2 * Copyright © 2017 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 shall be included
12 * in all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
15 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
20 * DEALINGS IN THE SOFTWARE.
26 * ============================= GENXML CODE =============================
27 * [This file is compiled once per generation.]
28 * =======================================================================
30 * This is the main state upload code.
32 * Gallium uses Constant State Objects, or CSOs, for most state. Large,
33 * complex, or highly reusable state can be created once, and bound and
34 * rebound multiple times. This is modeled with the pipe->create_*_state()
35 * and pipe->bind_*_state() hooks. Highly dynamic or inexpensive state is
36 * streamed out on the fly, via pipe->set_*_state() hooks.
38 * OpenGL involves frequently mutating context state, which is mirrored in
39 * core Mesa by highly mutable data structures. However, most applications
40 * typically draw the same things over and over - from frame to frame, most
41 * of the same objects are still visible and need to be redrawn. So, rather
42 * than inventing new state all the time, applications usually mutate to swap
43 * between known states that we've seen before.
45 * Gallium isolates us from this mutation by tracking API state, and
46 * distilling it into a set of Constant State Objects, or CSOs. Large,
47 * complex, or typically reusable state can be created once, then reused
48 * multiple times. Drivers can create and store their own associated data.
49 * This create/bind model corresponds to the pipe->create_*_state() and
50 * pipe->bind_*_state() driver hooks.
52 * Some state is cheap to create, or expected to be highly dynamic. Rather
53 * than creating and caching piles of CSOs for these, Gallium simply streams
54 * them out, via the pipe->set_*_state() driver hooks.
56 * To reduce draw time overhead, we try to compute as much state at create
57 * time as possible. Wherever possible, we translate the Gallium pipe state
58 * to 3DSTATE commands, and store those commands in the CSO. At draw time,
59 * we can simply memcpy them into a batch buffer.
61 * No hardware matches the abstraction perfectly, so some commands require
62 * information from multiple CSOs. In this case, we can store two copies
63 * of the packet (one in each CSO), and simply | together their DWords at
64 * draw time. Sometimes the second set is trivial (one or two fields), so
65 * we simply pack it at draw time.
67 * There are two main components in the file below. First, the CSO hooks
68 * create/bind/track state. The second are the draw-time upload functions,
69 * iris_upload_render_state() and iris_upload_compute_state(), which read
70 * the context state and emit the commands into the actual batch.
81 #define __gen_validate_value(x) VALGRIND_CHECK_MEM_IS_DEFINED(&(x), sizeof(x))
87 #include "pipe/p_defines.h"
88 #include "pipe/p_state.h"
89 #include "pipe/p_context.h"
90 #include "pipe/p_screen.h"
91 #include "util/u_dual_blend.h"
92 #include "util/u_inlines.h"
93 #include "util/u_format.h"
94 #include "util/u_framebuffer.h"
95 #include "util/u_transfer.h"
96 #include "util/u_upload_mgr.h"
97 #include "util/u_viewport.h"
98 #include "drm-uapi/i915_drm.h"
100 #include "intel/compiler/brw_compiler.h"
101 #include "intel/common/gen_l3_config.h"
102 #include "intel/common/gen_sample_positions.h"
103 #include "iris_batch.h"
104 #include "iris_context.h"
105 #include "iris_defines.h"
106 #include "iris_pipe.h"
107 #include "iris_resource.h"
109 #include "iris_genx_macros.h"
110 #include "intel/common/gen_guardband.h"
113 #define MOCS_PTE 0x18
116 #define MOCS_PTE (1 << 1)
117 #define MOCS_WB (2 << 1)
121 mocs(const struct iris_bo
*bo
)
123 return bo
&& bo
->external
? MOCS_PTE
: MOCS_WB
;
127 * Statically assert that PIPE_* enums match the hardware packets.
128 * (As long as they match, we don't need to translate them.)
130 UNUSED
static void pipe_asserts()
132 #define PIPE_ASSERT(x) STATIC_ASSERT((int)x)
134 /* pipe_logicop happens to match the hardware. */
135 PIPE_ASSERT(PIPE_LOGICOP_CLEAR
== LOGICOP_CLEAR
);
136 PIPE_ASSERT(PIPE_LOGICOP_NOR
== LOGICOP_NOR
);
137 PIPE_ASSERT(PIPE_LOGICOP_AND_INVERTED
== LOGICOP_AND_INVERTED
);
138 PIPE_ASSERT(PIPE_LOGICOP_COPY_INVERTED
== LOGICOP_COPY_INVERTED
);
139 PIPE_ASSERT(PIPE_LOGICOP_AND_REVERSE
== LOGICOP_AND_REVERSE
);
140 PIPE_ASSERT(PIPE_LOGICOP_INVERT
== LOGICOP_INVERT
);
141 PIPE_ASSERT(PIPE_LOGICOP_XOR
== LOGICOP_XOR
);
142 PIPE_ASSERT(PIPE_LOGICOP_NAND
== LOGICOP_NAND
);
143 PIPE_ASSERT(PIPE_LOGICOP_AND
== LOGICOP_AND
);
144 PIPE_ASSERT(PIPE_LOGICOP_EQUIV
== LOGICOP_EQUIV
);
145 PIPE_ASSERT(PIPE_LOGICOP_NOOP
== LOGICOP_NOOP
);
146 PIPE_ASSERT(PIPE_LOGICOP_OR_INVERTED
== LOGICOP_OR_INVERTED
);
147 PIPE_ASSERT(PIPE_LOGICOP_COPY
== LOGICOP_COPY
);
148 PIPE_ASSERT(PIPE_LOGICOP_OR_REVERSE
== LOGICOP_OR_REVERSE
);
149 PIPE_ASSERT(PIPE_LOGICOP_OR
== LOGICOP_OR
);
150 PIPE_ASSERT(PIPE_LOGICOP_SET
== LOGICOP_SET
);
152 /* pipe_blend_func happens to match the hardware. */
153 PIPE_ASSERT(PIPE_BLENDFACTOR_ONE
== BLENDFACTOR_ONE
);
154 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_COLOR
== BLENDFACTOR_SRC_COLOR
);
155 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_ALPHA
== BLENDFACTOR_SRC_ALPHA
);
156 PIPE_ASSERT(PIPE_BLENDFACTOR_DST_ALPHA
== BLENDFACTOR_DST_ALPHA
);
157 PIPE_ASSERT(PIPE_BLENDFACTOR_DST_COLOR
== BLENDFACTOR_DST_COLOR
);
158 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
== BLENDFACTOR_SRC_ALPHA_SATURATE
);
159 PIPE_ASSERT(PIPE_BLENDFACTOR_CONST_COLOR
== BLENDFACTOR_CONST_COLOR
);
160 PIPE_ASSERT(PIPE_BLENDFACTOR_CONST_ALPHA
== BLENDFACTOR_CONST_ALPHA
);
161 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC1_COLOR
== BLENDFACTOR_SRC1_COLOR
);
162 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC1_ALPHA
== BLENDFACTOR_SRC1_ALPHA
);
163 PIPE_ASSERT(PIPE_BLENDFACTOR_ZERO
== BLENDFACTOR_ZERO
);
164 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC_COLOR
== BLENDFACTOR_INV_SRC_COLOR
);
165 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC_ALPHA
== BLENDFACTOR_INV_SRC_ALPHA
);
166 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_DST_ALPHA
== BLENDFACTOR_INV_DST_ALPHA
);
167 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_DST_COLOR
== BLENDFACTOR_INV_DST_COLOR
);
168 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_CONST_COLOR
== BLENDFACTOR_INV_CONST_COLOR
);
169 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_CONST_ALPHA
== BLENDFACTOR_INV_CONST_ALPHA
);
170 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC1_COLOR
== BLENDFACTOR_INV_SRC1_COLOR
);
171 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC1_ALPHA
== BLENDFACTOR_INV_SRC1_ALPHA
);
173 /* pipe_blend_func happens to match the hardware. */
174 PIPE_ASSERT(PIPE_BLEND_ADD
== BLENDFUNCTION_ADD
);
175 PIPE_ASSERT(PIPE_BLEND_SUBTRACT
== BLENDFUNCTION_SUBTRACT
);
176 PIPE_ASSERT(PIPE_BLEND_REVERSE_SUBTRACT
== BLENDFUNCTION_REVERSE_SUBTRACT
);
177 PIPE_ASSERT(PIPE_BLEND_MIN
== BLENDFUNCTION_MIN
);
178 PIPE_ASSERT(PIPE_BLEND_MAX
== BLENDFUNCTION_MAX
);
180 /* pipe_stencil_op happens to match the hardware. */
181 PIPE_ASSERT(PIPE_STENCIL_OP_KEEP
== STENCILOP_KEEP
);
182 PIPE_ASSERT(PIPE_STENCIL_OP_ZERO
== STENCILOP_ZERO
);
183 PIPE_ASSERT(PIPE_STENCIL_OP_REPLACE
== STENCILOP_REPLACE
);
184 PIPE_ASSERT(PIPE_STENCIL_OP_INCR
== STENCILOP_INCRSAT
);
185 PIPE_ASSERT(PIPE_STENCIL_OP_DECR
== STENCILOP_DECRSAT
);
186 PIPE_ASSERT(PIPE_STENCIL_OP_INCR_WRAP
== STENCILOP_INCR
);
187 PIPE_ASSERT(PIPE_STENCIL_OP_DECR_WRAP
== STENCILOP_DECR
);
188 PIPE_ASSERT(PIPE_STENCIL_OP_INVERT
== STENCILOP_INVERT
);
190 /* pipe_sprite_coord_mode happens to match 3DSTATE_SBE */
191 PIPE_ASSERT(PIPE_SPRITE_COORD_UPPER_LEFT
== UPPERLEFT
);
192 PIPE_ASSERT(PIPE_SPRITE_COORD_LOWER_LEFT
== LOWERLEFT
);
197 translate_prim_type(enum pipe_prim_type prim
, uint8_t verts_per_patch
)
199 static const unsigned map
[] = {
200 [PIPE_PRIM_POINTS
] = _3DPRIM_POINTLIST
,
201 [PIPE_PRIM_LINES
] = _3DPRIM_LINELIST
,
202 [PIPE_PRIM_LINE_LOOP
] = _3DPRIM_LINELOOP
,
203 [PIPE_PRIM_LINE_STRIP
] = _3DPRIM_LINESTRIP
,
204 [PIPE_PRIM_TRIANGLES
] = _3DPRIM_TRILIST
,
205 [PIPE_PRIM_TRIANGLE_STRIP
] = _3DPRIM_TRISTRIP
,
206 [PIPE_PRIM_TRIANGLE_FAN
] = _3DPRIM_TRIFAN
,
207 [PIPE_PRIM_QUADS
] = _3DPRIM_QUADLIST
,
208 [PIPE_PRIM_QUAD_STRIP
] = _3DPRIM_QUADSTRIP
,
209 [PIPE_PRIM_POLYGON
] = _3DPRIM_POLYGON
,
210 [PIPE_PRIM_LINES_ADJACENCY
] = _3DPRIM_LINELIST_ADJ
,
211 [PIPE_PRIM_LINE_STRIP_ADJACENCY
] = _3DPRIM_LINESTRIP_ADJ
,
212 [PIPE_PRIM_TRIANGLES_ADJACENCY
] = _3DPRIM_TRILIST_ADJ
,
213 [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
] = _3DPRIM_TRISTRIP_ADJ
,
214 [PIPE_PRIM_PATCHES
] = _3DPRIM_PATCHLIST_1
- 1,
217 return map
[prim
] + (prim
== PIPE_PRIM_PATCHES
? verts_per_patch
: 0);
221 translate_compare_func(enum pipe_compare_func pipe_func
)
223 static const unsigned map
[] = {
224 [PIPE_FUNC_NEVER
] = COMPAREFUNCTION_NEVER
,
225 [PIPE_FUNC_LESS
] = COMPAREFUNCTION_LESS
,
226 [PIPE_FUNC_EQUAL
] = COMPAREFUNCTION_EQUAL
,
227 [PIPE_FUNC_LEQUAL
] = COMPAREFUNCTION_LEQUAL
,
228 [PIPE_FUNC_GREATER
] = COMPAREFUNCTION_GREATER
,
229 [PIPE_FUNC_NOTEQUAL
] = COMPAREFUNCTION_NOTEQUAL
,
230 [PIPE_FUNC_GEQUAL
] = COMPAREFUNCTION_GEQUAL
,
231 [PIPE_FUNC_ALWAYS
] = COMPAREFUNCTION_ALWAYS
,
233 return map
[pipe_func
];
237 translate_shadow_func(enum pipe_compare_func pipe_func
)
239 /* Gallium specifies the result of shadow comparisons as:
241 * 1 if ref <op> texel,
246 * 0 if texel <op> ref,
249 * So we need to flip the operator and also negate.
251 static const unsigned map
[] = {
252 [PIPE_FUNC_NEVER
] = PREFILTEROPALWAYS
,
253 [PIPE_FUNC_LESS
] = PREFILTEROPLEQUAL
,
254 [PIPE_FUNC_EQUAL
] = PREFILTEROPNOTEQUAL
,
255 [PIPE_FUNC_LEQUAL
] = PREFILTEROPLESS
,
256 [PIPE_FUNC_GREATER
] = PREFILTEROPGEQUAL
,
257 [PIPE_FUNC_NOTEQUAL
] = PREFILTEROPEQUAL
,
258 [PIPE_FUNC_GEQUAL
] = PREFILTEROPGREATER
,
259 [PIPE_FUNC_ALWAYS
] = PREFILTEROPNEVER
,
261 return map
[pipe_func
];
265 translate_cull_mode(unsigned pipe_face
)
267 static const unsigned map
[4] = {
268 [PIPE_FACE_NONE
] = CULLMODE_NONE
,
269 [PIPE_FACE_FRONT
] = CULLMODE_FRONT
,
270 [PIPE_FACE_BACK
] = CULLMODE_BACK
,
271 [PIPE_FACE_FRONT_AND_BACK
] = CULLMODE_BOTH
,
273 return map
[pipe_face
];
277 translate_fill_mode(unsigned pipe_polymode
)
279 static const unsigned map
[4] = {
280 [PIPE_POLYGON_MODE_FILL
] = FILL_MODE_SOLID
,
281 [PIPE_POLYGON_MODE_LINE
] = FILL_MODE_WIREFRAME
,
282 [PIPE_POLYGON_MODE_POINT
] = FILL_MODE_POINT
,
283 [PIPE_POLYGON_MODE_FILL_RECTANGLE
] = FILL_MODE_SOLID
,
285 return map
[pipe_polymode
];
289 translate_mip_filter(enum pipe_tex_mipfilter pipe_mip
)
291 static const unsigned map
[] = {
292 [PIPE_TEX_MIPFILTER_NEAREST
] = MIPFILTER_NEAREST
,
293 [PIPE_TEX_MIPFILTER_LINEAR
] = MIPFILTER_LINEAR
,
294 [PIPE_TEX_MIPFILTER_NONE
] = MIPFILTER_NONE
,
296 return map
[pipe_mip
];
300 translate_wrap(unsigned pipe_wrap
)
302 static const unsigned map
[] = {
303 [PIPE_TEX_WRAP_REPEAT
] = TCM_WRAP
,
304 [PIPE_TEX_WRAP_CLAMP
] = TCM_HALF_BORDER
,
305 [PIPE_TEX_WRAP_CLAMP_TO_EDGE
] = TCM_CLAMP
,
306 [PIPE_TEX_WRAP_CLAMP_TO_BORDER
] = TCM_CLAMP_BORDER
,
307 [PIPE_TEX_WRAP_MIRROR_REPEAT
] = TCM_MIRROR
,
308 [PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
] = TCM_MIRROR_ONCE
,
310 /* These are unsupported. */
311 [PIPE_TEX_WRAP_MIRROR_CLAMP
] = -1,
312 [PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
] = -1,
314 return map
[pipe_wrap
];
318 * Allocate space for some indirect state.
320 * Return a pointer to the map (to fill it out) and a state ref (for
321 * referring to the state in GPU commands).
324 upload_state(struct u_upload_mgr
*uploader
,
325 struct iris_state_ref
*ref
,
330 u_upload_alloc(uploader
, 0, size
, alignment
, &ref
->offset
, &ref
->res
, &p
);
335 * Stream out temporary/short-lived state.
337 * This allocates space, pins the BO, and includes the BO address in the
338 * returned offset (which works because all state lives in 32-bit memory
342 stream_state(struct iris_batch
*batch
,
343 struct u_upload_mgr
*uploader
,
344 struct pipe_resource
**out_res
,
347 uint32_t *out_offset
)
351 u_upload_alloc(uploader
, 0, size
, alignment
, out_offset
, out_res
, &ptr
);
353 struct iris_bo
*bo
= iris_resource_bo(*out_res
);
354 iris_use_pinned_bo(batch
, bo
, false);
356 *out_offset
+= iris_bo_offset_from_base_address(bo
);
358 iris_record_state_size(batch
->state_sizes
, *out_offset
, size
);
364 * stream_state() + memcpy.
367 emit_state(struct iris_batch
*batch
,
368 struct u_upload_mgr
*uploader
,
369 struct pipe_resource
**out_res
,
376 stream_state(batch
, uploader
, out_res
, size
, alignment
, &offset
);
379 memcpy(map
, data
, size
);
385 * Did field 'x' change between 'old_cso' and 'new_cso'?
387 * (If so, we may want to set some dirty flags.)
389 #define cso_changed(x) (!old_cso || (old_cso->x != new_cso->x))
390 #define cso_changed_memcmp(x) \
391 (!old_cso || memcmp(old_cso->x, new_cso->x, sizeof(old_cso->x)) != 0)
394 flush_for_state_base_change(struct iris_batch
*batch
)
396 /* Flush before emitting STATE_BASE_ADDRESS.
398 * This isn't documented anywhere in the PRM. However, it seems to be
399 * necessary prior to changing the surface state base adress. We've
400 * seen issues in Vulkan where we get GPU hangs when using multi-level
401 * command buffers which clear depth, reset state base address, and then
404 * Normally, in GL, we would trust the kernel to do sufficient stalls
405 * and flushes prior to executing our batch. However, it doesn't seem
406 * as if the kernel's flushing is always sufficient and we don't want to
409 * We make this an end-of-pipe sync instead of a normal flush because we
410 * do not know the current status of the GPU. On Haswell at least,
411 * having a fast-clear operation in flight at the same time as a normal
412 * rendering operation can cause hangs. Since the kernel's flushing is
413 * insufficient, we need to ensure that any rendering operations from
414 * other processes are definitely complete before we try to do our own
415 * rendering. It's a bit of a big hammer but it appears to work.
417 iris_emit_end_of_pipe_sync(batch
,
418 "change STATE_BASE_ADDRESS",
419 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
420 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
421 PIPE_CONTROL_DATA_CACHE_FLUSH
);
425 _iris_emit_lri(struct iris_batch
*batch
, uint32_t reg
, uint32_t val
)
427 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_IMM
), lri
) {
428 lri
.RegisterOffset
= reg
;
432 #define iris_emit_lri(b, r, v) _iris_emit_lri(b, GENX(r##_num), v)
435 _iris_emit_lrr(struct iris_batch
*batch
, uint32_t dst
, uint32_t src
)
437 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_REG
), lrr
) {
438 lrr
.SourceRegisterAddress
= src
;
439 lrr
.DestinationRegisterAddress
= dst
;
444 emit_pipeline_select(struct iris_batch
*batch
, uint32_t pipeline
)
446 #if GEN_GEN >= 8 && GEN_GEN < 10
447 /* From the Broadwell PRM, Volume 2a: Instructions, PIPELINE_SELECT:
449 * Software must clear the COLOR_CALC_STATE Valid field in
450 * 3DSTATE_CC_STATE_POINTERS command prior to send a PIPELINE_SELECT
451 * with Pipeline Select set to GPGPU.
453 * The internal hardware docs recommend the same workaround for Gen9
456 if (pipeline
== GPGPU
)
457 iris_emit_cmd(batch
, GENX(3DSTATE_CC_STATE_POINTERS
), t
);
461 /* From "BXML » GT » MI » vol1a GPU Overview » [Instruction]
462 * PIPELINE_SELECT [DevBWR+]":
466 * Software must ensure all the write caches are flushed through a
467 * stalling PIPE_CONTROL command followed by another PIPE_CONTROL
468 * command to invalidate read only caches prior to programming
469 * MI_PIPELINE_SELECT command to change the Pipeline Select Mode."
471 iris_emit_pipe_control_flush(batch
,
472 "workaround: PIPELINE_SELECT flushes (1/2)",
473 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
474 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
475 PIPE_CONTROL_DATA_CACHE_FLUSH
|
476 PIPE_CONTROL_CS_STALL
);
478 iris_emit_pipe_control_flush(batch
,
479 "workaround: PIPELINE_SELECT flushes (2/2)",
480 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
481 PIPE_CONTROL_CONST_CACHE_INVALIDATE
|
482 PIPE_CONTROL_STATE_CACHE_INVALIDATE
|
483 PIPE_CONTROL_INSTRUCTION_INVALIDATE
);
485 iris_emit_cmd(batch
, GENX(PIPELINE_SELECT
), sel
) {
489 sel
.PipelineSelection
= pipeline
;
494 init_glk_barrier_mode(struct iris_batch
*batch
, uint32_t value
)
499 * "This chicken bit works around a hardware issue with barrier
500 * logic encountered when switching between GPGPU and 3D pipelines.
501 * To workaround the issue, this mode bit should be set after a
502 * pipeline is selected."
505 iris_pack_state(GENX(SLICE_COMMON_ECO_CHICKEN1
), ®_val
, reg
) {
506 reg
.GLKBarrierMode
= value
;
507 reg
.GLKBarrierModeMask
= 1;
509 iris_emit_lri(batch
, SLICE_COMMON_ECO_CHICKEN1
, reg_val
);
514 init_state_base_address(struct iris_batch
*batch
)
516 flush_for_state_base_change(batch
);
518 /* We program most base addresses once at context initialization time.
519 * Each base address points at a 4GB memory zone, and never needs to
520 * change. See iris_bufmgr.h for a description of the memory zones.
522 * The one exception is Surface State Base Address, which needs to be
523 * updated occasionally. See iris_binder.c for the details there.
525 iris_emit_cmd(batch
, GENX(STATE_BASE_ADDRESS
), sba
) {
526 sba
.GeneralStateMOCS
= MOCS_WB
;
527 sba
.StatelessDataPortAccessMOCS
= MOCS_WB
;
528 sba
.DynamicStateMOCS
= MOCS_WB
;
529 sba
.IndirectObjectMOCS
= MOCS_WB
;
530 sba
.InstructionMOCS
= MOCS_WB
;
531 sba
.SurfaceStateMOCS
= MOCS_WB
;
533 sba
.GeneralStateBaseAddressModifyEnable
= true;
534 sba
.DynamicStateBaseAddressModifyEnable
= true;
535 sba
.IndirectObjectBaseAddressModifyEnable
= true;
536 sba
.InstructionBaseAddressModifyEnable
= true;
537 sba
.GeneralStateBufferSizeModifyEnable
= true;
538 sba
.DynamicStateBufferSizeModifyEnable
= true;
540 sba
.BindlessSurfaceStateBaseAddressModifyEnable
= true;
541 sba
.BindlessSurfaceStateMOCS
= MOCS_WB
;
543 sba
.IndirectObjectBufferSizeModifyEnable
= true;
544 sba
.InstructionBuffersizeModifyEnable
= true;
546 sba
.InstructionBaseAddress
= ro_bo(NULL
, IRIS_MEMZONE_SHADER_START
);
547 sba
.DynamicStateBaseAddress
= ro_bo(NULL
, IRIS_MEMZONE_DYNAMIC_START
);
549 sba
.GeneralStateBufferSize
= 0xfffff;
550 sba
.IndirectObjectBufferSize
= 0xfffff;
551 sba
.InstructionBufferSize
= 0xfffff;
552 sba
.DynamicStateBufferSize
= 0xfffff;
557 iris_emit_l3_config(struct iris_batch
*batch
, const struct gen_l3_config
*cfg
,
558 bool has_slm
, bool wants_dc_cache
)
561 iris_pack_state(GENX(L3CNTLREG
), ®_val
, reg
) {
562 reg
.SLMEnable
= has_slm
;
564 /* WA_1406697149: Bit 9 "Error Detection Behavior Control" must be set
565 * in L3CNTLREG register. The default setting of the bit is not the
566 * desirable behavior.
568 reg
.ErrorDetectionBehaviorControl
= true;
569 reg
.UseFullWays
= true;
571 reg
.URBAllocation
= cfg
->n
[GEN_L3P_URB
];
572 reg
.ROAllocation
= cfg
->n
[GEN_L3P_RO
];
573 reg
.DCAllocation
= cfg
->n
[GEN_L3P_DC
];
574 reg
.AllAllocation
= cfg
->n
[GEN_L3P_ALL
];
576 iris_emit_lri(batch
, L3CNTLREG
, reg_val
);
580 iris_emit_default_l3_config(struct iris_batch
*batch
,
581 const struct gen_device_info
*devinfo
,
584 bool wants_dc_cache
= true;
585 bool has_slm
= compute
;
586 const struct gen_l3_weights w
=
587 gen_get_default_l3_weights(devinfo
, wants_dc_cache
, has_slm
);
588 const struct gen_l3_config
*cfg
= gen_get_l3_config(devinfo
, w
);
589 iris_emit_l3_config(batch
, cfg
, has_slm
, wants_dc_cache
);
592 #if GEN_GEN == 9 || GEN_GEN == 10
594 iris_enable_obj_preemption(struct iris_batch
*batch
, bool enable
)
598 /* A fixed function pipe flush is required before modifying this field */
599 iris_emit_end_of_pipe_sync(batch
, enable
? "enable preemption"
600 : "disable preemption",
601 PIPE_CONTROL_RENDER_TARGET_FLUSH
);
603 /* enable object level preemption */
604 iris_pack_state(GENX(CS_CHICKEN1
), ®_val
, reg
) {
605 reg
.ReplayMode
= enable
;
606 reg
.ReplayModeMask
= true;
608 iris_emit_lri(batch
, CS_CHICKEN1
, reg_val
);
614 iris_upload_slice_hashing_state(struct iris_batch
*batch
)
616 const struct gen_device_info
*devinfo
= &batch
->screen
->devinfo
;
617 int subslices_delta
=
618 devinfo
->ppipe_subslices
[0] - devinfo
->ppipe_subslices
[1];
619 if (subslices_delta
== 0)
622 struct iris_context
*ice
= NULL
;
623 ice
= container_of(batch
, ice
, batches
[IRIS_BATCH_RENDER
]);
624 assert(&ice
->batches
[IRIS_BATCH_RENDER
] == batch
);
626 unsigned size
= GENX(SLICE_HASH_TABLE_length
) * 4;
627 uint32_t hash_address
;
628 struct pipe_resource
*tmp
= NULL
;
630 stream_state(batch
, ice
->state
.dynamic_uploader
, &tmp
,
631 size
, 64, &hash_address
);
632 pipe_resource_reference(&tmp
, NULL
);
634 struct GENX(SLICE_HASH_TABLE
) table0
= {
636 { 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1 },
637 { 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1 },
638 { 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0 },
639 { 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1 },
640 { 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1 },
641 { 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0 },
642 { 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1 },
643 { 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1 },
644 { 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0 },
645 { 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1 },
646 { 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1 },
647 { 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0 },
648 { 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1 },
649 { 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1 },
650 { 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0 },
651 { 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1 }
655 struct GENX(SLICE_HASH_TABLE
) table1
= {
657 { 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0 },
658 { 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0 },
659 { 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1 },
660 { 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0 },
661 { 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0 },
662 { 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1 },
663 { 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0 },
664 { 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0 },
665 { 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1 },
666 { 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0 },
667 { 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0 },
668 { 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1 },
669 { 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0 },
670 { 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0 },
671 { 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1 },
672 { 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0 }
676 const struct GENX(SLICE_HASH_TABLE
) *table
=
677 subslices_delta
< 0 ? &table0
: &table1
;
678 GENX(SLICE_HASH_TABLE_pack
)(NULL
, map
, table
);
680 iris_emit_cmd(batch
, GENX(3DSTATE_SLICE_TABLE_STATE_POINTERS
), ptr
) {
681 ptr
.SliceHashStatePointerValid
= true;
682 ptr
.SliceHashTableStatePointer
= hash_address
;
685 iris_emit_cmd(batch
, GENX(3DSTATE_3D_MODE
), mode
) {
686 mode
.SliceHashingTableEnable
= true;
692 * Upload the initial GPU state for a render context.
694 * This sets some invariant state that needs to be programmed a particular
695 * way, but we never actually change.
698 iris_init_render_context(struct iris_screen
*screen
,
699 struct iris_batch
*batch
,
700 struct iris_vtable
*vtbl
,
701 struct pipe_debug_callback
*dbg
)
703 UNUSED
const struct gen_device_info
*devinfo
= &screen
->devinfo
;
706 emit_pipeline_select(batch
, _3D
);
708 iris_emit_default_l3_config(batch
, devinfo
, false);
710 init_state_base_address(batch
);
713 iris_pack_state(GENX(CS_DEBUG_MODE2
), ®_val
, reg
) {
714 reg
.CONSTANT_BUFFERAddressOffsetDisable
= true;
715 reg
.CONSTANT_BUFFERAddressOffsetDisableMask
= true;
717 iris_emit_lri(batch
, CS_DEBUG_MODE2
, reg_val
);
719 iris_pack_state(GENX(INSTPM
), ®_val
, reg
) {
720 reg
.CONSTANT_BUFFERAddressOffsetDisable
= true;
721 reg
.CONSTANT_BUFFERAddressOffsetDisableMask
= true;
723 iris_emit_lri(batch
, INSTPM
, reg_val
);
727 iris_pack_state(GENX(CACHE_MODE_1
), ®_val
, reg
) {
728 reg
.FloatBlendOptimizationEnable
= true;
729 reg
.FloatBlendOptimizationEnableMask
= true;
730 reg
.PartialResolveDisableInVC
= true;
731 reg
.PartialResolveDisableInVCMask
= true;
733 iris_emit_lri(batch
, CACHE_MODE_1
, reg_val
);
735 if (devinfo
->is_geminilake
)
736 init_glk_barrier_mode(batch
, GLK_BARRIER_MODE_3D_HULL
);
740 iris_pack_state(GENX(SAMPLER_MODE
), ®_val
, reg
) {
741 reg
.HeaderlessMessageforPreemptableContexts
= 1;
742 reg
.HeaderlessMessageforPreemptableContextsMask
= 1;
744 iris_emit_lri(batch
, SAMPLER_MODE
, reg_val
);
746 /* Bit 1 must be set in HALF_SLICE_CHICKEN7. */
747 iris_pack_state(GENX(HALF_SLICE_CHICKEN7
), ®_val
, reg
) {
748 reg
.EnabledTexelOffsetPrecisionFix
= 1;
749 reg
.EnabledTexelOffsetPrecisionFixMask
= 1;
751 iris_emit_lri(batch
, HALF_SLICE_CHICKEN7
, reg_val
);
753 iris_pack_state(GENX(SLICE_COMMON_ECO_CHICKEN1
), ®_val
, reg
) {
754 reg
.StateCacheRedirectToCSSectionEnable
= true;
755 reg
.StateCacheRedirectToCSSectionEnableMask
= true;
757 iris_emit_lri(batch
, SLICE_COMMON_ECO_CHICKEN1
, reg_val
);
759 /* Hardware specification recommends disabling repacking for the
760 * compatibility with decompression mechanism in display controller.
762 if (devinfo
->disable_ccs_repack
) {
763 iris_pack_state(GENX(CACHE_MODE_0
), ®_val
, reg
) {
764 reg
.DisableRepackingforCompression
= true;
765 reg
.DisableRepackingforCompressionMask
= true;
767 iris_emit_lri(batch
, CACHE_MODE_0
, reg_val
);
770 iris_upload_slice_hashing_state(batch
);
773 /* 3DSTATE_DRAWING_RECTANGLE is non-pipelined, so we want to avoid
774 * changing it dynamically. We set it to the maximum size here, and
775 * instead include the render target dimensions in the viewport, so
776 * viewport extents clipping takes care of pruning stray geometry.
778 iris_emit_cmd(batch
, GENX(3DSTATE_DRAWING_RECTANGLE
), rect
) {
779 rect
.ClippedDrawingRectangleXMax
= UINT16_MAX
;
780 rect
.ClippedDrawingRectangleYMax
= UINT16_MAX
;
783 /* Set the initial MSAA sample positions. */
784 iris_emit_cmd(batch
, GENX(3DSTATE_SAMPLE_PATTERN
), pat
) {
785 GEN_SAMPLE_POS_1X(pat
._1xSample
);
786 GEN_SAMPLE_POS_2X(pat
._2xSample
);
787 GEN_SAMPLE_POS_4X(pat
._4xSample
);
788 GEN_SAMPLE_POS_8X(pat
._8xSample
);
790 GEN_SAMPLE_POS_16X(pat
._16xSample
);
794 /* Use the legacy AA line coverage computation. */
795 iris_emit_cmd(batch
, GENX(3DSTATE_AA_LINE_PARAMETERS
), foo
);
797 /* Disable chromakeying (it's for media) */
798 iris_emit_cmd(batch
, GENX(3DSTATE_WM_CHROMAKEY
), foo
);
800 /* We want regular rendering, not special HiZ operations. */
801 iris_emit_cmd(batch
, GENX(3DSTATE_WM_HZ_OP
), foo
);
803 /* No polygon stippling offsets are necessary. */
804 /* TODO: may need to set an offset for origin-UL framebuffers */
805 iris_emit_cmd(batch
, GENX(3DSTATE_POLY_STIPPLE_OFFSET
), foo
);
807 /* Set a static partitioning of the push constant area. */
808 /* TODO: this may be a bad idea...could starve the push ringbuffers... */
809 for (int i
= 0; i
<= MESA_SHADER_FRAGMENT
; i
++) {
810 iris_emit_cmd(batch
, GENX(3DSTATE_PUSH_CONSTANT_ALLOC_VS
), alloc
) {
811 alloc
._3DCommandSubOpcode
= 18 + i
;
812 alloc
.ConstantBufferOffset
= 6 * i
;
813 alloc
.ConstantBufferSize
= i
== MESA_SHADER_FRAGMENT
? 8 : 6;
818 /* Gen11+ is enabled for us by the kernel. */
819 iris_enable_obj_preemption(batch
, true);
824 iris_init_compute_context(struct iris_screen
*screen
,
825 struct iris_batch
*batch
,
826 struct iris_vtable
*vtbl
,
827 struct pipe_debug_callback
*dbg
)
829 UNUSED
const struct gen_device_info
*devinfo
= &screen
->devinfo
;
831 emit_pipeline_select(batch
, GPGPU
);
833 iris_emit_default_l3_config(batch
, devinfo
, true);
835 init_state_base_address(batch
);
838 if (devinfo
->is_geminilake
)
839 init_glk_barrier_mode(batch
, GLK_BARRIER_MODE_GPGPU
);
843 struct iris_vertex_buffer_state
{
844 /** The VERTEX_BUFFER_STATE hardware structure. */
845 uint32_t state
[GENX(VERTEX_BUFFER_STATE_length
)];
847 /** The resource to source vertex data from. */
848 struct pipe_resource
*resource
;
851 struct iris_depth_buffer_state
{
852 /* Depth/HiZ/Stencil related hardware packets. */
853 uint32_t packets
[GENX(3DSTATE_DEPTH_BUFFER_length
) +
854 GENX(3DSTATE_STENCIL_BUFFER_length
) +
855 GENX(3DSTATE_HIER_DEPTH_BUFFER_length
) +
856 GENX(3DSTATE_CLEAR_PARAMS_length
)];
860 * Generation-specific context state (ice->state.genx->...).
862 * Most state can go in iris_context directly, but these encode hardware
863 * packets which vary by generation.
865 struct iris_genx_state
{
866 struct iris_vertex_buffer_state vertex_buffers
[33];
867 uint32_t last_index_buffer
[GENX(3DSTATE_INDEX_BUFFER_length
)];
869 struct iris_depth_buffer_state depth_buffer
;
871 uint32_t so_buffers
[4 * GENX(3DSTATE_SO_BUFFER_length
)];
874 /* Is object level preemption enabled? */
875 bool object_preemption
;
880 struct brw_image_param image_param
[PIPE_MAX_SHADER_IMAGES
];
882 } shaders
[MESA_SHADER_STAGES
];
886 * The pipe->set_blend_color() driver hook.
888 * This corresponds to our COLOR_CALC_STATE.
891 iris_set_blend_color(struct pipe_context
*ctx
,
892 const struct pipe_blend_color
*state
)
894 struct iris_context
*ice
= (struct iris_context
*) ctx
;
896 /* Our COLOR_CALC_STATE is exactly pipe_blend_color, so just memcpy */
897 memcpy(&ice
->state
.blend_color
, state
, sizeof(struct pipe_blend_color
));
898 ice
->state
.dirty
|= IRIS_DIRTY_COLOR_CALC_STATE
;
902 * Gallium CSO for blend state (see pipe_blend_state).
904 struct iris_blend_state
{
905 /** Partial 3DSTATE_PS_BLEND */
906 uint32_t ps_blend
[GENX(3DSTATE_PS_BLEND_length
)];
908 /** Partial BLEND_STATE */
909 uint32_t blend_state
[GENX(BLEND_STATE_length
) +
910 BRW_MAX_DRAW_BUFFERS
* GENX(BLEND_STATE_ENTRY_length
)];
912 bool alpha_to_coverage
; /* for shader key */
914 /** Bitfield of whether blending is enabled for RT[i] - for aux resolves */
915 uint8_t blend_enables
;
917 /** Bitfield of whether color writes are enabled for RT[i] */
918 uint8_t color_write_enables
;
920 /** Does RT[0] use dual color blending? */
921 bool dual_color_blending
;
924 static enum pipe_blendfactor
925 fix_blendfactor(enum pipe_blendfactor f
, bool alpha_to_one
)
928 if (f
== PIPE_BLENDFACTOR_SRC1_ALPHA
)
929 return PIPE_BLENDFACTOR_ONE
;
931 if (f
== PIPE_BLENDFACTOR_INV_SRC1_ALPHA
)
932 return PIPE_BLENDFACTOR_ZERO
;
939 * The pipe->create_blend_state() driver hook.
941 * Translates a pipe_blend_state into iris_blend_state.
944 iris_create_blend_state(struct pipe_context
*ctx
,
945 const struct pipe_blend_state
*state
)
947 struct iris_blend_state
*cso
= malloc(sizeof(struct iris_blend_state
));
948 uint32_t *blend_entry
= cso
->blend_state
+ GENX(BLEND_STATE_length
);
950 cso
->blend_enables
= 0;
951 cso
->color_write_enables
= 0;
952 STATIC_ASSERT(BRW_MAX_DRAW_BUFFERS
<= 8);
954 cso
->alpha_to_coverage
= state
->alpha_to_coverage
;
956 bool indep_alpha_blend
= false;
958 for (int i
= 0; i
< BRW_MAX_DRAW_BUFFERS
; i
++) {
959 const struct pipe_rt_blend_state
*rt
=
960 &state
->rt
[state
->independent_blend_enable
? i
: 0];
962 enum pipe_blendfactor src_rgb
=
963 fix_blendfactor(rt
->rgb_src_factor
, state
->alpha_to_one
);
964 enum pipe_blendfactor src_alpha
=
965 fix_blendfactor(rt
->alpha_src_factor
, state
->alpha_to_one
);
966 enum pipe_blendfactor dst_rgb
=
967 fix_blendfactor(rt
->rgb_dst_factor
, state
->alpha_to_one
);
968 enum pipe_blendfactor dst_alpha
=
969 fix_blendfactor(rt
->alpha_dst_factor
, state
->alpha_to_one
);
971 if (rt
->rgb_func
!= rt
->alpha_func
||
972 src_rgb
!= src_alpha
|| dst_rgb
!= dst_alpha
)
973 indep_alpha_blend
= true;
975 if (rt
->blend_enable
)
976 cso
->blend_enables
|= 1u << i
;
979 cso
->color_write_enables
|= 1u << i
;
981 iris_pack_state(GENX(BLEND_STATE_ENTRY
), blend_entry
, be
) {
982 be
.LogicOpEnable
= state
->logicop_enable
;
983 be
.LogicOpFunction
= state
->logicop_func
;
985 be
.PreBlendSourceOnlyClampEnable
= false;
986 be
.ColorClampRange
= COLORCLAMP_RTFORMAT
;
987 be
.PreBlendColorClampEnable
= true;
988 be
.PostBlendColorClampEnable
= true;
990 be
.ColorBufferBlendEnable
= rt
->blend_enable
;
992 be
.ColorBlendFunction
= rt
->rgb_func
;
993 be
.AlphaBlendFunction
= rt
->alpha_func
;
994 be
.SourceBlendFactor
= src_rgb
;
995 be
.SourceAlphaBlendFactor
= src_alpha
;
996 be
.DestinationBlendFactor
= dst_rgb
;
997 be
.DestinationAlphaBlendFactor
= dst_alpha
;
999 be
.WriteDisableRed
= !(rt
->colormask
& PIPE_MASK_R
);
1000 be
.WriteDisableGreen
= !(rt
->colormask
& PIPE_MASK_G
);
1001 be
.WriteDisableBlue
= !(rt
->colormask
& PIPE_MASK_B
);
1002 be
.WriteDisableAlpha
= !(rt
->colormask
& PIPE_MASK_A
);
1004 blend_entry
+= GENX(BLEND_STATE_ENTRY_length
);
1007 iris_pack_command(GENX(3DSTATE_PS_BLEND
), cso
->ps_blend
, pb
) {
1008 /* pb.HasWriteableRT is filled in at draw time.
1009 * pb.AlphaTestEnable is filled in at draw time.
1011 * pb.ColorBufferBlendEnable is filled in at draw time so we can avoid
1012 * setting it when dual color blending without an appropriate shader.
1015 pb
.AlphaToCoverageEnable
= state
->alpha_to_coverage
;
1016 pb
.IndependentAlphaBlendEnable
= indep_alpha_blend
;
1018 pb
.SourceBlendFactor
=
1019 fix_blendfactor(state
->rt
[0].rgb_src_factor
, state
->alpha_to_one
);
1020 pb
.SourceAlphaBlendFactor
=
1021 fix_blendfactor(state
->rt
[0].alpha_src_factor
, state
->alpha_to_one
);
1022 pb
.DestinationBlendFactor
=
1023 fix_blendfactor(state
->rt
[0].rgb_dst_factor
, state
->alpha_to_one
);
1024 pb
.DestinationAlphaBlendFactor
=
1025 fix_blendfactor(state
->rt
[0].alpha_dst_factor
, state
->alpha_to_one
);
1028 iris_pack_state(GENX(BLEND_STATE
), cso
->blend_state
, bs
) {
1029 bs
.AlphaToCoverageEnable
= state
->alpha_to_coverage
;
1030 bs
.IndependentAlphaBlendEnable
= indep_alpha_blend
;
1031 bs
.AlphaToOneEnable
= state
->alpha_to_one
;
1032 bs
.AlphaToCoverageDitherEnable
= state
->alpha_to_coverage
;
1033 bs
.ColorDitherEnable
= state
->dither
;
1034 /* bl.AlphaTestEnable and bs.AlphaTestFunction are filled in later. */
1037 cso
->dual_color_blending
= util_blend_state_is_dual(state
, 0);
1043 * The pipe->bind_blend_state() driver hook.
1045 * Bind a blending CSO and flag related dirty bits.
1048 iris_bind_blend_state(struct pipe_context
*ctx
, void *state
)
1050 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1051 struct iris_blend_state
*cso
= state
;
1053 ice
->state
.cso_blend
= cso
;
1054 ice
->state
.blend_enables
= cso
? cso
->blend_enables
: 0;
1056 ice
->state
.dirty
|= IRIS_DIRTY_PS_BLEND
;
1057 ice
->state
.dirty
|= IRIS_DIRTY_BLEND_STATE
;
1058 ice
->state
.dirty
|= IRIS_DIRTY_RENDER_RESOLVES_AND_FLUSHES
;
1059 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_BLEND
];
1063 * Return true if the FS writes to any color outputs which are not disabled
1064 * via color masking.
1067 has_writeable_rt(const struct iris_blend_state
*cso_blend
,
1068 const struct shader_info
*fs_info
)
1073 unsigned rt_outputs
= fs_info
->outputs_written
>> FRAG_RESULT_DATA0
;
1075 if (fs_info
->outputs_written
& BITFIELD64_BIT(FRAG_RESULT_COLOR
))
1076 rt_outputs
= (1 << BRW_MAX_DRAW_BUFFERS
) - 1;
1078 return cso_blend
->color_write_enables
& rt_outputs
;
1082 * Gallium CSO for depth, stencil, and alpha testing state.
1084 struct iris_depth_stencil_alpha_state
{
1085 /** Partial 3DSTATE_WM_DEPTH_STENCIL. */
1086 uint32_t wmds
[GENX(3DSTATE_WM_DEPTH_STENCIL_length
)];
1088 /** Outbound to BLEND_STATE, 3DSTATE_PS_BLEND, COLOR_CALC_STATE. */
1089 struct pipe_alpha_state alpha
;
1091 /** Outbound to resolve and cache set tracking. */
1092 bool depth_writes_enabled
;
1093 bool stencil_writes_enabled
;
1097 * The pipe->create_depth_stencil_alpha_state() driver hook.
1099 * We encode most of 3DSTATE_WM_DEPTH_STENCIL, and just save off the alpha
1100 * testing state since we need pieces of it in a variety of places.
1103 iris_create_zsa_state(struct pipe_context
*ctx
,
1104 const struct pipe_depth_stencil_alpha_state
*state
)
1106 struct iris_depth_stencil_alpha_state
*cso
=
1107 malloc(sizeof(struct iris_depth_stencil_alpha_state
));
1109 bool two_sided_stencil
= state
->stencil
[1].enabled
;
1111 cso
->alpha
= state
->alpha
;
1112 cso
->depth_writes_enabled
= state
->depth
.writemask
;
1113 cso
->stencil_writes_enabled
=
1114 state
->stencil
[0].writemask
!= 0 ||
1115 (two_sided_stencil
&& state
->stencil
[1].writemask
!= 0);
1117 /* The state tracker needs to optimize away EQUAL writes for us. */
1118 assert(!(state
->depth
.func
== PIPE_FUNC_EQUAL
&& state
->depth
.writemask
));
1120 iris_pack_command(GENX(3DSTATE_WM_DEPTH_STENCIL
), cso
->wmds
, wmds
) {
1121 wmds
.StencilFailOp
= state
->stencil
[0].fail_op
;
1122 wmds
.StencilPassDepthFailOp
= state
->stencil
[0].zfail_op
;
1123 wmds
.StencilPassDepthPassOp
= state
->stencil
[0].zpass_op
;
1124 wmds
.StencilTestFunction
=
1125 translate_compare_func(state
->stencil
[0].func
);
1126 wmds
.BackfaceStencilFailOp
= state
->stencil
[1].fail_op
;
1127 wmds
.BackfaceStencilPassDepthFailOp
= state
->stencil
[1].zfail_op
;
1128 wmds
.BackfaceStencilPassDepthPassOp
= state
->stencil
[1].zpass_op
;
1129 wmds
.BackfaceStencilTestFunction
=
1130 translate_compare_func(state
->stencil
[1].func
);
1131 wmds
.DepthTestFunction
= translate_compare_func(state
->depth
.func
);
1132 wmds
.DoubleSidedStencilEnable
= two_sided_stencil
;
1133 wmds
.StencilTestEnable
= state
->stencil
[0].enabled
;
1134 wmds
.StencilBufferWriteEnable
=
1135 state
->stencil
[0].writemask
!= 0 ||
1136 (two_sided_stencil
&& state
->stencil
[1].writemask
!= 0);
1137 wmds
.DepthTestEnable
= state
->depth
.enabled
;
1138 wmds
.DepthBufferWriteEnable
= state
->depth
.writemask
;
1139 wmds
.StencilTestMask
= state
->stencil
[0].valuemask
;
1140 wmds
.StencilWriteMask
= state
->stencil
[0].writemask
;
1141 wmds
.BackfaceStencilTestMask
= state
->stencil
[1].valuemask
;
1142 wmds
.BackfaceStencilWriteMask
= state
->stencil
[1].writemask
;
1143 /* wmds.[Backface]StencilReferenceValue are merged later */
1150 * The pipe->bind_depth_stencil_alpha_state() driver hook.
1152 * Bind a depth/stencil/alpha CSO and flag related dirty bits.
1155 iris_bind_zsa_state(struct pipe_context
*ctx
, void *state
)
1157 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1158 struct iris_depth_stencil_alpha_state
*old_cso
= ice
->state
.cso_zsa
;
1159 struct iris_depth_stencil_alpha_state
*new_cso
= state
;
1162 if (cso_changed(alpha
.ref_value
))
1163 ice
->state
.dirty
|= IRIS_DIRTY_COLOR_CALC_STATE
;
1165 if (cso_changed(alpha
.enabled
))
1166 ice
->state
.dirty
|= IRIS_DIRTY_PS_BLEND
| IRIS_DIRTY_BLEND_STATE
;
1168 if (cso_changed(alpha
.func
))
1169 ice
->state
.dirty
|= IRIS_DIRTY_BLEND_STATE
;
1171 if (cso_changed(depth_writes_enabled
))
1172 ice
->state
.dirty
|= IRIS_DIRTY_RENDER_RESOLVES_AND_FLUSHES
;
1174 ice
->state
.depth_writes_enabled
= new_cso
->depth_writes_enabled
;
1175 ice
->state
.stencil_writes_enabled
= new_cso
->stencil_writes_enabled
;
1178 ice
->state
.cso_zsa
= new_cso
;
1179 ice
->state
.dirty
|= IRIS_DIRTY_CC_VIEWPORT
;
1180 ice
->state
.dirty
|= IRIS_DIRTY_WM_DEPTH_STENCIL
;
1181 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_DEPTH_STENCIL_ALPHA
];
1185 * Gallium CSO for rasterizer state.
1187 struct iris_rasterizer_state
{
1188 uint32_t sf
[GENX(3DSTATE_SF_length
)];
1189 uint32_t clip
[GENX(3DSTATE_CLIP_length
)];
1190 uint32_t raster
[GENX(3DSTATE_RASTER_length
)];
1191 uint32_t wm
[GENX(3DSTATE_WM_length
)];
1192 uint32_t line_stipple
[GENX(3DSTATE_LINE_STIPPLE_length
)];
1194 uint8_t num_clip_plane_consts
;
1195 bool clip_halfz
; /* for CC_VIEWPORT */
1196 bool depth_clip_near
; /* for CC_VIEWPORT */
1197 bool depth_clip_far
; /* for CC_VIEWPORT */
1198 bool flatshade
; /* for shader state */
1199 bool flatshade_first
; /* for stream output */
1200 bool clamp_fragment_color
; /* for shader state */
1201 bool light_twoside
; /* for shader state */
1202 bool rasterizer_discard
; /* for 3DSTATE_STREAMOUT and 3DSTATE_CLIP */
1203 bool half_pixel_center
; /* for 3DSTATE_MULTISAMPLE */
1204 bool line_stipple_enable
;
1205 bool poly_stipple_enable
;
1207 bool force_persample_interp
;
1208 bool conservative_rasterization
;
1209 bool fill_mode_point_or_line
;
1210 enum pipe_sprite_coord_mode sprite_coord_mode
; /* PIPE_SPRITE_* */
1211 uint16_t sprite_coord_enable
;
1215 get_line_width(const struct pipe_rasterizer_state
*state
)
1217 float line_width
= state
->line_width
;
1219 /* From the OpenGL 4.4 spec:
1221 * "The actual width of non-antialiased lines is determined by rounding
1222 * the supplied width to the nearest integer, then clamping it to the
1223 * implementation-dependent maximum non-antialiased line width."
1225 if (!state
->multisample
&& !state
->line_smooth
)
1226 line_width
= roundf(state
->line_width
);
1228 if (!state
->multisample
&& state
->line_smooth
&& line_width
< 1.5f
) {
1229 /* For 1 pixel line thickness or less, the general anti-aliasing
1230 * algorithm gives up, and a garbage line is generated. Setting a
1231 * Line Width of 0.0 specifies the rasterization of the "thinnest"
1232 * (one-pixel-wide), non-antialiased lines.
1234 * Lines rendered with zero Line Width are rasterized using the
1235 * "Grid Intersection Quantization" rules as specified by the
1236 * "Zero-Width (Cosmetic) Line Rasterization" section of the docs.
1245 * The pipe->create_rasterizer_state() driver hook.
1248 iris_create_rasterizer_state(struct pipe_context
*ctx
,
1249 const struct pipe_rasterizer_state
*state
)
1251 struct iris_rasterizer_state
*cso
=
1252 malloc(sizeof(struct iris_rasterizer_state
));
1254 cso
->multisample
= state
->multisample
;
1255 cso
->force_persample_interp
= state
->force_persample_interp
;
1256 cso
->clip_halfz
= state
->clip_halfz
;
1257 cso
->depth_clip_near
= state
->depth_clip_near
;
1258 cso
->depth_clip_far
= state
->depth_clip_far
;
1259 cso
->flatshade
= state
->flatshade
;
1260 cso
->flatshade_first
= state
->flatshade_first
;
1261 cso
->clamp_fragment_color
= state
->clamp_fragment_color
;
1262 cso
->light_twoside
= state
->light_twoside
;
1263 cso
->rasterizer_discard
= state
->rasterizer_discard
;
1264 cso
->half_pixel_center
= state
->half_pixel_center
;
1265 cso
->sprite_coord_mode
= state
->sprite_coord_mode
;
1266 cso
->sprite_coord_enable
= state
->sprite_coord_enable
;
1267 cso
->line_stipple_enable
= state
->line_stipple_enable
;
1268 cso
->poly_stipple_enable
= state
->poly_stipple_enable
;
1269 cso
->conservative_rasterization
=
1270 state
->conservative_raster_mode
== PIPE_CONSERVATIVE_RASTER_POST_SNAP
;
1272 cso
->fill_mode_point_or_line
=
1273 state
->fill_front
== PIPE_POLYGON_MODE_LINE
||
1274 state
->fill_front
== PIPE_POLYGON_MODE_POINT
||
1275 state
->fill_back
== PIPE_POLYGON_MODE_LINE
||
1276 state
->fill_back
== PIPE_POLYGON_MODE_POINT
;
1278 if (state
->clip_plane_enable
!= 0)
1279 cso
->num_clip_plane_consts
= util_logbase2(state
->clip_plane_enable
) + 1;
1281 cso
->num_clip_plane_consts
= 0;
1283 float line_width
= get_line_width(state
);
1285 iris_pack_command(GENX(3DSTATE_SF
), cso
->sf
, sf
) {
1286 sf
.StatisticsEnable
= true;
1287 sf
.AALineDistanceMode
= AALINEDISTANCE_TRUE
;
1288 sf
.LineEndCapAntialiasingRegionWidth
=
1289 state
->line_smooth
? _10pixels
: _05pixels
;
1290 sf
.LastPixelEnable
= state
->line_last_pixel
;
1291 sf
.LineWidth
= line_width
;
1292 sf
.SmoothPointEnable
= (state
->point_smooth
|| state
->multisample
) &&
1293 !state
->point_quad_rasterization
;
1294 sf
.PointWidthSource
= state
->point_size_per_vertex
? Vertex
: State
;
1295 sf
.PointWidth
= state
->point_size
;
1297 if (state
->flatshade_first
) {
1298 sf
.TriangleFanProvokingVertexSelect
= 1;
1300 sf
.TriangleStripListProvokingVertexSelect
= 2;
1301 sf
.TriangleFanProvokingVertexSelect
= 2;
1302 sf
.LineStripListProvokingVertexSelect
= 1;
1306 iris_pack_command(GENX(3DSTATE_RASTER
), cso
->raster
, rr
) {
1307 rr
.FrontWinding
= state
->front_ccw
? CounterClockwise
: Clockwise
;
1308 rr
.CullMode
= translate_cull_mode(state
->cull_face
);
1309 rr
.FrontFaceFillMode
= translate_fill_mode(state
->fill_front
);
1310 rr
.BackFaceFillMode
= translate_fill_mode(state
->fill_back
);
1311 rr
.DXMultisampleRasterizationEnable
= state
->multisample
;
1312 rr
.GlobalDepthOffsetEnableSolid
= state
->offset_tri
;
1313 rr
.GlobalDepthOffsetEnableWireframe
= state
->offset_line
;
1314 rr
.GlobalDepthOffsetEnablePoint
= state
->offset_point
;
1315 rr
.GlobalDepthOffsetConstant
= state
->offset_units
* 2;
1316 rr
.GlobalDepthOffsetScale
= state
->offset_scale
;
1317 rr
.GlobalDepthOffsetClamp
= state
->offset_clamp
;
1318 rr
.SmoothPointEnable
= state
->point_smooth
;
1319 rr
.AntialiasingEnable
= state
->line_smooth
;
1320 rr
.ScissorRectangleEnable
= state
->scissor
;
1322 rr
.ViewportZNearClipTestEnable
= state
->depth_clip_near
;
1323 rr
.ViewportZFarClipTestEnable
= state
->depth_clip_far
;
1324 rr
.ConservativeRasterizationEnable
=
1325 cso
->conservative_rasterization
;
1327 rr
.ViewportZClipTestEnable
= (state
->depth_clip_near
|| state
->depth_clip_far
);
1331 iris_pack_command(GENX(3DSTATE_CLIP
), cso
->clip
, cl
) {
1332 /* cl.NonPerspectiveBarycentricEnable is filled in at draw time from
1333 * the FS program; cl.ForceZeroRTAIndexEnable is filled in from the FB.
1335 cl
.EarlyCullEnable
= true;
1336 cl
.UserClipDistanceClipTestEnableBitmask
= state
->clip_plane_enable
;
1337 cl
.ForceUserClipDistanceClipTestEnableBitmask
= true;
1338 cl
.APIMode
= state
->clip_halfz
? APIMODE_D3D
: APIMODE_OGL
;
1339 cl
.GuardbandClipTestEnable
= true;
1340 cl
.ClipEnable
= true;
1341 cl
.MinimumPointWidth
= 0.125;
1342 cl
.MaximumPointWidth
= 255.875;
1344 if (state
->flatshade_first
) {
1345 cl
.TriangleFanProvokingVertexSelect
= 1;
1347 cl
.TriangleStripListProvokingVertexSelect
= 2;
1348 cl
.TriangleFanProvokingVertexSelect
= 2;
1349 cl
.LineStripListProvokingVertexSelect
= 1;
1353 iris_pack_command(GENX(3DSTATE_WM
), cso
->wm
, wm
) {
1354 /* wm.BarycentricInterpolationMode and wm.EarlyDepthStencilControl are
1355 * filled in at draw time from the FS program.
1357 wm
.LineAntialiasingRegionWidth
= _10pixels
;
1358 wm
.LineEndCapAntialiasingRegionWidth
= _05pixels
;
1359 wm
.PointRasterizationRule
= RASTRULE_UPPER_RIGHT
;
1360 wm
.LineStippleEnable
= state
->line_stipple_enable
;
1361 wm
.PolygonStippleEnable
= state
->poly_stipple_enable
;
1364 /* Remap from 0..255 back to 1..256 */
1365 const unsigned line_stipple_factor
= state
->line_stipple_factor
+ 1;
1367 iris_pack_command(GENX(3DSTATE_LINE_STIPPLE
), cso
->line_stipple
, line
) {
1368 line
.LineStipplePattern
= state
->line_stipple_pattern
;
1369 line
.LineStippleInverseRepeatCount
= 1.0f
/ line_stipple_factor
;
1370 line
.LineStippleRepeatCount
= line_stipple_factor
;
1377 * The pipe->bind_rasterizer_state() driver hook.
1379 * Bind a rasterizer CSO and flag related dirty bits.
1382 iris_bind_rasterizer_state(struct pipe_context
*ctx
, void *state
)
1384 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1385 struct iris_rasterizer_state
*old_cso
= ice
->state
.cso_rast
;
1386 struct iris_rasterizer_state
*new_cso
= state
;
1389 /* Try to avoid re-emitting 3DSTATE_LINE_STIPPLE, it's non-pipelined */
1390 if (cso_changed_memcmp(line_stipple
))
1391 ice
->state
.dirty
|= IRIS_DIRTY_LINE_STIPPLE
;
1393 if (cso_changed(half_pixel_center
))
1394 ice
->state
.dirty
|= IRIS_DIRTY_MULTISAMPLE
;
1396 if (cso_changed(line_stipple_enable
) || cso_changed(poly_stipple_enable
))
1397 ice
->state
.dirty
|= IRIS_DIRTY_WM
;
1399 if (cso_changed(rasterizer_discard
))
1400 ice
->state
.dirty
|= IRIS_DIRTY_STREAMOUT
| IRIS_DIRTY_CLIP
;
1402 if (cso_changed(flatshade_first
))
1403 ice
->state
.dirty
|= IRIS_DIRTY_STREAMOUT
;
1405 if (cso_changed(depth_clip_near
) || cso_changed(depth_clip_far
) ||
1406 cso_changed(clip_halfz
))
1407 ice
->state
.dirty
|= IRIS_DIRTY_CC_VIEWPORT
;
1409 if (cso_changed(sprite_coord_enable
) ||
1410 cso_changed(sprite_coord_mode
) ||
1411 cso_changed(light_twoside
))
1412 ice
->state
.dirty
|= IRIS_DIRTY_SBE
;
1414 if (cso_changed(conservative_rasterization
))
1415 ice
->state
.dirty
|= IRIS_DIRTY_FS
;
1418 ice
->state
.cso_rast
= new_cso
;
1419 ice
->state
.dirty
|= IRIS_DIRTY_RASTER
;
1420 ice
->state
.dirty
|= IRIS_DIRTY_CLIP
;
1421 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_RASTERIZER
];
1425 * Return true if the given wrap mode requires the border color to exist.
1427 * (We can skip uploading it if the sampler isn't going to use it.)
1430 wrap_mode_needs_border_color(unsigned wrap_mode
)
1432 return wrap_mode
== TCM_CLAMP_BORDER
|| wrap_mode
== TCM_HALF_BORDER
;
1436 * Gallium CSO for sampler state.
1438 struct iris_sampler_state
{
1439 union pipe_color_union border_color
;
1440 bool needs_border_color
;
1442 uint32_t sampler_state
[GENX(SAMPLER_STATE_length
)];
1446 * The pipe->create_sampler_state() driver hook.
1448 * We fill out SAMPLER_STATE (except for the border color pointer), and
1449 * store that on the CPU. It doesn't make sense to upload it to a GPU
1450 * buffer object yet, because 3DSTATE_SAMPLER_STATE_POINTERS requires
1451 * all bound sampler states to be in contiguous memor.
1454 iris_create_sampler_state(struct pipe_context
*ctx
,
1455 const struct pipe_sampler_state
*state
)
1457 struct iris_sampler_state
*cso
= CALLOC_STRUCT(iris_sampler_state
);
1462 STATIC_ASSERT(PIPE_TEX_FILTER_NEAREST
== MAPFILTER_NEAREST
);
1463 STATIC_ASSERT(PIPE_TEX_FILTER_LINEAR
== MAPFILTER_LINEAR
);
1465 unsigned wrap_s
= translate_wrap(state
->wrap_s
);
1466 unsigned wrap_t
= translate_wrap(state
->wrap_t
);
1467 unsigned wrap_r
= translate_wrap(state
->wrap_r
);
1469 memcpy(&cso
->border_color
, &state
->border_color
, sizeof(cso
->border_color
));
1471 cso
->needs_border_color
= wrap_mode_needs_border_color(wrap_s
) ||
1472 wrap_mode_needs_border_color(wrap_t
) ||
1473 wrap_mode_needs_border_color(wrap_r
);
1475 float min_lod
= state
->min_lod
;
1476 unsigned mag_img_filter
= state
->mag_img_filter
;
1478 // XXX: explain this code ported from ilo...I don't get it at all...
1479 if (state
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
&&
1480 state
->min_lod
> 0.0f
) {
1482 mag_img_filter
= state
->min_img_filter
;
1485 iris_pack_state(GENX(SAMPLER_STATE
), cso
->sampler_state
, samp
) {
1486 samp
.TCXAddressControlMode
= wrap_s
;
1487 samp
.TCYAddressControlMode
= wrap_t
;
1488 samp
.TCZAddressControlMode
= wrap_r
;
1489 samp
.CubeSurfaceControlMode
= state
->seamless_cube_map
;
1490 samp
.NonnormalizedCoordinateEnable
= !state
->normalized_coords
;
1491 samp
.MinModeFilter
= state
->min_img_filter
;
1492 samp
.MagModeFilter
= mag_img_filter
;
1493 samp
.MipModeFilter
= translate_mip_filter(state
->min_mip_filter
);
1494 samp
.MaximumAnisotropy
= RATIO21
;
1496 if (state
->max_anisotropy
>= 2) {
1497 if (state
->min_img_filter
== PIPE_TEX_FILTER_LINEAR
) {
1498 samp
.MinModeFilter
= MAPFILTER_ANISOTROPIC
;
1499 samp
.AnisotropicAlgorithm
= EWAApproximation
;
1502 if (state
->mag_img_filter
== PIPE_TEX_FILTER_LINEAR
)
1503 samp
.MagModeFilter
= MAPFILTER_ANISOTROPIC
;
1505 samp
.MaximumAnisotropy
=
1506 MIN2((state
->max_anisotropy
- 2) / 2, RATIO161
);
1509 /* Set address rounding bits if not using nearest filtering. */
1510 if (state
->min_img_filter
!= PIPE_TEX_FILTER_NEAREST
) {
1511 samp
.UAddressMinFilterRoundingEnable
= true;
1512 samp
.VAddressMinFilterRoundingEnable
= true;
1513 samp
.RAddressMinFilterRoundingEnable
= true;
1516 if (state
->mag_img_filter
!= PIPE_TEX_FILTER_NEAREST
) {
1517 samp
.UAddressMagFilterRoundingEnable
= true;
1518 samp
.VAddressMagFilterRoundingEnable
= true;
1519 samp
.RAddressMagFilterRoundingEnable
= true;
1522 if (state
->compare_mode
== PIPE_TEX_COMPARE_R_TO_TEXTURE
)
1523 samp
.ShadowFunction
= translate_shadow_func(state
->compare_func
);
1525 const float hw_max_lod
= GEN_GEN
>= 7 ? 14 : 13;
1527 samp
.LODPreClampMode
= CLAMP_MODE_OGL
;
1528 samp
.MinLOD
= CLAMP(min_lod
, 0, hw_max_lod
);
1529 samp
.MaxLOD
= CLAMP(state
->max_lod
, 0, hw_max_lod
);
1530 samp
.TextureLODBias
= CLAMP(state
->lod_bias
, -16, 15);
1532 /* .BorderColorPointer is filled in by iris_bind_sampler_states. */
1539 * The pipe->bind_sampler_states() driver hook.
1542 iris_bind_sampler_states(struct pipe_context
*ctx
,
1543 enum pipe_shader_type p_stage
,
1544 unsigned start
, unsigned count
,
1547 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1548 gl_shader_stage stage
= stage_from_pipe(p_stage
);
1549 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
1551 assert(start
+ count
<= IRIS_MAX_TEXTURE_SAMPLERS
);
1553 for (int i
= 0; i
< count
; i
++) {
1554 shs
->samplers
[start
+ i
] = states
[i
];
1557 ice
->state
.dirty
|= IRIS_DIRTY_SAMPLER_STATES_VS
<< stage
;
1561 * Upload the sampler states into a contiguous area of GPU memory, for
1562 * for 3DSTATE_SAMPLER_STATE_POINTERS_*.
1564 * Also fill out the border color state pointers.
1567 iris_upload_sampler_states(struct iris_context
*ice
, gl_shader_stage stage
)
1569 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
1570 const struct shader_info
*info
= iris_get_shader_info(ice
, stage
);
1572 /* We assume the state tracker will call pipe->bind_sampler_states()
1573 * if the program's number of textures changes.
1575 unsigned count
= info
? util_last_bit(info
->textures_used
) : 0;
1580 /* Assemble the SAMPLER_STATEs into a contiguous table that lives
1581 * in the dynamic state memory zone, so we can point to it via the
1582 * 3DSTATE_SAMPLER_STATE_POINTERS_* commands.
1584 unsigned size
= count
* 4 * GENX(SAMPLER_STATE_length
);
1586 upload_state(ice
->state
.dynamic_uploader
, &shs
->sampler_table
, size
, 32);
1590 struct pipe_resource
*res
= shs
->sampler_table
.res
;
1591 shs
->sampler_table
.offset
+=
1592 iris_bo_offset_from_base_address(iris_resource_bo(res
));
1594 iris_record_state_size(ice
->state
.sizes
, shs
->sampler_table
.offset
, size
);
1596 /* Make sure all land in the same BO */
1597 iris_border_color_pool_reserve(ice
, IRIS_MAX_TEXTURE_SAMPLERS
);
1599 ice
->state
.need_border_colors
&= ~(1 << stage
);
1601 for (int i
= 0; i
< count
; i
++) {
1602 struct iris_sampler_state
*state
= shs
->samplers
[i
];
1603 struct iris_sampler_view
*tex
= shs
->textures
[i
];
1606 memset(map
, 0, 4 * GENX(SAMPLER_STATE_length
));
1607 } else if (!state
->needs_border_color
) {
1608 memcpy(map
, state
->sampler_state
, 4 * GENX(SAMPLER_STATE_length
));
1610 ice
->state
.need_border_colors
|= 1 << stage
;
1612 /* We may need to swizzle the border color for format faking.
1613 * A/LA formats are faked as R/RG with 000R or R00G swizzles.
1614 * This means we need to move the border color's A channel into
1615 * the R or G channels so that those read swizzles will move it
1618 union pipe_color_union
*color
= &state
->border_color
;
1619 union pipe_color_union tmp
;
1621 enum pipe_format internal_format
= tex
->res
->internal_format
;
1623 if (util_format_is_alpha(internal_format
)) {
1624 unsigned char swz
[4] = {
1625 PIPE_SWIZZLE_W
, PIPE_SWIZZLE_0
,
1626 PIPE_SWIZZLE_0
, PIPE_SWIZZLE_0
1628 util_format_apply_color_swizzle(&tmp
, color
, swz
, true);
1630 } else if (util_format_is_luminance_alpha(internal_format
) &&
1631 internal_format
!= PIPE_FORMAT_L8A8_SRGB
) {
1632 unsigned char swz
[4] = {
1633 PIPE_SWIZZLE_X
, PIPE_SWIZZLE_W
,
1634 PIPE_SWIZZLE_0
, PIPE_SWIZZLE_0
1636 util_format_apply_color_swizzle(&tmp
, color
, swz
, true);
1641 /* Stream out the border color and merge the pointer. */
1642 uint32_t offset
= iris_upload_border_color(ice
, color
);
1644 uint32_t dynamic
[GENX(SAMPLER_STATE_length
)];
1645 iris_pack_state(GENX(SAMPLER_STATE
), dynamic
, dyns
) {
1646 dyns
.BorderColorPointer
= offset
;
1649 for (uint32_t j
= 0; j
< GENX(SAMPLER_STATE_length
); j
++)
1650 map
[j
] = state
->sampler_state
[j
] | dynamic
[j
];
1653 map
+= GENX(SAMPLER_STATE_length
);
1657 static enum isl_channel_select
1658 fmt_swizzle(const struct iris_format_info
*fmt
, enum pipe_swizzle swz
)
1661 case PIPE_SWIZZLE_X
: return fmt
->swizzle
.r
;
1662 case PIPE_SWIZZLE_Y
: return fmt
->swizzle
.g
;
1663 case PIPE_SWIZZLE_Z
: return fmt
->swizzle
.b
;
1664 case PIPE_SWIZZLE_W
: return fmt
->swizzle
.a
;
1665 case PIPE_SWIZZLE_1
: return SCS_ONE
;
1666 case PIPE_SWIZZLE_0
: return SCS_ZERO
;
1667 default: unreachable("invalid swizzle");
1672 fill_buffer_surface_state(struct isl_device
*isl_dev
,
1673 struct iris_resource
*res
,
1675 enum isl_format format
,
1676 struct isl_swizzle swizzle
,
1680 const struct isl_format_layout
*fmtl
= isl_format_get_layout(format
);
1681 const unsigned cpp
= format
== ISL_FORMAT_RAW
? 1 : fmtl
->bpb
/ 8;
1683 /* The ARB_texture_buffer_specification says:
1685 * "The number of texels in the buffer texture's texel array is given by
1687 * floor(<buffer_size> / (<components> * sizeof(<base_type>)),
1689 * where <buffer_size> is the size of the buffer object, in basic
1690 * machine units and <components> and <base_type> are the element count
1691 * and base data type for elements, as specified in Table X.1. The
1692 * number of texels in the texel array is then clamped to the
1693 * implementation-dependent limit MAX_TEXTURE_BUFFER_SIZE_ARB."
1695 * We need to clamp the size in bytes to MAX_TEXTURE_BUFFER_SIZE * stride,
1696 * so that when ISL divides by stride to obtain the number of texels, that
1697 * texel count is clamped to MAX_TEXTURE_BUFFER_SIZE.
1699 unsigned final_size
=
1700 MIN3(size
, res
->bo
->size
- res
->offset
- offset
,
1701 IRIS_MAX_TEXTURE_BUFFER_SIZE
* cpp
);
1703 isl_buffer_fill_state(isl_dev
, map
,
1704 .address
= res
->bo
->gtt_offset
+ res
->offset
+ offset
,
1705 .size_B
= final_size
,
1709 .mocs
= mocs(res
->bo
));
1712 #define SURFACE_STATE_ALIGNMENT 64
1715 * Allocate several contiguous SURFACE_STATE structures, one for each
1716 * supported auxiliary surface mode.
1719 alloc_surface_states(struct u_upload_mgr
*mgr
,
1720 struct iris_state_ref
*ref
,
1721 unsigned aux_usages
)
1723 const unsigned surf_size
= 4 * GENX(RENDER_SURFACE_STATE_length
);
1725 /* If this changes, update this to explicitly align pointers */
1726 STATIC_ASSERT(surf_size
== SURFACE_STATE_ALIGNMENT
);
1728 assert(aux_usages
!= 0);
1731 upload_state(mgr
, ref
, util_bitcount(aux_usages
) * surf_size
,
1732 SURFACE_STATE_ALIGNMENT
);
1734 ref
->offset
+= iris_bo_offset_from_base_address(iris_resource_bo(ref
->res
));
1741 * Return an ISL surface for use with non-coherent render target reads.
1743 * In a few complex cases, we can't use the SURFACE_STATE for normal render
1744 * target writes. We need to make a separate one for sampling which refers
1745 * to the single slice of the texture being read.
1748 get_rt_read_isl_surf(const struct gen_device_info
*devinfo
,
1749 struct iris_resource
*res
,
1750 enum pipe_texture_target target
,
1751 struct isl_view
*view
,
1752 uint32_t *tile_x_sa
,
1753 uint32_t *tile_y_sa
,
1754 struct isl_surf
*surf
)
1759 const enum isl_dim_layout dim_layout
=
1760 iris_get_isl_dim_layout(devinfo
, res
->surf
.tiling
, target
);
1762 surf
->dim
= target_to_isl_surf_dim(target
);
1764 if (surf
->dim_layout
== dim_layout
)
1767 /* The layout of the specified texture target is not compatible with the
1768 * actual layout of the miptree structure in memory -- You're entering
1769 * dangerous territory, this can only possibly work if you only intended
1770 * to access a single level and slice of the texture, and the hardware
1771 * supports the tile offset feature in order to allow non-tile-aligned
1772 * base offsets, since we'll have to point the hardware to the first
1773 * texel of the level instead of relying on the usual base level/layer
1776 assert(view
->levels
== 1 && view
->array_len
== 1);
1777 assert(*tile_x_sa
== 0 && *tile_y_sa
== 0);
1779 res
->offset
+= iris_resource_get_tile_offsets(res
, view
->base_level
,
1780 view
->base_array_layer
,
1781 tile_x_sa
, tile_y_sa
);
1782 const unsigned l
= view
->base_level
;
1784 surf
->logical_level0_px
.width
= minify(surf
->logical_level0_px
.width
, l
);
1785 surf
->logical_level0_px
.height
= surf
->dim
<= ISL_SURF_DIM_1D
? 1 :
1786 minify(surf
->logical_level0_px
.height
, l
);
1787 surf
->logical_level0_px
.depth
= surf
->dim
<= ISL_SURF_DIM_2D
? 1 :
1788 minify(surf
->logical_level0_px
.depth
, l
);
1790 surf
->logical_level0_px
.array_len
= 1;
1792 surf
->dim_layout
= dim_layout
;
1794 view
->base_level
= 0;
1795 view
->base_array_layer
= 0;
1800 fill_surface_state(struct isl_device
*isl_dev
,
1802 struct iris_resource
*res
,
1803 struct isl_surf
*surf
,
1804 struct isl_view
*view
,
1809 struct isl_surf_fill_state_info f
= {
1812 .mocs
= mocs(res
->bo
),
1813 .address
= res
->bo
->gtt_offset
+ res
->offset
,
1814 .x_offset_sa
= tile_x_sa
,
1815 .y_offset_sa
= tile_y_sa
,
1818 assert(!iris_resource_unfinished_aux_import(res
));
1820 if (aux_usage
!= ISL_AUX_USAGE_NONE
) {
1821 f
.aux_surf
= &res
->aux
.surf
;
1822 f
.aux_usage
= aux_usage
;
1823 f
.aux_address
= res
->aux
.bo
->gtt_offset
+ res
->aux
.offset
;
1825 struct iris_bo
*clear_bo
= NULL
;
1826 uint64_t clear_offset
= 0;
1828 iris_resource_get_clear_color(res
, &clear_bo
, &clear_offset
);
1830 f
.clear_address
= clear_bo
->gtt_offset
+ clear_offset
;
1831 f
.use_clear_address
= isl_dev
->info
->gen
> 9;
1835 isl_surf_fill_state_s(isl_dev
, map
, &f
);
1839 * The pipe->create_sampler_view() driver hook.
1841 static struct pipe_sampler_view
*
1842 iris_create_sampler_view(struct pipe_context
*ctx
,
1843 struct pipe_resource
*tex
,
1844 const struct pipe_sampler_view
*tmpl
)
1846 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1847 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
1848 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1849 struct iris_sampler_view
*isv
= calloc(1, sizeof(struct iris_sampler_view
));
1854 /* initialize base object */
1856 isv
->base
.context
= ctx
;
1857 isv
->base
.texture
= NULL
;
1858 pipe_reference_init(&isv
->base
.reference
, 1);
1859 pipe_resource_reference(&isv
->base
.texture
, tex
);
1861 if (util_format_is_depth_or_stencil(tmpl
->format
)) {
1862 struct iris_resource
*zres
, *sres
;
1863 const struct util_format_description
*desc
=
1864 util_format_description(tmpl
->format
);
1866 iris_get_depth_stencil_resources(tex
, &zres
, &sres
);
1868 tex
= util_format_has_depth(desc
) ? &zres
->base
: &sres
->base
;
1871 isv
->res
= (struct iris_resource
*) tex
;
1873 void *map
= alloc_surface_states(ice
->state
.surface_uploader
,
1874 &isv
->surface_state
,
1875 isv
->res
->aux
.sampler_usages
);
1879 isl_surf_usage_flags_t usage
= ISL_SURF_USAGE_TEXTURE_BIT
;
1881 if (isv
->base
.target
== PIPE_TEXTURE_CUBE
||
1882 isv
->base
.target
== PIPE_TEXTURE_CUBE_ARRAY
)
1883 usage
|= ISL_SURF_USAGE_CUBE_BIT
;
1885 const struct iris_format_info fmt
=
1886 iris_format_for_usage(devinfo
, tmpl
->format
, usage
);
1888 isv
->clear_color
= isv
->res
->aux
.clear_color
;
1890 isv
->view
= (struct isl_view
) {
1892 .swizzle
= (struct isl_swizzle
) {
1893 .r
= fmt_swizzle(&fmt
, tmpl
->swizzle_r
),
1894 .g
= fmt_swizzle(&fmt
, tmpl
->swizzle_g
),
1895 .b
= fmt_swizzle(&fmt
, tmpl
->swizzle_b
),
1896 .a
= fmt_swizzle(&fmt
, tmpl
->swizzle_a
),
1901 /* Fill out SURFACE_STATE for this view. */
1902 if (tmpl
->target
!= PIPE_BUFFER
) {
1903 isv
->view
.base_level
= tmpl
->u
.tex
.first_level
;
1904 isv
->view
.levels
= tmpl
->u
.tex
.last_level
- tmpl
->u
.tex
.first_level
+ 1;
1905 // XXX: do I need to port f9fd0cf4790cb2a530e75d1a2206dbb9d8af7cb2?
1906 isv
->view
.base_array_layer
= tmpl
->u
.tex
.first_layer
;
1907 isv
->view
.array_len
=
1908 tmpl
->u
.tex
.last_layer
- tmpl
->u
.tex
.first_layer
+ 1;
1910 if (iris_resource_unfinished_aux_import(isv
->res
))
1911 iris_resource_finish_aux_import(&screen
->base
, isv
->res
);
1913 unsigned aux_modes
= isv
->res
->aux
.sampler_usages
;
1915 enum isl_aux_usage aux_usage
= u_bit_scan(&aux_modes
);
1917 /* If we have a multisampled depth buffer, do not create a sampler
1918 * surface state with HiZ.
1920 fill_surface_state(&screen
->isl_dev
, map
, isv
->res
, &isv
->res
->surf
,
1921 &isv
->view
, aux_usage
, 0, 0);
1923 map
+= SURFACE_STATE_ALIGNMENT
;
1926 fill_buffer_surface_state(&screen
->isl_dev
, isv
->res
, map
,
1927 isv
->view
.format
, isv
->view
.swizzle
,
1928 tmpl
->u
.buf
.offset
, tmpl
->u
.buf
.size
);
1935 iris_sampler_view_destroy(struct pipe_context
*ctx
,
1936 struct pipe_sampler_view
*state
)
1938 struct iris_sampler_view
*isv
= (void *) state
;
1939 pipe_resource_reference(&state
->texture
, NULL
);
1940 pipe_resource_reference(&isv
->surface_state
.res
, NULL
);
1945 * The pipe->create_surface() driver hook.
1947 * In Gallium nomenclature, "surfaces" are a view of a resource that
1948 * can be bound as a render target or depth/stencil buffer.
1950 static struct pipe_surface
*
1951 iris_create_surface(struct pipe_context
*ctx
,
1952 struct pipe_resource
*tex
,
1953 const struct pipe_surface
*tmpl
)
1955 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1956 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
1957 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1959 isl_surf_usage_flags_t usage
= 0;
1961 usage
= ISL_SURF_USAGE_STORAGE_BIT
;
1962 else if (util_format_is_depth_or_stencil(tmpl
->format
))
1963 usage
= ISL_SURF_USAGE_DEPTH_BIT
;
1965 usage
= ISL_SURF_USAGE_RENDER_TARGET_BIT
;
1967 const struct iris_format_info fmt
=
1968 iris_format_for_usage(devinfo
, tmpl
->format
, usage
);
1970 if ((usage
& ISL_SURF_USAGE_RENDER_TARGET_BIT
) &&
1971 !isl_format_supports_rendering(devinfo
, fmt
.fmt
)) {
1972 /* Framebuffer validation will reject this invalid case, but it
1973 * hasn't had the opportunity yet. In the meantime, we need to
1974 * avoid hitting ISL asserts about unsupported formats below.
1979 struct iris_surface
*surf
= calloc(1, sizeof(struct iris_surface
));
1980 struct pipe_surface
*psurf
= &surf
->base
;
1981 struct iris_resource
*res
= (struct iris_resource
*) tex
;
1986 pipe_reference_init(&psurf
->reference
, 1);
1987 pipe_resource_reference(&psurf
->texture
, tex
);
1988 psurf
->context
= ctx
;
1989 psurf
->format
= tmpl
->format
;
1990 psurf
->width
= tex
->width0
;
1991 psurf
->height
= tex
->height0
;
1992 psurf
->texture
= tex
;
1993 psurf
->u
.tex
.first_layer
= tmpl
->u
.tex
.first_layer
;
1994 psurf
->u
.tex
.last_layer
= tmpl
->u
.tex
.last_layer
;
1995 psurf
->u
.tex
.level
= tmpl
->u
.tex
.level
;
1997 uint32_t array_len
= tmpl
->u
.tex
.last_layer
- tmpl
->u
.tex
.first_layer
+ 1;
1999 struct isl_view
*view
= &surf
->view
;
2000 *view
= (struct isl_view
) {
2002 .base_level
= tmpl
->u
.tex
.level
,
2004 .base_array_layer
= tmpl
->u
.tex
.first_layer
,
2005 .array_len
= array_len
,
2006 .swizzle
= ISL_SWIZZLE_IDENTITY
,
2011 enum pipe_texture_target target
= (tex
->target
== PIPE_TEXTURE_3D
&&
2012 array_len
== 1) ? PIPE_TEXTURE_2D
:
2013 tex
->target
== PIPE_TEXTURE_1D_ARRAY
?
2014 PIPE_TEXTURE_2D_ARRAY
: tex
->target
;
2016 struct isl_view
*read_view
= &surf
->read_view
;
2017 *read_view
= (struct isl_view
) {
2019 .base_level
= tmpl
->u
.tex
.level
,
2021 .base_array_layer
= tmpl
->u
.tex
.first_layer
,
2022 .array_len
= array_len
,
2023 .swizzle
= ISL_SWIZZLE_IDENTITY
,
2024 .usage
= ISL_SURF_USAGE_TEXTURE_BIT
,
2028 surf
->clear_color
= res
->aux
.clear_color
;
2030 /* Bail early for depth/stencil - we don't want SURFACE_STATE for them. */
2031 if (res
->surf
.usage
& (ISL_SURF_USAGE_DEPTH_BIT
|
2032 ISL_SURF_USAGE_STENCIL_BIT
))
2036 void *map
= alloc_surface_states(ice
->state
.surface_uploader
,
2037 &surf
->surface_state
,
2038 res
->aux
.possible_usages
);
2039 if (!unlikely(map
)) {
2040 pipe_resource_reference(&surf
->surface_state
.res
, NULL
);
2045 void *map_read
= alloc_surface_states(ice
->state
.surface_uploader
,
2046 &surf
->surface_state_read
,
2047 res
->aux
.possible_usages
);
2048 if (!unlikely(map_read
)) {
2049 pipe_resource_reference(&surf
->surface_state_read
.res
, NULL
);
2054 if (!isl_format_is_compressed(res
->surf
.format
)) {
2055 if (iris_resource_unfinished_aux_import(res
))
2056 iris_resource_finish_aux_import(&screen
->base
, res
);
2058 /* This is a normal surface. Fill out a SURFACE_STATE for each possible
2059 * auxiliary surface mode and return the pipe_surface.
2061 unsigned aux_modes
= res
->aux
.possible_usages
;
2064 uint32_t offset
= res
->offset
;
2066 enum isl_aux_usage aux_usage
= u_bit_scan(&aux_modes
);
2067 fill_surface_state(&screen
->isl_dev
, map
, res
, &res
->surf
,
2068 view
, aux_usage
, 0, 0);
2069 map
+= SURFACE_STATE_ALIGNMENT
;
2072 struct isl_surf surf
;
2073 uint32_t tile_x_sa
= 0, tile_y_sa
= 0;
2074 get_rt_read_isl_surf(devinfo
, res
, target
, read_view
,
2075 &tile_x_sa
, &tile_y_sa
, &surf
);
2076 fill_surface_state(&screen
->isl_dev
, map_read
, res
, &surf
, read_view
,
2077 aux_usage
, tile_x_sa
, tile_y_sa
);
2078 /* Restore offset because we change offset in case of handling
2079 * non_coherent fb fetch
2081 res
->offset
= offset
;
2082 map_read
+= SURFACE_STATE_ALIGNMENT
;
2089 /* The resource has a compressed format, which is not renderable, but we
2090 * have a renderable view format. We must be attempting to upload blocks
2091 * of compressed data via an uncompressed view.
2093 * In this case, we can assume there are no auxiliary buffers, a single
2094 * miplevel, and that the resource is single-sampled. Gallium may try
2095 * and create an uncompressed view with multiple layers, however.
2097 assert(!isl_format_is_compressed(fmt
.fmt
));
2098 assert(res
->aux
.possible_usages
== 1 << ISL_AUX_USAGE_NONE
);
2099 assert(res
->surf
.samples
== 1);
2100 assert(view
->levels
== 1);
2102 struct isl_surf isl_surf
;
2103 uint32_t offset_B
= 0, tile_x_sa
= 0, tile_y_sa
= 0;
2105 if (view
->base_level
> 0) {
2106 /* We can't rely on the hardware's miplevel selection with such
2107 * a substantial lie about the format, so we select a single image
2108 * using the Tile X/Y Offset fields. In this case, we can't handle
2109 * multiple array slices.
2111 * On Broadwell, HALIGN and VALIGN are specified in pixels and are
2112 * hard-coded to align to exactly the block size of the compressed
2113 * texture. This means that, when reinterpreted as a non-compressed
2114 * texture, the tile offsets may be anything and we can't rely on
2117 * Return NULL to force the state tracker to take fallback paths.
2119 if (view
->array_len
> 1 || GEN_GEN
== 8)
2122 const bool is_3d
= res
->surf
.dim
== ISL_SURF_DIM_3D
;
2123 isl_surf_get_image_surf(&screen
->isl_dev
, &res
->surf
,
2125 is_3d
? 0 : view
->base_array_layer
,
2126 is_3d
? view
->base_array_layer
: 0,
2128 &offset_B
, &tile_x_sa
, &tile_y_sa
);
2130 /* We use address and tile offsets to access a single level/layer
2131 * as a subimage, so reset level/layer so it doesn't offset again.
2133 view
->base_array_layer
= 0;
2134 view
->base_level
= 0;
2136 /* Level 0 doesn't require tile offsets, and the hardware can find
2137 * array slices using QPitch even with the format override, so we
2138 * can allow layers in this case. Copy the original ISL surface.
2140 memcpy(&isl_surf
, &res
->surf
, sizeof(isl_surf
));
2143 /* Scale down the image dimensions by the block size. */
2144 const struct isl_format_layout
*fmtl
=
2145 isl_format_get_layout(res
->surf
.format
);
2146 isl_surf
.format
= fmt
.fmt
;
2147 isl_surf
.logical_level0_px
= isl_surf_get_logical_level0_el(&isl_surf
);
2148 isl_surf
.phys_level0_sa
= isl_surf_get_phys_level0_el(&isl_surf
);
2149 tile_x_sa
/= fmtl
->bw
;
2150 tile_y_sa
/= fmtl
->bh
;
2152 psurf
->width
= isl_surf
.logical_level0_px
.width
;
2153 psurf
->height
= isl_surf
.logical_level0_px
.height
;
2155 struct isl_surf_fill_state_info f
= {
2158 .mocs
= mocs(res
->bo
),
2159 .address
= res
->bo
->gtt_offset
+ offset_B
,
2160 .x_offset_sa
= tile_x_sa
,
2161 .y_offset_sa
= tile_y_sa
,
2164 isl_surf_fill_state_s(&screen
->isl_dev
, map
, &f
);
2170 fill_default_image_param(struct brw_image_param
*param
)
2172 memset(param
, 0, sizeof(*param
));
2173 /* Set the swizzling shifts to all-ones to effectively disable swizzling --
2174 * See emit_address_calculation() in brw_fs_surface_builder.cpp for a more
2175 * detailed explanation of these parameters.
2177 param
->swizzling
[0] = 0xff;
2178 param
->swizzling
[1] = 0xff;
2182 fill_buffer_image_param(struct brw_image_param
*param
,
2183 enum pipe_format pfmt
,
2186 const unsigned cpp
= util_format_get_blocksize(pfmt
);
2188 fill_default_image_param(param
);
2189 param
->size
[0] = size
/ cpp
;
2190 param
->stride
[0] = cpp
;
2193 #define isl_surf_fill_image_param(x, ...)
2194 #define fill_default_image_param(x, ...)
2195 #define fill_buffer_image_param(x, ...)
2199 * The pipe->set_shader_images() driver hook.
2202 iris_set_shader_images(struct pipe_context
*ctx
,
2203 enum pipe_shader_type p_stage
,
2204 unsigned start_slot
, unsigned count
,
2205 const struct pipe_image_view
*p_images
)
2207 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2208 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
2209 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
2210 gl_shader_stage stage
= stage_from_pipe(p_stage
);
2211 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
2213 struct iris_genx_state
*genx
= ice
->state
.genx
;
2214 struct brw_image_param
*image_params
= genx
->shaders
[stage
].image_param
;
2217 shs
->bound_image_views
&= ~u_bit_consecutive(start_slot
, count
);
2219 for (unsigned i
= 0; i
< count
; i
++) {
2220 struct iris_image_view
*iv
= &shs
->image
[start_slot
+ i
];
2222 if (p_images
&& p_images
[i
].resource
) {
2223 const struct pipe_image_view
*img
= &p_images
[i
];
2224 struct iris_resource
*res
= (void *) img
->resource
;
2227 alloc_surface_states(ice
->state
.surface_uploader
,
2228 &iv
->surface_state
, 1 << ISL_AUX_USAGE_NONE
);
2232 util_copy_image_view(&iv
->base
, img
);
2234 shs
->bound_image_views
|= 1 << (start_slot
+ i
);
2236 res
->bind_history
|= PIPE_BIND_SHADER_IMAGE
;
2238 isl_surf_usage_flags_t usage
= ISL_SURF_USAGE_STORAGE_BIT
;
2239 enum isl_format isl_fmt
=
2240 iris_format_for_usage(devinfo
, img
->format
, usage
).fmt
;
2242 bool untyped_fallback
= false;
2244 if (img
->shader_access
& PIPE_IMAGE_ACCESS_READ
) {
2245 /* On Gen8, try to use typed surfaces reads (which support a
2246 * limited number of formats), and if not possible, fall back
2249 untyped_fallback
= GEN_GEN
== 8 &&
2250 !isl_has_matching_typed_storage_image_format(devinfo
, isl_fmt
);
2252 if (untyped_fallback
)
2253 isl_fmt
= ISL_FORMAT_RAW
;
2255 isl_fmt
= isl_lower_storage_image_format(devinfo
, isl_fmt
);
2258 if (res
->base
.target
!= PIPE_BUFFER
) {
2259 struct isl_view view
= {
2261 .base_level
= img
->u
.tex
.level
,
2263 .base_array_layer
= img
->u
.tex
.first_layer
,
2264 .array_len
= img
->u
.tex
.last_layer
- img
->u
.tex
.first_layer
+ 1,
2265 .swizzle
= ISL_SWIZZLE_IDENTITY
,
2269 if (untyped_fallback
) {
2270 fill_buffer_surface_state(&screen
->isl_dev
, res
, map
,
2271 isl_fmt
, ISL_SWIZZLE_IDENTITY
,
2274 /* Images don't support compression */
2275 unsigned aux_modes
= 1 << ISL_AUX_USAGE_NONE
;
2277 enum isl_aux_usage usage
= u_bit_scan(&aux_modes
);
2279 fill_surface_state(&screen
->isl_dev
, map
, res
, &res
->surf
,
2280 &view
, usage
, 0, 0);
2282 map
+= SURFACE_STATE_ALIGNMENT
;
2286 isl_surf_fill_image_param(&screen
->isl_dev
,
2287 &image_params
[start_slot
+ i
],
2290 util_range_add(&res
->valid_buffer_range
, img
->u
.buf
.offset
,
2291 img
->u
.buf
.offset
+ img
->u
.buf
.size
);
2293 fill_buffer_surface_state(&screen
->isl_dev
, res
, map
,
2294 isl_fmt
, ISL_SWIZZLE_IDENTITY
,
2295 img
->u
.buf
.offset
, img
->u
.buf
.size
);
2296 fill_buffer_image_param(&image_params
[start_slot
+ i
],
2297 img
->format
, img
->u
.buf
.size
);
2300 pipe_resource_reference(&iv
->base
.resource
, NULL
);
2301 pipe_resource_reference(&iv
->surface_state
.res
, NULL
);
2302 fill_default_image_param(&image_params
[start_slot
+ i
]);
2306 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_VS
<< stage
;
2308 stage
== MESA_SHADER_COMPUTE
? IRIS_DIRTY_COMPUTE_RESOLVES_AND_FLUSHES
2309 : IRIS_DIRTY_RENDER_RESOLVES_AND_FLUSHES
;
2311 /* Broadwell also needs brw_image_params re-uploaded */
2313 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_VS
<< stage
;
2314 shs
->sysvals_need_upload
= true;
2320 * The pipe->set_sampler_views() driver hook.
2323 iris_set_sampler_views(struct pipe_context
*ctx
,
2324 enum pipe_shader_type p_stage
,
2325 unsigned start
, unsigned count
,
2326 struct pipe_sampler_view
**views
)
2328 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2329 gl_shader_stage stage
= stage_from_pipe(p_stage
);
2330 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
2332 shs
->bound_sampler_views
&= ~u_bit_consecutive(start
, count
);
2334 for (unsigned i
= 0; i
< count
; i
++) {
2335 struct pipe_sampler_view
*pview
= views
? views
[i
] : NULL
;
2336 pipe_sampler_view_reference((struct pipe_sampler_view
**)
2337 &shs
->textures
[start
+ i
], pview
);
2338 struct iris_sampler_view
*view
= (void *) pview
;
2340 view
->res
->bind_history
|= PIPE_BIND_SAMPLER_VIEW
;
2341 shs
->bound_sampler_views
|= 1 << (start
+ i
);
2345 ice
->state
.dirty
|= (IRIS_DIRTY_BINDINGS_VS
<< stage
);
2347 stage
== MESA_SHADER_COMPUTE
? IRIS_DIRTY_COMPUTE_RESOLVES_AND_FLUSHES
2348 : IRIS_DIRTY_RENDER_RESOLVES_AND_FLUSHES
;
2352 * The pipe->set_tess_state() driver hook.
2355 iris_set_tess_state(struct pipe_context
*ctx
,
2356 const float default_outer_level
[4],
2357 const float default_inner_level
[2])
2359 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2360 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_TESS_CTRL
];
2362 memcpy(&ice
->state
.default_outer_level
[0], &default_outer_level
[0], 4 * sizeof(float));
2363 memcpy(&ice
->state
.default_inner_level
[0], &default_inner_level
[0], 2 * sizeof(float));
2365 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_TCS
;
2366 shs
->sysvals_need_upload
= true;
2370 iris_surface_destroy(struct pipe_context
*ctx
, struct pipe_surface
*p_surf
)
2372 struct iris_surface
*surf
= (void *) p_surf
;
2373 pipe_resource_reference(&p_surf
->texture
, NULL
);
2374 pipe_resource_reference(&surf
->surface_state
.res
, NULL
);
2375 pipe_resource_reference(&surf
->surface_state_read
.res
, NULL
);
2380 iris_set_clip_state(struct pipe_context
*ctx
,
2381 const struct pipe_clip_state
*state
)
2383 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2384 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_VERTEX
];
2385 struct iris_shader_state
*gshs
= &ice
->state
.shaders
[MESA_SHADER_GEOMETRY
];
2386 struct iris_shader_state
*tshs
= &ice
->state
.shaders
[MESA_SHADER_TESS_EVAL
];
2388 memcpy(&ice
->state
.clip_planes
, state
, sizeof(*state
));
2390 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_VS
| IRIS_DIRTY_CONSTANTS_GS
|
2391 IRIS_DIRTY_CONSTANTS_TES
;
2392 shs
->sysvals_need_upload
= true;
2393 gshs
->sysvals_need_upload
= true;
2394 tshs
->sysvals_need_upload
= true;
2398 * The pipe->set_polygon_stipple() driver hook.
2401 iris_set_polygon_stipple(struct pipe_context
*ctx
,
2402 const struct pipe_poly_stipple
*state
)
2404 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2405 memcpy(&ice
->state
.poly_stipple
, state
, sizeof(*state
));
2406 ice
->state
.dirty
|= IRIS_DIRTY_POLYGON_STIPPLE
;
2410 * The pipe->set_sample_mask() driver hook.
2413 iris_set_sample_mask(struct pipe_context
*ctx
, unsigned sample_mask
)
2415 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2417 /* We only support 16x MSAA, so we have 16 bits of sample maks.
2418 * st/mesa may pass us 0xffffffff though, meaning "enable all samples".
2420 ice
->state
.sample_mask
= sample_mask
& 0xffff;
2421 ice
->state
.dirty
|= IRIS_DIRTY_SAMPLE_MASK
;
2425 * The pipe->set_scissor_states() driver hook.
2427 * This corresponds to our SCISSOR_RECT state structures. It's an
2428 * exact match, so we just store them, and memcpy them out later.
2431 iris_set_scissor_states(struct pipe_context
*ctx
,
2432 unsigned start_slot
,
2433 unsigned num_scissors
,
2434 const struct pipe_scissor_state
*rects
)
2436 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2438 for (unsigned i
= 0; i
< num_scissors
; i
++) {
2439 if (rects
[i
].minx
== rects
[i
].maxx
|| rects
[i
].miny
== rects
[i
].maxy
) {
2440 /* If the scissor was out of bounds and got clamped to 0 width/height
2441 * at the bounds, the subtraction of 1 from maximums could produce a
2442 * negative number and thus not clip anything. Instead, just provide
2443 * a min > max scissor inside the bounds, which produces the expected
2446 ice
->state
.scissors
[start_slot
+ i
] = (struct pipe_scissor_state
) {
2447 .minx
= 1, .maxx
= 0, .miny
= 1, .maxy
= 0,
2450 ice
->state
.scissors
[start_slot
+ i
] = (struct pipe_scissor_state
) {
2451 .minx
= rects
[i
].minx
, .miny
= rects
[i
].miny
,
2452 .maxx
= rects
[i
].maxx
- 1, .maxy
= rects
[i
].maxy
- 1,
2457 ice
->state
.dirty
|= IRIS_DIRTY_SCISSOR_RECT
;
2461 * The pipe->set_stencil_ref() driver hook.
2463 * This is added to 3DSTATE_WM_DEPTH_STENCIL dynamically at draw time.
2466 iris_set_stencil_ref(struct pipe_context
*ctx
,
2467 const struct pipe_stencil_ref
*state
)
2469 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2470 memcpy(&ice
->state
.stencil_ref
, state
, sizeof(*state
));
2472 ice
->state
.dirty
|= IRIS_DIRTY_COLOR_CALC_STATE
;
2474 ice
->state
.dirty
|= IRIS_DIRTY_WM_DEPTH_STENCIL
;
2478 viewport_extent(const struct pipe_viewport_state
*state
, int axis
, float sign
)
2480 return copysignf(state
->scale
[axis
], sign
) + state
->translate
[axis
];
2484 * The pipe->set_viewport_states() driver hook.
2486 * This corresponds to our SF_CLIP_VIEWPORT states. We can't calculate
2487 * the guardband yet, as we need the framebuffer dimensions, but we can
2488 * at least fill out the rest.
2491 iris_set_viewport_states(struct pipe_context
*ctx
,
2492 unsigned start_slot
,
2494 const struct pipe_viewport_state
*states
)
2496 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2498 memcpy(&ice
->state
.viewports
[start_slot
], states
, sizeof(*states
) * count
);
2500 ice
->state
.dirty
|= IRIS_DIRTY_SF_CL_VIEWPORT
;
2502 if (ice
->state
.cso_rast
&& (!ice
->state
.cso_rast
->depth_clip_near
||
2503 !ice
->state
.cso_rast
->depth_clip_far
))
2504 ice
->state
.dirty
|= IRIS_DIRTY_CC_VIEWPORT
;
2508 * The pipe->set_framebuffer_state() driver hook.
2510 * Sets the current draw FBO, including color render targets, depth,
2511 * and stencil buffers.
2514 iris_set_framebuffer_state(struct pipe_context
*ctx
,
2515 const struct pipe_framebuffer_state
*state
)
2517 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2518 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
2519 struct isl_device
*isl_dev
= &screen
->isl_dev
;
2520 struct pipe_framebuffer_state
*cso
= &ice
->state
.framebuffer
;
2521 struct iris_resource
*zres
;
2522 struct iris_resource
*stencil_res
;
2524 unsigned samples
= util_framebuffer_get_num_samples(state
);
2525 unsigned layers
= util_framebuffer_get_num_layers(state
);
2527 if (cso
->samples
!= samples
) {
2528 ice
->state
.dirty
|= IRIS_DIRTY_MULTISAMPLE
;
2530 /* We need to toggle 3DSTATE_PS::32 Pixel Dispatch Enable */
2531 if (GEN_GEN
>= 9 && (cso
->samples
== 16 || samples
== 16))
2532 ice
->state
.dirty
|= IRIS_DIRTY_FS
;
2535 if (cso
->nr_cbufs
!= state
->nr_cbufs
) {
2536 ice
->state
.dirty
|= IRIS_DIRTY_BLEND_STATE
;
2539 if ((cso
->layers
== 0) != (layers
== 0)) {
2540 ice
->state
.dirty
|= IRIS_DIRTY_CLIP
;
2543 if (cso
->width
!= state
->width
|| cso
->height
!= state
->height
) {
2544 ice
->state
.dirty
|= IRIS_DIRTY_SF_CL_VIEWPORT
;
2547 if (cso
->zsbuf
|| state
->zsbuf
) {
2548 ice
->state
.dirty
|= IRIS_DIRTY_DEPTH_BUFFER
;
2551 util_copy_framebuffer_state(cso
, state
);
2552 cso
->samples
= samples
;
2553 cso
->layers
= layers
;
2555 struct iris_depth_buffer_state
*cso_z
= &ice
->state
.genx
->depth_buffer
;
2557 struct isl_view view
= {
2560 .base_array_layer
= 0,
2562 .swizzle
= ISL_SWIZZLE_IDENTITY
,
2565 struct isl_depth_stencil_hiz_emit_info info
= { .view
= &view
};
2568 iris_get_depth_stencil_resources(cso
->zsbuf
->texture
, &zres
,
2571 view
.base_level
= cso
->zsbuf
->u
.tex
.level
;
2572 view
.base_array_layer
= cso
->zsbuf
->u
.tex
.first_layer
;
2574 cso
->zsbuf
->u
.tex
.last_layer
- cso
->zsbuf
->u
.tex
.first_layer
+ 1;
2577 view
.usage
|= ISL_SURF_USAGE_DEPTH_BIT
;
2579 info
.depth_surf
= &zres
->surf
;
2580 info
.depth_address
= zres
->bo
->gtt_offset
+ zres
->offset
;
2581 info
.mocs
= mocs(zres
->bo
);
2583 view
.format
= zres
->surf
.format
;
2585 if (iris_resource_level_has_hiz(zres
, view
.base_level
)) {
2586 info
.hiz_usage
= ISL_AUX_USAGE_HIZ
;
2587 info
.hiz_surf
= &zres
->aux
.surf
;
2588 info
.hiz_address
= zres
->aux
.bo
->gtt_offset
+ zres
->aux
.offset
;
2593 view
.usage
|= ISL_SURF_USAGE_STENCIL_BIT
;
2594 info
.stencil_surf
= &stencil_res
->surf
;
2595 info
.stencil_address
= stencil_res
->bo
->gtt_offset
+ stencil_res
->offset
;
2597 view
.format
= stencil_res
->surf
.format
;
2598 info
.mocs
= mocs(stencil_res
->bo
);
2603 isl_emit_depth_stencil_hiz_s(isl_dev
, cso_z
->packets
, &info
);
2605 /* Make a null surface for unbound buffers */
2606 void *null_surf_map
=
2607 upload_state(ice
->state
.surface_uploader
, &ice
->state
.null_fb
,
2608 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
2609 isl_null_fill_state(&screen
->isl_dev
, null_surf_map
,
2610 isl_extent3d(MAX2(cso
->width
, 1),
2611 MAX2(cso
->height
, 1),
2612 cso
->layers
? cso
->layers
: 1));
2613 ice
->state
.null_fb
.offset
+=
2614 iris_bo_offset_from_base_address(iris_resource_bo(ice
->state
.null_fb
.res
));
2616 /* Render target change */
2617 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_FS
;
2619 ice
->state
.dirty
|= IRIS_DIRTY_RENDER_RESOLVES_AND_FLUSHES
;
2621 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_FRAMEBUFFER
];
2624 // XXX: we may want to flag IRIS_DIRTY_MULTISAMPLE (or SAMPLE_MASK?)
2625 // XXX: see commit 979fc1bc9bcc64027ff2cfafd285676f31b930a6
2627 /* The PIPE_CONTROL command description says:
2629 * "Whenever a Binding Table Index (BTI) used by a Render Target Message
2630 * points to a different RENDER_SURFACE_STATE, SW must issue a Render
2631 * Target Cache Flush by enabling this bit. When render target flush
2632 * is set due to new association of BTI, PS Scoreboard Stall bit must
2633 * be set in this packet."
2635 // XXX: does this need to happen at 3DSTATE_BTP_PS time?
2636 iris_emit_pipe_control_flush(&ice
->batches
[IRIS_BATCH_RENDER
],
2637 "workaround: RT BTI change [draw]",
2638 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
2639 PIPE_CONTROL_STALL_AT_SCOREBOARD
);
2644 * The pipe->set_constant_buffer() driver hook.
2646 * This uploads any constant data in user buffers, and references
2647 * any UBO resources containing constant data.
2650 iris_set_constant_buffer(struct pipe_context
*ctx
,
2651 enum pipe_shader_type p_stage
, unsigned index
,
2652 const struct pipe_constant_buffer
*input
)
2654 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2655 gl_shader_stage stage
= stage_from_pipe(p_stage
);
2656 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
2657 struct pipe_shader_buffer
*cbuf
= &shs
->constbuf
[index
];
2659 if (input
&& input
->buffer_size
&& (input
->buffer
|| input
->user_buffer
)) {
2660 shs
->bound_cbufs
|= 1u << index
;
2662 if (input
->user_buffer
) {
2664 pipe_resource_reference(&cbuf
->buffer
, NULL
);
2665 u_upload_alloc(ice
->ctx
.const_uploader
, 0, input
->buffer_size
, 64,
2666 &cbuf
->buffer_offset
, &cbuf
->buffer
, (void **) &map
);
2668 if (!cbuf
->buffer
) {
2669 /* Allocation was unsuccessful - just unbind */
2670 iris_set_constant_buffer(ctx
, p_stage
, index
, NULL
);
2675 memcpy(map
, input
->user_buffer
, input
->buffer_size
);
2676 } else if (input
->buffer
) {
2677 pipe_resource_reference(&cbuf
->buffer
, input
->buffer
);
2679 cbuf
->buffer_offset
= input
->buffer_offset
;
2681 MIN2(input
->buffer_size
,
2682 iris_resource_bo(cbuf
->buffer
)->size
- cbuf
->buffer_offset
);
2685 struct iris_resource
*res
= (void *) cbuf
->buffer
;
2686 res
->bind_history
|= PIPE_BIND_CONSTANT_BUFFER
;
2688 iris_upload_ubo_ssbo_surf_state(ice
, cbuf
,
2689 &shs
->constbuf_surf_state
[index
],
2692 shs
->bound_cbufs
&= ~(1u << index
);
2693 pipe_resource_reference(&cbuf
->buffer
, NULL
);
2694 pipe_resource_reference(&shs
->constbuf_surf_state
[index
].res
, NULL
);
2697 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_VS
<< stage
;
2698 // XXX: maybe not necessary all the time...?
2699 // XXX: we need 3DS_BTP to commit these changes, and if we fell back to
2700 // XXX: pull model we may need actual new bindings...
2701 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_VS
<< stage
;
2705 upload_sysvals(struct iris_context
*ice
,
2706 gl_shader_stage stage
)
2708 UNUSED
struct iris_genx_state
*genx
= ice
->state
.genx
;
2709 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
2711 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
2712 if (!shader
|| shader
->num_system_values
== 0)
2715 assert(shader
->num_cbufs
> 0);
2717 unsigned sysval_cbuf_index
= shader
->num_cbufs
- 1;
2718 struct pipe_shader_buffer
*cbuf
= &shs
->constbuf
[sysval_cbuf_index
];
2719 unsigned upload_size
= shader
->num_system_values
* sizeof(uint32_t);
2720 uint32_t *map
= NULL
;
2722 assert(sysval_cbuf_index
< PIPE_MAX_CONSTANT_BUFFERS
);
2723 u_upload_alloc(ice
->ctx
.const_uploader
, 0, upload_size
, 64,
2724 &cbuf
->buffer_offset
, &cbuf
->buffer
, (void **) &map
);
2726 for (int i
= 0; i
< shader
->num_system_values
; i
++) {
2727 uint32_t sysval
= shader
->system_values
[i
];
2730 if (BRW_PARAM_DOMAIN(sysval
) == BRW_PARAM_DOMAIN_IMAGE
) {
2732 unsigned img
= BRW_PARAM_IMAGE_IDX(sysval
);
2733 unsigned offset
= BRW_PARAM_IMAGE_OFFSET(sysval
);
2734 struct brw_image_param
*param
=
2735 &genx
->shaders
[stage
].image_param
[img
];
2737 assert(offset
< sizeof(struct brw_image_param
));
2738 value
= ((uint32_t *) param
)[offset
];
2740 } else if (sysval
== BRW_PARAM_BUILTIN_ZERO
) {
2742 } else if (BRW_PARAM_BUILTIN_IS_CLIP_PLANE(sysval
)) {
2743 int plane
= BRW_PARAM_BUILTIN_CLIP_PLANE_IDX(sysval
);
2744 int comp
= BRW_PARAM_BUILTIN_CLIP_PLANE_COMP(sysval
);
2745 value
= fui(ice
->state
.clip_planes
.ucp
[plane
][comp
]);
2746 } else if (sysval
== BRW_PARAM_BUILTIN_PATCH_VERTICES_IN
) {
2747 if (stage
== MESA_SHADER_TESS_CTRL
) {
2748 value
= ice
->state
.vertices_per_patch
;
2750 assert(stage
== MESA_SHADER_TESS_EVAL
);
2751 const struct shader_info
*tcs_info
=
2752 iris_get_shader_info(ice
, MESA_SHADER_TESS_CTRL
);
2754 value
= tcs_info
->tess
.tcs_vertices_out
;
2756 value
= ice
->state
.vertices_per_patch
;
2758 } else if (sysval
>= BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_X
&&
2759 sysval
<= BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_W
) {
2760 unsigned i
= sysval
- BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_X
;
2761 value
= fui(ice
->state
.default_outer_level
[i
]);
2762 } else if (sysval
== BRW_PARAM_BUILTIN_TESS_LEVEL_INNER_X
) {
2763 value
= fui(ice
->state
.default_inner_level
[0]);
2764 } else if (sysval
== BRW_PARAM_BUILTIN_TESS_LEVEL_INNER_Y
) {
2765 value
= fui(ice
->state
.default_inner_level
[1]);
2767 assert(!"unhandled system value");
2773 cbuf
->buffer_size
= upload_size
;
2774 iris_upload_ubo_ssbo_surf_state(ice
, cbuf
,
2775 &shs
->constbuf_surf_state
[sysval_cbuf_index
], false);
2777 shs
->sysvals_need_upload
= false;
2781 * The pipe->set_shader_buffers() driver hook.
2783 * This binds SSBOs and ABOs. Unfortunately, we need to stream out
2784 * SURFACE_STATE here, as the buffer offset may change each time.
2787 iris_set_shader_buffers(struct pipe_context
*ctx
,
2788 enum pipe_shader_type p_stage
,
2789 unsigned start_slot
, unsigned count
,
2790 const struct pipe_shader_buffer
*buffers
,
2791 unsigned writable_bitmask
)
2793 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2794 gl_shader_stage stage
= stage_from_pipe(p_stage
);
2795 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
2797 unsigned modified_bits
= u_bit_consecutive(start_slot
, count
);
2799 shs
->bound_ssbos
&= ~modified_bits
;
2800 shs
->writable_ssbos
&= ~modified_bits
;
2801 shs
->writable_ssbos
|= writable_bitmask
<< start_slot
;
2803 for (unsigned i
= 0; i
< count
; i
++) {
2804 if (buffers
&& buffers
[i
].buffer
) {
2805 struct iris_resource
*res
= (void *) buffers
[i
].buffer
;
2806 struct pipe_shader_buffer
*ssbo
= &shs
->ssbo
[start_slot
+ i
];
2807 struct iris_state_ref
*surf_state
=
2808 &shs
->ssbo_surf_state
[start_slot
+ i
];
2809 pipe_resource_reference(&ssbo
->buffer
, &res
->base
);
2810 ssbo
->buffer_offset
= buffers
[i
].buffer_offset
;
2812 MIN2(buffers
[i
].buffer_size
, res
->bo
->size
- ssbo
->buffer_offset
);
2814 shs
->bound_ssbos
|= 1 << (start_slot
+ i
);
2816 iris_upload_ubo_ssbo_surf_state(ice
, ssbo
, surf_state
, true);
2818 res
->bind_history
|= PIPE_BIND_SHADER_BUFFER
;
2820 util_range_add(&res
->valid_buffer_range
, ssbo
->buffer_offset
,
2821 ssbo
->buffer_offset
+ ssbo
->buffer_size
);
2823 pipe_resource_reference(&shs
->ssbo
[start_slot
+ i
].buffer
, NULL
);
2824 pipe_resource_reference(&shs
->ssbo_surf_state
[start_slot
+ i
].res
,
2829 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_VS
<< stage
;
2833 iris_delete_state(struct pipe_context
*ctx
, void *state
)
2839 * The pipe->set_vertex_buffers() driver hook.
2841 * This translates pipe_vertex_buffer to our 3DSTATE_VERTEX_BUFFERS packet.
2844 iris_set_vertex_buffers(struct pipe_context
*ctx
,
2845 unsigned start_slot
, unsigned count
,
2846 const struct pipe_vertex_buffer
*buffers
)
2848 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2849 struct iris_genx_state
*genx
= ice
->state
.genx
;
2851 ice
->state
.bound_vertex_buffers
&= ~u_bit_consecutive64(start_slot
, count
);
2853 for (unsigned i
= 0; i
< count
; i
++) {
2854 const struct pipe_vertex_buffer
*buffer
= buffers
? &buffers
[i
] : NULL
;
2855 struct iris_vertex_buffer_state
*state
=
2856 &genx
->vertex_buffers
[start_slot
+ i
];
2859 pipe_resource_reference(&state
->resource
, NULL
);
2863 /* We may see user buffers that are NULL bindings. */
2864 assert(!(buffer
->is_user_buffer
&& buffer
->buffer
.user
!= NULL
));
2866 pipe_resource_reference(&state
->resource
, buffer
->buffer
.resource
);
2867 struct iris_resource
*res
= (void *) state
->resource
;
2870 ice
->state
.bound_vertex_buffers
|= 1ull << (start_slot
+ i
);
2871 res
->bind_history
|= PIPE_BIND_VERTEX_BUFFER
;
2874 iris_pack_state(GENX(VERTEX_BUFFER_STATE
), state
->state
, vb
) {
2875 vb
.VertexBufferIndex
= start_slot
+ i
;
2876 vb
.AddressModifyEnable
= true;
2877 vb
.BufferPitch
= buffer
->stride
;
2879 vb
.BufferSize
= res
->bo
->size
- (int) buffer
->buffer_offset
;
2880 vb
.BufferStartingAddress
=
2881 ro_bo(NULL
, res
->bo
->gtt_offset
+ (int) buffer
->buffer_offset
);
2882 vb
.MOCS
= mocs(res
->bo
);
2884 vb
.NullVertexBuffer
= true;
2889 ice
->state
.dirty
|= IRIS_DIRTY_VERTEX_BUFFERS
;
2893 * Gallium CSO for vertex elements.
2895 struct iris_vertex_element_state
{
2896 uint32_t vertex_elements
[1 + 33 * GENX(VERTEX_ELEMENT_STATE_length
)];
2897 uint32_t vf_instancing
[33 * GENX(3DSTATE_VF_INSTANCING_length
)];
2898 uint32_t edgeflag_ve
[GENX(VERTEX_ELEMENT_STATE_length
)];
2899 uint32_t edgeflag_vfi
[GENX(3DSTATE_VF_INSTANCING_length
)];
2904 * The pipe->create_vertex_elements() driver hook.
2906 * This translates pipe_vertex_element to our 3DSTATE_VERTEX_ELEMENTS
2907 * and 3DSTATE_VF_INSTANCING commands. The vertex_elements and vf_instancing
2908 * arrays are ready to be emitted at draw time if no EdgeFlag or SGVs are
2909 * needed. In these cases we will need information available at draw time.
2910 * We setup edgeflag_ve and edgeflag_vfi as alternatives last
2911 * 3DSTATE_VERTEX_ELEMENT and 3DSTATE_VF_INSTANCING that can be used at
2912 * draw time if we detect that EdgeFlag is needed by the Vertex Shader.
2915 iris_create_vertex_elements(struct pipe_context
*ctx
,
2917 const struct pipe_vertex_element
*state
)
2919 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
2920 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
2921 struct iris_vertex_element_state
*cso
=
2922 malloc(sizeof(struct iris_vertex_element_state
));
2926 iris_pack_command(GENX(3DSTATE_VERTEX_ELEMENTS
), cso
->vertex_elements
, ve
) {
2928 1 + GENX(VERTEX_ELEMENT_STATE_length
) * MAX2(count
, 1) - 2;
2931 uint32_t *ve_pack_dest
= &cso
->vertex_elements
[1];
2932 uint32_t *vfi_pack_dest
= cso
->vf_instancing
;
2935 iris_pack_state(GENX(VERTEX_ELEMENT_STATE
), ve_pack_dest
, ve
) {
2937 ve
.SourceElementFormat
= ISL_FORMAT_R32G32B32A32_FLOAT
;
2938 ve
.Component0Control
= VFCOMP_STORE_0
;
2939 ve
.Component1Control
= VFCOMP_STORE_0
;
2940 ve
.Component2Control
= VFCOMP_STORE_0
;
2941 ve
.Component3Control
= VFCOMP_STORE_1_FP
;
2944 iris_pack_command(GENX(3DSTATE_VF_INSTANCING
), vfi_pack_dest
, vi
) {
2948 for (int i
= 0; i
< count
; i
++) {
2949 const struct iris_format_info fmt
=
2950 iris_format_for_usage(devinfo
, state
[i
].src_format
, 0);
2951 unsigned comp
[4] = { VFCOMP_STORE_SRC
, VFCOMP_STORE_SRC
,
2952 VFCOMP_STORE_SRC
, VFCOMP_STORE_SRC
};
2954 switch (isl_format_get_num_channels(fmt
.fmt
)) {
2955 case 0: comp
[0] = VFCOMP_STORE_0
; /* fallthrough */
2956 case 1: comp
[1] = VFCOMP_STORE_0
; /* fallthrough */
2957 case 2: comp
[2] = VFCOMP_STORE_0
; /* fallthrough */
2959 comp
[3] = isl_format_has_int_channel(fmt
.fmt
) ? VFCOMP_STORE_1_INT
2960 : VFCOMP_STORE_1_FP
;
2963 iris_pack_state(GENX(VERTEX_ELEMENT_STATE
), ve_pack_dest
, ve
) {
2964 ve
.EdgeFlagEnable
= false;
2965 ve
.VertexBufferIndex
= state
[i
].vertex_buffer_index
;
2967 ve
.SourceElementOffset
= state
[i
].src_offset
;
2968 ve
.SourceElementFormat
= fmt
.fmt
;
2969 ve
.Component0Control
= comp
[0];
2970 ve
.Component1Control
= comp
[1];
2971 ve
.Component2Control
= comp
[2];
2972 ve
.Component3Control
= comp
[3];
2975 iris_pack_command(GENX(3DSTATE_VF_INSTANCING
), vfi_pack_dest
, vi
) {
2976 vi
.VertexElementIndex
= i
;
2977 vi
.InstancingEnable
= state
[i
].instance_divisor
> 0;
2978 vi
.InstanceDataStepRate
= state
[i
].instance_divisor
;
2981 ve_pack_dest
+= GENX(VERTEX_ELEMENT_STATE_length
);
2982 vfi_pack_dest
+= GENX(3DSTATE_VF_INSTANCING_length
);
2985 /* An alternative version of the last VE and VFI is stored so it
2986 * can be used at draw time in case Vertex Shader uses EdgeFlag
2989 const unsigned edgeflag_index
= count
- 1;
2990 const struct iris_format_info fmt
=
2991 iris_format_for_usage(devinfo
, state
[edgeflag_index
].src_format
, 0);
2992 iris_pack_state(GENX(VERTEX_ELEMENT_STATE
), cso
->edgeflag_ve
, ve
) {
2993 ve
.EdgeFlagEnable
= true ;
2994 ve
.VertexBufferIndex
= state
[edgeflag_index
].vertex_buffer_index
;
2996 ve
.SourceElementOffset
= state
[edgeflag_index
].src_offset
;
2997 ve
.SourceElementFormat
= fmt
.fmt
;
2998 ve
.Component0Control
= VFCOMP_STORE_SRC
;
2999 ve
.Component1Control
= VFCOMP_STORE_0
;
3000 ve
.Component2Control
= VFCOMP_STORE_0
;
3001 ve
.Component3Control
= VFCOMP_STORE_0
;
3003 iris_pack_command(GENX(3DSTATE_VF_INSTANCING
), cso
->edgeflag_vfi
, vi
) {
3004 /* The vi.VertexElementIndex of the EdgeFlag Vertex Element is filled
3005 * at draw time, as it should change if SGVs are emitted.
3007 vi
.InstancingEnable
= state
[edgeflag_index
].instance_divisor
> 0;
3008 vi
.InstanceDataStepRate
= state
[edgeflag_index
].instance_divisor
;
3016 * The pipe->bind_vertex_elements_state() driver hook.
3019 iris_bind_vertex_elements_state(struct pipe_context
*ctx
, void *state
)
3021 struct iris_context
*ice
= (struct iris_context
*) ctx
;
3022 struct iris_vertex_element_state
*old_cso
= ice
->state
.cso_vertex_elements
;
3023 struct iris_vertex_element_state
*new_cso
= state
;
3025 /* 3DSTATE_VF_SGVs overrides the last VE, so if the count is changing,
3026 * we need to re-emit it to ensure we're overriding the right one.
3028 if (new_cso
&& cso_changed(count
))
3029 ice
->state
.dirty
|= IRIS_DIRTY_VF_SGVS
;
3031 ice
->state
.cso_vertex_elements
= state
;
3032 ice
->state
.dirty
|= IRIS_DIRTY_VERTEX_ELEMENTS
;
3036 * The pipe->create_stream_output_target() driver hook.
3038 * "Target" here refers to a destination buffer. We translate this into
3039 * a 3DSTATE_SO_BUFFER packet. We can handle most fields, but don't yet
3040 * know which buffer this represents, or whether we ought to zero the
3041 * write-offsets, or append. Those are handled in the set() hook.
3043 static struct pipe_stream_output_target
*
3044 iris_create_stream_output_target(struct pipe_context
*ctx
,
3045 struct pipe_resource
*p_res
,
3046 unsigned buffer_offset
,
3047 unsigned buffer_size
)
3049 struct iris_resource
*res
= (void *) p_res
;
3050 struct iris_stream_output_target
*cso
= calloc(1, sizeof(*cso
));
3054 res
->bind_history
|= PIPE_BIND_STREAM_OUTPUT
;
3056 pipe_reference_init(&cso
->base
.reference
, 1);
3057 pipe_resource_reference(&cso
->base
.buffer
, p_res
);
3058 cso
->base
.buffer_offset
= buffer_offset
;
3059 cso
->base
.buffer_size
= buffer_size
;
3060 cso
->base
.context
= ctx
;
3062 util_range_add(&res
->valid_buffer_range
, buffer_offset
,
3063 buffer_offset
+ buffer_size
);
3065 upload_state(ctx
->stream_uploader
, &cso
->offset
, sizeof(uint32_t), 4);
3071 iris_stream_output_target_destroy(struct pipe_context
*ctx
,
3072 struct pipe_stream_output_target
*state
)
3074 struct iris_stream_output_target
*cso
= (void *) state
;
3076 pipe_resource_reference(&cso
->base
.buffer
, NULL
);
3077 pipe_resource_reference(&cso
->offset
.res
, NULL
);
3083 * The pipe->set_stream_output_targets() driver hook.
3085 * At this point, we know which targets are bound to a particular index,
3086 * and also whether we want to append or start over. We can finish the
3087 * 3DSTATE_SO_BUFFER packets we started earlier.
3090 iris_set_stream_output_targets(struct pipe_context
*ctx
,
3091 unsigned num_targets
,
3092 struct pipe_stream_output_target
**targets
,
3093 const unsigned *offsets
)
3095 struct iris_context
*ice
= (struct iris_context
*) ctx
;
3096 struct iris_genx_state
*genx
= ice
->state
.genx
;
3097 uint32_t *so_buffers
= genx
->so_buffers
;
3099 const bool active
= num_targets
> 0;
3100 if (ice
->state
.streamout_active
!= active
) {
3101 ice
->state
.streamout_active
= active
;
3102 ice
->state
.dirty
|= IRIS_DIRTY_STREAMOUT
;
3104 /* We only emit 3DSTATE_SO_DECL_LIST when streamout is active, because
3105 * it's a non-pipelined command. If we're switching streamout on, we
3106 * may have missed emitting it earlier, so do so now. (We're already
3107 * taking a stall to update 3DSTATE_SO_BUFFERS anyway...)
3110 ice
->state
.dirty
|= IRIS_DIRTY_SO_DECL_LIST
;
3113 for (int i
= 0; i
< PIPE_MAX_SO_BUFFERS
; i
++) {
3114 struct iris_stream_output_target
*tgt
=
3115 (void *) ice
->state
.so_target
[i
];
3117 struct iris_resource
*res
= (void *) tgt
->base
.buffer
;
3119 flush
|= iris_flush_bits_for_history(res
);
3120 iris_dirty_for_history(ice
, res
);
3123 iris_emit_pipe_control_flush(&ice
->batches
[IRIS_BATCH_RENDER
],
3124 "make streamout results visible", flush
);
3128 for (int i
= 0; i
< 4; i
++) {
3129 pipe_so_target_reference(&ice
->state
.so_target
[i
],
3130 i
< num_targets
? targets
[i
] : NULL
);
3133 /* No need to update 3DSTATE_SO_BUFFER unless SOL is active. */
3137 for (unsigned i
= 0; i
< 4; i
++,
3138 so_buffers
+= GENX(3DSTATE_SO_BUFFER_length
)) {
3140 struct iris_stream_output_target
*tgt
= (void *) ice
->state
.so_target
[i
];
3141 unsigned offset
= offsets
[i
];
3144 iris_pack_command(GENX(3DSTATE_SO_BUFFER
), so_buffers
, sob
)
3145 sob
.SOBufferIndex
= i
;
3149 struct iris_resource
*res
= (void *) tgt
->base
.buffer
;
3151 /* Note that offsets[i] will either be 0, causing us to zero
3152 * the value in the buffer, or 0xFFFFFFFF, which happens to mean
3153 * "continue appending at the existing offset."
3155 assert(offset
== 0 || offset
== 0xFFFFFFFF);
3157 /* We might be called by Begin (offset = 0), Pause, then Resume
3158 * (offset = 0xFFFFFFFF) before ever drawing (where these commands
3159 * will actually be sent to the GPU). In this case, we don't want
3160 * to append - we still want to do our initial zeroing.
3165 iris_pack_command(GENX(3DSTATE_SO_BUFFER
), so_buffers
, sob
) {
3166 sob
.SurfaceBaseAddress
=
3167 rw_bo(NULL
, res
->bo
->gtt_offset
+ tgt
->base
.buffer_offset
);
3168 sob
.SOBufferEnable
= true;
3169 sob
.StreamOffsetWriteEnable
= true;
3170 sob
.StreamOutputBufferOffsetAddressEnable
= true;
3171 sob
.MOCS
= mocs(res
->bo
);
3173 sob
.SurfaceSize
= MAX2(tgt
->base
.buffer_size
/ 4, 1) - 1;
3175 sob
.SOBufferIndex
= i
;
3176 sob
.StreamOffset
= offset
;
3177 sob
.StreamOutputBufferOffsetAddress
=
3178 rw_bo(NULL
, iris_resource_bo(tgt
->offset
.res
)->gtt_offset
+
3179 tgt
->offset
.offset
);
3183 ice
->state
.dirty
|= IRIS_DIRTY_SO_BUFFERS
;
3187 * An iris-vtable helper for encoding the 3DSTATE_SO_DECL_LIST and
3188 * 3DSTATE_STREAMOUT packets.
3190 * 3DSTATE_SO_DECL_LIST is a list of shader outputs we want the streamout
3191 * hardware to record. We can create it entirely based on the shader, with
3192 * no dynamic state dependencies.
3194 * 3DSTATE_STREAMOUT is an annoying mix of shader-based information and
3195 * state-based settings. We capture the shader-related ones here, and merge
3196 * the rest in at draw time.
3199 iris_create_so_decl_list(const struct pipe_stream_output_info
*info
,
3200 const struct brw_vue_map
*vue_map
)
3202 struct GENX(SO_DECL
) so_decl
[MAX_VERTEX_STREAMS
][128];
3203 int buffer_mask
[MAX_VERTEX_STREAMS
] = {0, 0, 0, 0};
3204 int next_offset
[MAX_VERTEX_STREAMS
] = {0, 0, 0, 0};
3205 int decls
[MAX_VERTEX_STREAMS
] = {0, 0, 0, 0};
3207 STATIC_ASSERT(ARRAY_SIZE(so_decl
[0]) >= MAX_PROGRAM_OUTPUTS
);
3209 memset(so_decl
, 0, sizeof(so_decl
));
3211 /* Construct the list of SO_DECLs to be emitted. The formatting of the
3212 * command feels strange -- each dword pair contains a SO_DECL per stream.
3214 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
3215 const struct pipe_stream_output
*output
= &info
->output
[i
];
3216 const int buffer
= output
->output_buffer
;
3217 const int varying
= output
->register_index
;
3218 const unsigned stream_id
= output
->stream
;
3219 assert(stream_id
< MAX_VERTEX_STREAMS
);
3221 buffer_mask
[stream_id
] |= 1 << buffer
;
3223 assert(vue_map
->varying_to_slot
[varying
] >= 0);
3225 /* Mesa doesn't store entries for gl_SkipComponents in the Outputs[]
3226 * array. Instead, it simply increments DstOffset for the following
3227 * input by the number of components that should be skipped.
3229 * Our hardware is unusual in that it requires us to program SO_DECLs
3230 * for fake "hole" components, rather than simply taking the offset
3231 * for each real varying. Each hole can have size 1, 2, 3, or 4; we
3232 * program as many size = 4 holes as we can, then a final hole to
3233 * accommodate the final 1, 2, or 3 remaining.
3235 int skip_components
= output
->dst_offset
- next_offset
[buffer
];
3237 while (skip_components
> 0) {
3238 so_decl
[stream_id
][decls
[stream_id
]++] = (struct GENX(SO_DECL
)) {
3240 .OutputBufferSlot
= output
->output_buffer
,
3241 .ComponentMask
= (1 << MIN2(skip_components
, 4)) - 1,
3243 skip_components
-= 4;
3246 next_offset
[buffer
] = output
->dst_offset
+ output
->num_components
;
3248 so_decl
[stream_id
][decls
[stream_id
]++] = (struct GENX(SO_DECL
)) {
3249 .OutputBufferSlot
= output
->output_buffer
,
3250 .RegisterIndex
= vue_map
->varying_to_slot
[varying
],
3252 ((1 << output
->num_components
) - 1) << output
->start_component
,
3255 if (decls
[stream_id
] > max_decls
)
3256 max_decls
= decls
[stream_id
];
3259 unsigned dwords
= GENX(3DSTATE_STREAMOUT_length
) + (3 + 2 * max_decls
);
3260 uint32_t *map
= ralloc_size(NULL
, sizeof(uint32_t) * dwords
);
3261 uint32_t *so_decl_map
= map
+ GENX(3DSTATE_STREAMOUT_length
);
3263 iris_pack_command(GENX(3DSTATE_STREAMOUT
), map
, sol
) {
3264 int urb_entry_read_offset
= 0;
3265 int urb_entry_read_length
= (vue_map
->num_slots
+ 1) / 2 -
3266 urb_entry_read_offset
;
3268 /* We always read the whole vertex. This could be reduced at some
3269 * point by reading less and offsetting the register index in the
3272 sol
.Stream0VertexReadOffset
= urb_entry_read_offset
;
3273 sol
.Stream0VertexReadLength
= urb_entry_read_length
- 1;
3274 sol
.Stream1VertexReadOffset
= urb_entry_read_offset
;
3275 sol
.Stream1VertexReadLength
= urb_entry_read_length
- 1;
3276 sol
.Stream2VertexReadOffset
= urb_entry_read_offset
;
3277 sol
.Stream2VertexReadLength
= urb_entry_read_length
- 1;
3278 sol
.Stream3VertexReadOffset
= urb_entry_read_offset
;
3279 sol
.Stream3VertexReadLength
= urb_entry_read_length
- 1;
3281 /* Set buffer pitches; 0 means unbound. */
3282 sol
.Buffer0SurfacePitch
= 4 * info
->stride
[0];
3283 sol
.Buffer1SurfacePitch
= 4 * info
->stride
[1];
3284 sol
.Buffer2SurfacePitch
= 4 * info
->stride
[2];
3285 sol
.Buffer3SurfacePitch
= 4 * info
->stride
[3];
3288 iris_pack_command(GENX(3DSTATE_SO_DECL_LIST
), so_decl_map
, list
) {
3289 list
.DWordLength
= 3 + 2 * max_decls
- 2;
3290 list
.StreamtoBufferSelects0
= buffer_mask
[0];
3291 list
.StreamtoBufferSelects1
= buffer_mask
[1];
3292 list
.StreamtoBufferSelects2
= buffer_mask
[2];
3293 list
.StreamtoBufferSelects3
= buffer_mask
[3];
3294 list
.NumEntries0
= decls
[0];
3295 list
.NumEntries1
= decls
[1];
3296 list
.NumEntries2
= decls
[2];
3297 list
.NumEntries3
= decls
[3];
3300 for (int i
= 0; i
< max_decls
; i
++) {
3301 iris_pack_state(GENX(SO_DECL_ENTRY
), so_decl_map
+ 3 + i
* 2, entry
) {
3302 entry
.Stream0Decl
= so_decl
[0][i
];
3303 entry
.Stream1Decl
= so_decl
[1][i
];
3304 entry
.Stream2Decl
= so_decl
[2][i
];
3305 entry
.Stream3Decl
= so_decl
[3][i
];
3313 iris_compute_sbe_urb_read_interval(uint64_t fs_input_slots
,
3314 const struct brw_vue_map
*last_vue_map
,
3315 bool two_sided_color
,
3316 unsigned *out_offset
,
3317 unsigned *out_length
)
3319 /* The compiler computes the first URB slot without considering COL/BFC
3320 * swizzling (because it doesn't know whether it's enabled), so we need
3321 * to do that here too. This may result in a smaller offset, which
3324 const unsigned first_slot
=
3325 brw_compute_first_urb_slot_required(fs_input_slots
, last_vue_map
);
3327 /* This becomes the URB read offset (counted in pairs of slots). */
3328 assert(first_slot
% 2 == 0);
3329 *out_offset
= first_slot
/ 2;
3331 /* We need to adjust the inputs read to account for front/back color
3332 * swizzling, as it can make the URB length longer.
3334 for (int c
= 0; c
<= 1; c
++) {
3335 if (fs_input_slots
& (VARYING_BIT_COL0
<< c
)) {
3336 /* If two sided color is enabled, the fragment shader's gl_Color
3337 * (COL0) input comes from either the gl_FrontColor (COL0) or
3338 * gl_BackColor (BFC0) input varyings. Mark BFC as used, too.
3340 if (two_sided_color
)
3341 fs_input_slots
|= (VARYING_BIT_BFC0
<< c
);
3343 /* If front color isn't written, we opt to give them back color
3344 * instead of an undefined value. Switch from COL to BFC.
3346 if (last_vue_map
->varying_to_slot
[VARYING_SLOT_COL0
+ c
] == -1) {
3347 fs_input_slots
&= ~(VARYING_BIT_COL0
<< c
);
3348 fs_input_slots
|= (VARYING_BIT_BFC0
<< c
);
3353 /* Compute the minimum URB Read Length necessary for the FS inputs.
3355 * From the Sandy Bridge PRM, Volume 2, Part 1, documentation for
3356 * 3DSTATE_SF DWord 1 bits 15:11, "Vertex URB Entry Read Length":
3358 * "This field should be set to the minimum length required to read the
3359 * maximum source attribute. The maximum source attribute is indicated
3360 * by the maximum value of the enabled Attribute # Source Attribute if
3361 * Attribute Swizzle Enable is set, Number of Output Attributes-1 if
3362 * enable is not set.
3363 * read_length = ceiling((max_source_attr + 1) / 2)
3365 * [errata] Corruption/Hang possible if length programmed larger than
3368 * Similar text exists for Ivy Bridge.
3370 * We find the last URB slot that's actually read by the FS.
3372 unsigned last_read_slot
= last_vue_map
->num_slots
- 1;
3373 while (last_read_slot
> first_slot
&& !(fs_input_slots
&
3374 (1ull << last_vue_map
->slot_to_varying
[last_read_slot
])))
3377 /* The URB read length is the difference of the two, counted in pairs. */
3378 *out_length
= DIV_ROUND_UP(last_read_slot
- first_slot
+ 1, 2);
3382 iris_emit_sbe_swiz(struct iris_batch
*batch
,
3383 const struct iris_context
*ice
,
3384 unsigned urb_read_offset
,
3385 unsigned sprite_coord_enables
)
3387 struct GENX(SF_OUTPUT_ATTRIBUTE_DETAIL
) attr_overrides
[16] = {};
3388 const struct brw_wm_prog_data
*wm_prog_data
= (void *)
3389 ice
->shaders
.prog
[MESA_SHADER_FRAGMENT
]->prog_data
;
3390 const struct brw_vue_map
*vue_map
= ice
->shaders
.last_vue_map
;
3391 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
3393 /* XXX: this should be generated when putting programs in place */
3395 for (int fs_attr
= 0; fs_attr
< VARYING_SLOT_MAX
; fs_attr
++) {
3396 const int input_index
= wm_prog_data
->urb_setup
[fs_attr
];
3397 if (input_index
< 0 || input_index
>= 16)
3400 struct GENX(SF_OUTPUT_ATTRIBUTE_DETAIL
) *attr
=
3401 &attr_overrides
[input_index
];
3402 int slot
= vue_map
->varying_to_slot
[fs_attr
];
3404 /* Viewport and Layer are stored in the VUE header. We need to override
3405 * them to zero if earlier stages didn't write them, as GL requires that
3406 * they read back as zero when not explicitly set.
3409 case VARYING_SLOT_VIEWPORT
:
3410 case VARYING_SLOT_LAYER
:
3411 attr
->ComponentOverrideX
= true;
3412 attr
->ComponentOverrideW
= true;
3413 attr
->ConstantSource
= CONST_0000
;
3415 if (!(vue_map
->slots_valid
& VARYING_BIT_LAYER
))
3416 attr
->ComponentOverrideY
= true;
3417 if (!(vue_map
->slots_valid
& VARYING_BIT_VIEWPORT
))
3418 attr
->ComponentOverrideZ
= true;
3421 case VARYING_SLOT_PRIMITIVE_ID
:
3422 /* Override if the previous shader stage didn't write gl_PrimitiveID. */
3424 attr
->ComponentOverrideX
= true;
3425 attr
->ComponentOverrideY
= true;
3426 attr
->ComponentOverrideZ
= true;
3427 attr
->ComponentOverrideW
= true;
3428 attr
->ConstantSource
= PRIM_ID
;
3436 if (sprite_coord_enables
& (1 << input_index
))
3439 /* If there was only a back color written but not front, use back
3440 * as the color instead of undefined.
3442 if (slot
== -1 && fs_attr
== VARYING_SLOT_COL0
)
3443 slot
= vue_map
->varying_to_slot
[VARYING_SLOT_BFC0
];
3444 if (slot
== -1 && fs_attr
== VARYING_SLOT_COL1
)
3445 slot
= vue_map
->varying_to_slot
[VARYING_SLOT_BFC1
];
3447 /* Not written by the previous stage - undefined. */
3449 attr
->ComponentOverrideX
= true;
3450 attr
->ComponentOverrideY
= true;
3451 attr
->ComponentOverrideZ
= true;
3452 attr
->ComponentOverrideW
= true;
3453 attr
->ConstantSource
= CONST_0001_FLOAT
;
3457 /* Compute the location of the attribute relative to the read offset,
3458 * which is counted in 256-bit increments (two 128-bit VUE slots).
3460 const int source_attr
= slot
- 2 * urb_read_offset
;
3461 assert(source_attr
>= 0 && source_attr
<= 32);
3462 attr
->SourceAttribute
= source_attr
;
3464 /* If we are doing two-sided color, and the VUE slot following this one
3465 * represents a back-facing color, then we need to instruct the SF unit
3466 * to do back-facing swizzling.
3468 if (cso_rast
->light_twoside
&&
3469 ((vue_map
->slot_to_varying
[slot
] == VARYING_SLOT_COL0
&&
3470 vue_map
->slot_to_varying
[slot
+1] == VARYING_SLOT_BFC0
) ||
3471 (vue_map
->slot_to_varying
[slot
] == VARYING_SLOT_COL1
&&
3472 vue_map
->slot_to_varying
[slot
+1] == VARYING_SLOT_BFC1
)))
3473 attr
->SwizzleSelect
= INPUTATTR_FACING
;
3476 iris_emit_cmd(batch
, GENX(3DSTATE_SBE_SWIZ
), sbes
) {
3477 for (int i
= 0; i
< 16; i
++)
3478 sbes
.Attribute
[i
] = attr_overrides
[i
];
3483 iris_calculate_point_sprite_overrides(const struct brw_wm_prog_data
*prog_data
,
3484 const struct iris_rasterizer_state
*cso
)
3486 unsigned overrides
= 0;
3488 if (prog_data
->urb_setup
[VARYING_SLOT_PNTC
] != -1)
3489 overrides
|= 1 << prog_data
->urb_setup
[VARYING_SLOT_PNTC
];
3491 for (int i
= 0; i
< 8; i
++) {
3492 if ((cso
->sprite_coord_enable
& (1 << i
)) &&
3493 prog_data
->urb_setup
[VARYING_SLOT_TEX0
+ i
] != -1)
3494 overrides
|= 1 << prog_data
->urb_setup
[VARYING_SLOT_TEX0
+ i
];
3501 iris_emit_sbe(struct iris_batch
*batch
, const struct iris_context
*ice
)
3503 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
3504 const struct brw_wm_prog_data
*wm_prog_data
= (void *)
3505 ice
->shaders
.prog
[MESA_SHADER_FRAGMENT
]->prog_data
;
3506 const struct shader_info
*fs_info
=
3507 iris_get_shader_info(ice
, MESA_SHADER_FRAGMENT
);
3509 unsigned urb_read_offset
, urb_read_length
;
3510 iris_compute_sbe_urb_read_interval(fs_info
->inputs_read
,
3511 ice
->shaders
.last_vue_map
,
3512 cso_rast
->light_twoside
,
3513 &urb_read_offset
, &urb_read_length
);
3515 unsigned sprite_coord_overrides
=
3516 iris_calculate_point_sprite_overrides(wm_prog_data
, cso_rast
);
3518 iris_emit_cmd(batch
, GENX(3DSTATE_SBE
), sbe
) {
3519 sbe
.AttributeSwizzleEnable
= true;
3520 sbe
.NumberofSFOutputAttributes
= wm_prog_data
->num_varying_inputs
;
3521 sbe
.PointSpriteTextureCoordinateOrigin
= cso_rast
->sprite_coord_mode
;
3522 sbe
.VertexURBEntryReadOffset
= urb_read_offset
;
3523 sbe
.VertexURBEntryReadLength
= urb_read_length
;
3524 sbe
.ForceVertexURBEntryReadOffset
= true;
3525 sbe
.ForceVertexURBEntryReadLength
= true;
3526 sbe
.ConstantInterpolationEnable
= wm_prog_data
->flat_inputs
;
3527 sbe
.PointSpriteTextureCoordinateEnable
= sprite_coord_overrides
;
3529 for (int i
= 0; i
< 32; i
++) {
3530 sbe
.AttributeActiveComponentFormat
[i
] = ACTIVE_COMPONENT_XYZW
;
3535 iris_emit_sbe_swiz(batch
, ice
, urb_read_offset
, sprite_coord_overrides
);
3538 /* ------------------------------------------------------------------- */
3541 * Populate VS program key fields based on the current state.
3544 iris_populate_vs_key(const struct iris_context
*ice
,
3545 const struct shader_info
*info
,
3546 gl_shader_stage last_stage
,
3547 struct brw_vs_prog_key
*key
)
3549 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
3551 if (info
->clip_distance_array_size
== 0 &&
3552 (info
->outputs_written
& (VARYING_BIT_POS
| VARYING_BIT_CLIP_VERTEX
)) &&
3553 last_stage
== MESA_SHADER_VERTEX
)
3554 key
->nr_userclip_plane_consts
= cso_rast
->num_clip_plane_consts
;
3558 * Populate TCS program key fields based on the current state.
3561 iris_populate_tcs_key(const struct iris_context
*ice
,
3562 struct brw_tcs_prog_key
*key
)
3567 * Populate TES program key fields based on the current state.
3570 iris_populate_tes_key(const struct iris_context
*ice
,
3571 const struct shader_info
*info
,
3572 gl_shader_stage last_stage
,
3573 struct brw_tes_prog_key
*key
)
3575 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
3577 if (info
->clip_distance_array_size
== 0 &&
3578 (info
->outputs_written
& (VARYING_BIT_POS
| VARYING_BIT_CLIP_VERTEX
)) &&
3579 last_stage
== MESA_SHADER_TESS_EVAL
)
3580 key
->nr_userclip_plane_consts
= cso_rast
->num_clip_plane_consts
;
3584 * Populate GS program key fields based on the current state.
3587 iris_populate_gs_key(const struct iris_context
*ice
,
3588 const struct shader_info
*info
,
3589 gl_shader_stage last_stage
,
3590 struct brw_gs_prog_key
*key
)
3592 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
3594 if (info
->clip_distance_array_size
== 0 &&
3595 (info
->outputs_written
& (VARYING_BIT_POS
| VARYING_BIT_CLIP_VERTEX
)) &&
3596 last_stage
== MESA_SHADER_GEOMETRY
)
3597 key
->nr_userclip_plane_consts
= cso_rast
->num_clip_plane_consts
;
3601 * Populate FS program key fields based on the current state.
3604 iris_populate_fs_key(const struct iris_context
*ice
,
3605 const struct shader_info
*info
,
3606 struct brw_wm_prog_key
*key
)
3608 struct iris_screen
*screen
= (void *) ice
->ctx
.screen
;
3609 const struct pipe_framebuffer_state
*fb
= &ice
->state
.framebuffer
;
3610 const struct iris_depth_stencil_alpha_state
*zsa
= ice
->state
.cso_zsa
;
3611 const struct iris_rasterizer_state
*rast
= ice
->state
.cso_rast
;
3612 const struct iris_blend_state
*blend
= ice
->state
.cso_blend
;
3614 key
->nr_color_regions
= fb
->nr_cbufs
;
3616 key
->clamp_fragment_color
= rast
->clamp_fragment_color
;
3618 key
->alpha_to_coverage
= blend
->alpha_to_coverage
;
3620 key
->alpha_test_replicate_alpha
= fb
->nr_cbufs
> 1 && zsa
->alpha
.enabled
;
3622 key
->flat_shade
= rast
->flatshade
&&
3623 (info
->inputs_read
& (VARYING_BIT_COL0
| VARYING_BIT_COL1
));
3625 key
->persample_interp
= rast
->force_persample_interp
;
3626 key
->multisample_fbo
= rast
->multisample
&& fb
->samples
> 1;
3628 key
->coherent_fb_fetch
= GEN_GEN
>= 9;
3630 key
->force_dual_color_blend
=
3631 screen
->driconf
.dual_color_blend_by_location
&&
3632 (blend
->blend_enables
& 1) && blend
->dual_color_blending
;
3634 /* TODO: Respect glHint for key->high_quality_derivatives */
3638 iris_populate_cs_key(const struct iris_context
*ice
,
3639 struct brw_cs_prog_key
*key
)
3644 KSP(const struct iris_compiled_shader
*shader
)
3646 struct iris_resource
*res
= (void *) shader
->assembly
.res
;
3647 return iris_bo_offset_from_base_address(res
->bo
) + shader
->assembly
.offset
;
3650 /* Gen11 workaround table #2056 WABTPPrefetchDisable suggests to disable
3651 * prefetching of binding tables in A0 and B0 steppings. XXX: Revisit
3652 * this WA on C0 stepping.
3654 * TODO: Fill out SamplerCount for prefetching?
3657 #define INIT_THREAD_DISPATCH_FIELDS(pkt, prefix, stage) \
3658 pkt.KernelStartPointer = KSP(shader); \
3659 pkt.BindingTableEntryCount = GEN_GEN == 11 ? 0 : \
3660 shader->bt.size_bytes / 4; \
3661 pkt.FloatingPointMode = prog_data->use_alt_mode; \
3663 pkt.DispatchGRFStartRegisterForURBData = \
3664 prog_data->dispatch_grf_start_reg; \
3665 pkt.prefix##URBEntryReadLength = vue_prog_data->urb_read_length; \
3666 pkt.prefix##URBEntryReadOffset = 0; \
3668 pkt.StatisticsEnable = true; \
3669 pkt.Enable = true; \
3671 if (prog_data->total_scratch) { \
3672 struct iris_bo *bo = \
3673 iris_get_scratch_space(ice, prog_data->total_scratch, stage); \
3674 uint32_t scratch_addr = bo->gtt_offset; \
3675 pkt.PerThreadScratchSpace = ffs(prog_data->total_scratch) - 11; \
3676 pkt.ScratchSpaceBasePointer = rw_bo(NULL, scratch_addr); \
3680 * Encode most of 3DSTATE_VS based on the compiled shader.
3683 iris_store_vs_state(struct iris_context
*ice
,
3684 const struct gen_device_info
*devinfo
,
3685 struct iris_compiled_shader
*shader
)
3687 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3688 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3690 iris_pack_command(GENX(3DSTATE_VS
), shader
->derived_data
, vs
) {
3691 INIT_THREAD_DISPATCH_FIELDS(vs
, Vertex
, MESA_SHADER_VERTEX
);
3692 vs
.MaximumNumberofThreads
= devinfo
->max_vs_threads
- 1;
3693 vs
.SIMD8DispatchEnable
= true;
3694 vs
.UserClipDistanceCullTestEnableBitmask
=
3695 vue_prog_data
->cull_distance_mask
;
3700 * Encode most of 3DSTATE_HS based on the compiled shader.
3703 iris_store_tcs_state(struct iris_context
*ice
,
3704 const struct gen_device_info
*devinfo
,
3705 struct iris_compiled_shader
*shader
)
3707 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3708 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3709 struct brw_tcs_prog_data
*tcs_prog_data
= (void *) prog_data
;
3711 iris_pack_command(GENX(3DSTATE_HS
), shader
->derived_data
, hs
) {
3712 INIT_THREAD_DISPATCH_FIELDS(hs
, Vertex
, MESA_SHADER_TESS_CTRL
);
3714 hs
.InstanceCount
= tcs_prog_data
->instances
- 1;
3715 hs
.MaximumNumberofThreads
= devinfo
->max_tcs_threads
- 1;
3716 hs
.IncludeVertexHandles
= true;
3719 hs
.DispatchMode
= vue_prog_data
->dispatch_mode
;
3720 hs
.IncludePrimitiveID
= tcs_prog_data
->include_primitive_id
;
3726 * Encode 3DSTATE_TE and most of 3DSTATE_DS based on the compiled shader.
3729 iris_store_tes_state(struct iris_context
*ice
,
3730 const struct gen_device_info
*devinfo
,
3731 struct iris_compiled_shader
*shader
)
3733 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3734 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3735 struct brw_tes_prog_data
*tes_prog_data
= (void *) prog_data
;
3737 uint32_t *te_state
= (void *) shader
->derived_data
;
3738 uint32_t *ds_state
= te_state
+ GENX(3DSTATE_TE_length
);
3740 iris_pack_command(GENX(3DSTATE_TE
), te_state
, te
) {
3741 te
.Partitioning
= tes_prog_data
->partitioning
;
3742 te
.OutputTopology
= tes_prog_data
->output_topology
;
3743 te
.TEDomain
= tes_prog_data
->domain
;
3745 te
.MaximumTessellationFactorOdd
= 63.0;
3746 te
.MaximumTessellationFactorNotOdd
= 64.0;
3749 iris_pack_command(GENX(3DSTATE_DS
), ds_state
, ds
) {
3750 INIT_THREAD_DISPATCH_FIELDS(ds
, Patch
, MESA_SHADER_TESS_EVAL
);
3752 ds
.DispatchMode
= DISPATCH_MODE_SIMD8_SINGLE_PATCH
;
3753 ds
.MaximumNumberofThreads
= devinfo
->max_tes_threads
- 1;
3754 ds
.ComputeWCoordinateEnable
=
3755 tes_prog_data
->domain
== BRW_TESS_DOMAIN_TRI
;
3757 ds
.UserClipDistanceCullTestEnableBitmask
=
3758 vue_prog_data
->cull_distance_mask
;
3764 * Encode most of 3DSTATE_GS based on the compiled shader.
3767 iris_store_gs_state(struct iris_context
*ice
,
3768 const struct gen_device_info
*devinfo
,
3769 struct iris_compiled_shader
*shader
)
3771 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3772 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3773 struct brw_gs_prog_data
*gs_prog_data
= (void *) prog_data
;
3775 iris_pack_command(GENX(3DSTATE_GS
), shader
->derived_data
, gs
) {
3776 INIT_THREAD_DISPATCH_FIELDS(gs
, Vertex
, MESA_SHADER_GEOMETRY
);
3778 gs
.OutputVertexSize
= gs_prog_data
->output_vertex_size_hwords
* 2 - 1;
3779 gs
.OutputTopology
= gs_prog_data
->output_topology
;
3780 gs
.ControlDataHeaderSize
=
3781 gs_prog_data
->control_data_header_size_hwords
;
3782 gs
.InstanceControl
= gs_prog_data
->invocations
- 1;
3783 gs
.DispatchMode
= DISPATCH_MODE_SIMD8
;
3784 gs
.IncludePrimitiveID
= gs_prog_data
->include_primitive_id
;
3785 gs
.ControlDataFormat
= gs_prog_data
->control_data_format
;
3786 gs
.ReorderMode
= TRAILING
;
3787 gs
.ExpectedVertexCount
= gs_prog_data
->vertices_in
;
3788 gs
.MaximumNumberofThreads
=
3789 GEN_GEN
== 8 ? (devinfo
->max_gs_threads
/ 2 - 1)
3790 : (devinfo
->max_gs_threads
- 1);
3792 if (gs_prog_data
->static_vertex_count
!= -1) {
3793 gs
.StaticOutput
= true;
3794 gs
.StaticOutputVertexCount
= gs_prog_data
->static_vertex_count
;
3796 gs
.IncludeVertexHandles
= vue_prog_data
->include_vue_handles
;
3798 gs
.UserClipDistanceCullTestEnableBitmask
=
3799 vue_prog_data
->cull_distance_mask
;
3801 const int urb_entry_write_offset
= 1;
3802 const uint32_t urb_entry_output_length
=
3803 DIV_ROUND_UP(vue_prog_data
->vue_map
.num_slots
, 2) -
3804 urb_entry_write_offset
;
3806 gs
.VertexURBEntryOutputReadOffset
= urb_entry_write_offset
;
3807 gs
.VertexURBEntryOutputLength
= MAX2(urb_entry_output_length
, 1);
3812 * Encode most of 3DSTATE_PS and 3DSTATE_PS_EXTRA based on the shader.
3815 iris_store_fs_state(struct iris_context
*ice
,
3816 const struct gen_device_info
*devinfo
,
3817 struct iris_compiled_shader
*shader
)
3819 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3820 struct brw_wm_prog_data
*wm_prog_data
= (void *) shader
->prog_data
;
3822 uint32_t *ps_state
= (void *) shader
->derived_data
;
3823 uint32_t *psx_state
= ps_state
+ GENX(3DSTATE_PS_length
);
3825 iris_pack_command(GENX(3DSTATE_PS
), ps_state
, ps
) {
3826 ps
.VectorMaskEnable
= true;
3827 // XXX: WABTPPrefetchDisable, see above, drop at C0
3828 ps
.BindingTableEntryCount
= GEN_GEN
== 11 ? 0 :
3829 shader
->bt
.size_bytes
/ 4;
3830 ps
.FloatingPointMode
= prog_data
->use_alt_mode
;
3831 ps
.MaximumNumberofThreadsPerPSD
= 64 - (GEN_GEN
== 8 ? 2 : 1);
3833 ps
.PushConstantEnable
= prog_data
->ubo_ranges
[0].length
> 0;
3835 /* From the documentation for this packet:
3836 * "If the PS kernel does not need the Position XY Offsets to
3837 * compute a Position Value, then this field should be programmed
3838 * to POSOFFSET_NONE."
3840 * "SW Recommendation: If the PS kernel needs the Position Offsets
3841 * to compute a Position XY value, this field should match Position
3842 * ZW Interpolation Mode to ensure a consistent position.xyzw
3845 * We only require XY sample offsets. So, this recommendation doesn't
3846 * look useful at the moment. We might need this in future.
3848 ps
.PositionXYOffsetSelect
=
3849 wm_prog_data
->uses_pos_offset
? POSOFFSET_SAMPLE
: POSOFFSET_NONE
;
3851 if (prog_data
->total_scratch
) {
3852 struct iris_bo
*bo
=
3853 iris_get_scratch_space(ice
, prog_data
->total_scratch
,
3854 MESA_SHADER_FRAGMENT
);
3855 uint32_t scratch_addr
= bo
->gtt_offset
;
3856 ps
.PerThreadScratchSpace
= ffs(prog_data
->total_scratch
) - 11;
3857 ps
.ScratchSpaceBasePointer
= rw_bo(NULL
, scratch_addr
);
3861 iris_pack_command(GENX(3DSTATE_PS_EXTRA
), psx_state
, psx
) {
3862 psx
.PixelShaderValid
= true;
3863 psx
.PixelShaderComputedDepthMode
= wm_prog_data
->computed_depth_mode
;
3864 psx
.PixelShaderKillsPixel
= wm_prog_data
->uses_kill
;
3865 psx
.AttributeEnable
= wm_prog_data
->num_varying_inputs
!= 0;
3866 psx
.PixelShaderUsesSourceDepth
= wm_prog_data
->uses_src_depth
;
3867 psx
.PixelShaderUsesSourceW
= wm_prog_data
->uses_src_w
;
3868 psx
.PixelShaderIsPerSample
= wm_prog_data
->persample_dispatch
;
3869 psx
.oMaskPresenttoRenderTarget
= wm_prog_data
->uses_omask
;
3872 psx
.PixelShaderPullsBary
= wm_prog_data
->pulls_bary
;
3873 psx
.PixelShaderComputesStencil
= wm_prog_data
->computed_stencil
;
3879 * Compute the size of the derived data (shader command packets).
3881 * This must match the data written by the iris_store_xs_state() functions.
3884 iris_store_cs_state(struct iris_context
*ice
,
3885 const struct gen_device_info
*devinfo
,
3886 struct iris_compiled_shader
*shader
)
3888 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3889 struct brw_cs_prog_data
*cs_prog_data
= (void *) shader
->prog_data
;
3890 void *map
= shader
->derived_data
;
3892 iris_pack_state(GENX(INTERFACE_DESCRIPTOR_DATA
), map
, desc
) {
3893 desc
.KernelStartPointer
= KSP(shader
);
3894 desc
.ConstantURBEntryReadLength
= cs_prog_data
->push
.per_thread
.regs
;
3895 desc
.NumberofThreadsinGPGPUThreadGroup
= cs_prog_data
->threads
;
3896 desc
.SharedLocalMemorySize
=
3897 encode_slm_size(GEN_GEN
, prog_data
->total_shared
);
3898 desc
.BarrierEnable
= cs_prog_data
->uses_barrier
;
3899 desc
.CrossThreadConstantDataReadLength
=
3900 cs_prog_data
->push
.cross_thread
.regs
;
3905 iris_derived_program_state_size(enum iris_program_cache_id cache_id
)
3907 assert(cache_id
<= IRIS_CACHE_BLORP
);
3909 static const unsigned dwords
[] = {
3910 [IRIS_CACHE_VS
] = GENX(3DSTATE_VS_length
),
3911 [IRIS_CACHE_TCS
] = GENX(3DSTATE_HS_length
),
3912 [IRIS_CACHE_TES
] = GENX(3DSTATE_TE_length
) + GENX(3DSTATE_DS_length
),
3913 [IRIS_CACHE_GS
] = GENX(3DSTATE_GS_length
),
3915 GENX(3DSTATE_PS_length
) + GENX(3DSTATE_PS_EXTRA_length
),
3916 [IRIS_CACHE_CS
] = GENX(INTERFACE_DESCRIPTOR_DATA_length
),
3917 [IRIS_CACHE_BLORP
] = 0,
3920 return sizeof(uint32_t) * dwords
[cache_id
];
3924 * Create any state packets corresponding to the given shader stage
3925 * (i.e. 3DSTATE_VS) and save them as "derived data" in the shader variant.
3926 * This means that we can look up a program in the in-memory cache and
3927 * get most of the state packet without having to reconstruct it.
3930 iris_store_derived_program_state(struct iris_context
*ice
,
3931 enum iris_program_cache_id cache_id
,
3932 struct iris_compiled_shader
*shader
)
3934 struct iris_screen
*screen
= (void *) ice
->ctx
.screen
;
3935 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
3939 iris_store_vs_state(ice
, devinfo
, shader
);
3941 case IRIS_CACHE_TCS
:
3942 iris_store_tcs_state(ice
, devinfo
, shader
);
3944 case IRIS_CACHE_TES
:
3945 iris_store_tes_state(ice
, devinfo
, shader
);
3948 iris_store_gs_state(ice
, devinfo
, shader
);
3951 iris_store_fs_state(ice
, devinfo
, shader
);
3954 iris_store_cs_state(ice
, devinfo
, shader
);
3955 case IRIS_CACHE_BLORP
:
3962 /* ------------------------------------------------------------------- */
3964 static const uint32_t push_constant_opcodes
[] = {
3965 [MESA_SHADER_VERTEX
] = 21,
3966 [MESA_SHADER_TESS_CTRL
] = 25, /* HS */
3967 [MESA_SHADER_TESS_EVAL
] = 26, /* DS */
3968 [MESA_SHADER_GEOMETRY
] = 22,
3969 [MESA_SHADER_FRAGMENT
] = 23,
3970 [MESA_SHADER_COMPUTE
] = 0,
3974 use_null_surface(struct iris_batch
*batch
, struct iris_context
*ice
)
3976 struct iris_bo
*state_bo
= iris_resource_bo(ice
->state
.unbound_tex
.res
);
3978 iris_use_pinned_bo(batch
, state_bo
, false);
3980 return ice
->state
.unbound_tex
.offset
;
3984 use_null_fb_surface(struct iris_batch
*batch
, struct iris_context
*ice
)
3986 /* If set_framebuffer_state() was never called, fall back to 1x1x1 */
3987 if (!ice
->state
.null_fb
.res
)
3988 return use_null_surface(batch
, ice
);
3990 struct iris_bo
*state_bo
= iris_resource_bo(ice
->state
.null_fb
.res
);
3992 iris_use_pinned_bo(batch
, state_bo
, false);
3994 return ice
->state
.null_fb
.offset
;
3998 surf_state_offset_for_aux(struct iris_resource
*res
,
4000 enum isl_aux_usage aux_usage
)
4002 return SURFACE_STATE_ALIGNMENT
*
4003 util_bitcount(aux_modes
& ((1 << aux_usage
) - 1));
4008 surf_state_update_clear_value(struct iris_batch
*batch
,
4009 struct iris_resource
*res
,
4010 struct iris_state_ref
*state
,
4012 enum isl_aux_usage aux_usage
)
4014 struct isl_device
*isl_dev
= &batch
->screen
->isl_dev
;
4015 struct iris_bo
*state_bo
= iris_resource_bo(state
->res
);
4016 uint64_t real_offset
= state
->offset
+ IRIS_MEMZONE_BINDER_START
;
4017 uint32_t offset_into_bo
= real_offset
- state_bo
->gtt_offset
;
4018 uint32_t clear_offset
= offset_into_bo
+
4019 isl_dev
->ss
.clear_value_offset
+
4020 surf_state_offset_for_aux(res
, aux_modes
, aux_usage
);
4021 uint32_t *color
= res
->aux
.clear_color
.u32
;
4023 assert(isl_dev
->ss
.clear_value_size
== 16);
4025 if (aux_usage
== ISL_AUX_USAGE_HIZ
) {
4026 iris_emit_pipe_control_write(batch
, "update fast clear value (Z)",
4027 PIPE_CONTROL_WRITE_IMMEDIATE
,
4028 state_bo
, clear_offset
, color
[0]);
4030 iris_emit_pipe_control_write(batch
, "update fast clear color (RG__)",
4031 PIPE_CONTROL_WRITE_IMMEDIATE
,
4032 state_bo
, clear_offset
,
4033 (uint64_t) color
[0] |
4034 (uint64_t) color
[1] << 32);
4035 iris_emit_pipe_control_write(batch
, "update fast clear color (__BA)",
4036 PIPE_CONTROL_WRITE_IMMEDIATE
,
4037 state_bo
, clear_offset
+ 8,
4038 (uint64_t) color
[2] |
4039 (uint64_t) color
[3] << 32);
4042 iris_emit_pipe_control_flush(batch
,
4043 "update fast clear: state cache invalidate",
4044 PIPE_CONTROL_FLUSH_ENABLE
|
4045 PIPE_CONTROL_STATE_CACHE_INVALIDATE
);
4050 update_clear_value(struct iris_context
*ice
,
4051 struct iris_batch
*batch
,
4052 struct iris_resource
*res
,
4053 struct iris_state_ref
*state
,
4054 unsigned all_aux_modes
,
4055 struct isl_view
*view
)
4057 UNUSED
struct isl_device
*isl_dev
= &batch
->screen
->isl_dev
;
4058 UNUSED
unsigned aux_modes
= all_aux_modes
;
4060 /* We only need to update the clear color in the surface state for gen8 and
4061 * gen9. Newer gens can read it directly from the clear color state buffer.
4064 /* Skip updating the ISL_AUX_USAGE_NONE surface state */
4065 aux_modes
&= ~(1 << ISL_AUX_USAGE_NONE
);
4068 enum isl_aux_usage aux_usage
= u_bit_scan(&aux_modes
);
4070 surf_state_update_clear_value(batch
, res
, state
, all_aux_modes
,
4074 pipe_resource_reference(&state
->res
, NULL
);
4076 void *map
= alloc_surface_states(ice
->state
.surface_uploader
,
4077 state
, all_aux_modes
);
4079 enum isl_aux_usage aux_usage
= u_bit_scan(&aux_modes
);
4080 fill_surface_state(isl_dev
, map
, res
, &res
->surf
, view
, aux_usage
, 0, 0);
4081 map
+= SURFACE_STATE_ALIGNMENT
;
4087 * Add a surface to the validation list, as well as the buffer containing
4088 * the corresponding SURFACE_STATE.
4090 * Returns the binding table entry (offset to SURFACE_STATE).
4093 use_surface(struct iris_context
*ice
,
4094 struct iris_batch
*batch
,
4095 struct pipe_surface
*p_surf
,
4097 enum isl_aux_usage aux_usage
,
4098 bool is_read_surface
)
4100 struct iris_surface
*surf
= (void *) p_surf
;
4101 struct iris_resource
*res
= (void *) p_surf
->texture
;
4102 uint32_t offset
= 0;
4104 iris_use_pinned_bo(batch
, iris_resource_bo(p_surf
->texture
), writeable
);
4105 if (GEN_GEN
== 8 && is_read_surface
) {
4106 iris_use_pinned_bo(batch
, iris_resource_bo(surf
->surface_state_read
.res
), false);
4108 iris_use_pinned_bo(batch
, iris_resource_bo(surf
->surface_state
.res
), false);
4112 iris_use_pinned_bo(batch
, res
->aux
.bo
, writeable
);
4113 if (res
->aux
.clear_color_bo
)
4114 iris_use_pinned_bo(batch
, res
->aux
.clear_color_bo
, false);
4116 if (memcmp(&res
->aux
.clear_color
, &surf
->clear_color
,
4117 sizeof(surf
->clear_color
)) != 0) {
4118 update_clear_value(ice
, batch
, res
, &surf
->surface_state
,
4119 res
->aux
.possible_usages
, &surf
->view
);
4121 update_clear_value(ice
, batch
, res
, &surf
->surface_state_read
,
4122 res
->aux
.possible_usages
, &surf
->read_view
);
4124 surf
->clear_color
= res
->aux
.clear_color
;
4128 offset
= (GEN_GEN
== 8 && is_read_surface
) ? surf
->surface_state_read
.offset
4129 : surf
->surface_state
.offset
;
4132 surf_state_offset_for_aux(res
, res
->aux
.possible_usages
, aux_usage
);
4136 use_sampler_view(struct iris_context
*ice
,
4137 struct iris_batch
*batch
,
4138 struct iris_sampler_view
*isv
)
4141 enum isl_aux_usage aux_usage
=
4142 iris_resource_texture_aux_usage(ice
, isv
->res
, isv
->view
.format
, 0);
4144 iris_use_pinned_bo(batch
, isv
->res
->bo
, false);
4145 iris_use_pinned_bo(batch
, iris_resource_bo(isv
->surface_state
.res
), false);
4147 if (isv
->res
->aux
.bo
) {
4148 iris_use_pinned_bo(batch
, isv
->res
->aux
.bo
, false);
4149 if (isv
->res
->aux
.clear_color_bo
)
4150 iris_use_pinned_bo(batch
, isv
->res
->aux
.clear_color_bo
, false);
4151 if (memcmp(&isv
->res
->aux
.clear_color
, &isv
->clear_color
,
4152 sizeof(isv
->clear_color
)) != 0) {
4153 update_clear_value(ice
, batch
, isv
->res
, &isv
->surface_state
,
4154 isv
->res
->aux
.sampler_usages
, &isv
->view
);
4155 isv
->clear_color
= isv
->res
->aux
.clear_color
;
4159 return isv
->surface_state
.offset
+
4160 surf_state_offset_for_aux(isv
->res
, isv
->res
->aux
.sampler_usages
,
4165 use_ubo_ssbo(struct iris_batch
*batch
,
4166 struct iris_context
*ice
,
4167 struct pipe_shader_buffer
*buf
,
4168 struct iris_state_ref
*surf_state
,
4172 return use_null_surface(batch
, ice
);
4174 iris_use_pinned_bo(batch
, iris_resource_bo(buf
->buffer
), writable
);
4175 iris_use_pinned_bo(batch
, iris_resource_bo(surf_state
->res
), false);
4177 return surf_state
->offset
;
4181 use_image(struct iris_batch
*batch
, struct iris_context
*ice
,
4182 struct iris_shader_state
*shs
, int i
)
4184 struct iris_image_view
*iv
= &shs
->image
[i
];
4185 struct iris_resource
*res
= (void *) iv
->base
.resource
;
4188 return use_null_surface(batch
, ice
);
4190 bool write
= iv
->base
.shader_access
& PIPE_IMAGE_ACCESS_WRITE
;
4192 iris_use_pinned_bo(batch
, res
->bo
, write
);
4193 iris_use_pinned_bo(batch
, iris_resource_bo(iv
->surface_state
.res
), false);
4196 iris_use_pinned_bo(batch
, res
->aux
.bo
, write
);
4198 return iv
->surface_state
.offset
;
4201 #define push_bt_entry(addr) \
4202 assert(addr >= binder_addr); \
4203 assert(s < shader->bt.size_bytes / sizeof(uint32_t)); \
4204 if (!pin_only) bt_map[s++] = (addr) - binder_addr;
4206 #define bt_assert(section) \
4207 if (!pin_only && shader->bt.used_mask[section] != 0) \
4208 assert(shader->bt.offsets[section] == s);
4211 * Populate the binding table for a given shader stage.
4213 * This fills out the table of pointers to surfaces required by the shader,
4214 * and also adds those buffers to the validation list so the kernel can make
4215 * resident before running our batch.
4218 iris_populate_binding_table(struct iris_context
*ice
,
4219 struct iris_batch
*batch
,
4220 gl_shader_stage stage
,
4223 const struct iris_binder
*binder
= &ice
->state
.binder
;
4224 struct iris_uncompiled_shader
*ish
= ice
->shaders
.uncompiled
[stage
];
4225 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
4229 struct iris_binding_table
*bt
= &shader
->bt
;
4230 UNUSED
struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
4231 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4232 uint32_t binder_addr
= binder
->bo
->gtt_offset
;
4234 uint32_t *bt_map
= binder
->map
+ binder
->bt_offset
[stage
];
4237 const struct shader_info
*info
= iris_get_shader_info(ice
, stage
);
4239 /* TCS passthrough doesn't need a binding table. */
4240 assert(stage
== MESA_SHADER_TESS_CTRL
);
4244 if (stage
== MESA_SHADER_COMPUTE
&&
4245 shader
->bt
.used_mask
[IRIS_SURFACE_GROUP_CS_WORK_GROUPS
]) {
4246 /* surface for gl_NumWorkGroups */
4247 struct iris_state_ref
*grid_data
= &ice
->state
.grid_size
;
4248 struct iris_state_ref
*grid_state
= &ice
->state
.grid_surf_state
;
4249 iris_use_pinned_bo(batch
, iris_resource_bo(grid_data
->res
), false);
4250 iris_use_pinned_bo(batch
, iris_resource_bo(grid_state
->res
), false);
4251 push_bt_entry(grid_state
->offset
);
4254 if (stage
== MESA_SHADER_FRAGMENT
) {
4255 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4256 /* Note that cso_fb->nr_cbufs == fs_key->nr_color_regions. */
4257 if (cso_fb
->nr_cbufs
) {
4258 for (unsigned i
= 0; i
< cso_fb
->nr_cbufs
; i
++) {
4260 if (cso_fb
->cbufs
[i
]) {
4261 addr
= use_surface(ice
, batch
, cso_fb
->cbufs
[i
], true,
4262 ice
->state
.draw_aux_usage
[i
], false);
4264 addr
= use_null_fb_surface(batch
, ice
);
4266 push_bt_entry(addr
);
4269 uint32_t addr
= use_null_fb_surface(batch
, ice
);
4270 push_bt_entry(addr
);
4274 #define foreach_surface_used(index, group) \
4276 for (int index = 0; index < bt->sizes[group]; index++) \
4277 if (iris_group_index_to_bti(bt, group, index) != \
4278 IRIS_SURFACE_NOT_USED)
4280 foreach_surface_used(i
, IRIS_SURFACE_GROUP_RENDER_TARGET_READ
) {
4281 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4283 if (cso_fb
->cbufs
[i
]) {
4284 addr
= use_surface(ice
, batch
, cso_fb
->cbufs
[i
],
4285 true, ice
->state
.draw_aux_usage
[i
], true);
4286 push_bt_entry(addr
);
4290 foreach_surface_used(i
, IRIS_SURFACE_GROUP_TEXTURE
) {
4291 struct iris_sampler_view
*view
= shs
->textures
[i
];
4292 uint32_t addr
= view
? use_sampler_view(ice
, batch
, view
)
4293 : use_null_surface(batch
, ice
);
4294 push_bt_entry(addr
);
4297 foreach_surface_used(i
, IRIS_SURFACE_GROUP_IMAGE
) {
4298 uint32_t addr
= use_image(batch
, ice
, shs
, i
);
4299 push_bt_entry(addr
);
4302 foreach_surface_used(i
, IRIS_SURFACE_GROUP_UBO
) {
4305 if (i
== bt
->sizes
[IRIS_SURFACE_GROUP_UBO
] - 1) {
4306 if (ish
->const_data
) {
4307 iris_use_pinned_bo(batch
, iris_resource_bo(ish
->const_data
), false);
4308 iris_use_pinned_bo(batch
, iris_resource_bo(ish
->const_data_state
.res
),
4310 addr
= ish
->const_data_state
.offset
;
4312 /* This can only happen with INTEL_DISABLE_COMPACT_BINDING_TABLE=1. */
4313 addr
= use_null_surface(batch
, ice
);
4316 addr
= use_ubo_ssbo(batch
, ice
, &shs
->constbuf
[i
],
4317 &shs
->constbuf_surf_state
[i
], false);
4320 push_bt_entry(addr
);
4323 foreach_surface_used(i
, IRIS_SURFACE_GROUP_SSBO
) {
4325 use_ubo_ssbo(batch
, ice
, &shs
->ssbo
[i
], &shs
->ssbo_surf_state
[i
],
4326 shs
->writable_ssbos
& (1u << i
));
4327 push_bt_entry(addr
);
4331 /* XXX: YUV surfaces not implemented yet */
4332 bt_assert(plane_start
[1], ...);
4333 bt_assert(plane_start
[2], ...);
4338 iris_use_optional_res(struct iris_batch
*batch
,
4339 struct pipe_resource
*res
,
4343 struct iris_bo
*bo
= iris_resource_bo(res
);
4344 iris_use_pinned_bo(batch
, bo
, writeable
);
4349 pin_depth_and_stencil_buffers(struct iris_batch
*batch
,
4350 struct pipe_surface
*zsbuf
,
4351 struct iris_depth_stencil_alpha_state
*cso_zsa
)
4356 struct iris_resource
*zres
, *sres
;
4357 iris_get_depth_stencil_resources(zsbuf
->texture
, &zres
, &sres
);
4360 iris_use_pinned_bo(batch
, zres
->bo
, cso_zsa
->depth_writes_enabled
);
4362 iris_use_pinned_bo(batch
, zres
->aux
.bo
,
4363 cso_zsa
->depth_writes_enabled
);
4368 iris_use_pinned_bo(batch
, sres
->bo
, cso_zsa
->stencil_writes_enabled
);
4372 /* ------------------------------------------------------------------- */
4375 * Pin any BOs which were installed by a previous batch, and restored
4376 * via the hardware logical context mechanism.
4378 * We don't need to re-emit all state every batch - the hardware context
4379 * mechanism will save and restore it for us. This includes pointers to
4380 * various BOs...which won't exist unless we ask the kernel to pin them
4381 * by adding them to the validation list.
4383 * We can skip buffers if we've re-emitted those packets, as we're
4384 * overwriting those stale pointers with new ones, and don't actually
4385 * refer to the old BOs.
4388 iris_restore_render_saved_bos(struct iris_context
*ice
,
4389 struct iris_batch
*batch
,
4390 const struct pipe_draw_info
*draw
)
4392 struct iris_genx_state
*genx
= ice
->state
.genx
;
4394 const uint64_t clean
= ~ice
->state
.dirty
;
4396 if (clean
& IRIS_DIRTY_CC_VIEWPORT
) {
4397 iris_use_optional_res(batch
, ice
->state
.last_res
.cc_vp
, false);
4400 if (clean
& IRIS_DIRTY_SF_CL_VIEWPORT
) {
4401 iris_use_optional_res(batch
, ice
->state
.last_res
.sf_cl_vp
, false);
4404 if (clean
& IRIS_DIRTY_BLEND_STATE
) {
4405 iris_use_optional_res(batch
, ice
->state
.last_res
.blend
, false);
4408 if (clean
& IRIS_DIRTY_COLOR_CALC_STATE
) {
4409 iris_use_optional_res(batch
, ice
->state
.last_res
.color_calc
, false);
4412 if (clean
& IRIS_DIRTY_SCISSOR_RECT
) {
4413 iris_use_optional_res(batch
, ice
->state
.last_res
.scissor
, false);
4416 if (ice
->state
.streamout_active
&& (clean
& IRIS_DIRTY_SO_BUFFERS
)) {
4417 for (int i
= 0; i
< 4; i
++) {
4418 struct iris_stream_output_target
*tgt
=
4419 (void *) ice
->state
.so_target
[i
];
4421 iris_use_pinned_bo(batch
, iris_resource_bo(tgt
->base
.buffer
),
4423 iris_use_pinned_bo(batch
, iris_resource_bo(tgt
->offset
.res
),
4429 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4430 if (!(clean
& (IRIS_DIRTY_CONSTANTS_VS
<< stage
)))
4433 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4434 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
4439 struct brw_stage_prog_data
*prog_data
= (void *) shader
->prog_data
;
4441 for (int i
= 0; i
< 4; i
++) {
4442 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[i
];
4444 if (range
->length
== 0)
4447 /* Range block is a binding table index, map back to UBO index. */
4448 unsigned block_index
= iris_bti_to_group_index(
4449 &shader
->bt
, IRIS_SURFACE_GROUP_UBO
, range
->block
);
4450 assert(block_index
!= IRIS_SURFACE_NOT_USED
);
4452 struct pipe_shader_buffer
*cbuf
= &shs
->constbuf
[block_index
];
4453 struct iris_resource
*res
= (void *) cbuf
->buffer
;
4456 iris_use_pinned_bo(batch
, res
->bo
, false);
4458 iris_use_pinned_bo(batch
, batch
->screen
->workaround_bo
, false);
4462 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4463 if (clean
& (IRIS_DIRTY_BINDINGS_VS
<< stage
)) {
4464 /* Re-pin any buffers referred to by the binding table. */
4465 iris_populate_binding_table(ice
, batch
, stage
, true);
4469 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4470 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4471 struct pipe_resource
*res
= shs
->sampler_table
.res
;
4473 iris_use_pinned_bo(batch
, iris_resource_bo(res
), false);
4476 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4477 if (clean
& (IRIS_DIRTY_VS
<< stage
)) {
4478 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
4481 struct iris_bo
*bo
= iris_resource_bo(shader
->assembly
.res
);
4482 iris_use_pinned_bo(batch
, bo
, false);
4484 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
4486 if (prog_data
->total_scratch
> 0) {
4487 struct iris_bo
*bo
=
4488 iris_get_scratch_space(ice
, prog_data
->total_scratch
, stage
);
4489 iris_use_pinned_bo(batch
, bo
, true);
4495 if ((clean
& IRIS_DIRTY_DEPTH_BUFFER
) &&
4496 (clean
& IRIS_DIRTY_WM_DEPTH_STENCIL
)) {
4497 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4498 pin_depth_and_stencil_buffers(batch
, cso_fb
->zsbuf
, ice
->state
.cso_zsa
);
4501 iris_use_optional_res(batch
, ice
->state
.last_res
.index_buffer
, false);
4503 if (clean
& IRIS_DIRTY_VERTEX_BUFFERS
) {
4504 uint64_t bound
= ice
->state
.bound_vertex_buffers
;
4506 const int i
= u_bit_scan64(&bound
);
4507 struct pipe_resource
*res
= genx
->vertex_buffers
[i
].resource
;
4508 iris_use_pinned_bo(batch
, iris_resource_bo(res
), false);
4514 iris_restore_compute_saved_bos(struct iris_context
*ice
,
4515 struct iris_batch
*batch
,
4516 const struct pipe_grid_info
*grid
)
4518 const uint64_t clean
= ~ice
->state
.dirty
;
4520 const int stage
= MESA_SHADER_COMPUTE
;
4521 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4523 if (clean
& IRIS_DIRTY_BINDINGS_CS
) {
4524 /* Re-pin any buffers referred to by the binding table. */
4525 iris_populate_binding_table(ice
, batch
, stage
, true);
4528 struct pipe_resource
*sampler_res
= shs
->sampler_table
.res
;
4530 iris_use_pinned_bo(batch
, iris_resource_bo(sampler_res
), false);
4532 if ((clean
& IRIS_DIRTY_SAMPLER_STATES_CS
) &&
4533 (clean
& IRIS_DIRTY_BINDINGS_CS
) &&
4534 (clean
& IRIS_DIRTY_CONSTANTS_CS
) &&
4535 (clean
& IRIS_DIRTY_CS
)) {
4536 iris_use_optional_res(batch
, ice
->state
.last_res
.cs_desc
, false);
4539 if (clean
& IRIS_DIRTY_CS
) {
4540 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
4543 struct iris_bo
*bo
= iris_resource_bo(shader
->assembly
.res
);
4544 iris_use_pinned_bo(batch
, bo
, false);
4546 struct iris_bo
*curbe_bo
=
4547 iris_resource_bo(ice
->state
.last_res
.cs_thread_ids
);
4548 iris_use_pinned_bo(batch
, curbe_bo
, false);
4550 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
4552 if (prog_data
->total_scratch
> 0) {
4553 struct iris_bo
*bo
=
4554 iris_get_scratch_space(ice
, prog_data
->total_scratch
, stage
);
4555 iris_use_pinned_bo(batch
, bo
, true);
4562 * Possibly emit STATE_BASE_ADDRESS to update Surface State Base Address.
4565 iris_update_surface_base_address(struct iris_batch
*batch
,
4566 struct iris_binder
*binder
)
4568 if (batch
->last_surface_base_address
== binder
->bo
->gtt_offset
)
4571 flush_for_state_base_change(batch
);
4573 iris_emit_cmd(batch
, GENX(STATE_BASE_ADDRESS
), sba
) {
4574 sba
.SurfaceStateBaseAddressModifyEnable
= true;
4575 sba
.SurfaceStateBaseAddress
= ro_bo(binder
->bo
, 0);
4577 /* The hardware appears to pay attention to the MOCS fields even
4578 * if you don't set the "Address Modify Enable" bit for the base.
4580 sba
.GeneralStateMOCS
= MOCS_WB
;
4581 sba
.StatelessDataPortAccessMOCS
= MOCS_WB
;
4582 sba
.DynamicStateMOCS
= MOCS_WB
;
4583 sba
.IndirectObjectMOCS
= MOCS_WB
;
4584 sba
.InstructionMOCS
= MOCS_WB
;
4585 sba
.SurfaceStateMOCS
= MOCS_WB
;
4587 sba
.BindlessSurfaceStateMOCS
= MOCS_WB
;
4591 batch
->last_surface_base_address
= binder
->bo
->gtt_offset
;
4595 iris_viewport_zmin_zmax(const struct pipe_viewport_state
*vp
, bool halfz
,
4596 bool window_space_position
, float *zmin
, float *zmax
)
4598 if (window_space_position
) {
4603 util_viewport_zmin_zmax(vp
, halfz
, zmin
, zmax
);
4607 iris_upload_dirty_render_state(struct iris_context
*ice
,
4608 struct iris_batch
*batch
,
4609 const struct pipe_draw_info
*draw
)
4611 const uint64_t dirty
= ice
->state
.dirty
;
4613 if (!(dirty
& IRIS_ALL_DIRTY_FOR_RENDER
))
4616 struct iris_genx_state
*genx
= ice
->state
.genx
;
4617 struct iris_binder
*binder
= &ice
->state
.binder
;
4618 struct brw_wm_prog_data
*wm_prog_data
= (void *)
4619 ice
->shaders
.prog
[MESA_SHADER_FRAGMENT
]->prog_data
;
4621 if (dirty
& IRIS_DIRTY_CC_VIEWPORT
) {
4622 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
4623 uint32_t cc_vp_address
;
4625 /* XXX: could avoid streaming for depth_clip [0,1] case. */
4626 uint32_t *cc_vp_map
=
4627 stream_state(batch
, ice
->state
.dynamic_uploader
,
4628 &ice
->state
.last_res
.cc_vp
,
4629 4 * ice
->state
.num_viewports
*
4630 GENX(CC_VIEWPORT_length
), 32, &cc_vp_address
);
4631 for (int i
= 0; i
< ice
->state
.num_viewports
; i
++) {
4633 iris_viewport_zmin_zmax(&ice
->state
.viewports
[i
], cso_rast
->clip_halfz
,
4634 ice
->state
.window_space_position
,
4636 if (cso_rast
->depth_clip_near
)
4638 if (cso_rast
->depth_clip_far
)
4641 iris_pack_state(GENX(CC_VIEWPORT
), cc_vp_map
, ccv
) {
4642 ccv
.MinimumDepth
= zmin
;
4643 ccv
.MaximumDepth
= zmax
;
4646 cc_vp_map
+= GENX(CC_VIEWPORT_length
);
4649 iris_emit_cmd(batch
, GENX(3DSTATE_VIEWPORT_STATE_POINTERS_CC
), ptr
) {
4650 ptr
.CCViewportPointer
= cc_vp_address
;
4654 if (dirty
& IRIS_DIRTY_SF_CL_VIEWPORT
) {
4655 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4656 uint32_t sf_cl_vp_address
;
4658 stream_state(batch
, ice
->state
.dynamic_uploader
,
4659 &ice
->state
.last_res
.sf_cl_vp
,
4660 4 * ice
->state
.num_viewports
*
4661 GENX(SF_CLIP_VIEWPORT_length
), 64, &sf_cl_vp_address
);
4663 for (unsigned i
= 0; i
< ice
->state
.num_viewports
; i
++) {
4664 const struct pipe_viewport_state
*state
= &ice
->state
.viewports
[i
];
4665 float gb_xmin
, gb_xmax
, gb_ymin
, gb_ymax
;
4667 float vp_xmin
= viewport_extent(state
, 0, -1.0f
);
4668 float vp_xmax
= viewport_extent(state
, 0, 1.0f
);
4669 float vp_ymin
= viewport_extent(state
, 1, -1.0f
);
4670 float vp_ymax
= viewport_extent(state
, 1, 1.0f
);
4672 gen_calculate_guardband_size(cso_fb
->width
, cso_fb
->height
,
4673 state
->scale
[0], state
->scale
[1],
4674 state
->translate
[0], state
->translate
[1],
4675 &gb_xmin
, &gb_xmax
, &gb_ymin
, &gb_ymax
);
4677 iris_pack_state(GENX(SF_CLIP_VIEWPORT
), vp_map
, vp
) {
4678 vp
.ViewportMatrixElementm00
= state
->scale
[0];
4679 vp
.ViewportMatrixElementm11
= state
->scale
[1];
4680 vp
.ViewportMatrixElementm22
= state
->scale
[2];
4681 vp
.ViewportMatrixElementm30
= state
->translate
[0];
4682 vp
.ViewportMatrixElementm31
= state
->translate
[1];
4683 vp
.ViewportMatrixElementm32
= state
->translate
[2];
4684 vp
.XMinClipGuardband
= gb_xmin
;
4685 vp
.XMaxClipGuardband
= gb_xmax
;
4686 vp
.YMinClipGuardband
= gb_ymin
;
4687 vp
.YMaxClipGuardband
= gb_ymax
;
4688 vp
.XMinViewPort
= MAX2(vp_xmin
, 0);
4689 vp
.XMaxViewPort
= MIN2(vp_xmax
, cso_fb
->width
) - 1;
4690 vp
.YMinViewPort
= MAX2(vp_ymin
, 0);
4691 vp
.YMaxViewPort
= MIN2(vp_ymax
, cso_fb
->height
) - 1;
4694 vp_map
+= GENX(SF_CLIP_VIEWPORT_length
);
4697 iris_emit_cmd(batch
, GENX(3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP
), ptr
) {
4698 ptr
.SFClipViewportPointer
= sf_cl_vp_address
;
4702 if (dirty
& IRIS_DIRTY_URB
) {
4705 for (int i
= MESA_SHADER_VERTEX
; i
<= MESA_SHADER_GEOMETRY
; i
++) {
4706 if (!ice
->shaders
.prog
[i
]) {
4709 struct brw_vue_prog_data
*vue_prog_data
=
4710 (void *) ice
->shaders
.prog
[i
]->prog_data
;
4711 size
[i
] = vue_prog_data
->urb_entry_size
;
4713 assert(size
[i
] != 0);
4716 genX(emit_urb_setup
)(ice
, batch
, size
,
4717 ice
->shaders
.prog
[MESA_SHADER_TESS_EVAL
] != NULL
,
4718 ice
->shaders
.prog
[MESA_SHADER_GEOMETRY
] != NULL
);
4721 if (dirty
& IRIS_DIRTY_BLEND_STATE
) {
4722 struct iris_blend_state
*cso_blend
= ice
->state
.cso_blend
;
4723 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4724 struct iris_depth_stencil_alpha_state
*cso_zsa
= ice
->state
.cso_zsa
;
4725 const int header_dwords
= GENX(BLEND_STATE_length
);
4727 /* Always write at least one BLEND_STATE - the final RT message will
4728 * reference BLEND_STATE[0] even if there aren't color writes. There
4729 * may still be alpha testing, computed depth, and so on.
4731 const int rt_dwords
=
4732 MAX2(cso_fb
->nr_cbufs
, 1) * GENX(BLEND_STATE_ENTRY_length
);
4734 uint32_t blend_offset
;
4735 uint32_t *blend_map
=
4736 stream_state(batch
, ice
->state
.dynamic_uploader
,
4737 &ice
->state
.last_res
.blend
,
4738 4 * (header_dwords
+ rt_dwords
), 64, &blend_offset
);
4740 uint32_t blend_state_header
;
4741 iris_pack_state(GENX(BLEND_STATE
), &blend_state_header
, bs
) {
4742 bs
.AlphaTestEnable
= cso_zsa
->alpha
.enabled
;
4743 bs
.AlphaTestFunction
= translate_compare_func(cso_zsa
->alpha
.func
);
4746 blend_map
[0] = blend_state_header
| cso_blend
->blend_state
[0];
4747 memcpy(&blend_map
[1], &cso_blend
->blend_state
[1], 4 * rt_dwords
);
4749 iris_emit_cmd(batch
, GENX(3DSTATE_BLEND_STATE_POINTERS
), ptr
) {
4750 ptr
.BlendStatePointer
= blend_offset
;
4751 ptr
.BlendStatePointerValid
= true;
4755 if (dirty
& IRIS_DIRTY_COLOR_CALC_STATE
) {
4756 struct iris_depth_stencil_alpha_state
*cso
= ice
->state
.cso_zsa
;
4758 struct pipe_stencil_ref
*p_stencil_refs
= &ice
->state
.stencil_ref
;
4762 stream_state(batch
, ice
->state
.dynamic_uploader
,
4763 &ice
->state
.last_res
.color_calc
,
4764 sizeof(uint32_t) * GENX(COLOR_CALC_STATE_length
),
4766 iris_pack_state(GENX(COLOR_CALC_STATE
), cc_map
, cc
) {
4767 cc
.AlphaTestFormat
= ALPHATEST_FLOAT32
;
4768 cc
.AlphaReferenceValueAsFLOAT32
= cso
->alpha
.ref_value
;
4769 cc
.BlendConstantColorRed
= ice
->state
.blend_color
.color
[0];
4770 cc
.BlendConstantColorGreen
= ice
->state
.blend_color
.color
[1];
4771 cc
.BlendConstantColorBlue
= ice
->state
.blend_color
.color
[2];
4772 cc
.BlendConstantColorAlpha
= ice
->state
.blend_color
.color
[3];
4774 cc
.StencilReferenceValue
= p_stencil_refs
->ref_value
[0];
4775 cc
.BackfaceStencilReferenceValue
= p_stencil_refs
->ref_value
[1];
4778 iris_emit_cmd(batch
, GENX(3DSTATE_CC_STATE_POINTERS
), ptr
) {
4779 ptr
.ColorCalcStatePointer
= cc_offset
;
4780 ptr
.ColorCalcStatePointerValid
= true;
4784 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4785 if (!(dirty
& (IRIS_DIRTY_CONSTANTS_VS
<< stage
)))
4788 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4789 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
4794 if (shs
->sysvals_need_upload
)
4795 upload_sysvals(ice
, stage
);
4797 struct brw_stage_prog_data
*prog_data
= (void *) shader
->prog_data
;
4799 iris_emit_cmd(batch
, GENX(3DSTATE_CONSTANT_VS
), pkt
) {
4800 pkt
._3DCommandSubOpcode
= push_constant_opcodes
[stage
];
4802 /* The Skylake PRM contains the following restriction:
4804 * "The driver must ensure The following case does not occur
4805 * without a flush to the 3D engine: 3DSTATE_CONSTANT_* with
4806 * buffer 3 read length equal to zero committed followed by a
4807 * 3DSTATE_CONSTANT_* with buffer 0 read length not equal to
4810 * To avoid this, we program the buffers in the highest slots.
4811 * This way, slot 0 is only used if slot 3 is also used.
4815 for (int i
= 3; i
>= 0; i
--) {
4816 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[i
];
4818 if (range
->length
== 0)
4821 /* Range block is a binding table index, map back to UBO index. */
4822 unsigned block_index
= iris_bti_to_group_index(
4823 &shader
->bt
, IRIS_SURFACE_GROUP_UBO
, range
->block
);
4824 assert(block_index
!= IRIS_SURFACE_NOT_USED
);
4826 struct pipe_shader_buffer
*cbuf
= &shs
->constbuf
[block_index
];
4827 struct iris_resource
*res
= (void *) cbuf
->buffer
;
4829 assert(cbuf
->buffer_offset
% 32 == 0);
4831 pkt
.ConstantBody
.ReadLength
[n
] = range
->length
;
4832 pkt
.ConstantBody
.Buffer
[n
] =
4833 res
? ro_bo(res
->bo
, range
->start
* 32 + cbuf
->buffer_offset
)
4834 : ro_bo(batch
->screen
->workaround_bo
, 0);
4841 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4842 if (dirty
& (IRIS_DIRTY_BINDINGS_VS
<< stage
)) {
4843 iris_emit_cmd(batch
, GENX(3DSTATE_BINDING_TABLE_POINTERS_VS
), ptr
) {
4844 ptr
._3DCommandSubOpcode
= 38 + stage
;
4845 ptr
.PointertoVSBindingTable
= binder
->bt_offset
[stage
];
4850 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4851 if (dirty
& (IRIS_DIRTY_BINDINGS_VS
<< stage
)) {
4852 iris_populate_binding_table(ice
, batch
, stage
, false);
4856 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4857 if (!(dirty
& (IRIS_DIRTY_SAMPLER_STATES_VS
<< stage
)) ||
4858 !ice
->shaders
.prog
[stage
])
4861 iris_upload_sampler_states(ice
, stage
);
4863 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4864 struct pipe_resource
*res
= shs
->sampler_table
.res
;
4866 iris_use_pinned_bo(batch
, iris_resource_bo(res
), false);
4868 iris_emit_cmd(batch
, GENX(3DSTATE_SAMPLER_STATE_POINTERS_VS
), ptr
) {
4869 ptr
._3DCommandSubOpcode
= 43 + stage
;
4870 ptr
.PointertoVSSamplerState
= shs
->sampler_table
.offset
;
4874 if (ice
->state
.need_border_colors
)
4875 iris_use_pinned_bo(batch
, ice
->state
.border_color_pool
.bo
, false);
4877 if (dirty
& IRIS_DIRTY_MULTISAMPLE
) {
4878 iris_emit_cmd(batch
, GENX(3DSTATE_MULTISAMPLE
), ms
) {
4880 ice
->state
.cso_rast
->half_pixel_center
? CENTER
: UL_CORNER
;
4881 if (ice
->state
.framebuffer
.samples
> 0)
4882 ms
.NumberofMultisamples
= ffs(ice
->state
.framebuffer
.samples
) - 1;
4886 if (dirty
& IRIS_DIRTY_SAMPLE_MASK
) {
4887 iris_emit_cmd(batch
, GENX(3DSTATE_SAMPLE_MASK
), ms
) {
4888 ms
.SampleMask
= ice
->state
.sample_mask
;
4892 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4893 if (!(dirty
& (IRIS_DIRTY_VS
<< stage
)))
4896 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
4899 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
4900 struct iris_resource
*cache
= (void *) shader
->assembly
.res
;
4901 iris_use_pinned_bo(batch
, cache
->bo
, false);
4903 if (prog_data
->total_scratch
> 0) {
4904 struct iris_bo
*bo
=
4905 iris_get_scratch_space(ice
, prog_data
->total_scratch
, stage
);
4906 iris_use_pinned_bo(batch
, bo
, true);
4909 if (stage
== MESA_SHADER_FRAGMENT
) {
4910 UNUSED
struct iris_rasterizer_state
*cso
= ice
->state
.cso_rast
;
4911 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4913 uint32_t ps_state
[GENX(3DSTATE_PS_length
)] = {0};
4914 iris_pack_command(GENX(3DSTATE_PS
), ps_state
, ps
) {
4915 ps
._8PixelDispatchEnable
= wm_prog_data
->dispatch_8
;
4916 ps
._16PixelDispatchEnable
= wm_prog_data
->dispatch_16
;
4917 ps
._32PixelDispatchEnable
= wm_prog_data
->dispatch_32
;
4919 /* The docs for 3DSTATE_PS::32 Pixel Dispatch Enable say:
4921 * "When NUM_MULTISAMPLES = 16 or FORCE_SAMPLE_COUNT = 16,
4922 * SIMD32 Dispatch must not be enabled for PER_PIXEL dispatch
4925 * 16x MSAA only exists on Gen9+, so we can skip this on Gen8.
4927 if (GEN_GEN
>= 9 && cso_fb
->samples
== 16 &&
4928 !wm_prog_data
->persample_dispatch
) {
4929 assert(ps
._8PixelDispatchEnable
|| ps
._16PixelDispatchEnable
);
4930 ps
._32PixelDispatchEnable
= false;
4933 ps
.DispatchGRFStartRegisterForConstantSetupData0
=
4934 brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data
, ps
, 0);
4935 ps
.DispatchGRFStartRegisterForConstantSetupData1
=
4936 brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data
, ps
, 1);
4937 ps
.DispatchGRFStartRegisterForConstantSetupData2
=
4938 brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data
, ps
, 2);
4940 ps
.KernelStartPointer0
= KSP(shader
) +
4941 brw_wm_prog_data_prog_offset(wm_prog_data
, ps
, 0);
4942 ps
.KernelStartPointer1
= KSP(shader
) +
4943 brw_wm_prog_data_prog_offset(wm_prog_data
, ps
, 1);
4944 ps
.KernelStartPointer2
= KSP(shader
) +
4945 brw_wm_prog_data_prog_offset(wm_prog_data
, ps
, 2);
4948 uint32_t psx_state
[GENX(3DSTATE_PS_EXTRA_length
)] = {0};
4949 iris_pack_command(GENX(3DSTATE_PS_EXTRA
), psx_state
, psx
) {
4951 if (!wm_prog_data
->uses_sample_mask
)
4952 psx
.InputCoverageMaskState
= ICMS_NONE
;
4953 else if (wm_prog_data
->post_depth_coverage
)
4954 psx
.InputCoverageMaskState
= ICMS_DEPTH_COVERAGE
;
4955 else if (wm_prog_data
->inner_coverage
&&
4956 cso
->conservative_rasterization
)
4957 psx
.InputCoverageMaskState
= ICMS_INNER_CONSERVATIVE
;
4959 psx
.InputCoverageMaskState
= ICMS_NORMAL
;
4961 psx
.PixelShaderUsesInputCoverageMask
=
4962 wm_prog_data
->uses_sample_mask
;
4966 uint32_t *shader_ps
= (uint32_t *) shader
->derived_data
;
4967 uint32_t *shader_psx
= shader_ps
+ GENX(3DSTATE_PS_length
);
4968 iris_emit_merge(batch
, shader_ps
, ps_state
,
4969 GENX(3DSTATE_PS_length
));
4970 iris_emit_merge(batch
, shader_psx
, psx_state
,
4971 GENX(3DSTATE_PS_EXTRA_length
));
4973 iris_batch_emit(batch
, shader
->derived_data
,
4974 iris_derived_program_state_size(stage
));
4977 if (stage
== MESA_SHADER_TESS_EVAL
) {
4978 iris_emit_cmd(batch
, GENX(3DSTATE_HS
), hs
);
4979 iris_emit_cmd(batch
, GENX(3DSTATE_TE
), te
);
4980 iris_emit_cmd(batch
, GENX(3DSTATE_DS
), ds
);
4981 } else if (stage
== MESA_SHADER_GEOMETRY
) {
4982 iris_emit_cmd(batch
, GENX(3DSTATE_GS
), gs
);
4987 if (ice
->state
.streamout_active
) {
4988 if (dirty
& IRIS_DIRTY_SO_BUFFERS
) {
4989 iris_batch_emit(batch
, genx
->so_buffers
,
4990 4 * 4 * GENX(3DSTATE_SO_BUFFER_length
));
4991 for (int i
= 0; i
< 4; i
++) {
4992 struct iris_stream_output_target
*tgt
=
4993 (void *) ice
->state
.so_target
[i
];
4996 iris_use_pinned_bo(batch
, iris_resource_bo(tgt
->base
.buffer
),
4998 iris_use_pinned_bo(batch
, iris_resource_bo(tgt
->offset
.res
),
5004 if ((dirty
& IRIS_DIRTY_SO_DECL_LIST
) && ice
->state
.streamout
) {
5005 uint32_t *decl_list
=
5006 ice
->state
.streamout
+ GENX(3DSTATE_STREAMOUT_length
);
5007 iris_batch_emit(batch
, decl_list
, 4 * ((decl_list
[0] & 0xff) + 2));
5010 if (dirty
& IRIS_DIRTY_STREAMOUT
) {
5011 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
5013 uint32_t dynamic_sol
[GENX(3DSTATE_STREAMOUT_length
)];
5014 iris_pack_command(GENX(3DSTATE_STREAMOUT
), dynamic_sol
, sol
) {
5015 sol
.SOFunctionEnable
= true;
5016 sol
.SOStatisticsEnable
= true;
5018 sol
.RenderingDisable
= cso_rast
->rasterizer_discard
&&
5019 !ice
->state
.prims_generated_query_active
;
5020 sol
.ReorderMode
= cso_rast
->flatshade_first
? LEADING
: TRAILING
;
5023 assert(ice
->state
.streamout
);
5025 iris_emit_merge(batch
, ice
->state
.streamout
, dynamic_sol
,
5026 GENX(3DSTATE_STREAMOUT_length
));
5029 if (dirty
& IRIS_DIRTY_STREAMOUT
) {
5030 iris_emit_cmd(batch
, GENX(3DSTATE_STREAMOUT
), sol
);
5034 if (dirty
& IRIS_DIRTY_CLIP
) {
5035 struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
5036 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
5038 bool gs_or_tes
= ice
->shaders
.prog
[MESA_SHADER_GEOMETRY
] ||
5039 ice
->shaders
.prog
[MESA_SHADER_TESS_EVAL
];
5040 bool points_or_lines
= cso_rast
->fill_mode_point_or_line
||
5041 (gs_or_tes
? ice
->shaders
.output_topology_is_points_or_lines
5042 : ice
->state
.prim_is_points_or_lines
);
5044 uint32_t dynamic_clip
[GENX(3DSTATE_CLIP_length
)];
5045 iris_pack_command(GENX(3DSTATE_CLIP
), &dynamic_clip
, cl
) {
5046 cl
.StatisticsEnable
= ice
->state
.statistics_counters_enabled
;
5047 if (cso_rast
->rasterizer_discard
)
5048 cl
.ClipMode
= CLIPMODE_REJECT_ALL
;
5049 else if (ice
->state
.window_space_position
)
5050 cl
.ClipMode
= CLIPMODE_ACCEPT_ALL
;
5052 cl
.ClipMode
= CLIPMODE_NORMAL
;
5054 cl
.PerspectiveDivideDisable
= ice
->state
.window_space_position
;
5055 cl
.ViewportXYClipTestEnable
= !points_or_lines
;
5057 if (wm_prog_data
->barycentric_interp_modes
&
5058 BRW_BARYCENTRIC_NONPERSPECTIVE_BITS
)
5059 cl
.NonPerspectiveBarycentricEnable
= true;
5061 cl
.ForceZeroRTAIndexEnable
= cso_fb
->layers
== 0;
5062 cl
.MaximumVPIndex
= ice
->state
.num_viewports
- 1;
5064 iris_emit_merge(batch
, cso_rast
->clip
, dynamic_clip
,
5065 ARRAY_SIZE(cso_rast
->clip
));
5068 if (dirty
& IRIS_DIRTY_RASTER
) {
5069 struct iris_rasterizer_state
*cso
= ice
->state
.cso_rast
;
5070 iris_batch_emit(batch
, cso
->raster
, sizeof(cso
->raster
));
5072 uint32_t dynamic_sf
[GENX(3DSTATE_SF_length
)];
5073 iris_pack_command(GENX(3DSTATE_SF
), &dynamic_sf
, sf
) {
5074 sf
.ViewportTransformEnable
= !ice
->state
.window_space_position
;
5076 iris_emit_merge(batch
, cso
->sf
, dynamic_sf
,
5077 ARRAY_SIZE(dynamic_sf
));
5080 if (dirty
& IRIS_DIRTY_WM
) {
5081 struct iris_rasterizer_state
*cso
= ice
->state
.cso_rast
;
5082 uint32_t dynamic_wm
[GENX(3DSTATE_WM_length
)];
5084 iris_pack_command(GENX(3DSTATE_WM
), &dynamic_wm
, wm
) {
5085 wm
.StatisticsEnable
= ice
->state
.statistics_counters_enabled
;
5087 wm
.BarycentricInterpolationMode
=
5088 wm_prog_data
->barycentric_interp_modes
;
5090 if (wm_prog_data
->early_fragment_tests
)
5091 wm
.EarlyDepthStencilControl
= EDSC_PREPS
;
5092 else if (wm_prog_data
->has_side_effects
)
5093 wm
.EarlyDepthStencilControl
= EDSC_PSEXEC
;
5095 /* We could skip this bit if color writes are enabled. */
5096 if (wm_prog_data
->has_side_effects
|| wm_prog_data
->uses_kill
)
5097 wm
.ForceThreadDispatchEnable
= ForceON
;
5099 iris_emit_merge(batch
, cso
->wm
, dynamic_wm
, ARRAY_SIZE(cso
->wm
));
5102 if (dirty
& IRIS_DIRTY_SBE
) {
5103 iris_emit_sbe(batch
, ice
);
5106 if (dirty
& IRIS_DIRTY_PS_BLEND
) {
5107 struct iris_blend_state
*cso_blend
= ice
->state
.cso_blend
;
5108 struct iris_depth_stencil_alpha_state
*cso_zsa
= ice
->state
.cso_zsa
;
5109 const struct shader_info
*fs_info
=
5110 iris_get_shader_info(ice
, MESA_SHADER_FRAGMENT
);
5112 uint32_t dynamic_pb
[GENX(3DSTATE_PS_BLEND_length
)];
5113 iris_pack_command(GENX(3DSTATE_PS_BLEND
), &dynamic_pb
, pb
) {
5114 pb
.HasWriteableRT
= has_writeable_rt(cso_blend
, fs_info
);
5115 pb
.AlphaTestEnable
= cso_zsa
->alpha
.enabled
;
5117 /* The dual source blending docs caution against using SRC1 factors
5118 * when the shader doesn't use a dual source render target write.
5119 * Empirically, this can lead to GPU hangs, and the results are
5120 * undefined anyway, so simply disable blending to avoid the hang.
5122 pb
.ColorBufferBlendEnable
= (cso_blend
->blend_enables
& 1) &&
5123 (!cso_blend
->dual_color_blending
|| wm_prog_data
->dual_src_blend
);
5126 iris_emit_merge(batch
, cso_blend
->ps_blend
, dynamic_pb
,
5127 ARRAY_SIZE(cso_blend
->ps_blend
));
5130 if (dirty
& IRIS_DIRTY_WM_DEPTH_STENCIL
) {
5131 struct iris_depth_stencil_alpha_state
*cso
= ice
->state
.cso_zsa
;
5133 struct pipe_stencil_ref
*p_stencil_refs
= &ice
->state
.stencil_ref
;
5134 uint32_t stencil_refs
[GENX(3DSTATE_WM_DEPTH_STENCIL_length
)];
5135 iris_pack_command(GENX(3DSTATE_WM_DEPTH_STENCIL
), &stencil_refs
, wmds
) {
5136 wmds
.StencilReferenceValue
= p_stencil_refs
->ref_value
[0];
5137 wmds
.BackfaceStencilReferenceValue
= p_stencil_refs
->ref_value
[1];
5139 iris_emit_merge(batch
, cso
->wmds
, stencil_refs
, ARRAY_SIZE(cso
->wmds
));
5141 iris_batch_emit(batch
, cso
->wmds
, sizeof(cso
->wmds
));
5145 if (dirty
& IRIS_DIRTY_SCISSOR_RECT
) {
5146 uint32_t scissor_offset
=
5147 emit_state(batch
, ice
->state
.dynamic_uploader
,
5148 &ice
->state
.last_res
.scissor
,
5149 ice
->state
.scissors
,
5150 sizeof(struct pipe_scissor_state
) *
5151 ice
->state
.num_viewports
, 32);
5153 iris_emit_cmd(batch
, GENX(3DSTATE_SCISSOR_STATE_POINTERS
), ptr
) {
5154 ptr
.ScissorRectPointer
= scissor_offset
;
5158 if (dirty
& IRIS_DIRTY_DEPTH_BUFFER
) {
5159 struct iris_depth_buffer_state
*cso_z
= &ice
->state
.genx
->depth_buffer
;
5161 /* Do not emit the clear params yets. We need to update the clear value
5164 uint32_t clear_length
= GENX(3DSTATE_CLEAR_PARAMS_length
) * 4;
5165 uint32_t cso_z_size
= sizeof(cso_z
->packets
) - clear_length
;
5166 iris_batch_emit(batch
, cso_z
->packets
, cso_z_size
);
5168 union isl_color_value clear_value
= { .f32
= { 0, } };
5170 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
5171 if (cso_fb
->zsbuf
) {
5172 struct iris_resource
*zres
, *sres
;
5173 iris_get_depth_stencil_resources(cso_fb
->zsbuf
->texture
,
5175 if (zres
&& zres
->aux
.bo
)
5176 clear_value
= iris_resource_get_clear_color(zres
, NULL
, NULL
);
5179 uint32_t clear_params
[GENX(3DSTATE_CLEAR_PARAMS_length
)];
5180 iris_pack_command(GENX(3DSTATE_CLEAR_PARAMS
), clear_params
, clear
) {
5181 clear
.DepthClearValueValid
= true;
5182 clear
.DepthClearValue
= clear_value
.f32
[0];
5184 iris_batch_emit(batch
, clear_params
, clear_length
);
5187 if (dirty
& (IRIS_DIRTY_DEPTH_BUFFER
| IRIS_DIRTY_WM_DEPTH_STENCIL
)) {
5188 /* Listen for buffer changes, and also write enable changes. */
5189 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
5190 pin_depth_and_stencil_buffers(batch
, cso_fb
->zsbuf
, ice
->state
.cso_zsa
);
5193 if (dirty
& IRIS_DIRTY_POLYGON_STIPPLE
) {
5194 iris_emit_cmd(batch
, GENX(3DSTATE_POLY_STIPPLE_PATTERN
), poly
) {
5195 for (int i
= 0; i
< 32; i
++) {
5196 poly
.PatternRow
[i
] = ice
->state
.poly_stipple
.stipple
[i
];
5201 if (dirty
& IRIS_DIRTY_LINE_STIPPLE
) {
5202 struct iris_rasterizer_state
*cso
= ice
->state
.cso_rast
;
5203 iris_batch_emit(batch
, cso
->line_stipple
, sizeof(cso
->line_stipple
));
5206 if (dirty
& IRIS_DIRTY_VF_TOPOLOGY
) {
5207 iris_emit_cmd(batch
, GENX(3DSTATE_VF_TOPOLOGY
), topo
) {
5208 topo
.PrimitiveTopologyType
=
5209 translate_prim_type(draw
->mode
, draw
->vertices_per_patch
);
5213 if (dirty
& IRIS_DIRTY_VERTEX_BUFFERS
) {
5214 int count
= util_bitcount64(ice
->state
.bound_vertex_buffers
);
5215 int dynamic_bound
= ice
->state
.bound_vertex_buffers
;
5217 if (ice
->state
.vs_uses_draw_params
) {
5218 if (ice
->draw
.draw_params_offset
== 0) {
5219 u_upload_data(ice
->ctx
.stream_uploader
, 0, sizeof(ice
->draw
.params
),
5220 4, &ice
->draw
.params
, &ice
->draw
.draw_params_offset
,
5221 &ice
->draw
.draw_params_res
);
5223 assert(ice
->draw
.draw_params_res
);
5225 struct iris_vertex_buffer_state
*state
=
5226 &(ice
->state
.genx
->vertex_buffers
[count
]);
5227 pipe_resource_reference(&state
->resource
, ice
->draw
.draw_params_res
);
5228 struct iris_resource
*res
= (void *) state
->resource
;
5230 iris_pack_state(GENX(VERTEX_BUFFER_STATE
), state
->state
, vb
) {
5231 vb
.VertexBufferIndex
= count
;
5232 vb
.AddressModifyEnable
= true;
5234 vb
.BufferSize
= res
->bo
->size
- ice
->draw
.draw_params_offset
;
5235 vb
.BufferStartingAddress
=
5236 ro_bo(NULL
, res
->bo
->gtt_offset
+
5237 (int) ice
->draw
.draw_params_offset
);
5238 vb
.MOCS
= mocs(res
->bo
);
5240 dynamic_bound
|= 1ull << count
;
5244 if (ice
->state
.vs_uses_derived_draw_params
) {
5245 u_upload_data(ice
->ctx
.stream_uploader
, 0,
5246 sizeof(ice
->draw
.derived_params
), 4,
5247 &ice
->draw
.derived_params
,
5248 &ice
->draw
.derived_draw_params_offset
,
5249 &ice
->draw
.derived_draw_params_res
);
5251 struct iris_vertex_buffer_state
*state
=
5252 &(ice
->state
.genx
->vertex_buffers
[count
]);
5253 pipe_resource_reference(&state
->resource
,
5254 ice
->draw
.derived_draw_params_res
);
5255 struct iris_resource
*res
= (void *) ice
->draw
.derived_draw_params_res
;
5257 iris_pack_state(GENX(VERTEX_BUFFER_STATE
), state
->state
, vb
) {
5258 vb
.VertexBufferIndex
= count
;
5259 vb
.AddressModifyEnable
= true;
5262 res
->bo
->size
- ice
->draw
.derived_draw_params_offset
;
5263 vb
.BufferStartingAddress
=
5264 ro_bo(NULL
, res
->bo
->gtt_offset
+
5265 (int) ice
->draw
.derived_draw_params_offset
);
5266 vb
.MOCS
= mocs(res
->bo
);
5268 dynamic_bound
|= 1ull << count
;
5273 /* The VF cache designers cut corners, and made the cache key's
5274 * <VertexBufferIndex, Memory Address> tuple only consider the bottom
5275 * 32 bits of the address. If you have two vertex buffers which get
5276 * placed exactly 4 GiB apart and use them in back-to-back draw calls,
5277 * you can get collisions (even within a single batch).
5279 * So, we need to do a VF cache invalidate if the buffer for a VB
5280 * slot slot changes [48:32] address bits from the previous time.
5282 unsigned flush_flags
= 0;
5284 uint64_t bound
= dynamic_bound
;
5286 const int i
= u_bit_scan64(&bound
);
5287 uint16_t high_bits
= 0;
5289 struct iris_resource
*res
=
5290 (void *) genx
->vertex_buffers
[i
].resource
;
5292 iris_use_pinned_bo(batch
, res
->bo
, false);
5294 high_bits
= res
->bo
->gtt_offset
>> 32ull;
5295 if (high_bits
!= ice
->state
.last_vbo_high_bits
[i
]) {
5296 flush_flags
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
|
5297 PIPE_CONTROL_CS_STALL
;
5298 ice
->state
.last_vbo_high_bits
[i
] = high_bits
;
5304 iris_emit_pipe_control_flush(batch
,
5305 "workaround: VF cache 32-bit key [VB]",
5309 const unsigned vb_dwords
= GENX(VERTEX_BUFFER_STATE_length
);
5312 iris_get_command_space(batch
, 4 * (1 + vb_dwords
* count
));
5313 _iris_pack_command(batch
, GENX(3DSTATE_VERTEX_BUFFERS
), map
, vb
) {
5314 vb
.DWordLength
= (vb_dwords
* count
+ 1) - 2;
5318 bound
= dynamic_bound
;
5320 const int i
= u_bit_scan64(&bound
);
5321 memcpy(map
, genx
->vertex_buffers
[i
].state
,
5322 sizeof(uint32_t) * vb_dwords
);
5328 if (dirty
& IRIS_DIRTY_VERTEX_ELEMENTS
) {
5329 struct iris_vertex_element_state
*cso
= ice
->state
.cso_vertex_elements
;
5330 const unsigned entries
= MAX2(cso
->count
, 1);
5331 if (!(ice
->state
.vs_needs_sgvs_element
||
5332 ice
->state
.vs_uses_derived_draw_params
||
5333 ice
->state
.vs_needs_edge_flag
)) {
5334 iris_batch_emit(batch
, cso
->vertex_elements
, sizeof(uint32_t) *
5335 (1 + entries
* GENX(VERTEX_ELEMENT_STATE_length
)));
5337 uint32_t dynamic_ves
[1 + 33 * GENX(VERTEX_ELEMENT_STATE_length
)];
5338 const unsigned dyn_count
= cso
->count
+
5339 ice
->state
.vs_needs_sgvs_element
+
5340 ice
->state
.vs_uses_derived_draw_params
;
5342 iris_pack_command(GENX(3DSTATE_VERTEX_ELEMENTS
),
5345 1 + GENX(VERTEX_ELEMENT_STATE_length
) * dyn_count
- 2;
5347 memcpy(&dynamic_ves
[1], &cso
->vertex_elements
[1],
5348 (cso
->count
- ice
->state
.vs_needs_edge_flag
) *
5349 GENX(VERTEX_ELEMENT_STATE_length
) * sizeof(uint32_t));
5350 uint32_t *ve_pack_dest
=
5351 &dynamic_ves
[1 + (cso
->count
- ice
->state
.vs_needs_edge_flag
) *
5352 GENX(VERTEX_ELEMENT_STATE_length
)];
5354 if (ice
->state
.vs_needs_sgvs_element
) {
5355 uint32_t base_ctrl
= ice
->state
.vs_uses_draw_params
?
5356 VFCOMP_STORE_SRC
: VFCOMP_STORE_0
;
5357 iris_pack_state(GENX(VERTEX_ELEMENT_STATE
), ve_pack_dest
, ve
) {
5359 ve
.VertexBufferIndex
=
5360 util_bitcount64(ice
->state
.bound_vertex_buffers
);
5361 ve
.SourceElementFormat
= ISL_FORMAT_R32G32_UINT
;
5362 ve
.Component0Control
= base_ctrl
;
5363 ve
.Component1Control
= base_ctrl
;
5364 ve
.Component2Control
= VFCOMP_STORE_0
;
5365 ve
.Component3Control
= VFCOMP_STORE_0
;
5367 ve_pack_dest
+= GENX(VERTEX_ELEMENT_STATE_length
);
5369 if (ice
->state
.vs_uses_derived_draw_params
) {
5370 iris_pack_state(GENX(VERTEX_ELEMENT_STATE
), ve_pack_dest
, ve
) {
5372 ve
.VertexBufferIndex
=
5373 util_bitcount64(ice
->state
.bound_vertex_buffers
) +
5374 ice
->state
.vs_uses_draw_params
;
5375 ve
.SourceElementFormat
= ISL_FORMAT_R32G32_UINT
;
5376 ve
.Component0Control
= VFCOMP_STORE_SRC
;
5377 ve
.Component1Control
= VFCOMP_STORE_SRC
;
5378 ve
.Component2Control
= VFCOMP_STORE_0
;
5379 ve
.Component3Control
= VFCOMP_STORE_0
;
5381 ve_pack_dest
+= GENX(VERTEX_ELEMENT_STATE_length
);
5383 if (ice
->state
.vs_needs_edge_flag
) {
5384 for (int i
= 0; i
< GENX(VERTEX_ELEMENT_STATE_length
); i
++)
5385 ve_pack_dest
[i
] = cso
->edgeflag_ve
[i
];
5388 iris_batch_emit(batch
, &dynamic_ves
, sizeof(uint32_t) *
5389 (1 + dyn_count
* GENX(VERTEX_ELEMENT_STATE_length
)));
5392 if (!ice
->state
.vs_needs_edge_flag
) {
5393 iris_batch_emit(batch
, cso
->vf_instancing
, sizeof(uint32_t) *
5394 entries
* GENX(3DSTATE_VF_INSTANCING_length
));
5396 assert(cso
->count
> 0);
5397 const unsigned edgeflag_index
= cso
->count
- 1;
5398 uint32_t dynamic_vfi
[33 * GENX(3DSTATE_VF_INSTANCING_length
)];
5399 memcpy(&dynamic_vfi
[0], cso
->vf_instancing
, edgeflag_index
*
5400 GENX(3DSTATE_VF_INSTANCING_length
) * sizeof(uint32_t));
5402 uint32_t *vfi_pack_dest
= &dynamic_vfi
[0] +
5403 edgeflag_index
* GENX(3DSTATE_VF_INSTANCING_length
);
5404 iris_pack_command(GENX(3DSTATE_VF_INSTANCING
), vfi_pack_dest
, vi
) {
5405 vi
.VertexElementIndex
= edgeflag_index
+
5406 ice
->state
.vs_needs_sgvs_element
+
5407 ice
->state
.vs_uses_derived_draw_params
;
5409 for (int i
= 0; i
< GENX(3DSTATE_VF_INSTANCING_length
); i
++)
5410 vfi_pack_dest
[i
] |= cso
->edgeflag_vfi
[i
];
5412 iris_batch_emit(batch
, &dynamic_vfi
[0], sizeof(uint32_t) *
5413 entries
* GENX(3DSTATE_VF_INSTANCING_length
));
5417 if (dirty
& IRIS_DIRTY_VF_SGVS
) {
5418 const struct brw_vs_prog_data
*vs_prog_data
= (void *)
5419 ice
->shaders
.prog
[MESA_SHADER_VERTEX
]->prog_data
;
5420 struct iris_vertex_element_state
*cso
= ice
->state
.cso_vertex_elements
;
5422 iris_emit_cmd(batch
, GENX(3DSTATE_VF_SGVS
), sgv
) {
5423 if (vs_prog_data
->uses_vertexid
) {
5424 sgv
.VertexIDEnable
= true;
5425 sgv
.VertexIDComponentNumber
= 2;
5426 sgv
.VertexIDElementOffset
=
5427 cso
->count
- ice
->state
.vs_needs_edge_flag
;
5430 if (vs_prog_data
->uses_instanceid
) {
5431 sgv
.InstanceIDEnable
= true;
5432 sgv
.InstanceIDComponentNumber
= 3;
5433 sgv
.InstanceIDElementOffset
=
5434 cso
->count
- ice
->state
.vs_needs_edge_flag
;
5439 if (dirty
& IRIS_DIRTY_VF
) {
5440 iris_emit_cmd(batch
, GENX(3DSTATE_VF
), vf
) {
5441 if (draw
->primitive_restart
) {
5442 vf
.IndexedDrawCutIndexEnable
= true;
5443 vf
.CutIndex
= draw
->restart_index
;
5448 if (dirty
& IRIS_DIRTY_VF_STATISTICS
) {
5449 iris_emit_cmd(batch
, GENX(3DSTATE_VF_STATISTICS
), vf
) {
5450 vf
.StatisticsEnable
= true;
5454 if (ice
->state
.current_hash_scale
!= 1)
5455 genX(emit_hashing_mode
)(ice
, batch
, UINT_MAX
, UINT_MAX
, 1);
5457 /* TODO: Gen8 PMA fix */
5461 iris_upload_render_state(struct iris_context
*ice
,
5462 struct iris_batch
*batch
,
5463 const struct pipe_draw_info
*draw
)
5465 bool use_predicate
= ice
->state
.predicate
== IRIS_PREDICATE_STATE_USE_BIT
;
5467 /* Always pin the binder. If we're emitting new binding table pointers,
5468 * we need it. If not, we're probably inheriting old tables via the
5469 * context, and need it anyway. Since true zero-bindings cases are
5470 * practically non-existent, just pin it and avoid last_res tracking.
5472 iris_use_pinned_bo(batch
, ice
->state
.binder
.bo
, false);
5474 if (!batch
->contains_draw
) {
5475 iris_restore_render_saved_bos(ice
, batch
, draw
);
5476 batch
->contains_draw
= true;
5479 iris_upload_dirty_render_state(ice
, batch
, draw
);
5481 if (draw
->index_size
> 0) {
5484 if (draw
->has_user_indices
) {
5485 u_upload_data(ice
->ctx
.stream_uploader
, 0,
5486 draw
->count
* draw
->index_size
, 4, draw
->index
.user
,
5487 &offset
, &ice
->state
.last_res
.index_buffer
);
5489 struct iris_resource
*res
= (void *) draw
->index
.resource
;
5490 res
->bind_history
|= PIPE_BIND_INDEX_BUFFER
;
5492 pipe_resource_reference(&ice
->state
.last_res
.index_buffer
,
5493 draw
->index
.resource
);
5497 struct iris_genx_state
*genx
= ice
->state
.genx
;
5498 struct iris_bo
*bo
= iris_resource_bo(ice
->state
.last_res
.index_buffer
);
5500 uint32_t ib_packet
[GENX(3DSTATE_INDEX_BUFFER_length
)];
5501 iris_pack_command(GENX(3DSTATE_INDEX_BUFFER
), ib_packet
, ib
) {
5502 ib
.IndexFormat
= draw
->index_size
>> 1;
5504 ib
.BufferSize
= bo
->size
- offset
;
5505 ib
.BufferStartingAddress
= ro_bo(NULL
, bo
->gtt_offset
+ offset
);
5508 if (memcmp(genx
->last_index_buffer
, ib_packet
, sizeof(ib_packet
)) != 0) {
5509 memcpy(genx
->last_index_buffer
, ib_packet
, sizeof(ib_packet
));
5510 iris_batch_emit(batch
, ib_packet
, sizeof(ib_packet
));
5511 iris_use_pinned_bo(batch
, bo
, false);
5514 /* The VF cache key only uses 32-bits, see vertex buffer comment above */
5515 uint16_t high_bits
= bo
->gtt_offset
>> 32ull;
5516 if (high_bits
!= ice
->state
.last_index_bo_high_bits
) {
5517 iris_emit_pipe_control_flush(batch
,
5518 "workaround: VF cache 32-bit key [IB]",
5519 PIPE_CONTROL_VF_CACHE_INVALIDATE
|
5520 PIPE_CONTROL_CS_STALL
);
5521 ice
->state
.last_index_bo_high_bits
= high_bits
;
5525 #define _3DPRIM_END_OFFSET 0x2420
5526 #define _3DPRIM_START_VERTEX 0x2430
5527 #define _3DPRIM_VERTEX_COUNT 0x2434
5528 #define _3DPRIM_INSTANCE_COUNT 0x2438
5529 #define _3DPRIM_START_INSTANCE 0x243C
5530 #define _3DPRIM_BASE_VERTEX 0x2440
5532 if (draw
->indirect
) {
5533 if (draw
->indirect
->indirect_draw_count
) {
5534 use_predicate
= true;
5536 struct iris_bo
*draw_count_bo
=
5537 iris_resource_bo(draw
->indirect
->indirect_draw_count
);
5538 unsigned draw_count_offset
=
5539 draw
->indirect
->indirect_draw_count_offset
;
5541 iris_emit_pipe_control_flush(batch
,
5542 "ensure indirect draw buffer is flushed",
5543 PIPE_CONTROL_FLUSH_ENABLE
);
5545 if (ice
->state
.predicate
== IRIS_PREDICATE_STATE_USE_BIT
) {
5546 struct gen_mi_builder b
;
5547 gen_mi_builder_init(&b
, batch
);
5549 /* comparison = draw id < draw count */
5550 struct gen_mi_value comparison
=
5551 gen_mi_ult(&b
, gen_mi_imm(draw
->drawid
),
5552 gen_mi_mem32(ro_bo(draw_count_bo
,
5553 draw_count_offset
)));
5555 /* predicate = comparison & conditional rendering predicate */
5556 gen_mi_store(&b
, gen_mi_reg32(MI_PREDICATE_RESULT
),
5557 gen_mi_iand(&b
, comparison
,
5558 gen_mi_reg32(CS_GPR(15))));
5560 uint32_t mi_predicate
;
5562 /* Upload the id of the current primitive to MI_PREDICATE_SRC1. */
5563 ice
->vtbl
.load_register_imm64(batch
, MI_PREDICATE_SRC1
,
5565 /* Upload the current draw count from the draw parameters buffer
5566 * to MI_PREDICATE_SRC0.
5568 ice
->vtbl
.load_register_mem32(batch
, MI_PREDICATE_SRC0
,
5569 draw_count_bo
, draw_count_offset
);
5570 /* Zero the top 32-bits of MI_PREDICATE_SRC0 */
5571 ice
->vtbl
.load_register_imm32(batch
, MI_PREDICATE_SRC0
+ 4, 0);
5573 if (draw
->drawid
== 0) {
5574 mi_predicate
= MI_PREDICATE
| MI_PREDICATE_LOADOP_LOADINV
|
5575 MI_PREDICATE_COMBINEOP_SET
|
5576 MI_PREDICATE_COMPAREOP_SRCS_EQUAL
;
5578 /* While draw_index < draw_count the predicate's result will be
5579 * (draw_index == draw_count) ^ TRUE = TRUE
5580 * When draw_index == draw_count the result is
5581 * (TRUE) ^ TRUE = FALSE
5582 * After this all results will be:
5583 * (FALSE) ^ FALSE = FALSE
5585 mi_predicate
= MI_PREDICATE
| MI_PREDICATE_LOADOP_LOAD
|
5586 MI_PREDICATE_COMBINEOP_XOR
|
5587 MI_PREDICATE_COMPAREOP_SRCS_EQUAL
;
5589 iris_batch_emit(batch
, &mi_predicate
, sizeof(uint32_t));
5592 struct iris_bo
*bo
= iris_resource_bo(draw
->indirect
->buffer
);
5595 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
5596 lrm
.RegisterAddress
= _3DPRIM_VERTEX_COUNT
;
5597 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 0);
5599 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
5600 lrm
.RegisterAddress
= _3DPRIM_INSTANCE_COUNT
;
5601 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 4);
5603 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
5604 lrm
.RegisterAddress
= _3DPRIM_START_VERTEX
;
5605 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 8);
5607 if (draw
->index_size
) {
5608 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
5609 lrm
.RegisterAddress
= _3DPRIM_BASE_VERTEX
;
5610 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 12);
5612 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
5613 lrm
.RegisterAddress
= _3DPRIM_START_INSTANCE
;
5614 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 16);
5617 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
5618 lrm
.RegisterAddress
= _3DPRIM_START_INSTANCE
;
5619 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 12);
5621 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_IMM
), lri
) {
5622 lri
.RegisterOffset
= _3DPRIM_BASE_VERTEX
;
5626 } else if (draw
->count_from_stream_output
) {
5627 struct iris_stream_output_target
*so
=
5628 (void *) draw
->count_from_stream_output
;
5630 /* XXX: Replace with actual cache tracking */
5631 iris_emit_pipe_control_flush(batch
,
5632 "draw count from stream output stall",
5633 PIPE_CONTROL_CS_STALL
);
5635 struct gen_mi_builder b
;
5636 gen_mi_builder_init(&b
, batch
);
5638 struct iris_address addr
=
5639 ro_bo(iris_resource_bo(so
->offset
.res
), so
->offset
.offset
);
5640 struct gen_mi_value offset
=
5641 gen_mi_iadd_imm(&b
, gen_mi_mem32(addr
), -so
->base
.buffer_offset
);
5643 gen_mi_store(&b
, gen_mi_reg32(_3DPRIM_VERTEX_COUNT
),
5644 gen_mi_udiv32_imm(&b
, offset
, so
->stride
));
5646 _iris_emit_lri(batch
, _3DPRIM_START_VERTEX
, 0);
5647 _iris_emit_lri(batch
, _3DPRIM_BASE_VERTEX
, 0);
5648 _iris_emit_lri(batch
, _3DPRIM_START_INSTANCE
, 0);
5649 _iris_emit_lri(batch
, _3DPRIM_INSTANCE_COUNT
, draw
->instance_count
);
5652 iris_emit_cmd(batch
, GENX(3DPRIMITIVE
), prim
) {
5653 prim
.VertexAccessType
= draw
->index_size
> 0 ? RANDOM
: SEQUENTIAL
;
5654 prim
.PredicateEnable
= use_predicate
;
5656 if (draw
->indirect
|| draw
->count_from_stream_output
) {
5657 prim
.IndirectParameterEnable
= true;
5659 prim
.StartInstanceLocation
= draw
->start_instance
;
5660 prim
.InstanceCount
= draw
->instance_count
;
5661 prim
.VertexCountPerInstance
= draw
->count
;
5663 prim
.StartVertexLocation
= draw
->start
;
5665 if (draw
->index_size
) {
5666 prim
.BaseVertexLocation
+= draw
->index_bias
;
5668 prim
.StartVertexLocation
+= draw
->index_bias
;
5675 iris_upload_compute_state(struct iris_context
*ice
,
5676 struct iris_batch
*batch
,
5677 const struct pipe_grid_info
*grid
)
5679 const uint64_t dirty
= ice
->state
.dirty
;
5680 struct iris_screen
*screen
= batch
->screen
;
5681 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
5682 struct iris_binder
*binder
= &ice
->state
.binder
;
5683 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_COMPUTE
];
5684 struct iris_compiled_shader
*shader
=
5685 ice
->shaders
.prog
[MESA_SHADER_COMPUTE
];
5686 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
5687 struct brw_cs_prog_data
*cs_prog_data
= (void *) prog_data
;
5689 /* Always pin the binder. If we're emitting new binding table pointers,
5690 * we need it. If not, we're probably inheriting old tables via the
5691 * context, and need it anyway. Since true zero-bindings cases are
5692 * practically non-existent, just pin it and avoid last_res tracking.
5694 iris_use_pinned_bo(batch
, ice
->state
.binder
.bo
, false);
5696 if ((dirty
& IRIS_DIRTY_CONSTANTS_CS
) && shs
->sysvals_need_upload
)
5697 upload_sysvals(ice
, MESA_SHADER_COMPUTE
);
5699 if (dirty
& IRIS_DIRTY_BINDINGS_CS
)
5700 iris_populate_binding_table(ice
, batch
, MESA_SHADER_COMPUTE
, false);
5702 if (dirty
& IRIS_DIRTY_SAMPLER_STATES_CS
)
5703 iris_upload_sampler_states(ice
, MESA_SHADER_COMPUTE
);
5705 iris_use_optional_res(batch
, shs
->sampler_table
.res
, false);
5706 iris_use_pinned_bo(batch
, iris_resource_bo(shader
->assembly
.res
), false);
5708 if (ice
->state
.need_border_colors
)
5709 iris_use_pinned_bo(batch
, ice
->state
.border_color_pool
.bo
, false);
5711 if (dirty
& IRIS_DIRTY_CS
) {
5712 /* The MEDIA_VFE_STATE documentation for Gen8+ says:
5714 * "A stalling PIPE_CONTROL is required before MEDIA_VFE_STATE unless
5715 * the only bits that are changed are scoreboard related: Scoreboard
5716 * Enable, Scoreboard Type, Scoreboard Mask, Scoreboard Delta. For
5717 * these scoreboard related states, a MEDIA_STATE_FLUSH is
5720 iris_emit_pipe_control_flush(batch
,
5721 "workaround: stall before MEDIA_VFE_STATE",
5722 PIPE_CONTROL_CS_STALL
);
5724 iris_emit_cmd(batch
, GENX(MEDIA_VFE_STATE
), vfe
) {
5725 if (prog_data
->total_scratch
) {
5726 struct iris_bo
*bo
=
5727 iris_get_scratch_space(ice
, prog_data
->total_scratch
,
5728 MESA_SHADER_COMPUTE
);
5729 vfe
.PerThreadScratchSpace
= ffs(prog_data
->total_scratch
) - 11;
5730 vfe
.ScratchSpaceBasePointer
= rw_bo(bo
, 0);
5733 vfe
.MaximumNumberofThreads
=
5734 devinfo
->max_cs_threads
* screen
->subslice_total
- 1;
5736 vfe
.ResetGatewayTimer
=
5737 Resettingrelativetimerandlatchingtheglobaltimestamp
;
5740 vfe
.BypassGatewayControl
= true;
5742 vfe
.NumberofURBEntries
= 2;
5743 vfe
.URBEntryAllocationSize
= 2;
5745 vfe
.CURBEAllocationSize
=
5746 ALIGN(cs_prog_data
->push
.per_thread
.regs
* cs_prog_data
->threads
+
5747 cs_prog_data
->push
.cross_thread
.regs
, 2);
5751 /* TODO: Combine subgroup-id with cbuf0 so we can push regular uniforms */
5752 if (dirty
& IRIS_DIRTY_CS
) {
5753 uint32_t curbe_data_offset
= 0;
5754 assert(cs_prog_data
->push
.cross_thread
.dwords
== 0 &&
5755 cs_prog_data
->push
.per_thread
.dwords
== 1 &&
5756 cs_prog_data
->base
.param
[0] == BRW_PARAM_BUILTIN_SUBGROUP_ID
);
5757 uint32_t *curbe_data_map
=
5758 stream_state(batch
, ice
->state
.dynamic_uploader
,
5759 &ice
->state
.last_res
.cs_thread_ids
,
5760 ALIGN(cs_prog_data
->push
.total
.size
, 64), 64,
5761 &curbe_data_offset
);
5762 assert(curbe_data_map
);
5763 memset(curbe_data_map
, 0x5a, ALIGN(cs_prog_data
->push
.total
.size
, 64));
5764 iris_fill_cs_push_const_buffer(cs_prog_data
, curbe_data_map
);
5766 iris_emit_cmd(batch
, GENX(MEDIA_CURBE_LOAD
), curbe
) {
5767 curbe
.CURBETotalDataLength
=
5768 ALIGN(cs_prog_data
->push
.total
.size
, 64);
5769 curbe
.CURBEDataStartAddress
= curbe_data_offset
;
5773 if (dirty
& (IRIS_DIRTY_SAMPLER_STATES_CS
|
5774 IRIS_DIRTY_BINDINGS_CS
|
5775 IRIS_DIRTY_CONSTANTS_CS
|
5777 uint32_t desc
[GENX(INTERFACE_DESCRIPTOR_DATA_length
)];
5779 iris_pack_state(GENX(INTERFACE_DESCRIPTOR_DATA
), desc
, idd
) {
5780 idd
.SamplerStatePointer
= shs
->sampler_table
.offset
;
5781 idd
.BindingTablePointer
= binder
->bt_offset
[MESA_SHADER_COMPUTE
];
5784 for (int i
= 0; i
< GENX(INTERFACE_DESCRIPTOR_DATA_length
); i
++)
5785 desc
[i
] |= ((uint32_t *) shader
->derived_data
)[i
];
5787 iris_emit_cmd(batch
, GENX(MEDIA_INTERFACE_DESCRIPTOR_LOAD
), load
) {
5788 load
.InterfaceDescriptorTotalLength
=
5789 GENX(INTERFACE_DESCRIPTOR_DATA_length
) * sizeof(uint32_t);
5790 load
.InterfaceDescriptorDataStartAddress
=
5791 emit_state(batch
, ice
->state
.dynamic_uploader
,
5792 &ice
->state
.last_res
.cs_desc
, desc
, sizeof(desc
), 64);
5796 uint32_t group_size
= grid
->block
[0] * grid
->block
[1] * grid
->block
[2];
5797 uint32_t remainder
= group_size
& (cs_prog_data
->simd_size
- 1);
5798 uint32_t right_mask
;
5801 right_mask
= ~0u >> (32 - remainder
);
5803 right_mask
= ~0u >> (32 - cs_prog_data
->simd_size
);
5805 #define GPGPU_DISPATCHDIMX 0x2500
5806 #define GPGPU_DISPATCHDIMY 0x2504
5807 #define GPGPU_DISPATCHDIMZ 0x2508
5809 if (grid
->indirect
) {
5810 struct iris_state_ref
*grid_size
= &ice
->state
.grid_size
;
5811 struct iris_bo
*bo
= iris_resource_bo(grid_size
->res
);
5812 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
5813 lrm
.RegisterAddress
= GPGPU_DISPATCHDIMX
;
5814 lrm
.MemoryAddress
= ro_bo(bo
, grid_size
->offset
+ 0);
5816 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
5817 lrm
.RegisterAddress
= GPGPU_DISPATCHDIMY
;
5818 lrm
.MemoryAddress
= ro_bo(bo
, grid_size
->offset
+ 4);
5820 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
5821 lrm
.RegisterAddress
= GPGPU_DISPATCHDIMZ
;
5822 lrm
.MemoryAddress
= ro_bo(bo
, grid_size
->offset
+ 8);
5826 iris_emit_cmd(batch
, GENX(GPGPU_WALKER
), ggw
) {
5827 ggw
.IndirectParameterEnable
= grid
->indirect
!= NULL
;
5828 ggw
.SIMDSize
= cs_prog_data
->simd_size
/ 16;
5829 ggw
.ThreadDepthCounterMaximum
= 0;
5830 ggw
.ThreadHeightCounterMaximum
= 0;
5831 ggw
.ThreadWidthCounterMaximum
= cs_prog_data
->threads
- 1;
5832 ggw
.ThreadGroupIDXDimension
= grid
->grid
[0];
5833 ggw
.ThreadGroupIDYDimension
= grid
->grid
[1];
5834 ggw
.ThreadGroupIDZDimension
= grid
->grid
[2];
5835 ggw
.RightExecutionMask
= right_mask
;
5836 ggw
.BottomExecutionMask
= 0xffffffff;
5839 iris_emit_cmd(batch
, GENX(MEDIA_STATE_FLUSH
), msf
);
5841 if (!batch
->contains_draw
) {
5842 iris_restore_compute_saved_bos(ice
, batch
, grid
);
5843 batch
->contains_draw
= true;
5848 * State module teardown.
5851 iris_destroy_state(struct iris_context
*ice
)
5853 struct iris_genx_state
*genx
= ice
->state
.genx
;
5855 pipe_resource_reference(&ice
->draw
.draw_params_res
, NULL
);
5856 pipe_resource_reference(&ice
->draw
.derived_draw_params_res
, NULL
);
5858 uint64_t bound_vbs
= ice
->state
.bound_vertex_buffers
;
5860 const int i
= u_bit_scan64(&bound_vbs
);
5861 pipe_resource_reference(&genx
->vertex_buffers
[i
].resource
, NULL
);
5863 free(ice
->state
.genx
);
5865 for (int i
= 0; i
< 4; i
++) {
5866 pipe_so_target_reference(&ice
->state
.so_target
[i
], NULL
);
5869 for (unsigned i
= 0; i
< ice
->state
.framebuffer
.nr_cbufs
; i
++) {
5870 pipe_surface_reference(&ice
->state
.framebuffer
.cbufs
[i
], NULL
);
5872 pipe_surface_reference(&ice
->state
.framebuffer
.zsbuf
, NULL
);
5874 for (int stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
5875 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
5876 pipe_resource_reference(&shs
->sampler_table
.res
, NULL
);
5877 for (int i
= 0; i
< PIPE_MAX_CONSTANT_BUFFERS
; i
++) {
5878 pipe_resource_reference(&shs
->constbuf
[i
].buffer
, NULL
);
5879 pipe_resource_reference(&shs
->constbuf_surf_state
[i
].res
, NULL
);
5881 for (int i
= 0; i
< PIPE_MAX_SHADER_IMAGES
; i
++) {
5882 pipe_resource_reference(&shs
->image
[i
].base
.resource
, NULL
);
5883 pipe_resource_reference(&shs
->image
[i
].surface_state
.res
, NULL
);
5885 for (int i
= 0; i
< PIPE_MAX_SHADER_BUFFERS
; i
++) {
5886 pipe_resource_reference(&shs
->ssbo
[i
].buffer
, NULL
);
5887 pipe_resource_reference(&shs
->ssbo_surf_state
[i
].res
, NULL
);
5889 for (int i
= 0; i
< IRIS_MAX_TEXTURE_SAMPLERS
; i
++) {
5890 pipe_sampler_view_reference((struct pipe_sampler_view
**)
5891 &shs
->textures
[i
], NULL
);
5895 pipe_resource_reference(&ice
->state
.grid_size
.res
, NULL
);
5896 pipe_resource_reference(&ice
->state
.grid_surf_state
.res
, NULL
);
5898 pipe_resource_reference(&ice
->state
.null_fb
.res
, NULL
);
5899 pipe_resource_reference(&ice
->state
.unbound_tex
.res
, NULL
);
5901 pipe_resource_reference(&ice
->state
.last_res
.cc_vp
, NULL
);
5902 pipe_resource_reference(&ice
->state
.last_res
.sf_cl_vp
, NULL
);
5903 pipe_resource_reference(&ice
->state
.last_res
.color_calc
, NULL
);
5904 pipe_resource_reference(&ice
->state
.last_res
.scissor
, NULL
);
5905 pipe_resource_reference(&ice
->state
.last_res
.blend
, NULL
);
5906 pipe_resource_reference(&ice
->state
.last_res
.index_buffer
, NULL
);
5907 pipe_resource_reference(&ice
->state
.last_res
.cs_thread_ids
, NULL
);
5908 pipe_resource_reference(&ice
->state
.last_res
.cs_desc
, NULL
);
5911 /* ------------------------------------------------------------------- */
5914 iris_rebind_buffer(struct iris_context
*ice
,
5915 struct iris_resource
*res
,
5916 uint64_t old_address
)
5918 struct pipe_context
*ctx
= &ice
->ctx
;
5919 struct iris_screen
*screen
= (void *) ctx
->screen
;
5920 struct iris_genx_state
*genx
= ice
->state
.genx
;
5922 assert(res
->base
.target
== PIPE_BUFFER
);
5924 /* Buffers can't be framebuffer attachments, nor display related,
5925 * and we don't have upstream Clover support.
5927 assert(!(res
->bind_history
& (PIPE_BIND_DEPTH_STENCIL
|
5928 PIPE_BIND_RENDER_TARGET
|
5929 PIPE_BIND_BLENDABLE
|
5930 PIPE_BIND_DISPLAY_TARGET
|
5932 PIPE_BIND_COMPUTE_RESOURCE
|
5933 PIPE_BIND_GLOBAL
)));
5935 if (res
->bind_history
& PIPE_BIND_VERTEX_BUFFER
) {
5936 uint64_t bound_vbs
= ice
->state
.bound_vertex_buffers
;
5938 const int i
= u_bit_scan64(&bound_vbs
);
5939 struct iris_vertex_buffer_state
*state
= &genx
->vertex_buffers
[i
];
5941 /* Update the CPU struct */
5942 STATIC_ASSERT(GENX(VERTEX_BUFFER_STATE_BufferStartingAddress_start
) == 32);
5943 STATIC_ASSERT(GENX(VERTEX_BUFFER_STATE_BufferStartingAddress_bits
) == 64);
5944 uint64_t *addr
= (uint64_t *) &state
->state
[1];
5946 if (*addr
== old_address
) {
5947 *addr
= res
->bo
->gtt_offset
;
5948 ice
->state
.dirty
|= IRIS_DIRTY_VERTEX_BUFFERS
;
5953 /* We don't need to handle PIPE_BIND_INDEX_BUFFER here: we re-emit
5954 * the 3DSTATE_INDEX_BUFFER packet whenever the address changes.
5956 * There is also no need to handle these:
5957 * - PIPE_BIND_COMMAND_ARGS_BUFFER (emitted for every indirect draw)
5958 * - PIPE_BIND_QUERY_BUFFER (no persistent state references)
5961 if (res
->bind_history
& PIPE_BIND_STREAM_OUTPUT
) {
5962 /* XXX: be careful about resetting vs appending... */
5966 for (int s
= MESA_SHADER_VERTEX
; s
< MESA_SHADER_STAGES
; s
++) {
5967 struct iris_shader_state
*shs
= &ice
->state
.shaders
[s
];
5968 enum pipe_shader_type p_stage
= stage_to_pipe(s
);
5970 if (res
->bind_history
& PIPE_BIND_CONSTANT_BUFFER
) {
5971 /* Skip constant buffer 0, it's for regular uniforms, not UBOs */
5972 uint32_t bound_cbufs
= shs
->bound_cbufs
& ~1u;
5973 while (bound_cbufs
) {
5974 const int i
= u_bit_scan(&bound_cbufs
);
5975 struct pipe_shader_buffer
*cbuf
= &shs
->constbuf
[i
];
5976 struct iris_state_ref
*surf_state
= &shs
->constbuf_surf_state
[i
];
5978 if (res
->bo
== iris_resource_bo(cbuf
->buffer
)) {
5979 iris_upload_ubo_ssbo_surf_state(ice
, cbuf
, surf_state
, false);
5980 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_VS
<< s
;
5985 if (res
->bind_history
& PIPE_BIND_SHADER_BUFFER
) {
5986 uint32_t bound_ssbos
= shs
->bound_ssbos
;
5987 while (bound_ssbos
) {
5988 const int i
= u_bit_scan(&bound_ssbos
);
5989 struct pipe_shader_buffer
*ssbo
= &shs
->ssbo
[i
];
5991 if (res
->bo
== iris_resource_bo(ssbo
->buffer
)) {
5992 struct pipe_shader_buffer buf
= {
5993 .buffer
= &res
->base
,
5994 .buffer_offset
= ssbo
->buffer_offset
,
5995 .buffer_size
= ssbo
->buffer_size
,
5997 iris_set_shader_buffers(ctx
, p_stage
, i
, 1, &buf
,
5998 (shs
->writable_ssbos
>> i
) & 1);
6003 if (res
->bind_history
& PIPE_BIND_SAMPLER_VIEW
) {
6004 uint32_t bound_sampler_views
= shs
->bound_sampler_views
;
6005 while (bound_sampler_views
) {
6006 const int i
= u_bit_scan(&bound_sampler_views
);
6007 struct iris_sampler_view
*isv
= shs
->textures
[i
];
6009 if (res
->bo
== iris_resource_bo(isv
->base
.texture
)) {
6010 void *map
= alloc_surface_states(ice
->state
.surface_uploader
,
6011 &isv
->surface_state
,
6012 isv
->res
->aux
.sampler_usages
);
6014 fill_buffer_surface_state(&screen
->isl_dev
, isv
->res
, map
,
6015 isv
->view
.format
, isv
->view
.swizzle
,
6016 isv
->base
.u
.buf
.offset
,
6017 isv
->base
.u
.buf
.size
);
6018 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_VS
<< s
;
6023 if (res
->bind_history
& PIPE_BIND_SHADER_IMAGE
) {
6024 uint32_t bound_image_views
= shs
->bound_image_views
;
6025 while (bound_image_views
) {
6026 const int i
= u_bit_scan(&bound_image_views
);
6027 struct iris_image_view
*iv
= &shs
->image
[i
];
6029 if (res
->bo
== iris_resource_bo(iv
->base
.resource
)) {
6030 iris_set_shader_images(ctx
, p_stage
, i
, 1, &iv
->base
);
6037 /* ------------------------------------------------------------------- */
6040 iris_load_register_reg32(struct iris_batch
*batch
, uint32_t dst
,
6043 _iris_emit_lrr(batch
, dst
, src
);
6047 iris_load_register_reg64(struct iris_batch
*batch
, uint32_t dst
,
6050 _iris_emit_lrr(batch
, dst
, src
);
6051 _iris_emit_lrr(batch
, dst
+ 4, src
+ 4);
6055 iris_load_register_imm32(struct iris_batch
*batch
, uint32_t reg
,
6058 _iris_emit_lri(batch
, reg
, val
);
6062 iris_load_register_imm64(struct iris_batch
*batch
, uint32_t reg
,
6065 _iris_emit_lri(batch
, reg
+ 0, val
& 0xffffffff);
6066 _iris_emit_lri(batch
, reg
+ 4, val
>> 32);
6070 * Emit MI_LOAD_REGISTER_MEM to load a 32-bit MMIO register from a buffer.
6073 iris_load_register_mem32(struct iris_batch
*batch
, uint32_t reg
,
6074 struct iris_bo
*bo
, uint32_t offset
)
6076 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
6077 lrm
.RegisterAddress
= reg
;
6078 lrm
.MemoryAddress
= ro_bo(bo
, offset
);
6083 * Load a 64-bit value from a buffer into a MMIO register via
6084 * two MI_LOAD_REGISTER_MEM commands.
6087 iris_load_register_mem64(struct iris_batch
*batch
, uint32_t reg
,
6088 struct iris_bo
*bo
, uint32_t offset
)
6090 iris_load_register_mem32(batch
, reg
+ 0, bo
, offset
+ 0);
6091 iris_load_register_mem32(batch
, reg
+ 4, bo
, offset
+ 4);
6095 iris_store_register_mem32(struct iris_batch
*batch
, uint32_t reg
,
6096 struct iris_bo
*bo
, uint32_t offset
,
6099 iris_emit_cmd(batch
, GENX(MI_STORE_REGISTER_MEM
), srm
) {
6100 srm
.RegisterAddress
= reg
;
6101 srm
.MemoryAddress
= rw_bo(bo
, offset
);
6102 srm
.PredicateEnable
= predicated
;
6107 iris_store_register_mem64(struct iris_batch
*batch
, uint32_t reg
,
6108 struct iris_bo
*bo
, uint32_t offset
,
6111 iris_store_register_mem32(batch
, reg
+ 0, bo
, offset
+ 0, predicated
);
6112 iris_store_register_mem32(batch
, reg
+ 4, bo
, offset
+ 4, predicated
);
6116 iris_store_data_imm32(struct iris_batch
*batch
,
6117 struct iris_bo
*bo
, uint32_t offset
,
6120 iris_emit_cmd(batch
, GENX(MI_STORE_DATA_IMM
), sdi
) {
6121 sdi
.Address
= rw_bo(bo
, offset
);
6122 sdi
.ImmediateData
= imm
;
6127 iris_store_data_imm64(struct iris_batch
*batch
,
6128 struct iris_bo
*bo
, uint32_t offset
,
6131 /* Can't use iris_emit_cmd because MI_STORE_DATA_IMM has a length of
6132 * 2 in genxml but it's actually variable length and we need 5 DWords.
6134 void *map
= iris_get_command_space(batch
, 4 * 5);
6135 _iris_pack_command(batch
, GENX(MI_STORE_DATA_IMM
), map
, sdi
) {
6136 sdi
.DWordLength
= 5 - 2;
6137 sdi
.Address
= rw_bo(bo
, offset
);
6138 sdi
.ImmediateData
= imm
;
6143 iris_copy_mem_mem(struct iris_batch
*batch
,
6144 struct iris_bo
*dst_bo
, uint32_t dst_offset
,
6145 struct iris_bo
*src_bo
, uint32_t src_offset
,
6148 /* MI_COPY_MEM_MEM operates on DWords. */
6149 assert(bytes
% 4 == 0);
6150 assert(dst_offset
% 4 == 0);
6151 assert(src_offset
% 4 == 0);
6153 for (unsigned i
= 0; i
< bytes
; i
+= 4) {
6154 iris_emit_cmd(batch
, GENX(MI_COPY_MEM_MEM
), cp
) {
6155 cp
.DestinationMemoryAddress
= rw_bo(dst_bo
, dst_offset
+ i
);
6156 cp
.SourceMemoryAddress
= ro_bo(src_bo
, src_offset
+ i
);
6161 /* ------------------------------------------------------------------- */
6164 flags_to_post_sync_op(uint32_t flags
)
6166 if (flags
& PIPE_CONTROL_WRITE_IMMEDIATE
)
6167 return WriteImmediateData
;
6169 if (flags
& PIPE_CONTROL_WRITE_DEPTH_COUNT
)
6170 return WritePSDepthCount
;
6172 if (flags
& PIPE_CONTROL_WRITE_TIMESTAMP
)
6173 return WriteTimestamp
;
6179 * Do the given flags have a Post Sync or LRI Post Sync operation?
6181 static enum pipe_control_flags
6182 get_post_sync_flags(enum pipe_control_flags flags
)
6184 flags
&= PIPE_CONTROL_WRITE_IMMEDIATE
|
6185 PIPE_CONTROL_WRITE_DEPTH_COUNT
|
6186 PIPE_CONTROL_WRITE_TIMESTAMP
|
6187 PIPE_CONTROL_LRI_POST_SYNC_OP
;
6189 /* Only one "Post Sync Op" is allowed, and it's mutually exclusive with
6190 * "LRI Post Sync Operation". So more than one bit set would be illegal.
6192 assert(util_bitcount(flags
) <= 1);
6197 #define IS_COMPUTE_PIPELINE(batch) (batch->name == IRIS_BATCH_COMPUTE)
6200 * Emit a series of PIPE_CONTROL commands, taking into account any
6201 * workarounds necessary to actually accomplish the caller's request.
6203 * Unless otherwise noted, spec quotations in this function come from:
6205 * Synchronization of the 3D Pipeline > PIPE_CONTROL Command > Programming
6206 * Restrictions for PIPE_CONTROL.
6208 * You should not use this function directly. Use the helpers in
6209 * iris_pipe_control.c instead, which may split the pipe control further.
6212 iris_emit_raw_pipe_control(struct iris_batch
*batch
,
6219 UNUSED
const struct gen_device_info
*devinfo
= &batch
->screen
->devinfo
;
6220 enum pipe_control_flags post_sync_flags
= get_post_sync_flags(flags
);
6221 enum pipe_control_flags non_lri_post_sync_flags
=
6222 post_sync_flags
& ~PIPE_CONTROL_LRI_POST_SYNC_OP
;
6224 /* Recursive PIPE_CONTROL workarounds --------------------------------
6225 * (http://knowyourmeme.com/memes/xzibit-yo-dawg)
6227 * We do these first because we want to look at the original operation,
6228 * rather than any workarounds we set.
6230 if (GEN_GEN
== 9 && (flags
& PIPE_CONTROL_VF_CACHE_INVALIDATE
)) {
6231 /* The PIPE_CONTROL "VF Cache Invalidation Enable" bit description
6232 * lists several workarounds:
6234 * "Project: SKL, KBL, BXT
6236 * If the VF Cache Invalidation Enable is set to a 1 in a
6237 * PIPE_CONTROL, a separate Null PIPE_CONTROL, all bitfields
6238 * sets to 0, with the VF Cache Invalidation Enable set to 0
6239 * needs to be sent prior to the PIPE_CONTROL with VF Cache
6240 * Invalidation Enable set to a 1."
6242 iris_emit_raw_pipe_control(batch
,
6243 "workaround: recursive VF cache invalidate",
6247 if (GEN_GEN
== 9 && IS_COMPUTE_PIPELINE(batch
) && post_sync_flags
) {
6248 /* Project: SKL / Argument: LRI Post Sync Operation [23]
6250 * "PIPECONTROL command with “Command Streamer Stall Enable” must be
6251 * programmed prior to programming a PIPECONTROL command with "LRI
6252 * Post Sync Operation" in GPGPU mode of operation (i.e when
6253 * PIPELINE_SELECT command is set to GPGPU mode of operation)."
6255 * The same text exists a few rows below for Post Sync Op.
6257 iris_emit_raw_pipe_control(batch
,
6258 "workaround: CS stall before gpgpu post-sync",
6259 PIPE_CONTROL_CS_STALL
, bo
, offset
, imm
);
6262 if (GEN_GEN
== 10 && (flags
& PIPE_CONTROL_RENDER_TARGET_FLUSH
)) {
6264 * "Before sending a PIPE_CONTROL command with bit 12 set, SW must issue
6265 * another PIPE_CONTROL with Render Target Cache Flush Enable (bit 12)
6266 * = 0 and Pipe Control Flush Enable (bit 7) = 1"
6268 iris_emit_raw_pipe_control(batch
,
6269 "workaround: PC flush before RT flush",
6270 PIPE_CONTROL_FLUSH_ENABLE
, bo
, offset
, imm
);
6273 /* "Flush Types" workarounds ---------------------------------------------
6274 * We do these now because they may add post-sync operations or CS stalls.
6277 if (GEN_GEN
< 11 && flags
& PIPE_CONTROL_VF_CACHE_INVALIDATE
) {
6278 /* Project: BDW, SKL+ (stopping at CNL) / Argument: VF Invalidate
6280 * "'Post Sync Operation' must be enabled to 'Write Immediate Data' or
6281 * 'Write PS Depth Count' or 'Write Timestamp'."
6284 flags
|= PIPE_CONTROL_WRITE_IMMEDIATE
;
6285 post_sync_flags
|= PIPE_CONTROL_WRITE_IMMEDIATE
;
6286 non_lri_post_sync_flags
|= PIPE_CONTROL_WRITE_IMMEDIATE
;
6287 bo
= batch
->screen
->workaround_bo
;
6291 /* #1130 from Gen10 workarounds page:
6293 * "Enable Depth Stall on every Post Sync Op if Render target Cache
6294 * Flush is not enabled in same PIPE CONTROL and Enable Pixel score
6295 * board stall if Render target cache flush is enabled."
6297 * Applicable to CNL B0 and C0 steppings only.
6299 * The wording here is unclear, and this workaround doesn't look anything
6300 * like the internal bug report recommendations, but leave it be for now...
6302 if (GEN_GEN
== 10) {
6303 if (flags
& PIPE_CONTROL_RENDER_TARGET_FLUSH
) {
6304 flags
|= PIPE_CONTROL_STALL_AT_SCOREBOARD
;
6305 } else if (flags
& non_lri_post_sync_flags
) {
6306 flags
|= PIPE_CONTROL_DEPTH_STALL
;
6310 if (flags
& PIPE_CONTROL_DEPTH_STALL
) {
6311 /* From the PIPE_CONTROL instruction table, bit 13 (Depth Stall Enable):
6313 * "This bit must be DISABLED for operations other than writing
6316 * This seems like nonsense. An Ivybridge workaround requires us to
6317 * emit a PIPE_CONTROL with a depth stall and write immediate post-sync
6318 * operation. Gen8+ requires us to emit depth stalls and depth cache
6319 * flushes together. So, it's hard to imagine this means anything other
6320 * than "we originally intended this to be used for PS_DEPTH_COUNT".
6322 * We ignore the supposed restriction and do nothing.
6326 if (flags
& (PIPE_CONTROL_RENDER_TARGET_FLUSH
|
6327 PIPE_CONTROL_STALL_AT_SCOREBOARD
)) {
6328 /* From the PIPE_CONTROL instruction table, bit 12 and bit 1:
6330 * "This bit must be DISABLED for End-of-pipe (Read) fences,
6331 * PS_DEPTH_COUNT or TIMESTAMP queries."
6333 * TODO: Implement end-of-pipe checking.
6335 assert(!(post_sync_flags
& (PIPE_CONTROL_WRITE_DEPTH_COUNT
|
6336 PIPE_CONTROL_WRITE_TIMESTAMP
)));
6339 if (GEN_GEN
< 11 && (flags
& PIPE_CONTROL_STALL_AT_SCOREBOARD
)) {
6340 /* From the PIPE_CONTROL instruction table, bit 1:
6342 * "This bit is ignored if Depth Stall Enable is set.
6343 * Further, the render cache is not flushed even if Write Cache
6344 * Flush Enable bit is set."
6346 * We assert that the caller doesn't do this combination, to try and
6347 * prevent mistakes. It shouldn't hurt the GPU, though.
6349 * We skip this check on Gen11+ as the "Stall at Pixel Scoreboard"
6350 * and "Render Target Flush" combo is explicitly required for BTI
6351 * update workarounds.
6353 assert(!(flags
& (PIPE_CONTROL_DEPTH_STALL
|
6354 PIPE_CONTROL_RENDER_TARGET_FLUSH
)));
6357 /* PIPE_CONTROL page workarounds ------------------------------------- */
6359 if (GEN_GEN
<= 8 && (flags
& PIPE_CONTROL_STATE_CACHE_INVALIDATE
)) {
6360 /* From the PIPE_CONTROL page itself:
6363 * Restriction: Pipe_control with CS-stall bit set must be issued
6364 * before a pipe-control command that has the State Cache
6365 * Invalidate bit set."
6367 flags
|= PIPE_CONTROL_CS_STALL
;
6370 if (flags
& PIPE_CONTROL_FLUSH_LLC
) {
6371 /* From the PIPE_CONTROL instruction table, bit 26 (Flush LLC):
6374 * SW must always program Post-Sync Operation to "Write Immediate
6375 * Data" when Flush LLC is set."
6377 * For now, we just require the caller to do it.
6379 assert(flags
& PIPE_CONTROL_WRITE_IMMEDIATE
);
6382 /* "Post-Sync Operation" workarounds -------------------------------- */
6384 /* Project: All / Argument: Global Snapshot Count Reset [19]
6386 * "This bit must not be exercised on any product.
6387 * Requires stall bit ([20] of DW1) set."
6389 * We don't use this, so we just assert that it isn't used. The
6390 * PIPE_CONTROL instruction page indicates that they intended this
6391 * as a debug feature and don't think it is useful in production,
6392 * but it may actually be usable, should we ever want to.
6394 assert((flags
& PIPE_CONTROL_GLOBAL_SNAPSHOT_COUNT_RESET
) == 0);
6396 if (flags
& (PIPE_CONTROL_MEDIA_STATE_CLEAR
|
6397 PIPE_CONTROL_INDIRECT_STATE_POINTERS_DISABLE
)) {
6398 /* Project: All / Arguments:
6400 * - Generic Media State Clear [16]
6401 * - Indirect State Pointers Disable [16]
6403 * "Requires stall bit ([20] of DW1) set."
6405 * Also, the PIPE_CONTROL instruction table, bit 16 (Generic Media
6406 * State Clear) says:
6408 * "PIPECONTROL command with “Command Streamer Stall Enable” must be
6409 * programmed prior to programming a PIPECONTROL command with "Media
6410 * State Clear" set in GPGPU mode of operation"
6412 * This is a subset of the earlier rule, so there's nothing to do.
6414 flags
|= PIPE_CONTROL_CS_STALL
;
6417 if (flags
& PIPE_CONTROL_STORE_DATA_INDEX
) {
6418 /* Project: All / Argument: Store Data Index
6420 * "Post-Sync Operation ([15:14] of DW1) must be set to something other
6423 * For now, we just assert that the caller does this. We might want to
6424 * automatically add a write to the workaround BO...
6426 assert(non_lri_post_sync_flags
!= 0);
6429 if (flags
& PIPE_CONTROL_SYNC_GFDT
) {
6430 /* Project: All / Argument: Sync GFDT
6432 * "Post-Sync Operation ([15:14] of DW1) must be set to something other
6433 * than '0' or 0x2520[13] must be set."
6435 * For now, we just assert that the caller does this.
6437 assert(non_lri_post_sync_flags
!= 0);
6440 if (flags
& PIPE_CONTROL_TLB_INVALIDATE
) {
6441 /* Project: IVB+ / Argument: TLB inv
6443 * "Requires stall bit ([20] of DW1) set."
6445 * Also, from the PIPE_CONTROL instruction table:
6448 * Post Sync Operation or CS stall must be set to ensure a TLB
6449 * invalidation occurs. Otherwise no cycle will occur to the TLB
6450 * cache to invalidate."
6452 * This is not a subset of the earlier rule, so there's nothing to do.
6454 flags
|= PIPE_CONTROL_CS_STALL
;
6457 if (GEN_GEN
== 9 && devinfo
->gt
== 4) {
6458 /* TODO: The big Skylake GT4 post sync op workaround */
6461 /* "GPGPU specific workarounds" (both post-sync and flush) ------------ */
6463 if (IS_COMPUTE_PIPELINE(batch
)) {
6464 if (GEN_GEN
>= 9 && (flags
& PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
)) {
6465 /* Project: SKL+ / Argument: Tex Invalidate
6466 * "Requires stall bit ([20] of DW) set for all GPGPU Workloads."
6468 flags
|= PIPE_CONTROL_CS_STALL
;
6471 if (GEN_GEN
== 8 && (post_sync_flags
||
6472 (flags
& (PIPE_CONTROL_NOTIFY_ENABLE
|
6473 PIPE_CONTROL_DEPTH_STALL
|
6474 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
6475 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
6476 PIPE_CONTROL_DATA_CACHE_FLUSH
)))) {
6477 /* Project: BDW / Arguments:
6479 * - LRI Post Sync Operation [23]
6480 * - Post Sync Op [15:14]
6482 * - Depth Stall [13]
6483 * - Render Target Cache Flush [12]
6484 * - Depth Cache Flush [0]
6485 * - DC Flush Enable [5]
6487 * "Requires stall bit ([20] of DW) set for all GPGPU and Media
6490 flags
|= PIPE_CONTROL_CS_STALL
;
6492 /* Also, from the PIPE_CONTROL instruction table, bit 20:
6495 * This bit must be always set when PIPE_CONTROL command is
6496 * programmed by GPGPU and MEDIA workloads, except for the cases
6497 * when only Read Only Cache Invalidation bits are set (State
6498 * Cache Invalidation Enable, Instruction cache Invalidation
6499 * Enable, Texture Cache Invalidation Enable, Constant Cache
6500 * Invalidation Enable). This is to WA FFDOP CG issue, this WA
6501 * need not implemented when FF_DOP_CG is disable via "Fixed
6502 * Function DOP Clock Gate Disable" bit in RC_PSMI_CTRL register."
6504 * It sounds like we could avoid CS stalls in some cases, but we
6505 * don't currently bother. This list isn't exactly the list above,
6511 /* "Stall" workarounds ----------------------------------------------
6512 * These have to come after the earlier ones because we may have added
6513 * some additional CS stalls above.
6516 if (GEN_GEN
< 9 && (flags
& PIPE_CONTROL_CS_STALL
)) {
6517 /* Project: PRE-SKL, VLV, CHV
6519 * "[All Stepping][All SKUs]:
6521 * One of the following must also be set:
6523 * - Render Target Cache Flush Enable ([12] of DW1)
6524 * - Depth Cache Flush Enable ([0] of DW1)
6525 * - Stall at Pixel Scoreboard ([1] of DW1)
6526 * - Depth Stall ([13] of DW1)
6527 * - Post-Sync Operation ([13] of DW1)
6528 * - DC Flush Enable ([5] of DW1)"
6530 * If we don't already have one of those bits set, we choose to add
6531 * "Stall at Pixel Scoreboard". Some of the other bits require a
6532 * CS stall as a workaround (see above), which would send us into
6533 * an infinite recursion of PIPE_CONTROLs. "Stall at Pixel Scoreboard"
6534 * appears to be safe, so we choose that.
6536 const uint32_t wa_bits
= PIPE_CONTROL_RENDER_TARGET_FLUSH
|
6537 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
6538 PIPE_CONTROL_WRITE_IMMEDIATE
|
6539 PIPE_CONTROL_WRITE_DEPTH_COUNT
|
6540 PIPE_CONTROL_WRITE_TIMESTAMP
|
6541 PIPE_CONTROL_STALL_AT_SCOREBOARD
|
6542 PIPE_CONTROL_DEPTH_STALL
|
6543 PIPE_CONTROL_DATA_CACHE_FLUSH
;
6544 if (!(flags
& wa_bits
))
6545 flags
|= PIPE_CONTROL_STALL_AT_SCOREBOARD
;
6548 /* Emit --------------------------------------------------------------- */
6550 if (INTEL_DEBUG
& DEBUG_PIPE_CONTROL
) {
6552 " PC [%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%"PRIx64
"]: %s\n",
6553 (flags
& PIPE_CONTROL_FLUSH_ENABLE
) ? "PipeCon " : "",
6554 (flags
& PIPE_CONTROL_CS_STALL
) ? "CS " : "",
6555 (flags
& PIPE_CONTROL_STALL_AT_SCOREBOARD
) ? "Scoreboard " : "",
6556 (flags
& PIPE_CONTROL_VF_CACHE_INVALIDATE
) ? "VF " : "",
6557 (flags
& PIPE_CONTROL_RENDER_TARGET_FLUSH
) ? "RT " : "",
6558 (flags
& PIPE_CONTROL_CONST_CACHE_INVALIDATE
) ? "Const " : "",
6559 (flags
& PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
) ? "TC " : "",
6560 (flags
& PIPE_CONTROL_DATA_CACHE_FLUSH
) ? "DC " : "",
6561 (flags
& PIPE_CONTROL_DEPTH_CACHE_FLUSH
) ? "ZFlush " : "",
6562 (flags
& PIPE_CONTROL_DEPTH_STALL
) ? "ZStall " : "",
6563 (flags
& PIPE_CONTROL_STATE_CACHE_INVALIDATE
) ? "State " : "",
6564 (flags
& PIPE_CONTROL_TLB_INVALIDATE
) ? "TLB " : "",
6565 (flags
& PIPE_CONTROL_INSTRUCTION_INVALIDATE
) ? "Inst " : "",
6566 (flags
& PIPE_CONTROL_MEDIA_STATE_CLEAR
) ? "MediaClear " : "",
6567 (flags
& PIPE_CONTROL_NOTIFY_ENABLE
) ? "Notify " : "",
6568 (flags
& PIPE_CONTROL_GLOBAL_SNAPSHOT_COUNT_RESET
) ?
6570 (flags
& PIPE_CONTROL_INDIRECT_STATE_POINTERS_DISABLE
) ?
6572 (flags
& PIPE_CONTROL_WRITE_IMMEDIATE
) ? "WriteImm " : "",
6573 (flags
& PIPE_CONTROL_WRITE_DEPTH_COUNT
) ? "WriteZCount " : "",
6574 (flags
& PIPE_CONTROL_WRITE_TIMESTAMP
) ? "WriteTimestamp " : "",
6578 iris_emit_cmd(batch
, GENX(PIPE_CONTROL
), pc
) {
6579 pc
.LRIPostSyncOperation
= NoLRIOperation
;
6580 pc
.PipeControlFlushEnable
= flags
& PIPE_CONTROL_FLUSH_ENABLE
;
6581 pc
.DCFlushEnable
= flags
& PIPE_CONTROL_DATA_CACHE_FLUSH
;
6582 pc
.StoreDataIndex
= 0;
6583 pc
.CommandStreamerStallEnable
= flags
& PIPE_CONTROL_CS_STALL
;
6584 pc
.GlobalSnapshotCountReset
=
6585 flags
& PIPE_CONTROL_GLOBAL_SNAPSHOT_COUNT_RESET
;
6586 pc
.TLBInvalidate
= flags
& PIPE_CONTROL_TLB_INVALIDATE
;
6587 pc
.GenericMediaStateClear
= flags
& PIPE_CONTROL_MEDIA_STATE_CLEAR
;
6588 pc
.StallAtPixelScoreboard
= flags
& PIPE_CONTROL_STALL_AT_SCOREBOARD
;
6589 pc
.RenderTargetCacheFlushEnable
=
6590 flags
& PIPE_CONTROL_RENDER_TARGET_FLUSH
;
6591 pc
.DepthCacheFlushEnable
= flags
& PIPE_CONTROL_DEPTH_CACHE_FLUSH
;
6592 pc
.StateCacheInvalidationEnable
=
6593 flags
& PIPE_CONTROL_STATE_CACHE_INVALIDATE
;
6594 pc
.VFCacheInvalidationEnable
= flags
& PIPE_CONTROL_VF_CACHE_INVALIDATE
;
6595 pc
.ConstantCacheInvalidationEnable
=
6596 flags
& PIPE_CONTROL_CONST_CACHE_INVALIDATE
;
6597 pc
.PostSyncOperation
= flags_to_post_sync_op(flags
);
6598 pc
.DepthStallEnable
= flags
& PIPE_CONTROL_DEPTH_STALL
;
6599 pc
.InstructionCacheInvalidateEnable
=
6600 flags
& PIPE_CONTROL_INSTRUCTION_INVALIDATE
;
6601 pc
.NotifyEnable
= flags
& PIPE_CONTROL_NOTIFY_ENABLE
;
6602 pc
.IndirectStatePointersDisable
=
6603 flags
& PIPE_CONTROL_INDIRECT_STATE_POINTERS_DISABLE
;
6604 pc
.TextureCacheInvalidationEnable
=
6605 flags
& PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
6606 pc
.Address
= rw_bo(bo
, offset
);
6607 pc
.ImmediateData
= imm
;
6612 genX(emit_urb_setup
)(struct iris_context
*ice
,
6613 struct iris_batch
*batch
,
6614 const unsigned size
[4],
6615 bool tess_present
, bool gs_present
)
6617 const struct gen_device_info
*devinfo
= &batch
->screen
->devinfo
;
6618 const unsigned push_size_kB
= 32;
6619 unsigned entries
[4];
6622 ice
->shaders
.last_vs_entry_size
= size
[MESA_SHADER_VERTEX
];
6624 gen_get_urb_config(devinfo
, 1024 * push_size_kB
,
6625 1024 * ice
->shaders
.urb_size
,
6626 tess_present
, gs_present
,
6627 size
, entries
, start
);
6629 for (int i
= MESA_SHADER_VERTEX
; i
<= MESA_SHADER_GEOMETRY
; i
++) {
6630 iris_emit_cmd(batch
, GENX(3DSTATE_URB_VS
), urb
) {
6631 urb
._3DCommandSubOpcode
+= i
;
6632 urb
.VSURBStartingAddress
= start
[i
];
6633 urb
.VSURBEntryAllocationSize
= size
[i
] - 1;
6634 urb
.VSNumberofURBEntries
= entries
[i
];
6641 * Preemption on Gen9 has to be enabled or disabled in various cases.
6643 * See these workarounds for preemption:
6644 * - WaDisableMidObjectPreemptionForGSLineStripAdj
6645 * - WaDisableMidObjectPreemptionForTrifanOrPolygon
6646 * - WaDisableMidObjectPreemptionForLineLoop
6649 * We don't put this in the vtable because it's only used on Gen9.
6652 gen9_toggle_preemption(struct iris_context
*ice
,
6653 struct iris_batch
*batch
,
6654 const struct pipe_draw_info
*draw
)
6656 struct iris_genx_state
*genx
= ice
->state
.genx
;
6657 bool object_preemption
= true;
6659 /* WaDisableMidObjectPreemptionForGSLineStripAdj
6661 * "WA: Disable mid-draw preemption when draw-call is a linestrip_adj
6662 * and GS is enabled."
6664 if (draw
->mode
== PIPE_PRIM_LINE_STRIP_ADJACENCY
&&
6665 ice
->shaders
.prog
[MESA_SHADER_GEOMETRY
])
6666 object_preemption
= false;
6668 /* WaDisableMidObjectPreemptionForTrifanOrPolygon
6670 * "TriFan miscompare in Execlist Preemption test. Cut index that is
6671 * on a previous context. End the previous, the resume another context
6672 * with a tri-fan or polygon, and the vertex count is corrupted. If we
6673 * prempt again we will cause corruption.
6675 * WA: Disable mid-draw preemption when draw-call has a tri-fan."
6677 if (draw
->mode
== PIPE_PRIM_TRIANGLE_FAN
)
6678 object_preemption
= false;
6680 /* WaDisableMidObjectPreemptionForLineLoop
6682 * "VF Stats Counters Missing a vertex when preemption enabled.
6684 * WA: Disable mid-draw preemption when the draw uses a lineloop
6687 if (draw
->mode
== PIPE_PRIM_LINE_LOOP
)
6688 object_preemption
= false;
6692 * "VF is corrupting GAFS data when preempted on an instance boundary
6693 * and replayed with instancing enabled.
6695 * WA: Disable preemption when using instanceing."
6697 if (draw
->instance_count
> 1)
6698 object_preemption
= false;
6700 if (genx
->object_preemption
!= object_preemption
) {
6701 iris_enable_obj_preemption(batch
, object_preemption
);
6702 genx
->object_preemption
= object_preemption
;
6708 iris_lost_genx_state(struct iris_context
*ice
, struct iris_batch
*batch
)
6710 struct iris_genx_state
*genx
= ice
->state
.genx
;
6712 memset(genx
->last_index_buffer
, 0, sizeof(genx
->last_index_buffer
));
6716 iris_emit_mi_report_perf_count(struct iris_batch
*batch
,
6718 uint32_t offset_in_bytes
,
6721 iris_emit_cmd(batch
, GENX(MI_REPORT_PERF_COUNT
), mi_rpc
) {
6722 mi_rpc
.MemoryAddress
= rw_bo(bo
, offset_in_bytes
);
6723 mi_rpc
.ReportID
= report_id
;
6728 * Update the pixel hashing modes that determine the balancing of PS threads
6729 * across subslices and slices.
6731 * \param width Width bound of the rendering area (already scaled down if \p
6732 * scale is greater than 1).
6733 * \param height Height bound of the rendering area (already scaled down if \p
6734 * scale is greater than 1).
6735 * \param scale The number of framebuffer samples that could potentially be
6736 * affected by an individual channel of the PS thread. This is
6737 * typically one for single-sampled rendering, but for operations
6738 * like CCS resolves and fast clears a single PS invocation may
6739 * update a huge number of pixels, in which case a finer
6740 * balancing is desirable in order to maximally utilize the
6741 * bandwidth available. UINT_MAX can be used as shorthand for
6742 * "finest hashing mode available".
6745 genX(emit_hashing_mode
)(struct iris_context
*ice
, struct iris_batch
*batch
,
6746 unsigned width
, unsigned height
, unsigned scale
)
6749 const struct gen_device_info
*devinfo
= &batch
->screen
->devinfo
;
6750 const unsigned slice_hashing
[] = {
6751 /* Because all Gen9 platforms with more than one slice require
6752 * three-way subslice hashing, a single "normal" 16x16 slice hashing
6753 * block is guaranteed to suffer from substantial imbalance, with one
6754 * subslice receiving twice as much work as the other two in the
6757 * The performance impact of that would be particularly severe when
6758 * three-way hashing is also in use for slice balancing (which is the
6759 * case for all Gen9 GT4 platforms), because one of the slices
6760 * receives one every three 16x16 blocks in either direction, which
6761 * is roughly the periodicity of the underlying subslice imbalance
6762 * pattern ("roughly" because in reality the hardware's
6763 * implementation of three-way hashing doesn't do exact modulo 3
6764 * arithmetic, which somewhat decreases the magnitude of this effect
6765 * in practice). This leads to a systematic subslice imbalance
6766 * within that slice regardless of the size of the primitive. The
6767 * 32x32 hashing mode guarantees that the subslice imbalance within a
6768 * single slice hashing block is minimal, largely eliminating this
6772 /* Finest slice hashing mode available. */
6775 const unsigned subslice_hashing
[] = {
6776 /* 16x16 would provide a slight cache locality benefit especially
6777 * visible in the sampler L1 cache efficiency of low-bandwidth
6778 * non-LLC platforms, but it comes at the cost of greater subslice
6779 * imbalance for primitives of dimensions approximately intermediate
6780 * between 16x4 and 16x16.
6783 /* Finest subslice hashing mode available. */
6786 /* Dimensions of the smallest hashing block of a given hashing mode. If
6787 * the rendering area is smaller than this there can't possibly be any
6788 * benefit from switching to this mode, so we optimize out the
6791 const unsigned min_size
[][2] = {
6795 const unsigned idx
= scale
> 1;
6797 if (width
> min_size
[idx
][0] || height
> min_size
[idx
][1]) {
6800 iris_pack_state(GENX(GT_MODE
), >_mode
, reg
) {
6801 reg
.SliceHashing
= (devinfo
->num_slices
> 1 ? slice_hashing
[idx
] : 0);
6802 reg
.SliceHashingMask
= (devinfo
->num_slices
> 1 ? -1 : 0);
6803 reg
.SubsliceHashing
= subslice_hashing
[idx
];
6804 reg
.SubsliceHashingMask
= -1;
6807 iris_emit_raw_pipe_control(batch
,
6808 "workaround: CS stall before GT_MODE LRI",
6809 PIPE_CONTROL_STALL_AT_SCOREBOARD
|
6810 PIPE_CONTROL_CS_STALL
,
6813 iris_emit_lri(batch
, GT_MODE
, gt_mode
);
6815 ice
->state
.current_hash_scale
= scale
;
6821 genX(init_state
)(struct iris_context
*ice
)
6823 struct pipe_context
*ctx
= &ice
->ctx
;
6824 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
6826 ctx
->create_blend_state
= iris_create_blend_state
;
6827 ctx
->create_depth_stencil_alpha_state
= iris_create_zsa_state
;
6828 ctx
->create_rasterizer_state
= iris_create_rasterizer_state
;
6829 ctx
->create_sampler_state
= iris_create_sampler_state
;
6830 ctx
->create_sampler_view
= iris_create_sampler_view
;
6831 ctx
->create_surface
= iris_create_surface
;
6832 ctx
->create_vertex_elements_state
= iris_create_vertex_elements
;
6833 ctx
->bind_blend_state
= iris_bind_blend_state
;
6834 ctx
->bind_depth_stencil_alpha_state
= iris_bind_zsa_state
;
6835 ctx
->bind_sampler_states
= iris_bind_sampler_states
;
6836 ctx
->bind_rasterizer_state
= iris_bind_rasterizer_state
;
6837 ctx
->bind_vertex_elements_state
= iris_bind_vertex_elements_state
;
6838 ctx
->delete_blend_state
= iris_delete_state
;
6839 ctx
->delete_depth_stencil_alpha_state
= iris_delete_state
;
6840 ctx
->delete_rasterizer_state
= iris_delete_state
;
6841 ctx
->delete_sampler_state
= iris_delete_state
;
6842 ctx
->delete_vertex_elements_state
= iris_delete_state
;
6843 ctx
->set_blend_color
= iris_set_blend_color
;
6844 ctx
->set_clip_state
= iris_set_clip_state
;
6845 ctx
->set_constant_buffer
= iris_set_constant_buffer
;
6846 ctx
->set_shader_buffers
= iris_set_shader_buffers
;
6847 ctx
->set_shader_images
= iris_set_shader_images
;
6848 ctx
->set_sampler_views
= iris_set_sampler_views
;
6849 ctx
->set_tess_state
= iris_set_tess_state
;
6850 ctx
->set_framebuffer_state
= iris_set_framebuffer_state
;
6851 ctx
->set_polygon_stipple
= iris_set_polygon_stipple
;
6852 ctx
->set_sample_mask
= iris_set_sample_mask
;
6853 ctx
->set_scissor_states
= iris_set_scissor_states
;
6854 ctx
->set_stencil_ref
= iris_set_stencil_ref
;
6855 ctx
->set_vertex_buffers
= iris_set_vertex_buffers
;
6856 ctx
->set_viewport_states
= iris_set_viewport_states
;
6857 ctx
->sampler_view_destroy
= iris_sampler_view_destroy
;
6858 ctx
->surface_destroy
= iris_surface_destroy
;
6859 ctx
->draw_vbo
= iris_draw_vbo
;
6860 ctx
->launch_grid
= iris_launch_grid
;
6861 ctx
->create_stream_output_target
= iris_create_stream_output_target
;
6862 ctx
->stream_output_target_destroy
= iris_stream_output_target_destroy
;
6863 ctx
->set_stream_output_targets
= iris_set_stream_output_targets
;
6865 ice
->vtbl
.destroy_state
= iris_destroy_state
;
6866 ice
->vtbl
.init_render_context
= iris_init_render_context
;
6867 ice
->vtbl
.init_compute_context
= iris_init_compute_context
;
6868 ice
->vtbl
.upload_render_state
= iris_upload_render_state
;
6869 ice
->vtbl
.update_surface_base_address
= iris_update_surface_base_address
;
6870 ice
->vtbl
.upload_compute_state
= iris_upload_compute_state
;
6871 ice
->vtbl
.emit_raw_pipe_control
= iris_emit_raw_pipe_control
;
6872 ice
->vtbl
.emit_mi_report_perf_count
= iris_emit_mi_report_perf_count
;
6873 ice
->vtbl
.rebind_buffer
= iris_rebind_buffer
;
6874 ice
->vtbl
.load_register_reg32
= iris_load_register_reg32
;
6875 ice
->vtbl
.load_register_reg64
= iris_load_register_reg64
;
6876 ice
->vtbl
.load_register_imm32
= iris_load_register_imm32
;
6877 ice
->vtbl
.load_register_imm64
= iris_load_register_imm64
;
6878 ice
->vtbl
.load_register_mem32
= iris_load_register_mem32
;
6879 ice
->vtbl
.load_register_mem64
= iris_load_register_mem64
;
6880 ice
->vtbl
.store_register_mem32
= iris_store_register_mem32
;
6881 ice
->vtbl
.store_register_mem64
= iris_store_register_mem64
;
6882 ice
->vtbl
.store_data_imm32
= iris_store_data_imm32
;
6883 ice
->vtbl
.store_data_imm64
= iris_store_data_imm64
;
6884 ice
->vtbl
.copy_mem_mem
= iris_copy_mem_mem
;
6885 ice
->vtbl
.derived_program_state_size
= iris_derived_program_state_size
;
6886 ice
->vtbl
.store_derived_program_state
= iris_store_derived_program_state
;
6887 ice
->vtbl
.create_so_decl_list
= iris_create_so_decl_list
;
6888 ice
->vtbl
.populate_vs_key
= iris_populate_vs_key
;
6889 ice
->vtbl
.populate_tcs_key
= iris_populate_tcs_key
;
6890 ice
->vtbl
.populate_tes_key
= iris_populate_tes_key
;
6891 ice
->vtbl
.populate_gs_key
= iris_populate_gs_key
;
6892 ice
->vtbl
.populate_fs_key
= iris_populate_fs_key
;
6893 ice
->vtbl
.populate_cs_key
= iris_populate_cs_key
;
6894 ice
->vtbl
.mocs
= mocs
;
6895 ice
->vtbl
.lost_genx_state
= iris_lost_genx_state
;
6897 ice
->state
.dirty
= ~0ull;
6899 ice
->state
.statistics_counters_enabled
= true;
6901 ice
->state
.sample_mask
= 0xffff;
6902 ice
->state
.num_viewports
= 1;
6903 ice
->state
.genx
= calloc(1, sizeof(struct iris_genx_state
));
6905 /* Make a 1x1x1 null surface for unbound textures */
6906 void *null_surf_map
=
6907 upload_state(ice
->state
.surface_uploader
, &ice
->state
.unbound_tex
,
6908 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
6909 isl_null_fill_state(&screen
->isl_dev
, null_surf_map
, isl_extent3d(1, 1, 1));
6910 ice
->state
.unbound_tex
.offset
+=
6911 iris_bo_offset_from_base_address(iris_resource_bo(ice
->state
.unbound_tex
.res
));
6913 /* Default all scissor rectangles to be empty regions. */
6914 for (int i
= 0; i
< IRIS_MAX_VIEWPORTS
; i
++) {
6915 ice
->state
.scissors
[i
] = (struct pipe_scissor_state
) {
6916 .minx
= 1, .maxx
= 0, .miny
= 1, .maxy
= 0,