wm
) {
1382 wm
.StatisticsEnable
= true;
1383 wm
.LineEndCapAntialiasingRegionWidth
= _05pixels
;
1384 wm
.LineAntialiasingRegionWidth
= _10pixels
;
1385 wm
.PointRasterizationRule
= RASTRULE_UPPER_RIGHT
;
1387 if (anv_pipeline_has_stage(pipeline
, MESA_SHADER_FRAGMENT
)) {
1388 if (wm_prog_data
->early_fragment_tests
) {
1389 wm
.EarlyDepthStencilControl
= EDSC_PREPS
;
1390 } else if (wm_prog_data
->has_side_effects
) {
1391 wm
.EarlyDepthStencilControl
= EDSC_PSEXEC
;
1393 wm
.EarlyDepthStencilControl
= EDSC_NORMAL
;
1396 wm
.BarycentricInterpolationMode
=
1397 wm_prog_data
->barycentric_interp_modes
;
1400 wm
.PixelShaderComputedDepthMode
= wm_prog_data
->computed_depth_mode
;
1401 wm
.PixelShaderUsesSourceDepth
= wm_prog_data
->uses_src_depth
;
1402 wm
.PixelShaderUsesSourceW
= wm_prog_data
->uses_src_w
;
1403 wm
.PixelShaderUsesInputCoverageMask
= wm_prog_data
->uses_sample_mask
;
1405 /* If the subpass has a depth or stencil self-dependency, then we
1406 * need to force the hardware to do the depth/stencil write *after*
1407 * fragment shader execution. Otherwise, the writes may hit memory
1408 * before we get around to fetching from the input attachment and we
1409 * may get the depth or stencil value from the current draw rather
1410 * than the previous one.
1412 wm
.PixelShaderKillsPixel
= subpass
->has_ds_self_dep
||
1413 wm_prog_data
->uses_kill
;
1415 if (wm
.PixelShaderComputedDepthMode
!= PSCDEPTH_OFF
||
1416 wm_prog_data
->has_side_effects
||
1417 wm
.PixelShaderKillsPixel
||
1418 has_color_buffer_write_enabled(pipeline
, blend
))
1419 wm
.ThreadDispatchEnable
= true;
1422 wm
.MultisampleRasterizationMode
= MSRASTMODE_ON_PATTERN
;
1423 if (wm_prog_data
->persample_dispatch
) {
1424 wm
.MultisampleDispatchMode
= MSDISPMODE_PERSAMPLE
;
1426 wm
.MultisampleDispatchMode
= MSDISPMODE_PERPIXEL
;
1429 wm
.MultisampleRasterizationMode
= MSRASTMODE_OFF_PIXEL
;
1430 wm
.MultisampleDispatchMode
= MSDISPMODE_PERSAMPLE
;
1438 is_dual_src_blend_factor(VkBlendFactor factor
)
1440 return factor
== VK_BLEND_FACTOR_SRC1_COLOR
||
1441 factor
== VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
||
1442 factor
== VK_BLEND_FACTOR_SRC1_ALPHA
||
1443 factor
== VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
;
1447 emit_3dstate_ps(struct anv_pipeline
*pipeline
,
1448 const VkPipelineColorBlendStateCreateInfo
*blend
)
1450 MAYBE_UNUSED
const struct gen_device_info
*devinfo
= &pipeline
->device
->info
;
1451 const struct anv_shader_bin
*fs_bin
=
1452 pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
1454 if (!anv_pipeline_has_stage(pipeline
, MESA_SHADER_FRAGMENT
)) {
1455 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_PS
), ps
) {
1457 /* Even if no fragments are ever dispatched, gen7 hardware hangs if
1458 * we don't at least set the maximum number of threads.
1460 ps
.MaximumNumberofThreads
= devinfo
->max_wm_threads
- 1;
1466 const struct brw_wm_prog_data
*wm_prog_data
= get_wm_prog_data(pipeline
);
1469 /* The hardware wedges if you have this bit set but don't turn on any dual
1470 * source blend factors.
1472 bool dual_src_blend
= false;
1473 if (wm_prog_data
->dual_src_blend
&& blend
) {
1474 for (uint32_t i
= 0; i
< blend
->attachmentCount
; i
++) {
1475 const VkPipelineColorBlendAttachmentState
*bstate
=
1476 &blend
->pAttachments
[i
];
1478 if (bstate
->blendEnable
&&
1479 (is_dual_src_blend_factor(bstate
->srcColorBlendFactor
) ||
1480 is_dual_src_blend_factor(bstate
->dstColorBlendFactor
) ||
1481 is_dual_src_blend_factor(bstate
->srcAlphaBlendFactor
) ||
1482 is_dual_src_blend_factor(bstate
->dstAlphaBlendFactor
))) {
1483 dual_src_blend
= true;
1490 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_PS
), ps
) {
1491 ps
.KernelStartPointer0
= fs_bin
->kernel
.offset
;
1492 ps
.KernelStartPointer1
= 0;
1493 ps
.KernelStartPointer2
= fs_bin
->kernel
.offset
+
1494 wm_prog_data
->prog_offset_2
;
1495 ps
._8PixelDispatchEnable
= wm_prog_data
->dispatch_8
;
1496 ps
._16PixelDispatchEnable
= wm_prog_data
->dispatch_16
;
1497 ps
._32PixelDispatchEnable
= false;
1499 ps
.SingleProgramFlow
= false;
1500 ps
.VectorMaskEnable
= true;
1501 ps
.SamplerCount
= get_sampler_count(fs_bin
);
1502 ps
.BindingTableEntryCount
= get_binding_table_entry_count(fs_bin
);
1503 ps
.PushConstantEnable
= wm_prog_data
->base
.nr_params
> 0 ||
1504 wm_prog_data
->base
.ubo_ranges
[0].length
;
1505 ps
.PositionXYOffsetSelect
= wm_prog_data
->uses_pos_offset
?
1506 POSOFFSET_SAMPLE
: POSOFFSET_NONE
;
1508 ps
.AttributeEnable
= wm_prog_data
->num_varying_inputs
> 0;
1509 ps
.oMaskPresenttoRenderTarget
= wm_prog_data
->uses_omask
;
1510 ps
.DualSourceBlendEnable
= dual_src_blend
;
1514 /* Haswell requires the sample mask to be set in this packet as well
1515 * as in 3DSTATE_SAMPLE_MASK; the values should match.
1517 ps
.SampleMask
= 0xff;
1521 ps
.MaximumNumberofThreadsPerPSD
= 64 - 1;
1523 ps
.MaximumNumberofThreadsPerPSD
= 64 - 2;
1525 ps
.MaximumNumberofThreads
= devinfo
->max_wm_threads
- 1;
1528 ps
.DispatchGRFStartRegisterForConstantSetupData0
=
1529 wm_prog_data
->base
.dispatch_grf_start_reg
;
1530 ps
.DispatchGRFStartRegisterForConstantSetupData1
= 0;
1531 ps
.DispatchGRFStartRegisterForConstantSetupData2
=
1532 wm_prog_data
->dispatch_grf_start_reg_2
;
1534 ps
.PerThreadScratchSpace
= get_scratch_space(fs_bin
);
1535 ps
.ScratchSpaceBasePointer
=
1536 get_scratch_address(pipeline
, MESA_SHADER_FRAGMENT
, fs_bin
);
1542 emit_3dstate_ps_extra(struct anv_pipeline
*pipeline
,
1543 struct anv_subpass
*subpass
,
1544 const VkPipelineColorBlendStateCreateInfo
*blend
)
1546 const struct brw_wm_prog_data
*wm_prog_data
= get_wm_prog_data(pipeline
);
1548 if (!anv_pipeline_has_stage(pipeline
, MESA_SHADER_FRAGMENT
)) {
1549 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_PS_EXTRA
), ps
);
1553 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_PS_EXTRA
), ps
) {
1554 ps
.PixelShaderValid
= true;
1555 ps
.AttributeEnable
= wm_prog_data
->num_varying_inputs
> 0;
1556 ps
.oMaskPresenttoRenderTarget
= wm_prog_data
->uses_omask
;
1557 ps
.PixelShaderIsPerSample
= wm_prog_data
->persample_dispatch
;
1558 ps
.PixelShaderComputedDepthMode
= wm_prog_data
->computed_depth_mode
;
1559 ps
.PixelShaderUsesSourceDepth
= wm_prog_data
->uses_src_depth
;
1560 ps
.PixelShaderUsesSourceW
= wm_prog_data
->uses_src_w
;
1562 /* If the subpass has a depth or stencil self-dependency, then we need
1563 * to force the hardware to do the depth/stencil write *after* fragment
1564 * shader execution. Otherwise, the writes may hit memory before we get
1565 * around to fetching from the input attachment and we may get the depth
1566 * or stencil value from the current draw rather than the previous one.
1568 ps
.PixelShaderKillsPixel
= subpass
->has_ds_self_dep
||
1569 wm_prog_data
->uses_kill
;
1571 /* The stricter cross-primitive coherency guarantees that the hardware
1572 * gives us with the "Accesses UAV" bit set for at least one shader stage
1573 * and the "UAV coherency required" bit set on the 3DPRIMITIVE command are
1574 * redundant within the current image, atomic counter and SSBO GL APIs,
1575 * which all have very loose ordering and coherency requirements and
1576 * generally rely on the application to insert explicit barriers when a
1577 * shader invocation is expected to see the memory writes performed by the
1578 * invocations of some previous primitive. Regardless of the value of
1579 * "UAV coherency required", the "Accesses UAV" bits will implicitly cause
1580 * an in most cases useless DC flush when the lowermost stage with the bit
1581 * set finishes execution.
1583 * It would be nice to disable it, but in some cases we can't because on
1584 * Gen8+ it also has an influence on rasterization via the PS UAV-only
1585 * signal (which could be set independently from the coherency mechanism
1586 * in the 3DSTATE_WM command on Gen7), and because in some cases it will
1587 * determine whether the hardware skips execution of the fragment shader
1588 * or not via the ThreadDispatchEnable signal. However if we know that
1589 * GEN8_PS_BLEND_HAS_WRITEABLE_RT is going to be set and
1590 * GEN8_PSX_PIXEL_SHADER_NO_RT_WRITE is not set it shouldn't make any
1591 * difference so we may just disable it here.
1593 * Gen8 hardware tries to compute ThreadDispatchEnable for us but doesn't
1594 * take into account KillPixels when no depth or stencil writes are
1595 * enabled. In order for occlusion queries to work correctly with no
1596 * attachments, we need to force-enable here.
1598 if ((wm_prog_data
->has_side_effects
|| wm_prog_data
->uses_kill
) &&
1599 !has_color_buffer_write_enabled(pipeline
, blend
))
1600 ps
.PixelShaderHasUAV
= true;
1603 ps
.PixelShaderPullsBary
= wm_prog_data
->pulls_bary
;
1604 ps
.InputCoverageMaskState
= wm_prog_data
->uses_sample_mask
?
1605 ICMS_INNER_CONSERVATIVE
: ICMS_NONE
;
1607 ps
.PixelShaderUsesInputCoverageMask
= wm_prog_data
->uses_sample_mask
;
1613 emit_3dstate_vf_topology(struct anv_pipeline
*pipeline
)
1615 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_VF_TOPOLOGY
), vft
) {
1616 vft
.PrimitiveTopologyType
= pipeline
->topology
;
1622 emit_3dstate_vf_statistics(struct anv_pipeline
*pipeline
)
1624 anv_batch_emit(&pipeline
->batch
, GENX(3DSTATE_VF_STATISTICS
), vfs
) {
1625 vfs
.StatisticsEnable
= true;
1630 compute_kill_pixel(struct anv_pipeline
*pipeline
,
1631 const VkPipelineMultisampleStateCreateInfo
*ms_info
,
1632 const struct anv_subpass
*subpass
)
1634 if (!anv_pipeline_has_stage(pipeline
, MESA_SHADER_FRAGMENT
)) {
1635 pipeline
->kill_pixel
= false;
1639 const struct brw_wm_prog_data
*wm_prog_data
= get_wm_prog_data(pipeline
);
1641 /* This computes the KillPixel portion of the computation for whether or
1642 * not we want to enable the PMA fix on gen8 or gen9. It's given by this
1643 * chunk of the giant formula:
1645 * (3DSTATE_PS_EXTRA::PixelShaderKillsPixels ||
1646 * 3DSTATE_PS_EXTRA::oMask Present to RenderTarget ||
1647 * 3DSTATE_PS_BLEND::AlphaToCoverageEnable ||
1648 * 3DSTATE_PS_BLEND::AlphaTestEnable ||
1649 * 3DSTATE_WM_CHROMAKEY::ChromaKeyKillEnable)
1651 * 3DSTATE_WM_CHROMAKEY::ChromaKeyKillEnable is always false and so is
1652 * 3DSTATE_PS_BLEND::AlphaTestEnable since Vulkan doesn't have a concept
1655 pipeline
->kill_pixel
=
1656 subpass
->has_ds_self_dep
|| wm_prog_data
->uses_kill
||
1657 wm_prog_data
->uses_omask
||
1658 (ms_info
&& ms_info
->alphaToCoverageEnable
);
1662 genX(graphics_pipeline_create
)(
1664 struct anv_pipeline_cache
* cache
,
1665 const VkGraphicsPipelineCreateInfo
* pCreateInfo
,
1666 const VkAllocationCallbacks
* pAllocator
,
1667 VkPipeline
* pPipeline
)
1669 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1670 ANV_FROM_HANDLE(anv_render_pass
, pass
, pCreateInfo
->renderPass
);
1671 struct anv_subpass
*subpass
= &pass
->subpasses
[pCreateInfo
->subpass
];
1672 struct anv_pipeline
*pipeline
;
1675 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO
);
1677 pipeline
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*pipeline
), 8,
1678 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1679 if (pipeline
== NULL
)
1680 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1682 result
= anv_pipeline_init(pipeline
, device
, cache
,
1683 pCreateInfo
, pAllocator
);
1684 if (result
!= VK_SUCCESS
) {
1685 vk_free2(&device
->alloc
, pAllocator
, pipeline
);
1689 assert(pCreateInfo
->pVertexInputState
);
1690 emit_vertex_input(pipeline
, pCreateInfo
->pVertexInputState
);
1691 assert(pCreateInfo
->pRasterizationState
);
1692 emit_rs_state(pipeline
, pCreateInfo
->pRasterizationState
,
1693 pCreateInfo
->pMultisampleState
, pass
, subpass
);
1694 emit_ms_state(pipeline
, pCreateInfo
->pMultisampleState
);
1695 emit_ds_state(pipeline
, pCreateInfo
->pDepthStencilState
, pass
, subpass
);
1696 emit_cb_state(pipeline
, pCreateInfo
->pColorBlendState
,
1697 pCreateInfo
->pMultisampleState
);
1698 compute_kill_pixel(pipeline
, pCreateInfo
->pMultisampleState
, subpass
);
1700 emit_urb_setup(pipeline
);
1702 emit_3dstate_clip(pipeline
, pCreateInfo
->pViewportState
,
1703 pCreateInfo
->pRasterizationState
);
1704 emit_3dstate_streamout(pipeline
, pCreateInfo
->pRasterizationState
);
1707 /* From gen7_vs_state.c */
1710 * From Graphics BSpec: 3D-Media-GPGPU Engine > 3D Pipeline Stages >
1711 * Geometry > Geometry Shader > State:
1713 * "Note: Because of corruption in IVB:GT2, software needs to flush the
1714 * whole fixed function pipeline when the GS enable changes value in
1717 * The hardware architects have clarified that in this context "flush the
1718 * whole fixed function pipeline" means to emit a PIPE_CONTROL with the "CS
1721 if (!device
->info
.is_haswell
&& !device
->info
.is_baytrail
)
1722 gen7_emit_vs_workaround_flush(brw
);
1725 emit_3dstate_vs(pipeline
);
1726 emit_3dstate_hs_te_ds(pipeline
, pCreateInfo
->pTessellationState
);
1727 emit_3dstate_gs(pipeline
);
1728 emit_3dstate_sbe(pipeline
);
1729 emit_3dstate_wm(pipeline
, subpass
, pCreateInfo
->pColorBlendState
,
1730 pCreateInfo
->pMultisampleState
);
1731 emit_3dstate_ps(pipeline
, pCreateInfo
->pColorBlendState
);
1733 emit_3dstate_ps_extra(pipeline
, subpass
, pCreateInfo
->pColorBlendState
);
1734 emit_3dstate_vf_topology(pipeline
);
1736 emit_3dstate_vf_statistics(pipeline
);
1738 *pPipeline
= anv_pipeline_to_handle(pipeline
);
1740 return pipeline
->batch
.status
;
1744 compute_pipeline_create(
1746 struct anv_pipeline_cache
* cache
,
1747 const VkComputePipelineCreateInfo
* pCreateInfo
,
1748 const VkAllocationCallbacks
* pAllocator
,
1749 VkPipeline
* pPipeline
)
1751 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1752 const struct anv_physical_device
*physical_device
=
1753 &device
->instance
->physicalDevice
;
1754 const struct gen_device_info
*devinfo
= &physical_device
->info
;
1755 struct anv_pipeline
*pipeline
;
1758 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO
);
1760 pipeline
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*pipeline
), 8,
1761 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1762 if (pipeline
== NULL
)
1763 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1765 pipeline
->device
= device
;
1767 pipeline
->blend_state
.map
= NULL
;
1769 result
= anv_reloc_list_init(&pipeline
->batch_relocs
,
1770 pAllocator
? pAllocator
: &device
->alloc
);
1771 if (result
!= VK_SUCCESS
) {
1772 vk_free2(&device
->alloc
, pAllocator
, pipeline
);
1775 pipeline
->batch
.next
= pipeline
->batch
.start
= pipeline
->batch_data
;
1776 pipeline
->batch
.end
= pipeline
->batch
.start
+ sizeof(pipeline
->batch_data
);
1777 pipeline
->batch
.relocs
= &pipeline
->batch_relocs
;
1778 pipeline
->batch
.status
= VK_SUCCESS
;
1780 /* When we free the pipeline, we detect stages based on the NULL status
1781 * of various prog_data pointers. Make them NULL by default.
1783 memset(pipeline
->shaders
, 0, sizeof(pipeline
->shaders
));
1785 pipeline
->active_stages
= 0;
1787 pipeline
->needs_data_cache
= false;
1789 assert(pCreateInfo
->stage
.stage
== VK_SHADER_STAGE_COMPUTE_BIT
);
1790 ANV_FROM_HANDLE(anv_shader_module
, module
, pCreateInfo
->stage
.module
);
1791 result
= anv_pipeline_compile_cs(pipeline
, cache
, pCreateInfo
, module
,
1792 pCreateInfo
->stage
.pName
,
1793 pCreateInfo
->stage
.pSpecializationInfo
);
1794 if (result
!= VK_SUCCESS
) {
1795 vk_free2(&device
->alloc
, pAllocator
, pipeline
);
1799 const struct brw_cs_prog_data
*cs_prog_data
= get_cs_prog_data(pipeline
);
1801 anv_pipeline_setup_l3_config(pipeline
, cs_prog_data
->base
.total_shared
> 0);
1803 uint32_t group_size
= cs_prog_data
->local_size
[0] *
1804 cs_prog_data
->local_size
[1] * cs_prog_data
->local_size
[2];
1805 uint32_t remainder
= group_size
& (cs_prog_data
->simd_size
- 1);
1808 pipeline
->cs_right_mask
= ~0u >> (32 - remainder
);
1810 pipeline
->cs_right_mask
= ~0u >> (32 - cs_prog_data
->simd_size
);
1812 const uint32_t vfe_curbe_allocation
=
1813 ALIGN(cs_prog_data
->push
.per_thread
.regs
* cs_prog_data
->threads
+
1814 cs_prog_data
->push
.cross_thread
.regs
, 2);
1816 const uint32_t subslices
= MAX2(physical_device
->subslice_total
, 1);
1818 const struct anv_shader_bin
*cs_bin
=
1819 pipeline
->shaders
[MESA_SHADER_COMPUTE
];
1821 anv_batch_emit(&pipeline
->batch
, GENX(MEDIA_VFE_STATE
), vfe
) {
1825 vfe
.GPGPUMode
= true;
1827 vfe
.MaximumNumberofThreads
=
1828 devinfo
->max_cs_threads
* subslices
- 1;
1829 vfe
.NumberofURBEntries
= GEN_GEN
<= 7 ? 0 : 2;
1831 vfe
.ResetGatewayTimer
= true;
1834 vfe
.BypassGatewayControl
= true;
1836 vfe
.URBEntryAllocationSize
= GEN_GEN
<= 7 ? 0 : 2;
1837 vfe
.CURBEAllocationSize
= vfe_curbe_allocation
;
1839 vfe
.PerThreadScratchSpace
= get_scratch_space(cs_bin
);
1840 vfe
.ScratchSpaceBasePointer
=
1841 get_scratch_address(pipeline
, MESA_SHADER_COMPUTE
, cs_bin
);
1844 struct GENX(INTERFACE_DESCRIPTOR_DATA
) desc
= {
1845 .KernelStartPointer
= cs_bin
->kernel
.offset
,
1847 .SamplerCount
= get_sampler_count(cs_bin
),
1848 .BindingTableEntryCount
= get_binding_table_entry_count(cs_bin
),
1849 .BarrierEnable
= cs_prog_data
->uses_barrier
,
1850 .SharedLocalMemorySize
=
1851 encode_slm_size(GEN_GEN
, cs_prog_data
->base
.total_shared
),
1854 .ConstantURBEntryReadOffset
= 0,
1856 .ConstantURBEntryReadLength
= cs_prog_data
->push
.per_thread
.regs
,
1857 #if GEN_GEN >= 8 || GEN_IS_HASWELL
1858 .CrossThreadConstantDataReadLength
=
1859 cs_prog_data
->push
.cross_thread
.regs
,
1862 .NumberofThreadsinGPGPUThreadGroup
= cs_prog_data
->threads
,
1864 GENX(INTERFACE_DESCRIPTOR_DATA_pack
)(NULL
,
1865 pipeline
->interface_descriptor_data
,
1868 *pPipeline
= anv_pipeline_to_handle(pipeline
);
1870 return pipeline
->batch
.status
;
1873 VkResult
genX(CreateGraphicsPipelines
)(
1875 VkPipelineCache pipelineCache
,
1877 const VkGraphicsPipelineCreateInfo
* pCreateInfos
,
1878 const VkAllocationCallbacks
* pAllocator
,
1879 VkPipeline
* pPipelines
)
1881 ANV_FROM_HANDLE(anv_pipeline_cache
, pipeline_cache
, pipelineCache
);
1883 VkResult result
= VK_SUCCESS
;
1886 for (i
= 0; i
< count
; i
++) {
1887 result
= genX(graphics_pipeline_create
)(_device
,
1890 pAllocator
, &pPipelines
[i
]);
1892 /* Bail out on the first error as it is not obvious what error should be
1893 * report upon 2 different failures. */
1894 if (result
!= VK_SUCCESS
)
1898 for (; i
< count
; i
++)
1899 pPipelines
[i
] = VK_NULL_HANDLE
;
1904 VkResult
genX(CreateComputePipelines
)(
1906 VkPipelineCache pipelineCache
,
1908 const VkComputePipelineCreateInfo
* pCreateInfos
,
1909 const VkAllocationCallbacks
* pAllocator
,
1910 VkPipeline
* pPipelines
)
1912 ANV_FROM_HANDLE(anv_pipeline_cache
, pipeline_cache
, pipelineCache
);
1914 VkResult result
= VK_SUCCESS
;
1917 for (i
= 0; i
< count
; i
++) {
1918 result
= compute_pipeline_create(_device
, pipeline_cache
,
1920 pAllocator
, &pPipelines
[i
]);
1922 /* Bail out on the first error as it is not obvious what error should be
1923 * report upon 2 different failures. */
1924 if (result
!= VK_SUCCESS
)
1928 for (; i
< count
; i
++)
1929 pPipelines
[i
] = VK_NULL_HANDLE
;