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
27 #include "anv_private.h"
29 #include "genxml/gen_macros.h"
30 #include "genxml/genX_pack.h"
33 genX(cmd_buffer_emit_state_base_address
)(struct anv_cmd_buffer
*cmd_buffer
)
35 struct anv_device
*device
= cmd_buffer
->device
;
36 struct anv_bo
*scratch_bo
= NULL
;
38 cmd_buffer
->state
.scratch_size
=
39 anv_block_pool_size(&device
->scratch_block_pool
);
40 if (cmd_buffer
->state
.scratch_size
> 0)
41 scratch_bo
= &device
->scratch_block_pool
.bo
;
43 /* XXX: Do we need this on more than just BDW? */
45 /* Emit a render target cache flush.
47 * This isn't documented anywhere in the PRM. However, it seems to be
48 * necessary prior to changing the surface state base adress. Without
49 * this, we get GPU hangs when using multi-level command buffers which
50 * clear depth, reset state base address, and then go render stuff.
52 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
53 pc
.RenderTargetCacheFlushEnable
= true;
57 anv_batch_emit(&cmd_buffer
->batch
, GENX(STATE_BASE_ADDRESS
), sba
) {
58 sba
.GeneralStateBaseAddress
= (struct anv_address
) { scratch_bo
, 0 };
59 sba
.GeneralStateMemoryObjectControlState
= GENX(MOCS
);
60 sba
.GeneralStateBaseAddressModifyEnable
= true;
62 sba
.SurfaceStateBaseAddress
=
63 anv_cmd_buffer_surface_base_address(cmd_buffer
);
64 sba
.SurfaceStateMemoryObjectControlState
= GENX(MOCS
);
65 sba
.SurfaceStateBaseAddressModifyEnable
= true;
67 sba
.DynamicStateBaseAddress
=
68 (struct anv_address
) { &device
->dynamic_state_block_pool
.bo
, 0 };
69 sba
.DynamicStateMemoryObjectControlState
= GENX(MOCS
),
70 sba
.DynamicStateBaseAddressModifyEnable
= true,
72 sba
.IndirectObjectBaseAddress
= (struct anv_address
) { NULL
, 0 };
73 sba
.IndirectObjectMemoryObjectControlState
= GENX(MOCS
);
74 sba
.IndirectObjectBaseAddressModifyEnable
= true;
76 sba
.InstructionBaseAddress
=
77 (struct anv_address
) { &device
->instruction_block_pool
.bo
, 0 };
78 sba
.InstructionMemoryObjectControlState
= GENX(MOCS
);
79 sba
.InstructionBaseAddressModifyEnable
= true;
82 /* Broadwell requires that we specify a buffer size for a bunch of
83 * these fields. However, since we will be growing the BO's live, we
84 * just set them all to the maximum.
86 sba
.GeneralStateBufferSize
= 0xfffff;
87 sba
.GeneralStateBufferSizeModifyEnable
= true;
88 sba
.DynamicStateBufferSize
= 0xfffff;
89 sba
.DynamicStateBufferSizeModifyEnable
= true;
90 sba
.IndirectObjectBufferSize
= 0xfffff;
91 sba
.IndirectObjectBufferSizeModifyEnable
= true;
92 sba
.InstructionBufferSize
= 0xfffff;
93 sba
.InstructionBuffersizeModifyEnable
= true;
97 /* After re-setting the surface state base address, we have to do some
98 * cache flusing so that the sampler engine will pick up the new
99 * SURFACE_STATE objects and binding tables. From the Broadwell PRM,
100 * Shared Function > 3D Sampler > State > State Caching (page 96):
102 * Coherency with system memory in the state cache, like the texture
103 * cache is handled partially by software. It is expected that the
104 * command stream or shader will issue Cache Flush operation or
105 * Cache_Flush sampler message to ensure that the L1 cache remains
106 * coherent with system memory.
110 * Whenever the value of the Dynamic_State_Base_Addr,
111 * Surface_State_Base_Addr are altered, the L1 state cache must be
112 * invalidated to ensure the new surface or sampler state is fetched
113 * from system memory.
115 * The PIPE_CONTROL command has a "State Cache Invalidation Enable" bit
116 * which, according the PIPE_CONTROL instruction documentation in the
119 * Setting this bit is independent of any other bit in this packet.
120 * This bit controls the invalidation of the L1 and L2 state caches
121 * at the top of the pipe i.e. at the parsing time.
123 * Unfortunately, experimentation seems to indicate that state cache
124 * invalidation through a PIPE_CONTROL does nothing whatsoever in
125 * regards to surface state and binding tables. In stead, it seems that
126 * invalidating the texture cache is what is actually needed.
128 * XXX: As far as we have been able to determine through
129 * experimentation, shows that flush the texture cache appears to be
130 * sufficient. The theory here is that all of the sampling/rendering
131 * units cache the binding table in the texture cache. However, we have
132 * yet to be able to actually confirm this.
134 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
135 pc
.TextureCacheInvalidationEnable
= true;
139 void genX(CmdPipelineBarrier
)(
140 VkCommandBuffer commandBuffer
,
141 VkPipelineStageFlags srcStageMask
,
142 VkPipelineStageFlags destStageMask
,
144 uint32_t memoryBarrierCount
,
145 const VkMemoryBarrier
* pMemoryBarriers
,
146 uint32_t bufferMemoryBarrierCount
,
147 const VkBufferMemoryBarrier
* pBufferMemoryBarriers
,
148 uint32_t imageMemoryBarrierCount
,
149 const VkImageMemoryBarrier
* pImageMemoryBarriers
)
151 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
154 /* XXX: Right now, we're really dumb and just flush whatever categories
155 * the app asks for. One of these days we may make this a bit better
156 * but right now that's all the hardware allows for in most areas.
158 VkAccessFlags src_flags
= 0;
159 VkAccessFlags dst_flags
= 0;
161 for (uint32_t i
= 0; i
< memoryBarrierCount
; i
++) {
162 src_flags
|= pMemoryBarriers
[i
].srcAccessMask
;
163 dst_flags
|= pMemoryBarriers
[i
].dstAccessMask
;
166 for (uint32_t i
= 0; i
< bufferMemoryBarrierCount
; i
++) {
167 src_flags
|= pBufferMemoryBarriers
[i
].srcAccessMask
;
168 dst_flags
|= pBufferMemoryBarriers
[i
].dstAccessMask
;
171 for (uint32_t i
= 0; i
< imageMemoryBarrierCount
; i
++) {
172 src_flags
|= pImageMemoryBarriers
[i
].srcAccessMask
;
173 dst_flags
|= pImageMemoryBarriers
[i
].dstAccessMask
;
176 /* Mask out the Source access flags we care about */
177 const uint32_t src_mask
=
178 VK_ACCESS_SHADER_WRITE_BIT
|
179 VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT
|
180 VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT
|
181 VK_ACCESS_TRANSFER_WRITE_BIT
;
183 src_flags
= src_flags
& src_mask
;
185 /* Mask out the destination access flags we care about */
186 const uint32_t dst_mask
=
187 VK_ACCESS_INDIRECT_COMMAND_READ_BIT
|
188 VK_ACCESS_INDEX_READ_BIT
|
189 VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT
|
190 VK_ACCESS_UNIFORM_READ_BIT
|
191 VK_ACCESS_SHADER_READ_BIT
|
192 VK_ACCESS_COLOR_ATTACHMENT_READ_BIT
|
193 VK_ACCESS_TRANSFER_READ_BIT
;
195 dst_flags
= dst_flags
& dst_mask
;
197 /* The src flags represent how things were used previously. This is
198 * what we use for doing flushes.
200 struct GENX(PIPE_CONTROL
) flush_cmd
= {
201 GENX(PIPE_CONTROL_header
),
202 .PostSyncOperation
= NoWrite
,
205 for_each_bit(b
, src_flags
) {
206 switch ((VkAccessFlagBits
)(1 << b
)) {
207 case VK_ACCESS_SHADER_WRITE_BIT
:
208 flush_cmd
.DCFlushEnable
= true;
210 case VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT
:
211 flush_cmd
.RenderTargetCacheFlushEnable
= true;
213 case VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT
:
214 flush_cmd
.DepthCacheFlushEnable
= true;
216 case VK_ACCESS_TRANSFER_WRITE_BIT
:
217 flush_cmd
.RenderTargetCacheFlushEnable
= true;
218 flush_cmd
.DepthCacheFlushEnable
= true;
221 unreachable("should've masked this out by now");
225 /* If we end up doing two PIPE_CONTROLs, the first, flusing one also has to
226 * stall and wait for the flushing to finish, so we don't re-dirty the
227 * caches with in-flight rendering after the second PIPE_CONTROL
232 flush_cmd
.CommandStreamerStallEnable
= true;
234 if (src_flags
&& dst_flags
) {
235 dw
= anv_batch_emit_dwords(&cmd_buffer
->batch
, GENX(PIPE_CONTROL_length
));
236 GENX(PIPE_CONTROL_pack
)(&cmd_buffer
->batch
, dw
, &flush_cmd
);
239 /* The dst flags represent how things will be used in the future. This
240 * is what we use for doing cache invalidations.
242 struct GENX(PIPE_CONTROL
) invalidate_cmd
= {
243 GENX(PIPE_CONTROL_header
),
244 .PostSyncOperation
= NoWrite
,
247 for_each_bit(b
, dst_flags
) {
248 switch ((VkAccessFlagBits
)(1 << b
)) {
249 case VK_ACCESS_INDIRECT_COMMAND_READ_BIT
:
250 case VK_ACCESS_INDEX_READ_BIT
:
251 case VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT
:
252 invalidate_cmd
.VFCacheInvalidationEnable
= true;
254 case VK_ACCESS_UNIFORM_READ_BIT
:
255 invalidate_cmd
.ConstantCacheInvalidationEnable
= true;
257 case VK_ACCESS_SHADER_READ_BIT
:
258 invalidate_cmd
.TextureCacheInvalidationEnable
= true;
260 case VK_ACCESS_COLOR_ATTACHMENT_READ_BIT
:
261 invalidate_cmd
.TextureCacheInvalidationEnable
= true;
263 case VK_ACCESS_TRANSFER_READ_BIT
:
264 invalidate_cmd
.TextureCacheInvalidationEnable
= true;
267 unreachable("should've masked this out by now");
272 dw
= anv_batch_emit_dwords(&cmd_buffer
->batch
, GENX(PIPE_CONTROL_length
));
273 GENX(PIPE_CONTROL_pack
)(&cmd_buffer
->batch
, dw
, &invalidate_cmd
);
278 cmd_buffer_flush_push_constants(struct anv_cmd_buffer
*cmd_buffer
)
280 static const uint32_t push_constant_opcodes
[] = {
281 [MESA_SHADER_VERTEX
] = 21,
282 [MESA_SHADER_TESS_CTRL
] = 25, /* HS */
283 [MESA_SHADER_TESS_EVAL
] = 26, /* DS */
284 [MESA_SHADER_GEOMETRY
] = 22,
285 [MESA_SHADER_FRAGMENT
] = 23,
286 [MESA_SHADER_COMPUTE
] = 0,
289 VkShaderStageFlags flushed
= 0;
291 anv_foreach_stage(stage
, cmd_buffer
->state
.push_constants_dirty
) {
292 if (stage
== MESA_SHADER_COMPUTE
)
295 struct anv_state state
= anv_cmd_buffer_push_constants(cmd_buffer
, stage
);
297 if (state
.offset
== 0) {
298 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_CONSTANT_VS
), c
)
299 c
._3DCommandSubOpcode
= push_constant_opcodes
[stage
];
301 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_CONSTANT_VS
), c
) {
302 c
._3DCommandSubOpcode
= push_constant_opcodes
[stage
],
303 c
.ConstantBody
= (struct GENX(3DSTATE_CONSTANT_BODY
)) {
305 .PointerToConstantBuffer2
= { &cmd_buffer
->device
->dynamic_state_block_pool
.bo
, state
.offset
},
306 .ConstantBuffer2ReadLength
= DIV_ROUND_UP(state
.alloc_size
, 32),
308 .PointerToConstantBuffer0
= { .offset
= state
.offset
},
309 .ConstantBuffer0ReadLength
= DIV_ROUND_UP(state
.alloc_size
, 32),
315 flushed
|= mesa_to_vk_shader_stage(stage
);
318 cmd_buffer
->state
.push_constants_dirty
&= ~VK_SHADER_STAGE_ALL_GRAPHICS
;
324 genX(cmd_buffer_flush_state
)(struct anv_cmd_buffer
*cmd_buffer
)
326 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.pipeline
;
329 uint32_t vb_emit
= cmd_buffer
->state
.vb_dirty
& pipeline
->vb_used
;
331 assert((pipeline
->active_stages
& VK_SHADER_STAGE_COMPUTE_BIT
) == 0);
333 genX(cmd_buffer_config_l3
)(cmd_buffer
, pipeline
);
335 genX(flush_pipeline_select_3d
)(cmd_buffer
);
338 const uint32_t num_buffers
= __builtin_popcount(vb_emit
);
339 const uint32_t num_dwords
= 1 + num_buffers
* 4;
341 p
= anv_batch_emitn(&cmd_buffer
->batch
, num_dwords
,
342 GENX(3DSTATE_VERTEX_BUFFERS
));
344 for_each_bit(vb
, vb_emit
) {
345 struct anv_buffer
*buffer
= cmd_buffer
->state
.vertex_bindings
[vb
].buffer
;
346 uint32_t offset
= cmd_buffer
->state
.vertex_bindings
[vb
].offset
;
348 struct GENX(VERTEX_BUFFER_STATE
) state
= {
349 .VertexBufferIndex
= vb
,
352 .MemoryObjectControlState
= GENX(MOCS
),
354 .BufferAccessType
= pipeline
->instancing_enable
[vb
] ? INSTANCEDATA
: VERTEXDATA
,
355 .InstanceDataStepRate
= 1,
356 .VertexBufferMemoryObjectControlState
= GENX(MOCS
),
359 .AddressModifyEnable
= true,
360 .BufferPitch
= pipeline
->binding_stride
[vb
],
361 .BufferStartingAddress
= { buffer
->bo
, buffer
->offset
+ offset
},
364 .BufferSize
= buffer
->size
- offset
366 .EndAddress
= { buffer
->bo
, buffer
->offset
+ buffer
->size
- 1},
370 GENX(VERTEX_BUFFER_STATE_pack
)(&cmd_buffer
->batch
, &p
[1 + i
* 4], &state
);
375 cmd_buffer
->state
.vb_dirty
&= ~vb_emit
;
377 if (cmd_buffer
->state
.dirty
& ANV_CMD_DIRTY_PIPELINE
) {
378 /* If somebody compiled a pipeline after starting a command buffer the
379 * scratch bo may have grown since we started this cmd buffer (and
380 * emitted STATE_BASE_ADDRESS). If we're binding that pipeline now,
381 * reemit STATE_BASE_ADDRESS so that we use the bigger scratch bo. */
382 if (cmd_buffer
->state
.scratch_size
< pipeline
->total_scratch
)
383 anv_cmd_buffer_emit_state_base_address(cmd_buffer
);
385 anv_batch_emit_batch(&cmd_buffer
->batch
, &pipeline
->batch
);
387 /* From the BDW PRM for 3DSTATE_PUSH_CONSTANT_ALLOC_VS:
389 * "The 3DSTATE_CONSTANT_VS must be reprogrammed prior to
390 * the next 3DPRIMITIVE command after programming the
391 * 3DSTATE_PUSH_CONSTANT_ALLOC_VS"
393 * Since 3DSTATE_PUSH_CONSTANT_ALLOC_VS is programmed as part of
394 * pipeline setup, we need to dirty push constants.
396 cmd_buffer
->state
.push_constants_dirty
|= VK_SHADER_STAGE_ALL_GRAPHICS
;
400 if (cmd_buffer
->state
.descriptors_dirty
& VK_SHADER_STAGE_VERTEX_BIT
||
401 cmd_buffer
->state
.push_constants_dirty
& VK_SHADER_STAGE_VERTEX_BIT
) {
402 /* From the IVB PRM Vol. 2, Part 1, Section 3.2.1:
404 * "A PIPE_CONTROL with Post-Sync Operation set to 1h and a depth
405 * stall needs to be sent just prior to any 3DSTATE_VS,
406 * 3DSTATE_URB_VS, 3DSTATE_CONSTANT_VS,
407 * 3DSTATE_BINDING_TABLE_POINTER_VS,
408 * 3DSTATE_SAMPLER_STATE_POINTER_VS command. Only one
409 * PIPE_CONTROL needs to be sent before any combination of VS
410 * associated 3DSTATE."
412 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
413 pc
.DepthStallEnable
= true;
414 pc
.PostSyncOperation
= WriteImmediateData
;
416 (struct anv_address
) { &cmd_buffer
->device
->workaround_bo
, 0 };
421 /* We emit the binding tables and sampler tables first, then emit push
422 * constants and then finally emit binding table and sampler table
423 * pointers. It has to happen in this order, since emitting the binding
424 * tables may change the push constants (in case of storage images). After
425 * emitting push constants, on SKL+ we have to emit the corresponding
426 * 3DSTATE_BINDING_TABLE_POINTER_* for the push constants to take effect.
429 if (cmd_buffer
->state
.descriptors_dirty
)
430 dirty
= gen7_cmd_buffer_flush_descriptor_sets(cmd_buffer
);
432 if (cmd_buffer
->state
.push_constants_dirty
) {
434 /* On Sky Lake and later, the binding table pointers commands are
435 * what actually flush the changes to push constant state so we need
436 * to dirty them so they get re-emitted below.
438 dirty
|= cmd_buffer_flush_push_constants(cmd_buffer
);
440 cmd_buffer_flush_push_constants(cmd_buffer
);
445 gen7_cmd_buffer_emit_descriptor_pointers(cmd_buffer
, dirty
);
447 if (cmd_buffer
->state
.dirty
& ANV_CMD_DIRTY_DYNAMIC_VIEWPORT
)
448 gen8_cmd_buffer_emit_viewport(cmd_buffer
);
450 if (cmd_buffer
->state
.dirty
& ANV_CMD_DIRTY_DYNAMIC_SCISSOR
)
451 gen7_cmd_buffer_emit_scissor(cmd_buffer
);
453 genX(cmd_buffer_flush_dynamic_state
)(cmd_buffer
);
457 emit_base_vertex_instance_bo(struct anv_cmd_buffer
*cmd_buffer
,
458 struct anv_bo
*bo
, uint32_t offset
)
460 uint32_t *p
= anv_batch_emitn(&cmd_buffer
->batch
, 5,
461 GENX(3DSTATE_VERTEX_BUFFERS
));
463 GENX(VERTEX_BUFFER_STATE_pack
)(&cmd_buffer
->batch
, p
+ 1,
464 &(struct GENX(VERTEX_BUFFER_STATE
)) {
465 .VertexBufferIndex
= 32, /* Reserved for this */
466 .AddressModifyEnable
= true,
469 .MemoryObjectControlState
= GENX(MOCS
),
470 .BufferStartingAddress
= { bo
, offset
},
473 .VertexBufferMemoryObjectControlState
= GENX(MOCS
),
474 .BufferStartingAddress
= { bo
, offset
},
475 .EndAddress
= { bo
, offset
+ 8 },
481 emit_base_vertex_instance(struct anv_cmd_buffer
*cmd_buffer
,
482 uint32_t base_vertex
, uint32_t base_instance
)
484 struct anv_state id_state
=
485 anv_cmd_buffer_alloc_dynamic_state(cmd_buffer
, 8, 4);
487 ((uint32_t *)id_state
.map
)[0] = base_vertex
;
488 ((uint32_t *)id_state
.map
)[1] = base_instance
;
490 if (!cmd_buffer
->device
->info
.has_llc
)
491 anv_state_clflush(id_state
);
493 emit_base_vertex_instance_bo(cmd_buffer
,
494 &cmd_buffer
->device
->dynamic_state_block_pool
.bo
, id_state
.offset
);
498 VkCommandBuffer commandBuffer
,
499 uint32_t vertexCount
,
500 uint32_t instanceCount
,
501 uint32_t firstVertex
,
502 uint32_t firstInstance
)
504 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
505 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.pipeline
;
506 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
508 genX(cmd_buffer_flush_state
)(cmd_buffer
);
510 if (vs_prog_data
->uses_basevertex
|| vs_prog_data
->uses_baseinstance
)
511 emit_base_vertex_instance(cmd_buffer
, firstVertex
, firstInstance
);
513 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
514 prim
.VertexAccessType
= SEQUENTIAL
;
515 prim
.PrimitiveTopologyType
= pipeline
->topology
;
516 prim
.VertexCountPerInstance
= vertexCount
;
517 prim
.StartVertexLocation
= firstVertex
;
518 prim
.InstanceCount
= instanceCount
;
519 prim
.StartInstanceLocation
= firstInstance
;
520 prim
.BaseVertexLocation
= 0;
524 void genX(CmdDrawIndexed
)(
525 VkCommandBuffer commandBuffer
,
527 uint32_t instanceCount
,
529 int32_t vertexOffset
,
530 uint32_t firstInstance
)
532 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
533 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.pipeline
;
534 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
536 genX(cmd_buffer_flush_state
)(cmd_buffer
);
538 if (vs_prog_data
->uses_basevertex
|| vs_prog_data
->uses_baseinstance
)
539 emit_base_vertex_instance(cmd_buffer
, vertexOffset
, firstInstance
);
541 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
542 prim
.VertexAccessType
= RANDOM
;
543 prim
.PrimitiveTopologyType
= pipeline
->topology
;
544 prim
.VertexCountPerInstance
= indexCount
;
545 prim
.StartVertexLocation
= firstIndex
;
546 prim
.InstanceCount
= instanceCount
;
547 prim
.StartInstanceLocation
= firstInstance
;
548 prim
.BaseVertexLocation
= vertexOffset
;
552 /* Auto-Draw / Indirect Registers */
553 #define GEN7_3DPRIM_END_OFFSET 0x2420
554 #define GEN7_3DPRIM_START_VERTEX 0x2430
555 #define GEN7_3DPRIM_VERTEX_COUNT 0x2434
556 #define GEN7_3DPRIM_INSTANCE_COUNT 0x2438
557 #define GEN7_3DPRIM_START_INSTANCE 0x243C
558 #define GEN7_3DPRIM_BASE_VERTEX 0x2440
561 emit_lrm(struct anv_batch
*batch
,
562 uint32_t reg
, struct anv_bo
*bo
, uint32_t offset
)
564 anv_batch_emit(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
565 lrm
.RegisterAddress
= reg
;
566 lrm
.MemoryAddress
= (struct anv_address
) { bo
, offset
};
571 emit_lri(struct anv_batch
*batch
, uint32_t reg
, uint32_t imm
)
573 anv_batch_emit(batch
, GENX(MI_LOAD_REGISTER_IMM
), lri
) {
574 lri
.RegisterOffset
= reg
;
579 void genX(CmdDrawIndirect
)(
580 VkCommandBuffer commandBuffer
,
586 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
587 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
588 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.pipeline
;
589 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
590 struct anv_bo
*bo
= buffer
->bo
;
591 uint32_t bo_offset
= buffer
->offset
+ offset
;
593 genX(cmd_buffer_flush_state
)(cmd_buffer
);
595 if (vs_prog_data
->uses_basevertex
|| vs_prog_data
->uses_baseinstance
)
596 emit_base_vertex_instance_bo(cmd_buffer
, bo
, bo_offset
+ 8);
598 emit_lrm(&cmd_buffer
->batch
, GEN7_3DPRIM_VERTEX_COUNT
, bo
, bo_offset
);
599 emit_lrm(&cmd_buffer
->batch
, GEN7_3DPRIM_INSTANCE_COUNT
, bo
, bo_offset
+ 4);
600 emit_lrm(&cmd_buffer
->batch
, GEN7_3DPRIM_START_VERTEX
, bo
, bo_offset
+ 8);
601 emit_lrm(&cmd_buffer
->batch
, GEN7_3DPRIM_START_INSTANCE
, bo
, bo_offset
+ 12);
602 emit_lri(&cmd_buffer
->batch
, GEN7_3DPRIM_BASE_VERTEX
, 0);
604 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
605 prim
.IndirectParameterEnable
= true;
606 prim
.VertexAccessType
= SEQUENTIAL
;
607 prim
.PrimitiveTopologyType
= pipeline
->topology
;
611 void genX(CmdDrawIndexedIndirect
)(
612 VkCommandBuffer commandBuffer
,
618 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
619 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
620 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.pipeline
;
621 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
622 struct anv_bo
*bo
= buffer
->bo
;
623 uint32_t bo_offset
= buffer
->offset
+ offset
;
625 genX(cmd_buffer_flush_state
)(cmd_buffer
);
627 /* TODO: We need to stomp base vertex to 0 somehow */
628 if (vs_prog_data
->uses_basevertex
|| vs_prog_data
->uses_baseinstance
)
629 emit_base_vertex_instance_bo(cmd_buffer
, bo
, bo_offset
+ 12);
631 emit_lrm(&cmd_buffer
->batch
, GEN7_3DPRIM_VERTEX_COUNT
, bo
, bo_offset
);
632 emit_lrm(&cmd_buffer
->batch
, GEN7_3DPRIM_INSTANCE_COUNT
, bo
, bo_offset
+ 4);
633 emit_lrm(&cmd_buffer
->batch
, GEN7_3DPRIM_START_VERTEX
, bo
, bo_offset
+ 8);
634 emit_lrm(&cmd_buffer
->batch
, GEN7_3DPRIM_BASE_VERTEX
, bo
, bo_offset
+ 12);
635 emit_lrm(&cmd_buffer
->batch
, GEN7_3DPRIM_START_INSTANCE
, bo
, bo_offset
+ 16);
637 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
638 prim
.IndirectParameterEnable
= true;
639 prim
.VertexAccessType
= RANDOM
;
640 prim
.PrimitiveTopologyType
= pipeline
->topology
;
647 verify_cmd_parser(const struct anv_device
*device
,
648 int required_version
,
649 const char *function
)
651 if (device
->instance
->physicalDevice
.cmd_parser_version
< required_version
) {
652 vk_errorf(VK_ERROR_FEATURE_NOT_PRESENT
,
653 "cmd parser version %d is required for %s",
654 required_version
, function
);
663 void genX(CmdDispatch
)(
664 VkCommandBuffer commandBuffer
,
669 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
670 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.compute_pipeline
;
671 const struct brw_cs_prog_data
*prog_data
= get_cs_prog_data(pipeline
);
673 if (prog_data
->uses_num_work_groups
) {
674 struct anv_state state
=
675 anv_cmd_buffer_alloc_dynamic_state(cmd_buffer
, 12, 4);
676 uint32_t *sizes
= state
.map
;
680 if (!cmd_buffer
->device
->info
.has_llc
)
681 anv_state_clflush(state
);
682 cmd_buffer
->state
.num_workgroups_offset
= state
.offset
;
683 cmd_buffer
->state
.num_workgroups_bo
=
684 &cmd_buffer
->device
->dynamic_state_block_pool
.bo
;
687 genX(cmd_buffer_flush_compute_state
)(cmd_buffer
);
689 anv_batch_emit(&cmd_buffer
->batch
, GENX(GPGPU_WALKER
), ggw
) {
690 ggw
.SIMDSize
= prog_data
->simd_size
/ 16;
691 ggw
.ThreadDepthCounterMaximum
= 0;
692 ggw
.ThreadHeightCounterMaximum
= 0;
693 ggw
.ThreadWidthCounterMaximum
= pipeline
->cs_thread_width_max
- 1;
694 ggw
.ThreadGroupIDXDimension
= x
;
695 ggw
.ThreadGroupIDYDimension
= y
;
696 ggw
.ThreadGroupIDZDimension
= z
;
697 ggw
.RightExecutionMask
= pipeline
->cs_right_mask
;
698 ggw
.BottomExecutionMask
= 0xffffffff;
701 anv_batch_emit(&cmd_buffer
->batch
, GENX(MEDIA_STATE_FLUSH
), msf
);
704 #define GPGPU_DISPATCHDIMX 0x2500
705 #define GPGPU_DISPATCHDIMY 0x2504
706 #define GPGPU_DISPATCHDIMZ 0x2508
708 #define MI_PREDICATE_SRC0 0x2400
709 #define MI_PREDICATE_SRC1 0x2408
711 void genX(CmdDispatchIndirect
)(
712 VkCommandBuffer commandBuffer
,
716 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
717 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
718 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.compute_pipeline
;
719 const struct brw_cs_prog_data
*prog_data
= get_cs_prog_data(pipeline
);
720 struct anv_bo
*bo
= buffer
->bo
;
721 uint32_t bo_offset
= buffer
->offset
+ offset
;
722 struct anv_batch
*batch
= &cmd_buffer
->batch
;
725 /* Linux 4.4 added command parser version 5 which allows the GPGPU
726 * indirect dispatch registers to be written.
728 if (!verify_cmd_parser(cmd_buffer
->device
, 5, "vkCmdDispatchIndirect"))
732 if (prog_data
->uses_num_work_groups
) {
733 cmd_buffer
->state
.num_workgroups_offset
= bo_offset
;
734 cmd_buffer
->state
.num_workgroups_bo
= bo
;
737 genX(cmd_buffer_flush_compute_state
)(cmd_buffer
);
739 emit_lrm(batch
, GPGPU_DISPATCHDIMX
, bo
, bo_offset
);
740 emit_lrm(batch
, GPGPU_DISPATCHDIMY
, bo
, bo_offset
+ 4);
741 emit_lrm(batch
, GPGPU_DISPATCHDIMZ
, bo
, bo_offset
+ 8);
744 /* Clear upper 32-bits of SRC0 and all 64-bits of SRC1 */
745 emit_lri(batch
, MI_PREDICATE_SRC0
+ 4, 0);
746 emit_lri(batch
, MI_PREDICATE_SRC1
+ 0, 0);
747 emit_lri(batch
, MI_PREDICATE_SRC1
+ 4, 0);
749 /* Load compute_dispatch_indirect_x_size into SRC0 */
750 emit_lrm(batch
, MI_PREDICATE_SRC0
, bo
, bo_offset
+ 0);
752 /* predicate = (compute_dispatch_indirect_x_size == 0); */
753 anv_batch_emit(batch
, GENX(MI_PREDICATE
), mip
) {
754 mip
.LoadOperation
= LOAD_LOAD
;
755 mip
.CombineOperation
= COMBINE_SET
;
756 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
759 /* Load compute_dispatch_indirect_y_size into SRC0 */
760 emit_lrm(batch
, MI_PREDICATE_SRC0
, bo
, bo_offset
+ 4);
762 /* predicate |= (compute_dispatch_indirect_y_size == 0); */
763 anv_batch_emit(batch
, GENX(MI_PREDICATE
), mip
) {
764 mip
.LoadOperation
= LOAD_LOAD
;
765 mip
.CombineOperation
= COMBINE_OR
;
766 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
769 /* Load compute_dispatch_indirect_z_size into SRC0 */
770 emit_lrm(batch
, MI_PREDICATE_SRC0
, bo
, bo_offset
+ 8);
772 /* predicate |= (compute_dispatch_indirect_z_size == 0); */
773 anv_batch_emit(batch
, GENX(MI_PREDICATE
), mip
) {
774 mip
.LoadOperation
= LOAD_LOAD
;
775 mip
.CombineOperation
= COMBINE_OR
;
776 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
779 /* predicate = !predicate; */
780 #define COMPARE_FALSE 1
781 anv_batch_emit(batch
, GENX(MI_PREDICATE
), mip
) {
782 mip
.LoadOperation
= LOAD_LOADINV
;
783 mip
.CombineOperation
= COMBINE_OR
;
784 mip
.CompareOperation
= COMPARE_FALSE
;
788 anv_batch_emit(batch
, GENX(GPGPU_WALKER
), ggw
) {
789 ggw
.IndirectParameterEnable
= true;
790 ggw
.PredicateEnable
= GEN_GEN
<= 7;
791 ggw
.SIMDSize
= prog_data
->simd_size
/ 16;
792 ggw
.ThreadDepthCounterMaximum
= 0;
793 ggw
.ThreadHeightCounterMaximum
= 0;
794 ggw
.ThreadWidthCounterMaximum
= pipeline
->cs_thread_width_max
- 1;
795 ggw
.RightExecutionMask
= pipeline
->cs_right_mask
;
796 ggw
.BottomExecutionMask
= 0xffffffff;
799 anv_batch_emit(batch
, GENX(MEDIA_STATE_FLUSH
), msf
);
803 flush_pipeline_before_pipeline_select(struct anv_cmd_buffer
*cmd_buffer
,
806 #if GEN_GEN >= 8 && GEN_GEN < 10
807 /* From the Broadwell PRM, Volume 2a: Instructions, PIPELINE_SELECT:
809 * Software must clear the COLOR_CALC_STATE Valid field in
810 * 3DSTATE_CC_STATE_POINTERS command prior to send a PIPELINE_SELECT
811 * with Pipeline Select set to GPGPU.
813 * The internal hardware docs recommend the same workaround for Gen9
816 if (pipeline
== GPGPU
)
817 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_CC_STATE_POINTERS
), t
);
819 /* From "BXML » GT » MI » vol1a GPU Overview » [Instruction]
820 * PIPELINE_SELECT [DevBWR+]":
824 * Software must ensure all the write caches are flushed through a
825 * stalling PIPE_CONTROL command followed by another PIPE_CONTROL
826 * command to invalidate read only caches prior to programming
827 * MI_PIPELINE_SELECT command to change the Pipeline Select Mode.
829 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
830 pc
.RenderTargetCacheFlushEnable
= true;
831 pc
.DepthCacheFlushEnable
= true;
832 pc
.DCFlushEnable
= true;
833 pc
.PostSyncOperation
= NoWrite
;
834 pc
.CommandStreamerStallEnable
= true;
837 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
838 pc
.TextureCacheInvalidationEnable
= true;
839 pc
.ConstantCacheInvalidationEnable
= true;
840 pc
.StateCacheInvalidationEnable
= true;
841 pc
.InstructionCacheInvalidateEnable
= true;
842 pc
.PostSyncOperation
= NoWrite
;
848 genX(flush_pipeline_select_3d
)(struct anv_cmd_buffer
*cmd_buffer
)
850 if (cmd_buffer
->state
.current_pipeline
!= _3D
) {
851 flush_pipeline_before_pipeline_select(cmd_buffer
, _3D
);
853 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPELINE_SELECT
), ps
) {
857 ps
.PipelineSelection
= _3D
;
860 cmd_buffer
->state
.current_pipeline
= _3D
;
865 genX(flush_pipeline_select_gpgpu
)(struct anv_cmd_buffer
*cmd_buffer
)
867 if (cmd_buffer
->state
.current_pipeline
!= GPGPU
) {
868 flush_pipeline_before_pipeline_select(cmd_buffer
, GPGPU
);
870 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPELINE_SELECT
), ps
) {
874 ps
.PipelineSelection
= GPGPU
;
877 cmd_buffer
->state
.current_pipeline
= GPGPU
;
882 genX(cmd_buffer_alloc_null_surface_state
)(struct anv_cmd_buffer
*cmd_buffer
,
883 struct anv_framebuffer
*fb
)
885 struct anv_state state
=
886 anv_state_stream_alloc(&cmd_buffer
->surface_state_stream
, 64, 64);
888 struct GENX(RENDER_SURFACE_STATE
) null_ss
= {
889 .SurfaceType
= SURFTYPE_NULL
,
890 .SurfaceArray
= fb
->layers
> 0,
891 .SurfaceFormat
= ISL_FORMAT_R8G8B8A8_UNORM
,
895 .TiledSurface
= true,
897 .Width
= fb
->width
- 1,
898 .Height
= fb
->height
- 1,
899 .Depth
= fb
->layers
- 1,
900 .RenderTargetViewExtent
= fb
->layers
- 1,
903 GENX(RENDER_SURFACE_STATE_pack
)(NULL
, state
.map
, &null_ss
);
905 if (!cmd_buffer
->device
->info
.has_llc
)
906 anv_state_clflush(state
);
912 cmd_buffer_emit_depth_stencil(struct anv_cmd_buffer
*cmd_buffer
)
914 struct anv_device
*device
= cmd_buffer
->device
;
915 const struct anv_framebuffer
*fb
= cmd_buffer
->state
.framebuffer
;
916 const struct anv_image_view
*iview
=
917 anv_cmd_buffer_get_depth_stencil_view(cmd_buffer
);
918 const struct anv_image
*image
= iview
? iview
->image
: NULL
;
919 const bool has_depth
= image
&& (image
->aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
);
920 const bool has_stencil
=
921 image
&& (image
->aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
);
923 /* FIXME: Implement the PMA stall W/A */
924 /* FIXME: Width and Height are wrong */
926 /* Emit 3DSTATE_DEPTH_BUFFER */
928 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_DEPTH_BUFFER
), db
) {
929 db
.SurfaceType
= SURFTYPE_2D
;
930 db
.DepthWriteEnable
= true;
931 db
.StencilWriteEnable
= has_stencil
;
932 db
.HierarchicalDepthBufferEnable
= false;
934 db
.SurfaceFormat
= isl_surf_get_depth_format(&device
->isl_dev
,
935 &image
->depth_surface
.isl
);
937 db
.SurfaceBaseAddress
= (struct anv_address
) {
939 .offset
= image
->offset
+ image
->depth_surface
.offset
,
941 db
.DepthBufferObjectControlState
= GENX(MOCS
),
943 db
.SurfacePitch
= image
->depth_surface
.isl
.row_pitch
- 1;
944 db
.Height
= fb
->height
- 1;
945 db
.Width
= fb
->width
- 1;
948 db
.MinimumArrayElement
= 0;
952 isl_surf_get_array_pitch_el_rows(&image
->depth_surface
.isl
) >> 2,
954 db
.RenderTargetViewExtent
= 1 - 1;
957 /* Even when no depth buffer is present, the hardware requires that
958 * 3DSTATE_DEPTH_BUFFER be programmed correctly. The Broadwell PRM says:
960 * If a null depth buffer is bound, the driver must instead bind depth as:
961 * 3DSTATE_DEPTH.SurfaceType = SURFTYPE_2D
962 * 3DSTATE_DEPTH.Width = 1
963 * 3DSTATE_DEPTH.Height = 1
964 * 3DSTATE_DEPTH.SuraceFormat = D16_UNORM
965 * 3DSTATE_DEPTH.SurfaceBaseAddress = 0
966 * 3DSTATE_DEPTH.HierarchicalDepthBufferEnable = 0
967 * 3DSTATE_WM_DEPTH_STENCIL.DepthTestEnable = 0
968 * 3DSTATE_WM_DEPTH_STENCIL.DepthBufferWriteEnable = 0
970 * The PRM is wrong, though. The width and height must be programmed to
971 * actual framebuffer's width and height, even when neither depth buffer
972 * nor stencil buffer is present. Also, D16_UNORM is not allowed to
973 * be combined with a stencil buffer so we use D32_FLOAT instead.
975 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_DEPTH_BUFFER
), db
) {
976 db
.SurfaceType
= SURFTYPE_2D
;
977 db
.SurfaceFormat
= D32_FLOAT
;
978 db
.Width
= fb
->width
- 1;
979 db
.Height
= fb
->height
- 1;
980 db
.StencilWriteEnable
= has_stencil
;
984 /* Emit 3DSTATE_STENCIL_BUFFER */
986 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_STENCIL_BUFFER
), sb
) {
987 #if GEN_GEN >= 8 || GEN_IS_HASWELL
988 sb
.StencilBufferEnable
= true,
990 sb
.StencilBufferObjectControlState
= GENX(MOCS
),
992 /* Stencil buffers have strange pitch. The PRM says:
994 * The pitch must be set to 2x the value computed based on width,
995 * as the stencil buffer is stored with two rows interleaved.
997 sb
.SurfacePitch
= 2 * image
->stencil_surface
.isl
.row_pitch
- 1,
1000 sb
.SurfaceQPitch
= isl_surf_get_array_pitch_el_rows(&image
->stencil_surface
.isl
) >> 2,
1002 sb
.SurfaceBaseAddress
= (struct anv_address
) {
1004 .offset
= image
->offset
+ image
->stencil_surface
.offset
,
1008 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_STENCIL_BUFFER
), sb
);
1011 /* Disable hierarchial depth buffers. */
1012 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_HIER_DEPTH_BUFFER
), hz
);
1014 /* Clear the clear params. */
1015 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_CLEAR_PARAMS
), cp
);
1019 * @see anv_cmd_buffer_set_subpass()
1022 genX(cmd_buffer_set_subpass
)(struct anv_cmd_buffer
*cmd_buffer
,
1023 struct anv_subpass
*subpass
)
1025 cmd_buffer
->state
.subpass
= subpass
;
1027 cmd_buffer
->state
.descriptors_dirty
|= VK_SHADER_STAGE_FRAGMENT_BIT
;
1029 cmd_buffer_emit_depth_stencil(cmd_buffer
);
1032 void genX(CmdBeginRenderPass
)(
1033 VkCommandBuffer commandBuffer
,
1034 const VkRenderPassBeginInfo
* pRenderPassBegin
,
1035 VkSubpassContents contents
)
1037 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
1038 ANV_FROM_HANDLE(anv_render_pass
, pass
, pRenderPassBegin
->renderPass
);
1039 ANV_FROM_HANDLE(anv_framebuffer
, framebuffer
, pRenderPassBegin
->framebuffer
);
1041 cmd_buffer
->state
.framebuffer
= framebuffer
;
1042 cmd_buffer
->state
.pass
= pass
;
1043 anv_cmd_state_setup_attachments(cmd_buffer
, pRenderPassBegin
);
1045 genX(flush_pipeline_select_3d
)(cmd_buffer
);
1047 const VkRect2D
*render_area
= &pRenderPassBegin
->renderArea
;
1049 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_DRAWING_RECTANGLE
), r
) {
1050 r
.ClippedDrawingRectangleYMin
= MAX2(render_area
->offset
.y
, 0);
1051 r
.ClippedDrawingRectangleXMin
= MAX2(render_area
->offset
.x
, 0);
1052 r
.ClippedDrawingRectangleYMax
=
1053 render_area
->offset
.y
+ render_area
->extent
.height
- 1;
1054 r
.ClippedDrawingRectangleXMax
=
1055 render_area
->offset
.x
+ render_area
->extent
.width
- 1;
1056 r
.DrawingRectangleOriginY
= 0;
1057 r
.DrawingRectangleOriginX
= 0;
1060 genX(cmd_buffer_set_subpass
)(cmd_buffer
, pass
->subpasses
);
1061 anv_cmd_buffer_clear_subpass(cmd_buffer
);
1064 void genX(CmdNextSubpass
)(
1065 VkCommandBuffer commandBuffer
,
1066 VkSubpassContents contents
)
1068 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
1070 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
1072 anv_cmd_buffer_resolve_subpass(cmd_buffer
);
1073 genX(cmd_buffer_set_subpass
)(cmd_buffer
, cmd_buffer
->state
.subpass
+ 1);
1074 anv_cmd_buffer_clear_subpass(cmd_buffer
);
1077 void genX(CmdEndRenderPass
)(
1078 VkCommandBuffer commandBuffer
)
1080 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
1082 anv_cmd_buffer_resolve_subpass(cmd_buffer
);
1086 emit_ps_depth_count(struct anv_batch
*batch
,
1087 struct anv_bo
*bo
, uint32_t offset
)
1089 anv_batch_emit(batch
, GENX(PIPE_CONTROL
), pc
) {
1090 pc
.DestinationAddressType
= DAT_PPGTT
;
1091 pc
.PostSyncOperation
= WritePSDepthCount
;
1092 pc
.DepthStallEnable
= true;
1093 pc
.Address
= (struct anv_address
) { bo
, offset
};
1098 emit_query_availability(struct anv_batch
*batch
,
1099 struct anv_bo
*bo
, uint32_t offset
)
1101 anv_batch_emit(batch
, GENX(PIPE_CONTROL
), pc
) {
1102 pc
.DestinationAddressType
= DAT_PPGTT
;
1103 pc
.PostSyncOperation
= WriteImmediateData
;
1104 pc
.Address
= (struct anv_address
) { bo
, offset
};
1105 pc
.ImmediateData
= 1;
1109 void genX(CmdBeginQuery
)(
1110 VkCommandBuffer commandBuffer
,
1111 VkQueryPool queryPool
,
1113 VkQueryControlFlags flags
)
1115 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
1116 ANV_FROM_HANDLE(anv_query_pool
, pool
, queryPool
);
1118 /* Workaround: When meta uses the pipeline with the VS disabled, it seems
1119 * that the pipelining of the depth write breaks. What we see is that
1120 * samples from the render pass clear leaks into the first query
1121 * immediately after the clear. Doing a pipecontrol with a post-sync
1122 * operation and DepthStallEnable seems to work around the issue.
1124 if (cmd_buffer
->state
.need_query_wa
) {
1125 cmd_buffer
->state
.need_query_wa
= false;
1126 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
1127 pc
.DepthCacheFlushEnable
= true;
1128 pc
.DepthStallEnable
= true;
1132 switch (pool
->type
) {
1133 case VK_QUERY_TYPE_OCCLUSION
:
1134 emit_ps_depth_count(&cmd_buffer
->batch
, &pool
->bo
,
1135 query
* sizeof(struct anv_query_pool_slot
));
1138 case VK_QUERY_TYPE_PIPELINE_STATISTICS
:
1144 void genX(CmdEndQuery
)(
1145 VkCommandBuffer commandBuffer
,
1146 VkQueryPool queryPool
,
1149 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
1150 ANV_FROM_HANDLE(anv_query_pool
, pool
, queryPool
);
1152 switch (pool
->type
) {
1153 case VK_QUERY_TYPE_OCCLUSION
:
1154 emit_ps_depth_count(&cmd_buffer
->batch
, &pool
->bo
,
1155 query
* sizeof(struct anv_query_pool_slot
) + 8);
1157 emit_query_availability(&cmd_buffer
->batch
, &pool
->bo
,
1158 query
* sizeof(struct anv_query_pool_slot
) + 16);
1161 case VK_QUERY_TYPE_PIPELINE_STATISTICS
:
1167 #define TIMESTAMP 0x2358
1169 void genX(CmdWriteTimestamp
)(
1170 VkCommandBuffer commandBuffer
,
1171 VkPipelineStageFlagBits pipelineStage
,
1172 VkQueryPool queryPool
,
1175 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
1176 ANV_FROM_HANDLE(anv_query_pool
, pool
, queryPool
);
1177 uint32_t offset
= query
* sizeof(struct anv_query_pool_slot
);
1179 assert(pool
->type
== VK_QUERY_TYPE_TIMESTAMP
);
1181 switch (pipelineStage
) {
1182 case VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT
:
1183 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_STORE_REGISTER_MEM
), srm
) {
1184 srm
.RegisterAddress
= TIMESTAMP
;
1185 srm
.MemoryAddress
= (struct anv_address
) { &pool
->bo
, offset
};
1187 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_STORE_REGISTER_MEM
), srm
) {
1188 srm
.RegisterAddress
= TIMESTAMP
+ 4;
1189 srm
.MemoryAddress
= (struct anv_address
) { &pool
->bo
, offset
+ 4 };
1194 /* Everything else is bottom-of-pipe */
1195 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
1196 pc
.DestinationAddressType
= DAT_PPGTT
,
1197 pc
.PostSyncOperation
= WriteTimestamp
,
1198 pc
.Address
= (struct anv_address
) { &pool
->bo
, offset
};
1203 emit_query_availability(&cmd_buffer
->batch
, &pool
->bo
, query
+ 16);
1206 #if GEN_GEN > 7 || GEN_IS_HASWELL
1208 #define alu_opcode(v) __gen_uint((v), 20, 31)
1209 #define alu_operand1(v) __gen_uint((v), 10, 19)
1210 #define alu_operand2(v) __gen_uint((v), 0, 9)
1211 #define alu(opcode, operand1, operand2) \
1212 alu_opcode(opcode) | alu_operand1(operand1) | alu_operand2(operand2)
1214 #define OPCODE_NOOP 0x000
1215 #define OPCODE_LOAD 0x080
1216 #define OPCODE_LOADINV 0x480
1217 #define OPCODE_LOAD0 0x081
1218 #define OPCODE_LOAD1 0x481
1219 #define OPCODE_ADD 0x100
1220 #define OPCODE_SUB 0x101
1221 #define OPCODE_AND 0x102
1222 #define OPCODE_OR 0x103
1223 #define OPCODE_XOR 0x104
1224 #define OPCODE_STORE 0x180
1225 #define OPCODE_STOREINV 0x580
1227 #define OPERAND_R0 0x00
1228 #define OPERAND_R1 0x01
1229 #define OPERAND_R2 0x02
1230 #define OPERAND_R3 0x03
1231 #define OPERAND_R4 0x04
1232 #define OPERAND_SRCA 0x20
1233 #define OPERAND_SRCB 0x21
1234 #define OPERAND_ACCU 0x31
1235 #define OPERAND_ZF 0x32
1236 #define OPERAND_CF 0x33
1238 #define CS_GPR(n) (0x2600 + (n) * 8)
1241 emit_load_alu_reg_u64(struct anv_batch
*batch
, uint32_t reg
,
1242 struct anv_bo
*bo
, uint32_t offset
)
1244 anv_batch_emit(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
1245 lrm
.RegisterAddress
= reg
,
1246 lrm
.MemoryAddress
= (struct anv_address
) { bo
, offset
};
1248 anv_batch_emit(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
1249 lrm
.RegisterAddress
= reg
+ 4;
1250 lrm
.MemoryAddress
= (struct anv_address
) { bo
, offset
+ 4 };
1255 store_query_result(struct anv_batch
*batch
, uint32_t reg
,
1256 struct anv_bo
*bo
, uint32_t offset
, VkQueryResultFlags flags
)
1258 anv_batch_emit(batch
, GENX(MI_STORE_REGISTER_MEM
), srm
) {
1259 srm
.RegisterAddress
= reg
;
1260 srm
.MemoryAddress
= (struct anv_address
) { bo
, offset
};
1263 if (flags
& VK_QUERY_RESULT_64_BIT
) {
1264 anv_batch_emit(batch
, GENX(MI_STORE_REGISTER_MEM
), srm
) {
1265 srm
.RegisterAddress
= reg
+ 4;
1266 srm
.MemoryAddress
= (struct anv_address
) { bo
, offset
+ 4 };
1271 void genX(CmdCopyQueryPoolResults
)(
1272 VkCommandBuffer commandBuffer
,
1273 VkQueryPool queryPool
,
1274 uint32_t firstQuery
,
1275 uint32_t queryCount
,
1276 VkBuffer destBuffer
,
1277 VkDeviceSize destOffset
,
1278 VkDeviceSize destStride
,
1279 VkQueryResultFlags flags
)
1281 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
1282 ANV_FROM_HANDLE(anv_query_pool
, pool
, queryPool
);
1283 ANV_FROM_HANDLE(anv_buffer
, buffer
, destBuffer
);
1284 uint32_t slot_offset
, dst_offset
;
1286 if (flags
& VK_QUERY_RESULT_WAIT_BIT
) {
1287 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
1288 pc
.CommandStreamerStallEnable
= true;
1289 pc
.StallAtPixelScoreboard
= true;
1293 dst_offset
= buffer
->offset
+ destOffset
;
1294 for (uint32_t i
= 0; i
< queryCount
; i
++) {
1296 slot_offset
= (firstQuery
+ i
) * sizeof(struct anv_query_pool_slot
);
1297 switch (pool
->type
) {
1298 case VK_QUERY_TYPE_OCCLUSION
:
1299 emit_load_alu_reg_u64(&cmd_buffer
->batch
,
1300 CS_GPR(0), &pool
->bo
, slot_offset
);
1301 emit_load_alu_reg_u64(&cmd_buffer
->batch
,
1302 CS_GPR(1), &pool
->bo
, slot_offset
+ 8);
1304 /* FIXME: We need to clamp the result for 32 bit. */
1306 uint32_t *dw
= anv_batch_emitn(&cmd_buffer
->batch
, 5, GENX(MI_MATH
));
1307 dw
[1] = alu(OPCODE_LOAD
, OPERAND_SRCA
, OPERAND_R1
);
1308 dw
[2] = alu(OPCODE_LOAD
, OPERAND_SRCB
, OPERAND_R0
);
1309 dw
[3] = alu(OPCODE_SUB
, 0, 0);
1310 dw
[4] = alu(OPCODE_STORE
, OPERAND_R2
, OPERAND_ACCU
);
1313 case VK_QUERY_TYPE_TIMESTAMP
:
1314 emit_load_alu_reg_u64(&cmd_buffer
->batch
,
1315 CS_GPR(2), &pool
->bo
, slot_offset
);
1319 unreachable("unhandled query type");
1322 store_query_result(&cmd_buffer
->batch
,
1323 CS_GPR(2), buffer
->bo
, dst_offset
, flags
);
1325 if (flags
& VK_QUERY_RESULT_WITH_AVAILABILITY_BIT
) {
1326 emit_load_alu_reg_u64(&cmd_buffer
->batch
, CS_GPR(0),
1327 &pool
->bo
, slot_offset
+ 16);
1328 if (flags
& VK_QUERY_RESULT_64_BIT
)
1329 store_query_result(&cmd_buffer
->batch
,
1330 CS_GPR(0), buffer
->bo
, dst_offset
+ 8, flags
);
1332 store_query_result(&cmd_buffer
->batch
,
1333 CS_GPR(0), buffer
->bo
, dst_offset
+ 4, flags
);
1336 dst_offset
+= destStride
;