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_inlines.h"
92 #include "util/u_format.h"
93 #include "util/u_framebuffer.h"
94 #include "util/u_transfer.h"
95 #include "util/u_upload_mgr.h"
96 #include "util/u_viewport.h"
99 #include "intel/compiler/brw_compiler.h"
100 #include "intel/common/gen_l3_config.h"
101 #include "intel/common/gen_sample_positions.h"
102 #include "iris_batch.h"
103 #include "iris_context.h"
104 #include "iris_defines.h"
105 #include "iris_pipe.h"
106 #include "iris_resource.h"
108 #define __gen_address_type struct iris_address
109 #define __gen_user_data struct iris_batch
111 #define ARRAY_BYTES(x) (sizeof(uint32_t) * ARRAY_SIZE(x))
114 __gen_combine_address(struct iris_batch
*batch
, void *location
,
115 struct iris_address addr
, uint32_t delta
)
117 uint64_t result
= addr
.offset
+ delta
;
120 iris_use_pinned_bo(batch
, addr
.bo
, addr
.write
);
121 /* Assume this is a general address, not relative to a base. */
122 result
+= addr
.bo
->gtt_offset
;
128 #define __genxml_cmd_length(cmd) cmd ## _length
129 #define __genxml_cmd_length_bias(cmd) cmd ## _length_bias
130 #define __genxml_cmd_header(cmd) cmd ## _header
131 #define __genxml_cmd_pack(cmd) cmd ## _pack
133 #define _iris_pack_command(batch, cmd, dst, name) \
134 for (struct cmd name = { __genxml_cmd_header(cmd) }, \
135 *_dst = (void *)(dst); __builtin_expect(_dst != NULL, 1); \
136 ({ __genxml_cmd_pack(cmd)(batch, (void *)_dst, &name); \
140 #define iris_pack_command(cmd, dst, name) \
141 _iris_pack_command(NULL, cmd, dst, name)
143 #define iris_pack_state(cmd, dst, name) \
144 for (struct cmd name = {}, \
145 *_dst = (void *)(dst); __builtin_expect(_dst != NULL, 1); \
146 __genxml_cmd_pack(cmd)(NULL, (void *)_dst, &name), \
149 #define iris_emit_cmd(batch, cmd, name) \
150 _iris_pack_command(batch, cmd, iris_get_command_space(batch, 4 * __genxml_cmd_length(cmd)), name)
152 #define iris_emit_merge(batch, dwords0, dwords1, num_dwords) \
154 uint32_t *dw = iris_get_command_space(batch, 4 * num_dwords); \
155 for (uint32_t i = 0; i < num_dwords; i++) \
156 dw[i] = (dwords0)[i] | (dwords1)[i]; \
157 VG(VALGRIND_CHECK_MEM_IS_DEFINED(dw, num_dwords)); \
160 #include "genxml/genX_pack.h"
161 #include "genxml/gen_macros.h"
162 #include "genxml/genX_bits.h"
165 #define MOCS_PTE 0x18
168 #define MOCS_PTE (1 << 1)
169 #define MOCS_WB (2 << 1)
173 mocs(struct iris_bo
*bo
)
175 return bo
&& bo
->external
? MOCS_PTE
: MOCS_WB
;
179 * Statically assert that PIPE_* enums match the hardware packets.
180 * (As long as they match, we don't need to translate them.)
182 UNUSED
static void pipe_asserts()
184 #define PIPE_ASSERT(x) STATIC_ASSERT((int)x)
186 /* pipe_logicop happens to match the hardware. */
187 PIPE_ASSERT(PIPE_LOGICOP_CLEAR
== LOGICOP_CLEAR
);
188 PIPE_ASSERT(PIPE_LOGICOP_NOR
== LOGICOP_NOR
);
189 PIPE_ASSERT(PIPE_LOGICOP_AND_INVERTED
== LOGICOP_AND_INVERTED
);
190 PIPE_ASSERT(PIPE_LOGICOP_COPY_INVERTED
== LOGICOP_COPY_INVERTED
);
191 PIPE_ASSERT(PIPE_LOGICOP_AND_REVERSE
== LOGICOP_AND_REVERSE
);
192 PIPE_ASSERT(PIPE_LOGICOP_INVERT
== LOGICOP_INVERT
);
193 PIPE_ASSERT(PIPE_LOGICOP_XOR
== LOGICOP_XOR
);
194 PIPE_ASSERT(PIPE_LOGICOP_NAND
== LOGICOP_NAND
);
195 PIPE_ASSERT(PIPE_LOGICOP_AND
== LOGICOP_AND
);
196 PIPE_ASSERT(PIPE_LOGICOP_EQUIV
== LOGICOP_EQUIV
);
197 PIPE_ASSERT(PIPE_LOGICOP_NOOP
== LOGICOP_NOOP
);
198 PIPE_ASSERT(PIPE_LOGICOP_OR_INVERTED
== LOGICOP_OR_INVERTED
);
199 PIPE_ASSERT(PIPE_LOGICOP_COPY
== LOGICOP_COPY
);
200 PIPE_ASSERT(PIPE_LOGICOP_OR_REVERSE
== LOGICOP_OR_REVERSE
);
201 PIPE_ASSERT(PIPE_LOGICOP_OR
== LOGICOP_OR
);
202 PIPE_ASSERT(PIPE_LOGICOP_SET
== LOGICOP_SET
);
204 /* pipe_blend_func happens to match the hardware. */
205 PIPE_ASSERT(PIPE_BLENDFACTOR_ONE
== BLENDFACTOR_ONE
);
206 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_COLOR
== BLENDFACTOR_SRC_COLOR
);
207 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_ALPHA
== BLENDFACTOR_SRC_ALPHA
);
208 PIPE_ASSERT(PIPE_BLENDFACTOR_DST_ALPHA
== BLENDFACTOR_DST_ALPHA
);
209 PIPE_ASSERT(PIPE_BLENDFACTOR_DST_COLOR
== BLENDFACTOR_DST_COLOR
);
210 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
== BLENDFACTOR_SRC_ALPHA_SATURATE
);
211 PIPE_ASSERT(PIPE_BLENDFACTOR_CONST_COLOR
== BLENDFACTOR_CONST_COLOR
);
212 PIPE_ASSERT(PIPE_BLENDFACTOR_CONST_ALPHA
== BLENDFACTOR_CONST_ALPHA
);
213 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC1_COLOR
== BLENDFACTOR_SRC1_COLOR
);
214 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC1_ALPHA
== BLENDFACTOR_SRC1_ALPHA
);
215 PIPE_ASSERT(PIPE_BLENDFACTOR_ZERO
== BLENDFACTOR_ZERO
);
216 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC_COLOR
== BLENDFACTOR_INV_SRC_COLOR
);
217 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC_ALPHA
== BLENDFACTOR_INV_SRC_ALPHA
);
218 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_DST_ALPHA
== BLENDFACTOR_INV_DST_ALPHA
);
219 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_DST_COLOR
== BLENDFACTOR_INV_DST_COLOR
);
220 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_CONST_COLOR
== BLENDFACTOR_INV_CONST_COLOR
);
221 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_CONST_ALPHA
== BLENDFACTOR_INV_CONST_ALPHA
);
222 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC1_COLOR
== BLENDFACTOR_INV_SRC1_COLOR
);
223 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC1_ALPHA
== BLENDFACTOR_INV_SRC1_ALPHA
);
225 /* pipe_blend_func happens to match the hardware. */
226 PIPE_ASSERT(PIPE_BLEND_ADD
== BLENDFUNCTION_ADD
);
227 PIPE_ASSERT(PIPE_BLEND_SUBTRACT
== BLENDFUNCTION_SUBTRACT
);
228 PIPE_ASSERT(PIPE_BLEND_REVERSE_SUBTRACT
== BLENDFUNCTION_REVERSE_SUBTRACT
);
229 PIPE_ASSERT(PIPE_BLEND_MIN
== BLENDFUNCTION_MIN
);
230 PIPE_ASSERT(PIPE_BLEND_MAX
== BLENDFUNCTION_MAX
);
232 /* pipe_stencil_op happens to match the hardware. */
233 PIPE_ASSERT(PIPE_STENCIL_OP_KEEP
== STENCILOP_KEEP
);
234 PIPE_ASSERT(PIPE_STENCIL_OP_ZERO
== STENCILOP_ZERO
);
235 PIPE_ASSERT(PIPE_STENCIL_OP_REPLACE
== STENCILOP_REPLACE
);
236 PIPE_ASSERT(PIPE_STENCIL_OP_INCR
== STENCILOP_INCRSAT
);
237 PIPE_ASSERT(PIPE_STENCIL_OP_DECR
== STENCILOP_DECRSAT
);
238 PIPE_ASSERT(PIPE_STENCIL_OP_INCR_WRAP
== STENCILOP_INCR
);
239 PIPE_ASSERT(PIPE_STENCIL_OP_DECR_WRAP
== STENCILOP_DECR
);
240 PIPE_ASSERT(PIPE_STENCIL_OP_INVERT
== STENCILOP_INVERT
);
242 /* pipe_sprite_coord_mode happens to match 3DSTATE_SBE */
243 PIPE_ASSERT(PIPE_SPRITE_COORD_UPPER_LEFT
== UPPERLEFT
);
244 PIPE_ASSERT(PIPE_SPRITE_COORD_LOWER_LEFT
== LOWERLEFT
);
249 translate_prim_type(enum pipe_prim_type prim
, uint8_t verts_per_patch
)
251 static const unsigned map
[] = {
252 [PIPE_PRIM_POINTS
] = _3DPRIM_POINTLIST
,
253 [PIPE_PRIM_LINES
] = _3DPRIM_LINELIST
,
254 [PIPE_PRIM_LINE_LOOP
] = _3DPRIM_LINELOOP
,
255 [PIPE_PRIM_LINE_STRIP
] = _3DPRIM_LINESTRIP
,
256 [PIPE_PRIM_TRIANGLES
] = _3DPRIM_TRILIST
,
257 [PIPE_PRIM_TRIANGLE_STRIP
] = _3DPRIM_TRISTRIP
,
258 [PIPE_PRIM_TRIANGLE_FAN
] = _3DPRIM_TRIFAN
,
259 [PIPE_PRIM_QUADS
] = _3DPRIM_QUADLIST
,
260 [PIPE_PRIM_QUAD_STRIP
] = _3DPRIM_QUADSTRIP
,
261 [PIPE_PRIM_POLYGON
] = _3DPRIM_POLYGON
,
262 [PIPE_PRIM_LINES_ADJACENCY
] = _3DPRIM_LINELIST_ADJ
,
263 [PIPE_PRIM_LINE_STRIP_ADJACENCY
] = _3DPRIM_LINESTRIP_ADJ
,
264 [PIPE_PRIM_TRIANGLES_ADJACENCY
] = _3DPRIM_TRILIST_ADJ
,
265 [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
] = _3DPRIM_TRISTRIP_ADJ
,
266 [PIPE_PRIM_PATCHES
] = _3DPRIM_PATCHLIST_1
- 1,
269 return map
[prim
] + (prim
== PIPE_PRIM_PATCHES
? verts_per_patch
: 0);
273 translate_compare_func(enum pipe_compare_func pipe_func
)
275 static const unsigned map
[] = {
276 [PIPE_FUNC_NEVER
] = COMPAREFUNCTION_NEVER
,
277 [PIPE_FUNC_LESS
] = COMPAREFUNCTION_LESS
,
278 [PIPE_FUNC_EQUAL
] = COMPAREFUNCTION_EQUAL
,
279 [PIPE_FUNC_LEQUAL
] = COMPAREFUNCTION_LEQUAL
,
280 [PIPE_FUNC_GREATER
] = COMPAREFUNCTION_GREATER
,
281 [PIPE_FUNC_NOTEQUAL
] = COMPAREFUNCTION_NOTEQUAL
,
282 [PIPE_FUNC_GEQUAL
] = COMPAREFUNCTION_GEQUAL
,
283 [PIPE_FUNC_ALWAYS
] = COMPAREFUNCTION_ALWAYS
,
285 return map
[pipe_func
];
289 translate_shadow_func(enum pipe_compare_func pipe_func
)
291 /* Gallium specifies the result of shadow comparisons as:
293 * 1 if ref <op> texel,
298 * 0 if texel <op> ref,
301 * So we need to flip the operator and also negate.
303 static const unsigned map
[] = {
304 [PIPE_FUNC_NEVER
] = PREFILTEROPALWAYS
,
305 [PIPE_FUNC_LESS
] = PREFILTEROPLEQUAL
,
306 [PIPE_FUNC_EQUAL
] = PREFILTEROPNOTEQUAL
,
307 [PIPE_FUNC_LEQUAL
] = PREFILTEROPLESS
,
308 [PIPE_FUNC_GREATER
] = PREFILTEROPGEQUAL
,
309 [PIPE_FUNC_NOTEQUAL
] = PREFILTEROPEQUAL
,
310 [PIPE_FUNC_GEQUAL
] = PREFILTEROPGREATER
,
311 [PIPE_FUNC_ALWAYS
] = PREFILTEROPNEVER
,
313 return map
[pipe_func
];
317 translate_cull_mode(unsigned pipe_face
)
319 static const unsigned map
[4] = {
320 [PIPE_FACE_NONE
] = CULLMODE_NONE
,
321 [PIPE_FACE_FRONT
] = CULLMODE_FRONT
,
322 [PIPE_FACE_BACK
] = CULLMODE_BACK
,
323 [PIPE_FACE_FRONT_AND_BACK
] = CULLMODE_BOTH
,
325 return map
[pipe_face
];
329 translate_fill_mode(unsigned pipe_polymode
)
331 static const unsigned map
[4] = {
332 [PIPE_POLYGON_MODE_FILL
] = FILL_MODE_SOLID
,
333 [PIPE_POLYGON_MODE_LINE
] = FILL_MODE_WIREFRAME
,
334 [PIPE_POLYGON_MODE_POINT
] = FILL_MODE_POINT
,
335 [PIPE_POLYGON_MODE_FILL_RECTANGLE
] = FILL_MODE_SOLID
,
337 return map
[pipe_polymode
];
341 translate_mip_filter(enum pipe_tex_mipfilter pipe_mip
)
343 static const unsigned map
[] = {
344 [PIPE_TEX_MIPFILTER_NEAREST
] = MIPFILTER_NEAREST
,
345 [PIPE_TEX_MIPFILTER_LINEAR
] = MIPFILTER_LINEAR
,
346 [PIPE_TEX_MIPFILTER_NONE
] = MIPFILTER_NONE
,
348 return map
[pipe_mip
];
352 translate_wrap(unsigned pipe_wrap
)
354 static const unsigned map
[] = {
355 [PIPE_TEX_WRAP_REPEAT
] = TCM_WRAP
,
356 [PIPE_TEX_WRAP_CLAMP
] = TCM_HALF_BORDER
,
357 [PIPE_TEX_WRAP_CLAMP_TO_EDGE
] = TCM_CLAMP
,
358 [PIPE_TEX_WRAP_CLAMP_TO_BORDER
] = TCM_CLAMP_BORDER
,
359 [PIPE_TEX_WRAP_MIRROR_REPEAT
] = TCM_MIRROR
,
360 [PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
] = TCM_MIRROR_ONCE
,
362 /* These are unsupported. */
363 [PIPE_TEX_WRAP_MIRROR_CLAMP
] = -1,
364 [PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
] = -1,
366 return map
[pipe_wrap
];
369 static struct iris_address
370 ro_bo(struct iris_bo
*bo
, uint64_t offset
)
372 /* CSOs must pass NULL for bo! Otherwise it will add the BO to the
373 * validation list at CSO creation time, instead of draw time.
375 return (struct iris_address
) { .bo
= bo
, .offset
= offset
};
378 static struct iris_address
379 rw_bo(struct iris_bo
*bo
, uint64_t offset
)
381 /* CSOs must pass NULL for bo! Otherwise it will add the BO to the
382 * validation list at CSO creation time, instead of draw time.
384 return (struct iris_address
) { .bo
= bo
, .offset
= offset
, .write
= true };
388 * Allocate space for some indirect state.
390 * Return a pointer to the map (to fill it out) and a state ref (for
391 * referring to the state in GPU commands).
394 upload_state(struct u_upload_mgr
*uploader
,
395 struct iris_state_ref
*ref
,
400 u_upload_alloc(uploader
, 0, size
, alignment
, &ref
->offset
, &ref
->res
, &p
);
405 * Stream out temporary/short-lived state.
407 * This allocates space, pins the BO, and includes the BO address in the
408 * returned offset (which works because all state lives in 32-bit memory
412 stream_state(struct iris_batch
*batch
,
413 struct u_upload_mgr
*uploader
,
414 struct pipe_resource
**out_res
,
417 uint32_t *out_offset
)
421 u_upload_alloc(uploader
, 0, size
, alignment
, out_offset
, out_res
, &ptr
);
423 struct iris_bo
*bo
= iris_resource_bo(*out_res
);
424 iris_use_pinned_bo(batch
, bo
, false);
426 *out_offset
+= iris_bo_offset_from_base_address(bo
);
432 * stream_state() + memcpy.
435 emit_state(struct iris_batch
*batch
,
436 struct u_upload_mgr
*uploader
,
437 struct pipe_resource
**out_res
,
444 stream_state(batch
, uploader
, out_res
, size
, alignment
, &offset
);
447 memcpy(map
, data
, size
);
453 * Did field 'x' change between 'old_cso' and 'new_cso'?
455 * (If so, we may want to set some dirty flags.)
457 #define cso_changed(x) (!old_cso || (old_cso->x != new_cso->x))
458 #define cso_changed_memcmp(x) \
459 (!old_cso || memcmp(old_cso->x, new_cso->x, sizeof(old_cso->x)) != 0)
462 flush_for_state_base_change(struct iris_batch
*batch
)
464 /* Flush before emitting STATE_BASE_ADDRESS.
466 * This isn't documented anywhere in the PRM. However, it seems to be
467 * necessary prior to changing the surface state base adress. We've
468 * seen issues in Vulkan where we get GPU hangs when using multi-level
469 * command buffers which clear depth, reset state base address, and then
472 * Normally, in GL, we would trust the kernel to do sufficient stalls
473 * and flushes prior to executing our batch. However, it doesn't seem
474 * as if the kernel's flushing is always sufficient and we don't want to
477 * We make this an end-of-pipe sync instead of a normal flush because we
478 * do not know the current status of the GPU. On Haswell at least,
479 * having a fast-clear operation in flight at the same time as a normal
480 * rendering operation can cause hangs. Since the kernel's flushing is
481 * insufficient, we need to ensure that any rendering operations from
482 * other processes are definitely complete before we try to do our own
483 * rendering. It's a bit of a big hammer but it appears to work.
485 iris_emit_end_of_pipe_sync(batch
,
486 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
487 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
488 PIPE_CONTROL_DATA_CACHE_FLUSH
);
492 _iris_emit_lri(struct iris_batch
*batch
, uint32_t reg
, uint32_t val
)
494 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_IMM
), lri
) {
495 lri
.RegisterOffset
= reg
;
499 #define iris_emit_lri(b, r, v) _iris_emit_lri(b, GENX(r##_num), v)
502 _iris_emit_lrr(struct iris_batch
*batch
, uint32_t dst
, uint32_t src
)
504 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_REG
), lrr
) {
505 lrr
.SourceRegisterAddress
= src
;
506 lrr
.DestinationRegisterAddress
= dst
;
511 emit_pipeline_select(struct iris_batch
*batch
, uint32_t pipeline
)
513 #if GEN_GEN >= 8 && GEN_GEN < 10
514 /* From the Broadwell PRM, Volume 2a: Instructions, PIPELINE_SELECT:
516 * Software must clear the COLOR_CALC_STATE Valid field in
517 * 3DSTATE_CC_STATE_POINTERS command prior to send a PIPELINE_SELECT
518 * with Pipeline Select set to GPGPU.
520 * The internal hardware docs recommend the same workaround for Gen9
523 if (pipeline
== GPGPU
)
524 iris_emit_cmd(batch
, GENX(3DSTATE_CC_STATE_POINTERS
), t
);
528 /* From "BXML » GT » MI » vol1a GPU Overview » [Instruction]
529 * PIPELINE_SELECT [DevBWR+]":
533 * Software must ensure all the write caches are flushed through a
534 * stalling PIPE_CONTROL command followed by another PIPE_CONTROL
535 * command to invalidate read only caches prior to programming
536 * MI_PIPELINE_SELECT command to change the Pipeline Select Mode."
538 iris_emit_pipe_control_flush(batch
,
539 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
540 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
541 PIPE_CONTROL_DATA_CACHE_FLUSH
|
542 PIPE_CONTROL_CS_STALL
);
544 iris_emit_pipe_control_flush(batch
,
545 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
546 PIPE_CONTROL_CONST_CACHE_INVALIDATE
|
547 PIPE_CONTROL_STATE_CACHE_INVALIDATE
|
548 PIPE_CONTROL_INSTRUCTION_INVALIDATE
);
550 iris_emit_cmd(batch
, GENX(PIPELINE_SELECT
), sel
) {
554 sel
.PipelineSelection
= pipeline
;
559 init_glk_barrier_mode(struct iris_batch
*batch
, uint32_t value
)
564 * "This chicken bit works around a hardware issue with barrier
565 * logic encountered when switching between GPGPU and 3D pipelines.
566 * To workaround the issue, this mode bit should be set after a
567 * pipeline is selected."
570 iris_pack_state(GENX(SLICE_COMMON_ECO_CHICKEN1
), ®_val
, reg
) {
571 reg
.GLKBarrierMode
= value
;
572 reg
.GLKBarrierModeMask
= 1;
574 iris_emit_lri(batch
, SLICE_COMMON_ECO_CHICKEN1
, reg_val
);
579 init_state_base_address(struct iris_batch
*batch
)
581 flush_for_state_base_change(batch
);
583 /* We program most base addresses once at context initialization time.
584 * Each base address points at a 4GB memory zone, and never needs to
585 * change. See iris_bufmgr.h for a description of the memory zones.
587 * The one exception is Surface State Base Address, which needs to be
588 * updated occasionally. See iris_binder.c for the details there.
590 iris_emit_cmd(batch
, GENX(STATE_BASE_ADDRESS
), sba
) {
591 sba
.GeneralStateMOCS
= MOCS_WB
;
592 sba
.StatelessDataPortAccessMOCS
= MOCS_WB
;
593 sba
.DynamicStateMOCS
= MOCS_WB
;
594 sba
.IndirectObjectMOCS
= MOCS_WB
;
595 sba
.InstructionMOCS
= MOCS_WB
;
597 sba
.GeneralStateBaseAddressModifyEnable
= true;
598 sba
.DynamicStateBaseAddressModifyEnable
= true;
599 sba
.IndirectObjectBaseAddressModifyEnable
= true;
600 sba
.InstructionBaseAddressModifyEnable
= true;
601 sba
.GeneralStateBufferSizeModifyEnable
= true;
602 sba
.DynamicStateBufferSizeModifyEnable
= true;
604 sba
.BindlessSurfaceStateBaseAddressModifyEnable
= true;
605 sba
.BindlessSurfaceStateMOCS
= MOCS_WB
;
607 sba
.IndirectObjectBufferSizeModifyEnable
= true;
608 sba
.InstructionBuffersizeModifyEnable
= true;
610 sba
.InstructionBaseAddress
= ro_bo(NULL
, IRIS_MEMZONE_SHADER_START
);
611 sba
.DynamicStateBaseAddress
= ro_bo(NULL
, IRIS_MEMZONE_DYNAMIC_START
);
613 sba
.GeneralStateBufferSize
= 0xfffff;
614 sba
.IndirectObjectBufferSize
= 0xfffff;
615 sba
.InstructionBufferSize
= 0xfffff;
616 sba
.DynamicStateBufferSize
= 0xfffff;
621 * Upload the initial GPU state for a render context.
623 * This sets some invariant state that needs to be programmed a particular
624 * way, but we never actually change.
627 iris_init_render_context(struct iris_screen
*screen
,
628 struct iris_batch
*batch
,
629 struct iris_vtable
*vtbl
,
630 struct pipe_debug_callback
*dbg
)
632 UNUSED
const struct gen_device_info
*devinfo
= &screen
->devinfo
;
635 emit_pipeline_select(batch
, _3D
);
637 init_state_base_address(batch
);
640 // XXX: INSTPM on Gen8
641 iris_pack_state(GENX(CS_DEBUG_MODE2
), ®_val
, reg
) {
642 reg
.CONSTANT_BUFFERAddressOffsetDisable
= true;
643 reg
.CONSTANT_BUFFERAddressOffsetDisableMask
= true;
645 iris_emit_lri(batch
, CS_DEBUG_MODE2
, reg_val
);
647 iris_pack_state(GENX(INSTPM
), ®_val
, reg
) {
648 reg
.CONSTANT_BUFFERAddressOffsetDisable
= true;
649 reg
.CONSTANT_BUFFERAddressOffsetDisableMask
= true;
651 iris_emit_lri(batch
, INSTPM
, reg_val
);
655 iris_pack_state(GENX(CACHE_MODE_1
), ®_val
, reg
) {
656 reg
.FloatBlendOptimizationEnable
= true;
657 reg
.FloatBlendOptimizationEnableMask
= true;
658 reg
.PartialResolveDisableInVC
= true;
659 reg
.PartialResolveDisableInVCMask
= true;
661 iris_emit_lri(batch
, CACHE_MODE_1
, reg_val
);
663 if (devinfo
->is_geminilake
)
664 init_glk_barrier_mode(batch
, GLK_BARRIER_MODE_3D_HULL
);
668 iris_pack_state(GENX(SAMPLER_MODE
), ®_val
, reg
) {
669 reg
.HeaderlessMessageforPreemptableContexts
= 1;
670 reg
.HeaderlessMessageforPreemptableContextsMask
= 1;
672 iris_emit_lri(batch
, SAMPLER_MODE
, reg_val
);
677 /* 3DSTATE_DRAWING_RECTANGLE is non-pipelined, so we want to avoid
678 * changing it dynamically. We set it to the maximum size here, and
679 * instead include the render target dimensions in the viewport, so
680 * viewport extents clipping takes care of pruning stray geometry.
682 iris_emit_cmd(batch
, GENX(3DSTATE_DRAWING_RECTANGLE
), rect
) {
683 rect
.ClippedDrawingRectangleXMax
= UINT16_MAX
;
684 rect
.ClippedDrawingRectangleYMax
= UINT16_MAX
;
687 /* Set the initial MSAA sample positions. */
688 iris_emit_cmd(batch
, GENX(3DSTATE_SAMPLE_PATTERN
), pat
) {
689 GEN_SAMPLE_POS_1X(pat
._1xSample
);
690 GEN_SAMPLE_POS_2X(pat
._2xSample
);
691 GEN_SAMPLE_POS_4X(pat
._4xSample
);
692 GEN_SAMPLE_POS_8X(pat
._8xSample
);
694 GEN_SAMPLE_POS_16X(pat
._16xSample
);
698 /* Use the legacy AA line coverage computation. */
699 iris_emit_cmd(batch
, GENX(3DSTATE_AA_LINE_PARAMETERS
), foo
);
701 /* Disable chromakeying (it's for media) */
702 iris_emit_cmd(batch
, GENX(3DSTATE_WM_CHROMAKEY
), foo
);
704 /* We want regular rendering, not special HiZ operations. */
705 iris_emit_cmd(batch
, GENX(3DSTATE_WM_HZ_OP
), foo
);
707 /* No polygon stippling offsets are necessary. */
708 // XXX: may need to set an offset for origin-UL framebuffers
709 iris_emit_cmd(batch
, GENX(3DSTATE_POLY_STIPPLE_OFFSET
), foo
);
711 /* Set a static partitioning of the push constant area. */
712 // XXX: this may be a bad idea...could starve the push ringbuffers...
713 for (int i
= 0; i
<= MESA_SHADER_FRAGMENT
; i
++) {
714 iris_emit_cmd(batch
, GENX(3DSTATE_PUSH_CONSTANT_ALLOC_VS
), alloc
) {
715 alloc
._3DCommandSubOpcode
= 18 + i
;
716 alloc
.ConstantBufferOffset
= 6 * i
;
717 alloc
.ConstantBufferSize
= i
== MESA_SHADER_FRAGMENT
? 8 : 6;
723 iris_init_compute_context(struct iris_screen
*screen
,
724 struct iris_batch
*batch
,
725 struct iris_vtable
*vtbl
,
726 struct pipe_debug_callback
*dbg
)
728 UNUSED
const struct gen_device_info
*devinfo
= &screen
->devinfo
;
730 emit_pipeline_select(batch
, GPGPU
);
732 const bool has_slm
= true;
733 const bool wants_dc_cache
= true;
735 const struct gen_l3_weights w
=
736 gen_get_default_l3_weights(devinfo
, wants_dc_cache
, has_slm
);
737 const struct gen_l3_config
*cfg
= gen_get_l3_config(devinfo
, w
);
740 iris_pack_state(GENX(L3CNTLREG
), ®_val
, reg
) {
741 reg
.SLMEnable
= has_slm
;
743 /* WA_1406697149: Bit 9 "Error Detection Behavior Control" must be set
744 * in L3CNTLREG register. The default setting of the bit is not the
745 * desirable behavior.
747 reg
.ErrorDetectionBehaviorControl
= true;
749 reg
.URBAllocation
= cfg
->n
[GEN_L3P_URB
];
750 reg
.ROAllocation
= cfg
->n
[GEN_L3P_RO
];
751 reg
.DCAllocation
= cfg
->n
[GEN_L3P_DC
];
752 reg
.AllAllocation
= cfg
->n
[GEN_L3P_ALL
];
754 iris_emit_lri(batch
, L3CNTLREG
, reg_val
);
756 init_state_base_address(batch
);
759 if (devinfo
->is_geminilake
)
760 init_glk_barrier_mode(batch
, GLK_BARRIER_MODE_GPGPU
);
764 struct iris_vertex_buffer_state
{
765 /** The VERTEX_BUFFER_STATE hardware structure. */
766 uint32_t state
[GENX(VERTEX_BUFFER_STATE_length
)];
768 /** The resource to source vertex data from. */
769 struct pipe_resource
*resource
;
772 struct iris_depth_buffer_state
{
773 /* Depth/HiZ/Stencil related hardware packets. */
774 uint32_t packets
[GENX(3DSTATE_DEPTH_BUFFER_length
) +
775 GENX(3DSTATE_STENCIL_BUFFER_length
) +
776 GENX(3DSTATE_HIER_DEPTH_BUFFER_length
) +
777 GENX(3DSTATE_CLEAR_PARAMS_length
)];
781 * Generation-specific context state (ice->state.genx->...).
783 * Most state can go in iris_context directly, but these encode hardware
784 * packets which vary by generation.
786 struct iris_genx_state
{
787 struct iris_vertex_buffer_state vertex_buffers
[33];
789 struct iris_depth_buffer_state depth_buffer
;
791 uint32_t so_buffers
[4 * GENX(3DSTATE_SO_BUFFER_length
)];
795 * The pipe->set_blend_color() driver hook.
797 * This corresponds to our COLOR_CALC_STATE.
800 iris_set_blend_color(struct pipe_context
*ctx
,
801 const struct pipe_blend_color
*state
)
803 struct iris_context
*ice
= (struct iris_context
*) ctx
;
805 /* Our COLOR_CALC_STATE is exactly pipe_blend_color, so just memcpy */
806 memcpy(&ice
->state
.blend_color
, state
, sizeof(struct pipe_blend_color
));
807 ice
->state
.dirty
|= IRIS_DIRTY_COLOR_CALC_STATE
;
811 * Gallium CSO for blend state (see pipe_blend_state).
813 struct iris_blend_state
{
814 /** Partial 3DSTATE_PS_BLEND */
815 uint32_t ps_blend
[GENX(3DSTATE_PS_BLEND_length
)];
817 /** Partial BLEND_STATE */
818 uint32_t blend_state
[GENX(BLEND_STATE_length
) +
819 BRW_MAX_DRAW_BUFFERS
* GENX(BLEND_STATE_ENTRY_length
)];
821 bool alpha_to_coverage
; /* for shader key */
823 /** Bitfield of whether blending is enabled for RT[i] - for aux resolves */
824 uint8_t blend_enables
;
827 static enum pipe_blendfactor
828 fix_blendfactor(enum pipe_blendfactor f
, bool alpha_to_one
)
831 if (f
== PIPE_BLENDFACTOR_SRC1_ALPHA
)
832 return PIPE_BLENDFACTOR_ONE
;
834 if (f
== PIPE_BLENDFACTOR_INV_SRC1_ALPHA
)
835 return PIPE_BLENDFACTOR_ZERO
;
842 * The pipe->create_blend_state() driver hook.
844 * Translates a pipe_blend_state into iris_blend_state.
847 iris_create_blend_state(struct pipe_context
*ctx
,
848 const struct pipe_blend_state
*state
)
850 struct iris_blend_state
*cso
= malloc(sizeof(struct iris_blend_state
));
851 uint32_t *blend_entry
= cso
->blend_state
+ GENX(BLEND_STATE_length
);
853 cso
->blend_enables
= 0;
854 STATIC_ASSERT(BRW_MAX_DRAW_BUFFERS
<= 8);
856 cso
->alpha_to_coverage
= state
->alpha_to_coverage
;
858 bool indep_alpha_blend
= false;
860 for (int i
= 0; i
< BRW_MAX_DRAW_BUFFERS
; i
++) {
861 const struct pipe_rt_blend_state
*rt
=
862 &state
->rt
[state
->independent_blend_enable
? i
: 0];
864 enum pipe_blendfactor src_rgb
=
865 fix_blendfactor(rt
->rgb_src_factor
, state
->alpha_to_one
);
866 enum pipe_blendfactor src_alpha
=
867 fix_blendfactor(rt
->alpha_src_factor
, state
->alpha_to_one
);
868 enum pipe_blendfactor dst_rgb
=
869 fix_blendfactor(rt
->rgb_dst_factor
, state
->alpha_to_one
);
870 enum pipe_blendfactor dst_alpha
=
871 fix_blendfactor(rt
->alpha_dst_factor
, state
->alpha_to_one
);
873 if (rt
->rgb_func
!= rt
->alpha_func
||
874 src_rgb
!= src_alpha
|| dst_rgb
!= dst_alpha
)
875 indep_alpha_blend
= true;
877 if (rt
->blend_enable
)
878 cso
->blend_enables
|= 1u << i
;
880 iris_pack_state(GENX(BLEND_STATE_ENTRY
), blend_entry
, be
) {
881 be
.LogicOpEnable
= state
->logicop_enable
;
882 be
.LogicOpFunction
= state
->logicop_func
;
884 be
.PreBlendSourceOnlyClampEnable
= false;
885 be
.ColorClampRange
= COLORCLAMP_RTFORMAT
;
886 be
.PreBlendColorClampEnable
= true;
887 be
.PostBlendColorClampEnable
= true;
889 be
.ColorBufferBlendEnable
= rt
->blend_enable
;
891 be
.ColorBlendFunction
= rt
->rgb_func
;
892 be
.AlphaBlendFunction
= rt
->alpha_func
;
893 be
.SourceBlendFactor
= src_rgb
;
894 be
.SourceAlphaBlendFactor
= src_alpha
;
895 be
.DestinationBlendFactor
= dst_rgb
;
896 be
.DestinationAlphaBlendFactor
= dst_alpha
;
898 be
.WriteDisableRed
= !(rt
->colormask
& PIPE_MASK_R
);
899 be
.WriteDisableGreen
= !(rt
->colormask
& PIPE_MASK_G
);
900 be
.WriteDisableBlue
= !(rt
->colormask
& PIPE_MASK_B
);
901 be
.WriteDisableAlpha
= !(rt
->colormask
& PIPE_MASK_A
);
903 blend_entry
+= GENX(BLEND_STATE_ENTRY_length
);
906 iris_pack_command(GENX(3DSTATE_PS_BLEND
), cso
->ps_blend
, pb
) {
907 /* pb.HasWriteableRT is filled in at draw time. */
908 /* pb.AlphaTestEnable is filled in at draw time. */
909 pb
.AlphaToCoverageEnable
= state
->alpha_to_coverage
;
910 pb
.IndependentAlphaBlendEnable
= indep_alpha_blend
;
912 pb
.ColorBufferBlendEnable
= state
->rt
[0].blend_enable
;
914 pb
.SourceBlendFactor
=
915 fix_blendfactor(state
->rt
[0].rgb_src_factor
, state
->alpha_to_one
);
916 pb
.SourceAlphaBlendFactor
=
917 fix_blendfactor(state
->rt
[0].alpha_src_factor
, state
->alpha_to_one
);
918 pb
.DestinationBlendFactor
=
919 fix_blendfactor(state
->rt
[0].rgb_dst_factor
, state
->alpha_to_one
);
920 pb
.DestinationAlphaBlendFactor
=
921 fix_blendfactor(state
->rt
[0].alpha_dst_factor
, state
->alpha_to_one
);
924 iris_pack_state(GENX(BLEND_STATE
), cso
->blend_state
, bs
) {
925 bs
.AlphaToCoverageEnable
= state
->alpha_to_coverage
;
926 bs
.IndependentAlphaBlendEnable
= indep_alpha_blend
;
927 bs
.AlphaToOneEnable
= state
->alpha_to_one
;
928 bs
.AlphaToCoverageDitherEnable
= state
->alpha_to_coverage
;
929 bs
.ColorDitherEnable
= state
->dither
;
930 /* bl.AlphaTestEnable and bs.AlphaTestFunction are filled in later. */
938 * The pipe->bind_blend_state() driver hook.
940 * Bind a blending CSO and flag related dirty bits.
943 iris_bind_blend_state(struct pipe_context
*ctx
, void *state
)
945 struct iris_context
*ice
= (struct iris_context
*) ctx
;
946 struct iris_blend_state
*cso
= state
;
948 ice
->state
.cso_blend
= cso
;
949 ice
->state
.blend_enables
= cso
? cso
->blend_enables
: 0;
951 ice
->state
.dirty
|= IRIS_DIRTY_PS_BLEND
;
952 ice
->state
.dirty
|= IRIS_DIRTY_BLEND_STATE
;
953 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_BLEND
];
957 * Gallium CSO for depth, stencil, and alpha testing state.
959 struct iris_depth_stencil_alpha_state
{
960 /** Partial 3DSTATE_WM_DEPTH_STENCIL. */
961 uint32_t wmds
[GENX(3DSTATE_WM_DEPTH_STENCIL_length
)];
963 /** Outbound to BLEND_STATE, 3DSTATE_PS_BLEND, COLOR_CALC_STATE. */
964 struct pipe_alpha_state alpha
;
966 /** Outbound to resolve and cache set tracking. */
967 bool depth_writes_enabled
;
968 bool stencil_writes_enabled
;
972 * The pipe->create_depth_stencil_alpha_state() driver hook.
974 * We encode most of 3DSTATE_WM_DEPTH_STENCIL, and just save off the alpha
975 * testing state since we need pieces of it in a variety of places.
978 iris_create_zsa_state(struct pipe_context
*ctx
,
979 const struct pipe_depth_stencil_alpha_state
*state
)
981 struct iris_depth_stencil_alpha_state
*cso
=
982 malloc(sizeof(struct iris_depth_stencil_alpha_state
));
984 bool two_sided_stencil
= state
->stencil
[1].enabled
;
986 cso
->alpha
= state
->alpha
;
987 cso
->depth_writes_enabled
= state
->depth
.writemask
;
988 cso
->stencil_writes_enabled
=
989 state
->stencil
[0].writemask
!= 0 ||
990 (two_sided_stencil
&& state
->stencil
[1].writemask
!= 1);
992 /* The state tracker needs to optimize away EQUAL writes for us. */
993 assert(!(state
->depth
.func
== PIPE_FUNC_EQUAL
&& state
->depth
.writemask
));
995 iris_pack_command(GENX(3DSTATE_WM_DEPTH_STENCIL
), cso
->wmds
, wmds
) {
996 wmds
.StencilFailOp
= state
->stencil
[0].fail_op
;
997 wmds
.StencilPassDepthFailOp
= state
->stencil
[0].zfail_op
;
998 wmds
.StencilPassDepthPassOp
= state
->stencil
[0].zpass_op
;
999 wmds
.StencilTestFunction
=
1000 translate_compare_func(state
->stencil
[0].func
);
1001 wmds
.BackfaceStencilFailOp
= state
->stencil
[1].fail_op
;
1002 wmds
.BackfaceStencilPassDepthFailOp
= state
->stencil
[1].zfail_op
;
1003 wmds
.BackfaceStencilPassDepthPassOp
= state
->stencil
[1].zpass_op
;
1004 wmds
.BackfaceStencilTestFunction
=
1005 translate_compare_func(state
->stencil
[1].func
);
1006 wmds
.DepthTestFunction
= translate_compare_func(state
->depth
.func
);
1007 wmds
.DoubleSidedStencilEnable
= two_sided_stencil
;
1008 wmds
.StencilTestEnable
= state
->stencil
[0].enabled
;
1009 wmds
.StencilBufferWriteEnable
=
1010 state
->stencil
[0].writemask
!= 0 ||
1011 (two_sided_stencil
&& state
->stencil
[1].writemask
!= 0);
1012 wmds
.DepthTestEnable
= state
->depth
.enabled
;
1013 wmds
.DepthBufferWriteEnable
= state
->depth
.writemask
;
1014 wmds
.StencilTestMask
= state
->stencil
[0].valuemask
;
1015 wmds
.StencilWriteMask
= state
->stencil
[0].writemask
;
1016 wmds
.BackfaceStencilTestMask
= state
->stencil
[1].valuemask
;
1017 wmds
.BackfaceStencilWriteMask
= state
->stencil
[1].writemask
;
1018 /* wmds.[Backface]StencilReferenceValue are merged later */
1025 * The pipe->bind_depth_stencil_alpha_state() driver hook.
1027 * Bind a depth/stencil/alpha CSO and flag related dirty bits.
1030 iris_bind_zsa_state(struct pipe_context
*ctx
, void *state
)
1032 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1033 struct iris_depth_stencil_alpha_state
*old_cso
= ice
->state
.cso_zsa
;
1034 struct iris_depth_stencil_alpha_state
*new_cso
= state
;
1037 if (cso_changed(alpha
.ref_value
))
1038 ice
->state
.dirty
|= IRIS_DIRTY_COLOR_CALC_STATE
;
1040 if (cso_changed(alpha
.enabled
))
1041 ice
->state
.dirty
|= IRIS_DIRTY_PS_BLEND
| IRIS_DIRTY_BLEND_STATE
;
1043 if (cso_changed(alpha
.func
))
1044 ice
->state
.dirty
|= IRIS_DIRTY_BLEND_STATE
;
1046 ice
->state
.depth_writes_enabled
= new_cso
->depth_writes_enabled
;
1047 ice
->state
.stencil_writes_enabled
= new_cso
->stencil_writes_enabled
;
1050 ice
->state
.cso_zsa
= new_cso
;
1051 ice
->state
.dirty
|= IRIS_DIRTY_CC_VIEWPORT
;
1052 ice
->state
.dirty
|= IRIS_DIRTY_WM_DEPTH_STENCIL
;
1053 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_DEPTH_STENCIL_ALPHA
];
1057 * Gallium CSO for rasterizer state.
1059 struct iris_rasterizer_state
{
1060 uint32_t sf
[GENX(3DSTATE_SF_length
)];
1061 uint32_t clip
[GENX(3DSTATE_CLIP_length
)];
1062 uint32_t raster
[GENX(3DSTATE_RASTER_length
)];
1063 uint32_t wm
[GENX(3DSTATE_WM_length
)];
1064 uint32_t line_stipple
[GENX(3DSTATE_LINE_STIPPLE_length
)];
1066 uint8_t num_clip_plane_consts
;
1067 bool clip_halfz
; /* for CC_VIEWPORT */
1068 bool depth_clip_near
; /* for CC_VIEWPORT */
1069 bool depth_clip_far
; /* for CC_VIEWPORT */
1070 bool flatshade
; /* for shader state */
1071 bool flatshade_first
; /* for stream output */
1072 bool clamp_fragment_color
; /* for shader state */
1073 bool light_twoside
; /* for shader state */
1074 bool rasterizer_discard
; /* for 3DSTATE_STREAMOUT and 3DSTATE_CLIP */
1075 bool half_pixel_center
; /* for 3DSTATE_MULTISAMPLE */
1076 bool line_stipple_enable
;
1077 bool poly_stipple_enable
;
1079 bool force_persample_interp
;
1080 enum pipe_sprite_coord_mode sprite_coord_mode
; /* PIPE_SPRITE_* */
1081 uint16_t sprite_coord_enable
;
1085 get_line_width(const struct pipe_rasterizer_state
*state
)
1087 float line_width
= state
->line_width
;
1089 /* From the OpenGL 4.4 spec:
1091 * "The actual width of non-antialiased lines is determined by rounding
1092 * the supplied width to the nearest integer, then clamping it to the
1093 * implementation-dependent maximum non-antialiased line width."
1095 if (!state
->multisample
&& !state
->line_smooth
)
1096 line_width
= roundf(state
->line_width
);
1098 if (!state
->multisample
&& state
->line_smooth
&& line_width
< 1.5f
) {
1099 /* For 1 pixel line thickness or less, the general anti-aliasing
1100 * algorithm gives up, and a garbage line is generated. Setting a
1101 * Line Width of 0.0 specifies the rasterization of the "thinnest"
1102 * (one-pixel-wide), non-antialiased lines.
1104 * Lines rendered with zero Line Width are rasterized using the
1105 * "Grid Intersection Quantization" rules as specified by the
1106 * "Zero-Width (Cosmetic) Line Rasterization" section of the docs.
1115 * The pipe->create_rasterizer_state() driver hook.
1118 iris_create_rasterizer_state(struct pipe_context
*ctx
,
1119 const struct pipe_rasterizer_state
*state
)
1121 struct iris_rasterizer_state
*cso
=
1122 malloc(sizeof(struct iris_rasterizer_state
));
1130 offset_units_unscaled
- cap
not exposed
1134 // XXX: it may make more sense just to store the pipe_rasterizer_state,
1135 // we're copying a lot of booleans here. But we don't need all of them...
1137 cso
->multisample
= state
->multisample
;
1138 cso
->force_persample_interp
= state
->force_persample_interp
;
1139 cso
->clip_halfz
= state
->clip_halfz
;
1140 cso
->depth_clip_near
= state
->depth_clip_near
;
1141 cso
->depth_clip_far
= state
->depth_clip_far
;
1142 cso
->flatshade
= state
->flatshade
;
1143 cso
->flatshade_first
= state
->flatshade_first
;
1144 cso
->clamp_fragment_color
= state
->clamp_fragment_color
;
1145 cso
->light_twoside
= state
->light_twoside
;
1146 cso
->rasterizer_discard
= state
->rasterizer_discard
;
1147 cso
->half_pixel_center
= state
->half_pixel_center
;
1148 cso
->sprite_coord_mode
= state
->sprite_coord_mode
;
1149 cso
->sprite_coord_enable
= state
->sprite_coord_enable
;
1150 cso
->line_stipple_enable
= state
->line_stipple_enable
;
1151 cso
->poly_stipple_enable
= state
->poly_stipple_enable
;
1153 if (state
->clip_plane_enable
!= 0)
1154 cso
->num_clip_plane_consts
= util_logbase2(state
->clip_plane_enable
) + 1;
1156 cso
->num_clip_plane_consts
= 0;
1158 float line_width
= get_line_width(state
);
1160 iris_pack_command(GENX(3DSTATE_SF
), cso
->sf
, sf
) {
1161 sf
.StatisticsEnable
= true;
1162 sf
.ViewportTransformEnable
= true;
1163 sf
.AALineDistanceMode
= AALINEDISTANCE_TRUE
;
1164 sf
.LineEndCapAntialiasingRegionWidth
=
1165 state
->line_smooth
? _10pixels
: _05pixels
;
1166 sf
.LastPixelEnable
= state
->line_last_pixel
;
1167 sf
.LineWidth
= line_width
;
1168 sf
.SmoothPointEnable
= (state
->point_smooth
|| state
->multisample
) &&
1169 !state
->point_quad_rasterization
;
1170 sf
.PointWidthSource
= state
->point_size_per_vertex
? Vertex
: State
;
1171 sf
.PointWidth
= state
->point_size
;
1173 if (state
->flatshade_first
) {
1174 sf
.TriangleFanProvokingVertexSelect
= 1;
1176 sf
.TriangleStripListProvokingVertexSelect
= 2;
1177 sf
.TriangleFanProvokingVertexSelect
= 2;
1178 sf
.LineStripListProvokingVertexSelect
= 1;
1182 iris_pack_command(GENX(3DSTATE_RASTER
), cso
->raster
, rr
) {
1183 rr
.FrontWinding
= state
->front_ccw
? CounterClockwise
: Clockwise
;
1184 rr
.CullMode
= translate_cull_mode(state
->cull_face
);
1185 rr
.FrontFaceFillMode
= translate_fill_mode(state
->fill_front
);
1186 rr
.BackFaceFillMode
= translate_fill_mode(state
->fill_back
);
1187 rr
.DXMultisampleRasterizationEnable
= state
->multisample
;
1188 rr
.GlobalDepthOffsetEnableSolid
= state
->offset_tri
;
1189 rr
.GlobalDepthOffsetEnableWireframe
= state
->offset_line
;
1190 rr
.GlobalDepthOffsetEnablePoint
= state
->offset_point
;
1191 rr
.GlobalDepthOffsetConstant
= state
->offset_units
* 2;
1192 rr
.GlobalDepthOffsetScale
= state
->offset_scale
;
1193 rr
.GlobalDepthOffsetClamp
= state
->offset_clamp
;
1194 rr
.SmoothPointEnable
= state
->point_smooth
;
1195 rr
.AntialiasingEnable
= state
->line_smooth
;
1196 rr
.ScissorRectangleEnable
= state
->scissor
;
1198 rr
.ViewportZNearClipTestEnable
= state
->depth_clip_near
;
1199 rr
.ViewportZFarClipTestEnable
= state
->depth_clip_far
;
1201 rr
.ViewportZClipTestEnable
= (state
->depth_clip_near
|| state
->depth_clip_far
);
1203 //rr.ConservativeRasterizationEnable = not yet supported by Gallium...
1206 iris_pack_command(GENX(3DSTATE_CLIP
), cso
->clip
, cl
) {
1207 /* cl.NonPerspectiveBarycentricEnable is filled in at draw time from
1208 * the FS program; cl.ForceZeroRTAIndexEnable is filled in from the FB.
1210 cl
.EarlyCullEnable
= true;
1211 cl
.UserClipDistanceClipTestEnableBitmask
= state
->clip_plane_enable
;
1212 cl
.ForceUserClipDistanceClipTestEnableBitmask
= true;
1213 cl
.APIMode
= state
->clip_halfz
? APIMODE_D3D
: APIMODE_OGL
;
1214 cl
.GuardbandClipTestEnable
= true;
1215 cl
.ClipEnable
= true;
1216 cl
.ViewportXYClipTestEnable
= state
->point_tri_clip
;
1217 cl
.MinimumPointWidth
= 0.125;
1218 cl
.MaximumPointWidth
= 255.875;
1220 if (state
->flatshade_first
) {
1221 cl
.TriangleFanProvokingVertexSelect
= 1;
1223 cl
.TriangleStripListProvokingVertexSelect
= 2;
1224 cl
.TriangleFanProvokingVertexSelect
= 2;
1225 cl
.LineStripListProvokingVertexSelect
= 1;
1229 iris_pack_command(GENX(3DSTATE_WM
), cso
->wm
, wm
) {
1230 /* wm.BarycentricInterpolationMode and wm.EarlyDepthStencilControl are
1231 * filled in at draw time from the FS program.
1233 wm
.LineAntialiasingRegionWidth
= _10pixels
;
1234 wm
.LineEndCapAntialiasingRegionWidth
= _05pixels
;
1235 wm
.PointRasterizationRule
= RASTRULE_UPPER_RIGHT
;
1236 wm
.LineStippleEnable
= state
->line_stipple_enable
;
1237 wm
.PolygonStippleEnable
= state
->poly_stipple_enable
;
1240 /* Remap from 0..255 back to 1..256 */
1241 const unsigned line_stipple_factor
= state
->line_stipple_factor
+ 1;
1243 iris_pack_command(GENX(3DSTATE_LINE_STIPPLE
), cso
->line_stipple
, line
) {
1244 line
.LineStipplePattern
= state
->line_stipple_pattern
;
1245 line
.LineStippleInverseRepeatCount
= 1.0f
/ line_stipple_factor
;
1246 line
.LineStippleRepeatCount
= line_stipple_factor
;
1253 * The pipe->bind_rasterizer_state() driver hook.
1255 * Bind a rasterizer CSO and flag related dirty bits.
1258 iris_bind_rasterizer_state(struct pipe_context
*ctx
, void *state
)
1260 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1261 struct iris_rasterizer_state
*old_cso
= ice
->state
.cso_rast
;
1262 struct iris_rasterizer_state
*new_cso
= state
;
1265 /* Try to avoid re-emitting 3DSTATE_LINE_STIPPLE, it's non-pipelined */
1266 if (cso_changed_memcmp(line_stipple
))
1267 ice
->state
.dirty
|= IRIS_DIRTY_LINE_STIPPLE
;
1269 if (cso_changed(half_pixel_center
))
1270 ice
->state
.dirty
|= IRIS_DIRTY_MULTISAMPLE
;
1272 if (cso_changed(line_stipple_enable
) || cso_changed(poly_stipple_enable
))
1273 ice
->state
.dirty
|= IRIS_DIRTY_WM
;
1275 if (cso_changed(rasterizer_discard
))
1276 ice
->state
.dirty
|= IRIS_DIRTY_STREAMOUT
| IRIS_DIRTY_CLIP
;
1278 if (cso_changed(flatshade_first
))
1279 ice
->state
.dirty
|= IRIS_DIRTY_STREAMOUT
;
1281 if (cso_changed(depth_clip_near
) || cso_changed(depth_clip_far
) ||
1282 cso_changed(clip_halfz
))
1283 ice
->state
.dirty
|= IRIS_DIRTY_CC_VIEWPORT
;
1285 if (cso_changed(sprite_coord_enable
) ||
1286 cso_changed(sprite_coord_mode
) ||
1287 cso_changed(light_twoside
))
1288 ice
->state
.dirty
|= IRIS_DIRTY_SBE
;
1291 ice
->state
.cso_rast
= new_cso
;
1292 ice
->state
.dirty
|= IRIS_DIRTY_RASTER
;
1293 ice
->state
.dirty
|= IRIS_DIRTY_CLIP
;
1294 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_RASTERIZER
];
1298 * Return true if the given wrap mode requires the border color to exist.
1300 * (We can skip uploading it if the sampler isn't going to use it.)
1303 wrap_mode_needs_border_color(unsigned wrap_mode
)
1305 return wrap_mode
== TCM_CLAMP_BORDER
|| wrap_mode
== TCM_HALF_BORDER
;
1309 * Gallium CSO for sampler state.
1311 struct iris_sampler_state
{
1312 union pipe_color_union border_color
;
1313 bool needs_border_color
;
1315 uint32_t sampler_state
[GENX(SAMPLER_STATE_length
)];
1319 * The pipe->create_sampler_state() driver hook.
1321 * We fill out SAMPLER_STATE (except for the border color pointer), and
1322 * store that on the CPU. It doesn't make sense to upload it to a GPU
1323 * buffer object yet, because 3DSTATE_SAMPLER_STATE_POINTERS requires
1324 * all bound sampler states to be in contiguous memor.
1327 iris_create_sampler_state(struct pipe_context
*ctx
,
1328 const struct pipe_sampler_state
*state
)
1330 struct iris_sampler_state
*cso
= CALLOC_STRUCT(iris_sampler_state
);
1335 STATIC_ASSERT(PIPE_TEX_FILTER_NEAREST
== MAPFILTER_NEAREST
);
1336 STATIC_ASSERT(PIPE_TEX_FILTER_LINEAR
== MAPFILTER_LINEAR
);
1338 unsigned wrap_s
= translate_wrap(state
->wrap_s
);
1339 unsigned wrap_t
= translate_wrap(state
->wrap_t
);
1340 unsigned wrap_r
= translate_wrap(state
->wrap_r
);
1342 memcpy(&cso
->border_color
, &state
->border_color
, sizeof(cso
->border_color
));
1344 cso
->needs_border_color
= wrap_mode_needs_border_color(wrap_s
) ||
1345 wrap_mode_needs_border_color(wrap_t
) ||
1346 wrap_mode_needs_border_color(wrap_r
);
1348 float min_lod
= state
->min_lod
;
1349 unsigned mag_img_filter
= state
->mag_img_filter
;
1351 // XXX: explain this code ported from ilo...I don't get it at all...
1352 if (state
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
&&
1353 state
->min_lod
> 0.0f
) {
1355 mag_img_filter
= state
->min_img_filter
;
1358 iris_pack_state(GENX(SAMPLER_STATE
), cso
->sampler_state
, samp
) {
1359 samp
.TCXAddressControlMode
= wrap_s
;
1360 samp
.TCYAddressControlMode
= wrap_t
;
1361 samp
.TCZAddressControlMode
= wrap_r
;
1362 samp
.CubeSurfaceControlMode
= state
->seamless_cube_map
;
1363 samp
.NonnormalizedCoordinateEnable
= !state
->normalized_coords
;
1364 samp
.MinModeFilter
= state
->min_img_filter
;
1365 samp
.MagModeFilter
= mag_img_filter
;
1366 samp
.MipModeFilter
= translate_mip_filter(state
->min_mip_filter
);
1367 samp
.MaximumAnisotropy
= RATIO21
;
1369 if (state
->max_anisotropy
>= 2) {
1370 if (state
->min_img_filter
== PIPE_TEX_FILTER_LINEAR
) {
1371 samp
.MinModeFilter
= MAPFILTER_ANISOTROPIC
;
1372 samp
.AnisotropicAlgorithm
= EWAApproximation
;
1375 if (state
->mag_img_filter
== PIPE_TEX_FILTER_LINEAR
)
1376 samp
.MagModeFilter
= MAPFILTER_ANISOTROPIC
;
1378 samp
.MaximumAnisotropy
=
1379 MIN2((state
->max_anisotropy
- 2) / 2, RATIO161
);
1382 /* Set address rounding bits if not using nearest filtering. */
1383 if (state
->min_img_filter
!= PIPE_TEX_FILTER_NEAREST
) {
1384 samp
.UAddressMinFilterRoundingEnable
= true;
1385 samp
.VAddressMinFilterRoundingEnable
= true;
1386 samp
.RAddressMinFilterRoundingEnable
= true;
1389 if (state
->mag_img_filter
!= PIPE_TEX_FILTER_NEAREST
) {
1390 samp
.UAddressMagFilterRoundingEnable
= true;
1391 samp
.VAddressMagFilterRoundingEnable
= true;
1392 samp
.RAddressMagFilterRoundingEnable
= true;
1395 if (state
->compare_mode
== PIPE_TEX_COMPARE_R_TO_TEXTURE
)
1396 samp
.ShadowFunction
= translate_shadow_func(state
->compare_func
);
1398 const float hw_max_lod
= GEN_GEN
>= 7 ? 14 : 13;
1400 samp
.LODPreClampMode
= CLAMP_MODE_OGL
;
1401 samp
.MinLOD
= CLAMP(min_lod
, 0, hw_max_lod
);
1402 samp
.MaxLOD
= CLAMP(state
->max_lod
, 0, hw_max_lod
);
1403 samp
.TextureLODBias
= CLAMP(state
->lod_bias
, -16, 15);
1405 /* .BorderColorPointer is filled in by iris_bind_sampler_states. */
1412 * The pipe->bind_sampler_states() driver hook.
1414 * Now that we know all the sampler states, we upload them all into a
1415 * contiguous area of GPU memory, for 3DSTATE_SAMPLER_STATE_POINTERS_*.
1416 * We also fill out the border color state pointers at this point.
1418 * We could defer this work to draw time, but we assume that binding
1419 * will be less frequent than drawing.
1421 // XXX: this may be a bad idea, need to make sure that st/mesa calls us
1422 // XXX: with the complete set of shaders. If it makes multiple calls to
1423 // XXX: things one at a time, we could waste a lot of time assembling things.
1424 // XXX: it doesn't even BUY us anything to do it here, because we only flag
1425 // XXX: IRIS_DIRTY_SAMPLER_STATE when this is called...
1427 iris_bind_sampler_states(struct pipe_context
*ctx
,
1428 enum pipe_shader_type p_stage
,
1429 unsigned start
, unsigned count
,
1432 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1433 gl_shader_stage stage
= stage_from_pipe(p_stage
);
1434 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
1436 assert(start
+ count
<= IRIS_MAX_TEXTURE_SAMPLERS
);
1438 for (int i
= 0; i
< count
; i
++) {
1439 shs
->samplers
[start
+ i
] = states
[i
];
1442 /* Assemble the SAMPLER_STATEs into a contiguous table that lives
1443 * in the dynamic state memory zone, so we can point to it via the
1444 * 3DSTATE_SAMPLER_STATE_POINTERS_* commands.
1447 upload_state(ice
->state
.dynamic_uploader
, &shs
->sampler_table
,
1448 count
* 4 * GENX(SAMPLER_STATE_length
), 32);
1452 struct pipe_resource
*res
= shs
->sampler_table
.res
;
1453 shs
->sampler_table
.offset
+=
1454 iris_bo_offset_from_base_address(iris_resource_bo(res
));
1456 /* Make sure all land in the same BO */
1457 iris_border_color_pool_reserve(ice
, IRIS_MAX_TEXTURE_SAMPLERS
);
1459 for (int i
= 0; i
< count
; i
++) {
1460 struct iris_sampler_state
*state
= shs
->samplers
[i
];
1463 memset(map
, 0, 4 * GENX(SAMPLER_STATE_length
));
1464 } else if (!state
->needs_border_color
) {
1465 memcpy(map
, state
->sampler_state
, 4 * GENX(SAMPLER_STATE_length
));
1467 ice
->state
.need_border_colors
= true;
1469 /* Stream out the border color and merge the pointer. */
1471 iris_upload_border_color(ice
, &state
->border_color
);
1473 uint32_t dynamic
[GENX(SAMPLER_STATE_length
)];
1474 iris_pack_state(GENX(SAMPLER_STATE
), dynamic
, dyns
) {
1475 dyns
.BorderColorPointer
= offset
;
1478 for (uint32_t j
= 0; j
< GENX(SAMPLER_STATE_length
); j
++)
1479 map
[j
] = state
->sampler_state
[j
] | dynamic
[j
];
1482 map
+= GENX(SAMPLER_STATE_length
);
1485 ice
->state
.dirty
|= IRIS_DIRTY_SAMPLER_STATES_VS
<< stage
;
1488 static enum isl_channel_select
1489 fmt_swizzle(const struct iris_format_info
*fmt
, enum pipe_swizzle swz
)
1492 case PIPE_SWIZZLE_X
: return fmt
->swizzle
.r
;
1493 case PIPE_SWIZZLE_Y
: return fmt
->swizzle
.g
;
1494 case PIPE_SWIZZLE_Z
: return fmt
->swizzle
.b
;
1495 case PIPE_SWIZZLE_W
: return fmt
->swizzle
.a
;
1496 case PIPE_SWIZZLE_1
: return SCS_ONE
;
1497 case PIPE_SWIZZLE_0
: return SCS_ZERO
;
1498 default: unreachable("invalid swizzle");
1503 fill_buffer_surface_state(struct isl_device
*isl_dev
,
1506 enum isl_format format
,
1510 const struct isl_format_layout
*fmtl
= isl_format_get_layout(format
);
1511 const unsigned cpp
= format
== ISL_FORMAT_RAW
? 1 : fmtl
->bpb
/ 8;
1513 /* The ARB_texture_buffer_specification says:
1515 * "The number of texels in the buffer texture's texel array is given by
1517 * floor(<buffer_size> / (<components> * sizeof(<base_type>)),
1519 * where <buffer_size> is the size of the buffer object, in basic
1520 * machine units and <components> and <base_type> are the element count
1521 * and base data type for elements, as specified in Table X.1. The
1522 * number of texels in the texel array is then clamped to the
1523 * implementation-dependent limit MAX_TEXTURE_BUFFER_SIZE_ARB."
1525 * We need to clamp the size in bytes to MAX_TEXTURE_BUFFER_SIZE * stride,
1526 * so that when ISL divides by stride to obtain the number of texels, that
1527 * texel count is clamped to MAX_TEXTURE_BUFFER_SIZE.
1529 unsigned final_size
=
1530 MIN3(size
, bo
->size
- offset
, IRIS_MAX_TEXTURE_BUFFER_SIZE
* cpp
);
1532 isl_buffer_fill_state(isl_dev
, map
,
1533 .address
= bo
->gtt_offset
+ offset
,
1534 .size_B
= final_size
,
1541 * Allocate a SURFACE_STATE structure.
1544 alloc_surface_states(struct u_upload_mgr
*mgr
,
1545 struct iris_state_ref
*ref
)
1547 const unsigned surf_size
= 4 * GENX(RENDER_SURFACE_STATE_length
);
1549 void *map
= upload_state(mgr
, ref
, surf_size
, 64);
1551 ref
->offset
+= iris_bo_offset_from_base_address(iris_resource_bo(ref
->res
));
1557 fill_surface_state(struct isl_device
*isl_dev
,
1559 struct iris_resource
*res
,
1560 struct isl_view
*view
)
1562 struct isl_surf_fill_state_info f
= {
1565 .mocs
= mocs(res
->bo
),
1566 .address
= res
->bo
->gtt_offset
,
1569 isl_surf_fill_state_s(isl_dev
, map
, &f
);
1573 * The pipe->create_sampler_view() driver hook.
1575 static struct pipe_sampler_view
*
1576 iris_create_sampler_view(struct pipe_context
*ctx
,
1577 struct pipe_resource
*tex
,
1578 const struct pipe_sampler_view
*tmpl
)
1580 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1581 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
1582 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1583 struct iris_sampler_view
*isv
= calloc(1, sizeof(struct iris_sampler_view
));
1588 /* initialize base object */
1590 isv
->base
.context
= ctx
;
1591 isv
->base
.texture
= NULL
;
1592 pipe_reference_init(&isv
->base
.reference
, 1);
1593 pipe_resource_reference(&isv
->base
.texture
, tex
);
1595 void *map
= alloc_surface_states(ice
->state
.surface_uploader
,
1596 &isv
->surface_state
);
1600 if (util_format_is_depth_or_stencil(tmpl
->format
)) {
1601 struct iris_resource
*zres
, *sres
;
1602 const struct util_format_description
*desc
=
1603 util_format_description(tmpl
->format
);
1605 iris_get_depth_stencil_resources(tex
, &zres
, &sres
);
1607 tex
= util_format_has_depth(desc
) ? &zres
->base
: &sres
->base
;
1610 isv
->res
= (struct iris_resource
*) tex
;
1612 isl_surf_usage_flags_t usage
= ISL_SURF_USAGE_TEXTURE_BIT
;
1614 if (isv
->base
.target
== PIPE_TEXTURE_CUBE
||
1615 isv
->base
.target
== PIPE_TEXTURE_CUBE_ARRAY
)
1616 usage
|= ISL_SURF_USAGE_CUBE_BIT
;
1618 const struct iris_format_info fmt
=
1619 iris_format_for_usage(devinfo
, tmpl
->format
, usage
);
1621 isv
->view
= (struct isl_view
) {
1623 .swizzle
= (struct isl_swizzle
) {
1624 .r
= fmt_swizzle(&fmt
, tmpl
->swizzle_r
),
1625 .g
= fmt_swizzle(&fmt
, tmpl
->swizzle_g
),
1626 .b
= fmt_swizzle(&fmt
, tmpl
->swizzle_b
),
1627 .a
= fmt_swizzle(&fmt
, tmpl
->swizzle_a
),
1632 /* Fill out SURFACE_STATE for this view. */
1633 if (tmpl
->target
!= PIPE_BUFFER
) {
1634 isv
->view
.base_level
= tmpl
->u
.tex
.first_level
;
1635 isv
->view
.levels
= tmpl
->u
.tex
.last_level
- tmpl
->u
.tex
.first_level
+ 1;
1636 // XXX: do I need to port f9fd0cf4790cb2a530e75d1a2206dbb9d8af7cb2?
1637 isv
->view
.base_array_layer
= tmpl
->u
.tex
.first_layer
;
1638 isv
->view
.array_len
=
1639 tmpl
->u
.tex
.last_layer
- tmpl
->u
.tex
.first_layer
+ 1;
1641 fill_surface_state(&screen
->isl_dev
, map
, isv
->res
, &isv
->view
);
1643 fill_buffer_surface_state(&screen
->isl_dev
, isv
->res
->bo
, map
,
1644 isv
->view
.format
, tmpl
->u
.buf
.offset
,
1652 iris_sampler_view_destroy(struct pipe_context
*ctx
,
1653 struct pipe_sampler_view
*state
)
1655 struct iris_sampler_view
*isv
= (void *) state
;
1656 pipe_resource_reference(&state
->texture
, NULL
);
1657 pipe_resource_reference(&isv
->surface_state
.res
, NULL
);
1662 * The pipe->create_surface() driver hook.
1664 * In Gallium nomenclature, "surfaces" are a view of a resource that
1665 * can be bound as a render target or depth/stencil buffer.
1667 static struct pipe_surface
*
1668 iris_create_surface(struct pipe_context
*ctx
,
1669 struct pipe_resource
*tex
,
1670 const struct pipe_surface
*tmpl
)
1672 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1673 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
1674 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1675 struct iris_surface
*surf
= calloc(1, sizeof(struct iris_surface
));
1676 struct pipe_surface
*psurf
= &surf
->base
;
1677 struct iris_resource
*res
= (struct iris_resource
*) tex
;
1682 pipe_reference_init(&psurf
->reference
, 1);
1683 pipe_resource_reference(&psurf
->texture
, tex
);
1684 psurf
->context
= ctx
;
1685 psurf
->format
= tmpl
->format
;
1686 psurf
->width
= tex
->width0
;
1687 psurf
->height
= tex
->height0
;
1688 psurf
->texture
= tex
;
1689 psurf
->u
.tex
.first_layer
= tmpl
->u
.tex
.first_layer
;
1690 psurf
->u
.tex
.last_layer
= tmpl
->u
.tex
.last_layer
;
1691 psurf
->u
.tex
.level
= tmpl
->u
.tex
.level
;
1693 isl_surf_usage_flags_t usage
= 0;
1695 usage
= ISL_SURF_USAGE_STORAGE_BIT
;
1696 else if (util_format_is_depth_or_stencil(tmpl
->format
))
1697 usage
= ISL_SURF_USAGE_DEPTH_BIT
;
1699 usage
= ISL_SURF_USAGE_RENDER_TARGET_BIT
;
1701 const struct iris_format_info fmt
=
1702 iris_format_for_usage(devinfo
, psurf
->format
, usage
);
1704 if ((usage
& ISL_SURF_USAGE_RENDER_TARGET_BIT
) &&
1705 !isl_format_supports_rendering(devinfo
, fmt
.fmt
)) {
1706 /* Framebuffer validation will reject this invalid case, but it
1707 * hasn't had the opportunity yet. In the meantime, we need to
1708 * avoid hitting ISL asserts about unsupported formats below.
1714 surf
->view
= (struct isl_view
) {
1716 .base_level
= tmpl
->u
.tex
.level
,
1718 .base_array_layer
= tmpl
->u
.tex
.first_layer
,
1719 .array_len
= tmpl
->u
.tex
.last_layer
- tmpl
->u
.tex
.first_layer
+ 1,
1720 .swizzle
= ISL_SWIZZLE_IDENTITY
,
1724 /* Bail early for depth/stencil - we don't want SURFACE_STATE for them. */
1725 if (res
->surf
.usage
& (ISL_SURF_USAGE_DEPTH_BIT
|
1726 ISL_SURF_USAGE_STENCIL_BIT
))
1730 void *map
= alloc_surface_states(ice
->state
.surface_uploader
,
1731 &surf
->surface_state
);
1735 fill_surface_state(&screen
->isl_dev
, map
, res
, &surf
->view
);
1742 fill_default_image_param(struct brw_image_param
*param
)
1744 memset(param
, 0, sizeof(*param
));
1745 /* Set the swizzling shifts to all-ones to effectively disable swizzling --
1746 * See emit_address_calculation() in brw_fs_surface_builder.cpp for a more
1747 * detailed explanation of these parameters.
1749 param
->swizzling
[0] = 0xff;
1750 param
->swizzling
[1] = 0xff;
1754 fill_buffer_image_param(struct brw_image_param
*param
,
1755 enum pipe_format pfmt
,
1758 const unsigned cpp
= util_format_get_blocksize(pfmt
);
1760 fill_default_image_param(param
);
1761 param
->size
[0] = size
/ cpp
;
1762 param
->stride
[0] = cpp
;
1765 #define isl_surf_fill_image_param(x, ...)
1766 #define fill_default_image_param(x, ...)
1767 #define fill_buffer_image_param(x, ...)
1771 * The pipe->set_shader_images() driver hook.
1774 iris_set_shader_images(struct pipe_context
*ctx
,
1775 enum pipe_shader_type p_stage
,
1776 unsigned start_slot
, unsigned count
,
1777 const struct pipe_image_view
*p_images
)
1779 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1780 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
1781 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1782 gl_shader_stage stage
= stage_from_pipe(p_stage
);
1783 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
1785 shs
->bound_image_views
&= ~u_bit_consecutive(start_slot
, count
);
1787 for (unsigned i
= 0; i
< count
; i
++) {
1788 if (p_images
&& p_images
[i
].resource
) {
1789 const struct pipe_image_view
*img
= &p_images
[i
];
1790 struct iris_resource
*res
= (void *) img
->resource
;
1791 pipe_resource_reference(&shs
->image
[start_slot
+ i
].res
, &res
->base
);
1793 shs
->bound_image_views
|= 1 << (start_slot
+ i
);
1795 res
->bind_history
|= PIPE_BIND_SHADER_IMAGE
;
1797 // XXX: these are not retained forever, use a separate uploader?
1799 alloc_surface_states(ice
->state
.surface_uploader
,
1800 &shs
->image
[start_slot
+ i
].surface_state
);
1801 if (!unlikely(map
)) {
1802 pipe_resource_reference(&shs
->image
[start_slot
+ i
].res
, NULL
);
1806 isl_surf_usage_flags_t usage
= ISL_SURF_USAGE_STORAGE_BIT
;
1807 enum isl_format isl_fmt
=
1808 iris_format_for_usage(devinfo
, img
->format
, usage
).fmt
;
1810 bool untyped_fallback
= false;
1812 if (img
->shader_access
& PIPE_IMAGE_ACCESS_READ
) {
1813 /* On Gen8, try to use typed surfaces reads (which support a
1814 * limited number of formats), and if not possible, fall back
1817 untyped_fallback
= GEN_GEN
== 8 &&
1818 !isl_has_matching_typed_storage_image_format(devinfo
, isl_fmt
);
1820 if (untyped_fallback
)
1821 isl_fmt
= ISL_FORMAT_RAW
;
1823 isl_fmt
= isl_lower_storage_image_format(devinfo
, isl_fmt
);
1826 shs
->image
[start_slot
+ i
].access
= img
->shader_access
;
1828 if (res
->base
.target
!= PIPE_BUFFER
) {
1829 struct isl_view view
= {
1831 .base_level
= img
->u
.tex
.level
,
1833 .base_array_layer
= img
->u
.tex
.first_layer
,
1834 .array_len
= img
->u
.tex
.last_layer
- img
->u
.tex
.first_layer
+ 1,
1835 .swizzle
= ISL_SWIZZLE_IDENTITY
,
1839 if (untyped_fallback
) {
1840 fill_buffer_surface_state(&screen
->isl_dev
, res
->bo
, map
,
1841 isl_fmt
, 0, res
->bo
->size
);
1843 fill_surface_state(&screen
->isl_dev
, map
, res
, &view
);
1846 isl_surf_fill_image_param(&screen
->isl_dev
,
1847 &shs
->image
[start_slot
+ i
].param
,
1850 fill_buffer_surface_state(&screen
->isl_dev
, res
->bo
, map
,
1851 isl_fmt
, img
->u
.buf
.offset
,
1853 fill_buffer_image_param(&shs
->image
[start_slot
+ i
].param
,
1854 img
->format
, img
->u
.buf
.size
);
1857 pipe_resource_reference(&shs
->image
[start_slot
+ i
].res
, NULL
);
1858 pipe_resource_reference(&shs
->image
[start_slot
+ i
].surface_state
.res
,
1860 fill_default_image_param(&shs
->image
[start_slot
+ i
].param
);
1864 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_VS
<< stage
;
1866 /* Broadwell also needs brw_image_params re-uploaded */
1868 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_VS
<< stage
;
1869 shs
->cbuf0_needs_upload
= true;
1875 * The pipe->set_sampler_views() driver hook.
1878 iris_set_sampler_views(struct pipe_context
*ctx
,
1879 enum pipe_shader_type p_stage
,
1880 unsigned start
, unsigned count
,
1881 struct pipe_sampler_view
**views
)
1883 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1884 gl_shader_stage stage
= stage_from_pipe(p_stage
);
1885 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
1887 shs
->bound_sampler_views
&= ~u_bit_consecutive(start
, count
);
1889 for (unsigned i
= 0; i
< count
; i
++) {
1890 pipe_sampler_view_reference((struct pipe_sampler_view
**)
1891 &shs
->textures
[start
+ i
], views
[i
]);
1892 struct iris_sampler_view
*view
= (void *) views
[i
];
1894 view
->res
->bind_history
|= PIPE_BIND_SAMPLER_VIEW
;
1895 shs
->bound_sampler_views
|= 1 << (start
+ i
);
1899 ice
->state
.dirty
|= (IRIS_DIRTY_BINDINGS_VS
<< stage
);
1903 * The pipe->set_tess_state() driver hook.
1906 iris_set_tess_state(struct pipe_context
*ctx
,
1907 const float default_outer_level
[4],
1908 const float default_inner_level
[2])
1910 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1912 memcpy(&ice
->state
.default_outer_level
[0], &default_outer_level
[0], 4 * sizeof(float));
1913 memcpy(&ice
->state
.default_inner_level
[0], &default_inner_level
[0], 2 * sizeof(float));
1915 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_TCS
;
1919 iris_surface_destroy(struct pipe_context
*ctx
, struct pipe_surface
*p_surf
)
1921 struct iris_surface
*surf
= (void *) p_surf
;
1922 pipe_resource_reference(&p_surf
->texture
, NULL
);
1923 pipe_resource_reference(&surf
->surface_state
.res
, NULL
);
1928 iris_set_clip_state(struct pipe_context
*ctx
,
1929 const struct pipe_clip_state
*state
)
1931 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1932 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_VERTEX
];
1934 memcpy(&ice
->state
.clip_planes
, state
, sizeof(*state
));
1936 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_VS
;
1937 shs
->cbuf0_needs_upload
= true;
1941 * The pipe->set_polygon_stipple() driver hook.
1944 iris_set_polygon_stipple(struct pipe_context
*ctx
,
1945 const struct pipe_poly_stipple
*state
)
1947 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1948 memcpy(&ice
->state
.poly_stipple
, state
, sizeof(*state
));
1949 ice
->state
.dirty
|= IRIS_DIRTY_POLYGON_STIPPLE
;
1953 * The pipe->set_sample_mask() driver hook.
1956 iris_set_sample_mask(struct pipe_context
*ctx
, unsigned sample_mask
)
1958 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1960 /* We only support 16x MSAA, so we have 16 bits of sample maks.
1961 * st/mesa may pass us 0xffffffff though, meaning "enable all samples".
1963 ice
->state
.sample_mask
= sample_mask
& 0xffff;
1964 ice
->state
.dirty
|= IRIS_DIRTY_SAMPLE_MASK
;
1968 * The pipe->set_scissor_states() driver hook.
1970 * This corresponds to our SCISSOR_RECT state structures. It's an
1971 * exact match, so we just store them, and memcpy them out later.
1974 iris_set_scissor_states(struct pipe_context
*ctx
,
1975 unsigned start_slot
,
1976 unsigned num_scissors
,
1977 const struct pipe_scissor_state
*rects
)
1979 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1981 for (unsigned i
= 0; i
< num_scissors
; i
++) {
1982 if (rects
[i
].minx
== rects
[i
].maxx
|| rects
[i
].miny
== rects
[i
].maxy
) {
1983 /* If the scissor was out of bounds and got clamped to 0 width/height
1984 * at the bounds, the subtraction of 1 from maximums could produce a
1985 * negative number and thus not clip anything. Instead, just provide
1986 * a min > max scissor inside the bounds, which produces the expected
1989 ice
->state
.scissors
[start_slot
+ i
] = (struct pipe_scissor_state
) {
1990 .minx
= 1, .maxx
= 0, .miny
= 1, .maxy
= 0,
1993 ice
->state
.scissors
[start_slot
+ i
] = (struct pipe_scissor_state
) {
1994 .minx
= rects
[i
].minx
, .miny
= rects
[i
].miny
,
1995 .maxx
= rects
[i
].maxx
- 1, .maxy
= rects
[i
].maxy
- 1,
2000 ice
->state
.dirty
|= IRIS_DIRTY_SCISSOR_RECT
;
2004 * The pipe->set_stencil_ref() driver hook.
2006 * This is added to 3DSTATE_WM_DEPTH_STENCIL dynamically at draw time.
2009 iris_set_stencil_ref(struct pipe_context
*ctx
,
2010 const struct pipe_stencil_ref
*state
)
2012 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2013 memcpy(&ice
->state
.stencil_ref
, state
, sizeof(*state
));
2015 ice
->state
.dirty
|= IRIS_DIRTY_COLOR_CALC_STATE
;
2017 ice
->state
.dirty
|= IRIS_DIRTY_WM_DEPTH_STENCIL
;
2021 viewport_extent(const struct pipe_viewport_state
*state
, int axis
, float sign
)
2023 return copysignf(state
->scale
[axis
], sign
) + state
->translate
[axis
];
2027 calculate_guardband_size(uint32_t fb_width
, uint32_t fb_height
,
2028 float m00
, float m11
, float m30
, float m31
,
2029 float *xmin
, float *xmax
,
2030 float *ymin
, float *ymax
)
2032 /* According to the "Vertex X,Y Clamping and Quantization" section of the
2033 * Strips and Fans documentation:
2035 * "The vertex X and Y screen-space coordinates are also /clamped/ to the
2036 * fixed-point "guardband" range supported by the rasterization hardware"
2040 * "In almost all circumstances, if an object’s vertices are actually
2041 * modified by this clamping (i.e., had X or Y coordinates outside of
2042 * the guardband extent the rendered object will not match the intended
2043 * result. Therefore software should take steps to ensure that this does
2044 * not happen - e.g., by clipping objects such that they do not exceed
2045 * these limits after the Drawing Rectangle is applied."
2047 * I believe the fundamental restriction is that the rasterizer (in
2048 * the SF/WM stages) have a limit on the number of pixels that can be
2049 * rasterized. We need to ensure any coordinates beyond the rasterizer
2050 * limit are handled by the clipper. So effectively that limit becomes
2051 * the clipper's guardband size.
2053 * It goes on to say:
2055 * "In addition, in order to be correctly rendered, objects must have a
2056 * screenspace bounding box not exceeding 8K in the X or Y direction.
2057 * This additional restriction must also be comprehended by software,
2058 * i.e., enforced by use of clipping."
2060 * This makes no sense. Gen7+ hardware supports 16K render targets,
2061 * and you definitely need to be able to draw polygons that fill the
2062 * surface. Our assumption is that the rasterizer was limited to 8K
2063 * on Sandybridge, which only supports 8K surfaces, and it was actually
2064 * increased to 16K on Ivybridge and later.
2066 * So, limit the guardband to 16K on Gen7+ and 8K on Sandybridge.
2068 const float gb_size
= GEN_GEN
>= 7 ? 16384.0f
: 8192.0f
;
2070 if (m00
!= 0 && m11
!= 0) {
2071 /* First, we compute the screen-space render area */
2072 const float ss_ra_xmin
= MIN3( 0, m30
+ m00
, m30
- m00
);
2073 const float ss_ra_xmax
= MAX3( fb_width
, m30
+ m00
, m30
- m00
);
2074 const float ss_ra_ymin
= MIN3( 0, m31
+ m11
, m31
- m11
);
2075 const float ss_ra_ymax
= MAX3(fb_height
, m31
+ m11
, m31
- m11
);
2077 /* We want the guardband to be centered on that */
2078 const float ss_gb_xmin
= (ss_ra_xmin
+ ss_ra_xmax
) / 2 - gb_size
;
2079 const float ss_gb_xmax
= (ss_ra_xmin
+ ss_ra_xmax
) / 2 + gb_size
;
2080 const float ss_gb_ymin
= (ss_ra_ymin
+ ss_ra_ymax
) / 2 - gb_size
;
2081 const float ss_gb_ymax
= (ss_ra_ymin
+ ss_ra_ymax
) / 2 + gb_size
;
2083 /* Now we need it in native device coordinates */
2084 const float ndc_gb_xmin
= (ss_gb_xmin
- m30
) / m00
;
2085 const float ndc_gb_xmax
= (ss_gb_xmax
- m30
) / m00
;
2086 const float ndc_gb_ymin
= (ss_gb_ymin
- m31
) / m11
;
2087 const float ndc_gb_ymax
= (ss_gb_ymax
- m31
) / m11
;
2089 /* Thanks to Y-flipping and ORIGIN_UPPER_LEFT, the Y coordinates may be
2090 * flipped upside-down. X should be fine though.
2092 assert(ndc_gb_xmin
<= ndc_gb_xmax
);
2093 *xmin
= ndc_gb_xmin
;
2094 *xmax
= ndc_gb_xmax
;
2095 *ymin
= MIN2(ndc_gb_ymin
, ndc_gb_ymax
);
2096 *ymax
= MAX2(ndc_gb_ymin
, ndc_gb_ymax
);
2098 /* The viewport scales to 0, so nothing will be rendered. */
2107 * The pipe->set_viewport_states() driver hook.
2109 * This corresponds to our SF_CLIP_VIEWPORT states. We can't calculate
2110 * the guardband yet, as we need the framebuffer dimensions, but we can
2111 * at least fill out the rest.
2114 iris_set_viewport_states(struct pipe_context
*ctx
,
2115 unsigned start_slot
,
2117 const struct pipe_viewport_state
*states
)
2119 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2121 memcpy(&ice
->state
.viewports
[start_slot
], states
, sizeof(*states
) * count
);
2123 ice
->state
.dirty
|= IRIS_DIRTY_SF_CL_VIEWPORT
;
2125 if (ice
->state
.cso_rast
&& (!ice
->state
.cso_rast
->depth_clip_near
||
2126 !ice
->state
.cso_rast
->depth_clip_far
))
2127 ice
->state
.dirty
|= IRIS_DIRTY_CC_VIEWPORT
;
2131 * The pipe->set_framebuffer_state() driver hook.
2133 * Sets the current draw FBO, including color render targets, depth,
2134 * and stencil buffers.
2137 iris_set_framebuffer_state(struct pipe_context
*ctx
,
2138 const struct pipe_framebuffer_state
*state
)
2140 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2141 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
2142 struct isl_device
*isl_dev
= &screen
->isl_dev
;
2143 struct pipe_framebuffer_state
*cso
= &ice
->state
.framebuffer
;
2144 struct iris_resource
*zres
;
2145 struct iris_resource
*stencil_res
;
2147 unsigned samples
= util_framebuffer_get_num_samples(state
);
2148 unsigned layers
= util_framebuffer_get_num_layers(state
);
2150 if (cso
->samples
!= samples
) {
2151 ice
->state
.dirty
|= IRIS_DIRTY_MULTISAMPLE
;
2154 if (cso
->nr_cbufs
!= state
->nr_cbufs
) {
2155 ice
->state
.dirty
|= IRIS_DIRTY_BLEND_STATE
;
2158 if ((cso
->layers
== 0) != (layers
== 0)) {
2159 ice
->state
.dirty
|= IRIS_DIRTY_CLIP
;
2162 if (cso
->width
!= state
->width
|| cso
->height
!= state
->height
) {
2163 ice
->state
.dirty
|= IRIS_DIRTY_SF_CL_VIEWPORT
;
2166 util_copy_framebuffer_state(cso
, state
);
2167 cso
->samples
= samples
;
2168 cso
->layers
= layers
;
2170 struct iris_depth_buffer_state
*cso_z
= &ice
->state
.genx
->depth_buffer
;
2172 struct isl_view view
= {
2175 .base_array_layer
= 0,
2177 .swizzle
= ISL_SWIZZLE_IDENTITY
,
2180 struct isl_depth_stencil_hiz_emit_info info
= { .view
= &view
};
2183 iris_get_depth_stencil_resources(cso
->zsbuf
->texture
, &zres
,
2186 view
.base_level
= cso
->zsbuf
->u
.tex
.level
;
2187 view
.base_array_layer
= cso
->zsbuf
->u
.tex
.first_layer
;
2189 cso
->zsbuf
->u
.tex
.last_layer
- cso
->zsbuf
->u
.tex
.first_layer
+ 1;
2192 view
.usage
|= ISL_SURF_USAGE_DEPTH_BIT
;
2194 info
.depth_surf
= &zres
->surf
;
2195 info
.depth_address
= zres
->bo
->gtt_offset
;
2196 info
.mocs
= mocs(zres
->bo
);
2198 view
.format
= zres
->surf
.format
;
2202 view
.usage
|= ISL_SURF_USAGE_STENCIL_BIT
;
2203 info
.stencil_surf
= &stencil_res
->surf
;
2204 info
.stencil_address
= stencil_res
->bo
->gtt_offset
;
2206 view
.format
= stencil_res
->surf
.format
;
2207 info
.mocs
= mocs(stencil_res
->bo
);
2212 isl_emit_depth_stencil_hiz_s(isl_dev
, cso_z
->packets
, &info
);
2214 /* Make a null surface for unbound buffers */
2215 void *null_surf_map
=
2216 upload_state(ice
->state
.surface_uploader
, &ice
->state
.null_fb
,
2217 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
2218 isl_null_fill_state(&screen
->isl_dev
, null_surf_map
,
2219 isl_extent3d(MAX2(cso
->width
, 1),
2220 MAX2(cso
->height
, 1),
2221 cso
->layers
? cso
->layers
: 1));
2222 ice
->state
.null_fb
.offset
+=
2223 iris_bo_offset_from_base_address(iris_resource_bo(ice
->state
.null_fb
.res
));
2225 ice
->state
.dirty
|= IRIS_DIRTY_DEPTH_BUFFER
;
2227 /* Render target change */
2228 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_FS
;
2230 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_FRAMEBUFFER
];
2233 // XXX: we may want to flag IRIS_DIRTY_MULTISAMPLE (or SAMPLE_MASK?)
2234 // XXX: see commit 979fc1bc9bcc64027ff2cfafd285676f31b930a6
2236 /* The PIPE_CONTROL command description says:
2238 * "Whenever a Binding Table Index (BTI) used by a Render Target Message
2239 * points to a different RENDER_SURFACE_STATE, SW must issue a Render
2240 * Target Cache Flush by enabling this bit. When render target flush
2241 * is set due to new association of BTI, PS Scoreboard Stall bit must
2242 * be set in this packet."
2244 // XXX: does this need to happen at 3DSTATE_BTP_PS time?
2245 iris_emit_pipe_control_flush(&ice
->batches
[IRIS_BATCH_RENDER
],
2246 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
2247 PIPE_CONTROL_STALL_AT_SCOREBOARD
);
2252 upload_ubo_surf_state(struct iris_context
*ice
,
2253 struct iris_const_buffer
*cbuf
,
2254 unsigned buffer_size
)
2256 struct pipe_context
*ctx
= &ice
->ctx
;
2257 struct iris_screen
*screen
= (struct iris_screen
*) ctx
->screen
;
2259 // XXX: these are not retained forever, use a separate uploader?
2261 upload_state(ice
->state
.surface_uploader
, &cbuf
->surface_state
,
2262 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
2263 if (!unlikely(map
)) {
2264 pipe_resource_reference(&cbuf
->data
.res
, NULL
);
2268 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
2269 struct iris_bo
*surf_bo
= iris_resource_bo(cbuf
->surface_state
.res
);
2270 cbuf
->surface_state
.offset
+= iris_bo_offset_from_base_address(surf_bo
);
2272 isl_buffer_fill_state(&screen
->isl_dev
, map
,
2273 .address
= res
->bo
->gtt_offset
+ cbuf
->data
.offset
,
2274 .size_B
= MIN2(buffer_size
,
2275 res
->bo
->size
- cbuf
->data
.offset
),
2276 .format
= ISL_FORMAT_R32G32B32A32_FLOAT
,
2278 .mocs
= mocs(res
->bo
))
2282 * The pipe->set_constant_buffer() driver hook.
2284 * This uploads any constant data in user buffers, and references
2285 * any UBO resources containing constant data.
2288 iris_set_constant_buffer(struct pipe_context
*ctx
,
2289 enum pipe_shader_type p_stage
, unsigned index
,
2290 const struct pipe_constant_buffer
*input
)
2292 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2293 gl_shader_stage stage
= stage_from_pipe(p_stage
);
2294 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
2295 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[index
];
2297 if (input
&& input
->buffer
) {
2300 pipe_resource_reference(&cbuf
->data
.res
, input
->buffer
);
2301 cbuf
->data
.offset
= input
->buffer_offset
;
2303 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
2304 res
->bind_history
|= PIPE_BIND_CONSTANT_BUFFER
;
2306 upload_ubo_surf_state(ice
, cbuf
, input
->buffer_size
);
2308 pipe_resource_reference(&cbuf
->data
.res
, NULL
);
2309 pipe_resource_reference(&cbuf
->surface_state
.res
, NULL
);
2314 memcpy(&shs
->cbuf0
, input
, sizeof(shs
->cbuf0
));
2316 memset(&shs
->cbuf0
, 0, sizeof(shs
->cbuf0
));
2318 shs
->cbuf0_needs_upload
= true;
2321 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_VS
<< stage
;
2322 // XXX: maybe not necessary all the time...?
2323 // XXX: we need 3DS_BTP to commit these changes, and if we fell back to
2324 // XXX: pull model we may need actual new bindings...
2325 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_VS
<< stage
;
2329 upload_uniforms(struct iris_context
*ice
,
2330 gl_shader_stage stage
)
2332 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
2333 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[0];
2334 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
2336 unsigned upload_size
= shader
->num_system_values
* sizeof(uint32_t) +
2337 shs
->cbuf0
.buffer_size
;
2339 if (upload_size
== 0)
2343 upload_state(ice
->ctx
.const_uploader
, &cbuf
->data
, upload_size
, 64);
2345 for (int i
= 0; i
< shader
->num_system_values
; i
++) {
2346 uint32_t sysval
= shader
->system_values
[i
];
2349 if (BRW_PARAM_DOMAIN(sysval
) == BRW_PARAM_DOMAIN_IMAGE
) {
2350 unsigned img
= BRW_PARAM_IMAGE_IDX(sysval
);
2351 unsigned offset
= BRW_PARAM_IMAGE_OFFSET(sysval
);
2352 struct brw_image_param
*param
= &shs
->image
[img
].param
;
2354 assert(offset
< sizeof(struct brw_image_param
));
2355 value
= ((uint32_t *) param
)[offset
];
2356 } else if (sysval
== BRW_PARAM_BUILTIN_ZERO
) {
2358 } else if (BRW_PARAM_BUILTIN_IS_CLIP_PLANE(sysval
)) {
2359 int plane
= BRW_PARAM_BUILTIN_CLIP_PLANE_IDX(sysval
);
2360 int comp
= BRW_PARAM_BUILTIN_CLIP_PLANE_COMP(sysval
);
2361 value
= fui(ice
->state
.clip_planes
.ucp
[plane
][comp
]);
2362 } else if (sysval
== BRW_PARAM_BUILTIN_PATCH_VERTICES_IN
) {
2363 if (stage
== MESA_SHADER_TESS_CTRL
) {
2364 value
= ice
->state
.vertices_per_patch
;
2366 assert(stage
== MESA_SHADER_TESS_EVAL
);
2367 const struct shader_info
*tcs_info
=
2368 iris_get_shader_info(ice
, MESA_SHADER_TESS_CTRL
);
2371 value
= tcs_info
->tess
.tcs_vertices_out
;
2374 assert(!"unhandled system value");
2380 if (shs
->cbuf0
.user_buffer
) {
2381 memcpy(map
, shs
->cbuf0
.user_buffer
, shs
->cbuf0
.buffer_size
);
2384 upload_ubo_surf_state(ice
, cbuf
, upload_size
);
2388 * The pipe->set_shader_buffers() driver hook.
2390 * This binds SSBOs and ABOs. Unfortunately, we need to stream out
2391 * SURFACE_STATE here, as the buffer offset may change each time.
2394 iris_set_shader_buffers(struct pipe_context
*ctx
,
2395 enum pipe_shader_type p_stage
,
2396 unsigned start_slot
, unsigned count
,
2397 const struct pipe_shader_buffer
*buffers
)
2399 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2400 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
2401 gl_shader_stage stage
= stage_from_pipe(p_stage
);
2402 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
2404 for (unsigned i
= 0; i
< count
; i
++) {
2405 if (buffers
&& buffers
[i
].buffer
) {
2406 const struct pipe_shader_buffer
*buffer
= &buffers
[i
];
2407 struct iris_resource
*res
= (void *) buffer
->buffer
;
2408 pipe_resource_reference(&shs
->ssbo
[start_slot
+ i
], &res
->base
);
2410 res
->bind_history
|= PIPE_BIND_SHADER_BUFFER
;
2412 // XXX: these are not retained forever, use a separate uploader?
2414 upload_state(ice
->state
.surface_uploader
,
2415 &shs
->ssbo_surface_state
[start_slot
+ i
],
2416 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
2417 if (!unlikely(map
)) {
2418 pipe_resource_reference(&shs
->ssbo
[start_slot
+ i
], NULL
);
2422 struct iris_bo
*surf_state_bo
=
2423 iris_resource_bo(shs
->ssbo_surface_state
[start_slot
+ i
].res
);
2424 shs
->ssbo_surface_state
[start_slot
+ i
].offset
+=
2425 iris_bo_offset_from_base_address(surf_state_bo
);
2427 isl_buffer_fill_state(&screen
->isl_dev
, map
,
2429 res
->bo
->gtt_offset
+ buffer
->buffer_offset
,
2431 MIN2(buffer
->buffer_size
,
2432 res
->bo
->size
- buffer
->buffer_offset
),
2433 .format
= ISL_FORMAT_RAW
,
2435 .mocs
= mocs(res
->bo
));
2437 pipe_resource_reference(&shs
->ssbo
[start_slot
+ i
], NULL
);
2438 pipe_resource_reference(&shs
->ssbo_surface_state
[start_slot
+ i
].res
,
2443 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_VS
<< stage
;
2447 iris_delete_state(struct pipe_context
*ctx
, void *state
)
2453 * The pipe->set_vertex_buffers() driver hook.
2455 * This translates pipe_vertex_buffer to our 3DSTATE_VERTEX_BUFFERS packet.
2458 iris_set_vertex_buffers(struct pipe_context
*ctx
,
2459 unsigned start_slot
, unsigned count
,
2460 const struct pipe_vertex_buffer
*buffers
)
2462 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2463 struct iris_genx_state
*genx
= ice
->state
.genx
;
2465 ice
->state
.bound_vertex_buffers
&= ~u_bit_consecutive64(start_slot
, count
);
2467 for (unsigned i
= 0; i
< count
; i
++) {
2468 const struct pipe_vertex_buffer
*buffer
= buffers
? &buffers
[i
] : NULL
;
2469 struct iris_vertex_buffer_state
*state
=
2470 &genx
->vertex_buffers
[start_slot
+ i
];
2473 pipe_resource_reference(&state
->resource
, NULL
);
2477 assert(!buffer
->is_user_buffer
);
2479 ice
->state
.bound_vertex_buffers
|= 1ull << (start_slot
+ i
);
2481 pipe_resource_reference(&state
->resource
, buffer
->buffer
.resource
);
2482 struct iris_resource
*res
= (void *) state
->resource
;
2485 res
->bind_history
|= PIPE_BIND_VERTEX_BUFFER
;
2487 iris_pack_state(GENX(VERTEX_BUFFER_STATE
), state
->state
, vb
) {
2488 vb
.VertexBufferIndex
= start_slot
+ i
;
2489 vb
.AddressModifyEnable
= true;
2490 vb
.BufferPitch
= buffer
->stride
;
2492 vb
.BufferSize
= res
->bo
->size
;
2493 vb
.BufferStartingAddress
=
2494 ro_bo(NULL
, res
->bo
->gtt_offset
+ (int) buffer
->buffer_offset
);
2495 vb
.MOCS
= mocs(res
->bo
);
2497 vb
.NullVertexBuffer
= true;
2502 ice
->state
.dirty
|= IRIS_DIRTY_VERTEX_BUFFERS
;
2506 * Gallium CSO for vertex elements.
2508 struct iris_vertex_element_state
{
2509 uint32_t vertex_elements
[1 + 33 * GENX(VERTEX_ELEMENT_STATE_length
)];
2510 uint32_t vf_instancing
[33 * GENX(3DSTATE_VF_INSTANCING_length
)];
2515 * The pipe->create_vertex_elements() driver hook.
2517 * This translates pipe_vertex_element to our 3DSTATE_VERTEX_ELEMENTS
2518 * and 3DSTATE_VF_INSTANCING commands. SGVs are handled at draw time.
2521 iris_create_vertex_elements(struct pipe_context
*ctx
,
2523 const struct pipe_vertex_element
*state
)
2525 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
2526 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
2527 struct iris_vertex_element_state
*cso
=
2528 malloc(sizeof(struct iris_vertex_element_state
));
2533 * - create edge flag one
2535 * - if those are necessary, use count + 1/2/3... OR in the length
2537 iris_pack_command(GENX(3DSTATE_VERTEX_ELEMENTS
), cso
->vertex_elements
, ve
) {
2539 1 + GENX(VERTEX_ELEMENT_STATE_length
) * MAX2(count
, 1) - 2;
2542 uint32_t *ve_pack_dest
= &cso
->vertex_elements
[1];
2543 uint32_t *vfi_pack_dest
= cso
->vf_instancing
;
2546 iris_pack_state(GENX(VERTEX_ELEMENT_STATE
), ve_pack_dest
, ve
) {
2548 ve
.SourceElementFormat
= ISL_FORMAT_R32G32B32A32_FLOAT
;
2549 ve
.Component0Control
= VFCOMP_STORE_0
;
2550 ve
.Component1Control
= VFCOMP_STORE_0
;
2551 ve
.Component2Control
= VFCOMP_STORE_0
;
2552 ve
.Component3Control
= VFCOMP_STORE_1_FP
;
2555 iris_pack_command(GENX(3DSTATE_VF_INSTANCING
), vfi_pack_dest
, vi
) {
2559 for (int i
= 0; i
< count
; i
++) {
2560 const struct iris_format_info fmt
=
2561 iris_format_for_usage(devinfo
, state
[i
].src_format
, 0);
2562 unsigned comp
[4] = { VFCOMP_STORE_SRC
, VFCOMP_STORE_SRC
,
2563 VFCOMP_STORE_SRC
, VFCOMP_STORE_SRC
};
2565 switch (isl_format_get_num_channels(fmt
.fmt
)) {
2566 case 0: comp
[0] = VFCOMP_STORE_0
;
2567 case 1: comp
[1] = VFCOMP_STORE_0
;
2568 case 2: comp
[2] = VFCOMP_STORE_0
;
2570 comp
[3] = isl_format_has_int_channel(fmt
.fmt
) ? VFCOMP_STORE_1_INT
2571 : VFCOMP_STORE_1_FP
;
2574 iris_pack_state(GENX(VERTEX_ELEMENT_STATE
), ve_pack_dest
, ve
) {
2575 ve
.VertexBufferIndex
= state
[i
].vertex_buffer_index
;
2577 ve
.SourceElementOffset
= state
[i
].src_offset
;
2578 ve
.SourceElementFormat
= fmt
.fmt
;
2579 ve
.Component0Control
= comp
[0];
2580 ve
.Component1Control
= comp
[1];
2581 ve
.Component2Control
= comp
[2];
2582 ve
.Component3Control
= comp
[3];
2585 iris_pack_command(GENX(3DSTATE_VF_INSTANCING
), vfi_pack_dest
, vi
) {
2586 vi
.VertexElementIndex
= i
;
2587 vi
.InstancingEnable
= state
[i
].instance_divisor
> 0;
2588 vi
.InstanceDataStepRate
= state
[i
].instance_divisor
;
2591 ve_pack_dest
+= GENX(VERTEX_ELEMENT_STATE_length
);
2592 vfi_pack_dest
+= GENX(3DSTATE_VF_INSTANCING_length
);
2599 * The pipe->bind_vertex_elements_state() driver hook.
2602 iris_bind_vertex_elements_state(struct pipe_context
*ctx
, void *state
)
2604 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2605 struct iris_vertex_element_state
*old_cso
= ice
->state
.cso_vertex_elements
;
2606 struct iris_vertex_element_state
*new_cso
= state
;
2608 /* 3DSTATE_VF_SGVs overrides the last VE, so if the count is changing,
2609 * we need to re-emit it to ensure we're overriding the right one.
2611 if (new_cso
&& cso_changed(count
))
2612 ice
->state
.dirty
|= IRIS_DIRTY_VF_SGVS
;
2614 ice
->state
.cso_vertex_elements
= state
;
2615 ice
->state
.dirty
|= IRIS_DIRTY_VERTEX_ELEMENTS
;
2619 * The pipe->create_stream_output_target() driver hook.
2621 * "Target" here refers to a destination buffer. We translate this into
2622 * a 3DSTATE_SO_BUFFER packet. We can handle most fields, but don't yet
2623 * know which buffer this represents, or whether we ought to zero the
2624 * write-offsets, or append. Those are handled in the set() hook.
2626 static struct pipe_stream_output_target
*
2627 iris_create_stream_output_target(struct pipe_context
*ctx
,
2628 struct pipe_resource
*p_res
,
2629 unsigned buffer_offset
,
2630 unsigned buffer_size
)
2632 struct iris_resource
*res
= (void *) p_res
;
2633 struct iris_stream_output_target
*cso
= calloc(1, sizeof(*cso
));
2637 res
->bind_history
|= PIPE_BIND_STREAM_OUTPUT
;
2639 pipe_reference_init(&cso
->base
.reference
, 1);
2640 pipe_resource_reference(&cso
->base
.buffer
, p_res
);
2641 cso
->base
.buffer_offset
= buffer_offset
;
2642 cso
->base
.buffer_size
= buffer_size
;
2643 cso
->base
.context
= ctx
;
2645 upload_state(ctx
->stream_uploader
, &cso
->offset
, sizeof(uint32_t), 4);
2651 iris_stream_output_target_destroy(struct pipe_context
*ctx
,
2652 struct pipe_stream_output_target
*state
)
2654 struct iris_stream_output_target
*cso
= (void *) state
;
2656 pipe_resource_reference(&cso
->base
.buffer
, NULL
);
2657 pipe_resource_reference(&cso
->offset
.res
, NULL
);
2663 * The pipe->set_stream_output_targets() driver hook.
2665 * At this point, we know which targets are bound to a particular index,
2666 * and also whether we want to append or start over. We can finish the
2667 * 3DSTATE_SO_BUFFER packets we started earlier.
2670 iris_set_stream_output_targets(struct pipe_context
*ctx
,
2671 unsigned num_targets
,
2672 struct pipe_stream_output_target
**targets
,
2673 const unsigned *offsets
)
2675 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2676 struct iris_genx_state
*genx
= ice
->state
.genx
;
2677 uint32_t *so_buffers
= genx
->so_buffers
;
2679 const bool active
= num_targets
> 0;
2680 if (ice
->state
.streamout_active
!= active
) {
2681 ice
->state
.streamout_active
= active
;
2682 ice
->state
.dirty
|= IRIS_DIRTY_STREAMOUT
;
2684 /* We only emit 3DSTATE_SO_DECL_LIST when streamout is active, because
2685 * it's a non-pipelined command. If we're switching streamout on, we
2686 * may have missed emitting it earlier, so do so now. (We're already
2687 * taking a stall to update 3DSTATE_SO_BUFFERS anyway...)
2690 ice
->state
.dirty
|= IRIS_DIRTY_SO_DECL_LIST
;
2693 for (int i
= 0; i
< 4; i
++) {
2694 pipe_so_target_reference(&ice
->state
.so_target
[i
],
2695 i
< num_targets
? targets
[i
] : NULL
);
2698 /* No need to update 3DSTATE_SO_BUFFER unless SOL is active. */
2702 for (unsigned i
= 0; i
< 4; i
++,
2703 so_buffers
+= GENX(3DSTATE_SO_BUFFER_length
)) {
2705 if (i
>= num_targets
|| !targets
[i
]) {
2706 iris_pack_command(GENX(3DSTATE_SO_BUFFER
), so_buffers
, sob
)
2707 sob
.SOBufferIndex
= i
;
2711 struct iris_stream_output_target
*tgt
= (void *) targets
[i
];
2712 struct iris_resource
*res
= (void *) tgt
->base
.buffer
;
2714 /* Note that offsets[i] will either be 0, causing us to zero
2715 * the value in the buffer, or 0xFFFFFFFF, which happens to mean
2716 * "continue appending at the existing offset."
2718 assert(offsets
[i
] == 0 || offsets
[i
] == 0xFFFFFFFF);
2720 iris_pack_command(GENX(3DSTATE_SO_BUFFER
), so_buffers
, sob
) {
2721 sob
.SurfaceBaseAddress
=
2722 rw_bo(NULL
, res
->bo
->gtt_offset
+ tgt
->base
.buffer_offset
);
2723 sob
.SOBufferEnable
= true;
2724 sob
.StreamOffsetWriteEnable
= true;
2725 sob
.StreamOutputBufferOffsetAddressEnable
= true;
2726 sob
.MOCS
= mocs(res
->bo
);
2728 sob
.SurfaceSize
= MAX2(tgt
->base
.buffer_size
/ 4, 1) - 1;
2730 sob
.SOBufferIndex
= i
;
2731 sob
.StreamOffset
= offsets
[i
];
2732 sob
.StreamOutputBufferOffsetAddress
=
2733 rw_bo(NULL
, iris_resource_bo(tgt
->offset
.res
)->gtt_offset
+
2734 tgt
->offset
.offset
);
2738 ice
->state
.dirty
|= IRIS_DIRTY_SO_BUFFERS
;
2742 * An iris-vtable helper for encoding the 3DSTATE_SO_DECL_LIST and
2743 * 3DSTATE_STREAMOUT packets.
2745 * 3DSTATE_SO_DECL_LIST is a list of shader outputs we want the streamout
2746 * hardware to record. We can create it entirely based on the shader, with
2747 * no dynamic state dependencies.
2749 * 3DSTATE_STREAMOUT is an annoying mix of shader-based information and
2750 * state-based settings. We capture the shader-related ones here, and merge
2751 * the rest in at draw time.
2754 iris_create_so_decl_list(const struct pipe_stream_output_info
*info
,
2755 const struct brw_vue_map
*vue_map
)
2757 struct GENX(SO_DECL
) so_decl
[MAX_VERTEX_STREAMS
][128];
2758 int buffer_mask
[MAX_VERTEX_STREAMS
] = {0, 0, 0, 0};
2759 int next_offset
[MAX_VERTEX_STREAMS
] = {0, 0, 0, 0};
2760 int decls
[MAX_VERTEX_STREAMS
] = {0, 0, 0, 0};
2762 STATIC_ASSERT(ARRAY_SIZE(so_decl
[0]) >= MAX_PROGRAM_OUTPUTS
);
2764 memset(so_decl
, 0, sizeof(so_decl
));
2766 /* Construct the list of SO_DECLs to be emitted. The formatting of the
2767 * command feels strange -- each dword pair contains a SO_DECL per stream.
2769 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
2770 const struct pipe_stream_output
*output
= &info
->output
[i
];
2771 const int buffer
= output
->output_buffer
;
2772 const int varying
= output
->register_index
;
2773 const unsigned stream_id
= output
->stream
;
2774 assert(stream_id
< MAX_VERTEX_STREAMS
);
2776 buffer_mask
[stream_id
] |= 1 << buffer
;
2778 assert(vue_map
->varying_to_slot
[varying
] >= 0);
2780 /* Mesa doesn't store entries for gl_SkipComponents in the Outputs[]
2781 * array. Instead, it simply increments DstOffset for the following
2782 * input by the number of components that should be skipped.
2784 * Our hardware is unusual in that it requires us to program SO_DECLs
2785 * for fake "hole" components, rather than simply taking the offset
2786 * for each real varying. Each hole can have size 1, 2, 3, or 4; we
2787 * program as many size = 4 holes as we can, then a final hole to
2788 * accommodate the final 1, 2, or 3 remaining.
2790 int skip_components
= output
->dst_offset
- next_offset
[buffer
];
2792 while (skip_components
> 0) {
2793 so_decl
[stream_id
][decls
[stream_id
]++] = (struct GENX(SO_DECL
)) {
2795 .OutputBufferSlot
= output
->output_buffer
,
2796 .ComponentMask
= (1 << MIN2(skip_components
, 4)) - 1,
2798 skip_components
-= 4;
2801 next_offset
[buffer
] = output
->dst_offset
+ output
->num_components
;
2803 so_decl
[stream_id
][decls
[stream_id
]++] = (struct GENX(SO_DECL
)) {
2804 .OutputBufferSlot
= output
->output_buffer
,
2805 .RegisterIndex
= vue_map
->varying_to_slot
[varying
],
2807 ((1 << output
->num_components
) - 1) << output
->start_component
,
2810 if (decls
[stream_id
] > max_decls
)
2811 max_decls
= decls
[stream_id
];
2814 unsigned dwords
= GENX(3DSTATE_STREAMOUT_length
) + (3 + 2 * max_decls
);
2815 uint32_t *map
= ralloc_size(NULL
, sizeof(uint32_t) * dwords
);
2816 uint32_t *so_decl_map
= map
+ GENX(3DSTATE_STREAMOUT_length
);
2818 iris_pack_command(GENX(3DSTATE_STREAMOUT
), map
, sol
) {
2819 int urb_entry_read_offset
= 0;
2820 int urb_entry_read_length
= (vue_map
->num_slots
+ 1) / 2 -
2821 urb_entry_read_offset
;
2823 /* We always read the whole vertex. This could be reduced at some
2824 * point by reading less and offsetting the register index in the
2827 sol
.Stream0VertexReadOffset
= urb_entry_read_offset
;
2828 sol
.Stream0VertexReadLength
= urb_entry_read_length
- 1;
2829 sol
.Stream1VertexReadOffset
= urb_entry_read_offset
;
2830 sol
.Stream1VertexReadLength
= urb_entry_read_length
- 1;
2831 sol
.Stream2VertexReadOffset
= urb_entry_read_offset
;
2832 sol
.Stream2VertexReadLength
= urb_entry_read_length
- 1;
2833 sol
.Stream3VertexReadOffset
= urb_entry_read_offset
;
2834 sol
.Stream3VertexReadLength
= urb_entry_read_length
- 1;
2836 /* Set buffer pitches; 0 means unbound. */
2837 sol
.Buffer0SurfacePitch
= 4 * info
->stride
[0];
2838 sol
.Buffer1SurfacePitch
= 4 * info
->stride
[1];
2839 sol
.Buffer2SurfacePitch
= 4 * info
->stride
[2];
2840 sol
.Buffer3SurfacePitch
= 4 * info
->stride
[3];
2843 iris_pack_command(GENX(3DSTATE_SO_DECL_LIST
), so_decl_map
, list
) {
2844 list
.DWordLength
= 3 + 2 * max_decls
- 2;
2845 list
.StreamtoBufferSelects0
= buffer_mask
[0];
2846 list
.StreamtoBufferSelects1
= buffer_mask
[1];
2847 list
.StreamtoBufferSelects2
= buffer_mask
[2];
2848 list
.StreamtoBufferSelects3
= buffer_mask
[3];
2849 list
.NumEntries0
= decls
[0];
2850 list
.NumEntries1
= decls
[1];
2851 list
.NumEntries2
= decls
[2];
2852 list
.NumEntries3
= decls
[3];
2855 for (int i
= 0; i
< max_decls
; i
++) {
2856 iris_pack_state(GENX(SO_DECL_ENTRY
), so_decl_map
+ 3 + i
* 2, entry
) {
2857 entry
.Stream0Decl
= so_decl
[0][i
];
2858 entry
.Stream1Decl
= so_decl
[1][i
];
2859 entry
.Stream2Decl
= so_decl
[2][i
];
2860 entry
.Stream3Decl
= so_decl
[3][i
];
2868 iris_compute_sbe_urb_read_interval(uint64_t fs_input_slots
,
2869 const struct brw_vue_map
*last_vue_map
,
2870 bool two_sided_color
,
2871 unsigned *out_offset
,
2872 unsigned *out_length
)
2874 /* The compiler computes the first URB slot without considering COL/BFC
2875 * swizzling (because it doesn't know whether it's enabled), so we need
2876 * to do that here too. This may result in a smaller offset, which
2879 const unsigned first_slot
=
2880 brw_compute_first_urb_slot_required(fs_input_slots
, last_vue_map
);
2882 /* This becomes the URB read offset (counted in pairs of slots). */
2883 assert(first_slot
% 2 == 0);
2884 *out_offset
= first_slot
/ 2;
2886 /* We need to adjust the inputs read to account for front/back color
2887 * swizzling, as it can make the URB length longer.
2889 for (int c
= 0; c
<= 1; c
++) {
2890 if (fs_input_slots
& (VARYING_BIT_COL0
<< c
)) {
2891 /* If two sided color is enabled, the fragment shader's gl_Color
2892 * (COL0) input comes from either the gl_FrontColor (COL0) or
2893 * gl_BackColor (BFC0) input varyings. Mark BFC as used, too.
2895 if (two_sided_color
)
2896 fs_input_slots
|= (VARYING_BIT_BFC0
<< c
);
2898 /* If front color isn't written, we opt to give them back color
2899 * instead of an undefined value. Switch from COL to BFC.
2901 if (last_vue_map
->varying_to_slot
[VARYING_SLOT_COL0
+ c
] == -1) {
2902 fs_input_slots
&= ~(VARYING_BIT_COL0
<< c
);
2903 fs_input_slots
|= (VARYING_BIT_BFC0
<< c
);
2908 /* Compute the minimum URB Read Length necessary for the FS inputs.
2910 * From the Sandy Bridge PRM, Volume 2, Part 1, documentation for
2911 * 3DSTATE_SF DWord 1 bits 15:11, "Vertex URB Entry Read Length":
2913 * "This field should be set to the minimum length required to read the
2914 * maximum source attribute. The maximum source attribute is indicated
2915 * by the maximum value of the enabled Attribute # Source Attribute if
2916 * Attribute Swizzle Enable is set, Number of Output Attributes-1 if
2917 * enable is not set.
2918 * read_length = ceiling((max_source_attr + 1) / 2)
2920 * [errata] Corruption/Hang possible if length programmed larger than
2923 * Similar text exists for Ivy Bridge.
2925 * We find the last URB slot that's actually read by the FS.
2927 unsigned last_read_slot
= last_vue_map
->num_slots
- 1;
2928 while (last_read_slot
> first_slot
&& !(fs_input_slots
&
2929 (1ull << last_vue_map
->slot_to_varying
[last_read_slot
])))
2932 /* The URB read length is the difference of the two, counted in pairs. */
2933 *out_length
= DIV_ROUND_UP(last_read_slot
- first_slot
+ 1, 2);
2937 iris_emit_sbe_swiz(struct iris_batch
*batch
,
2938 const struct iris_context
*ice
,
2939 unsigned urb_read_offset
,
2940 unsigned sprite_coord_enables
)
2942 struct GENX(SF_OUTPUT_ATTRIBUTE_DETAIL
) attr_overrides
[16] = {};
2943 const struct brw_wm_prog_data
*wm_prog_data
= (void *)
2944 ice
->shaders
.prog
[MESA_SHADER_FRAGMENT
]->prog_data
;
2945 const struct brw_vue_map
*vue_map
= ice
->shaders
.last_vue_map
;
2946 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
2948 /* XXX: this should be generated when putting programs in place */
2950 for (int fs_attr
= 0; fs_attr
< VARYING_SLOT_MAX
; fs_attr
++) {
2951 const int input_index
= wm_prog_data
->urb_setup
[fs_attr
];
2952 if (input_index
< 0 || input_index
>= 16)
2955 struct GENX(SF_OUTPUT_ATTRIBUTE_DETAIL
) *attr
=
2956 &attr_overrides
[input_index
];
2957 int slot
= vue_map
->varying_to_slot
[fs_attr
];
2959 /* Viewport and Layer are stored in the VUE header. We need to override
2960 * them to zero if earlier stages didn't write them, as GL requires that
2961 * they read back as zero when not explicitly set.
2964 case VARYING_SLOT_VIEWPORT
:
2965 case VARYING_SLOT_LAYER
:
2966 attr
->ComponentOverrideX
= true;
2967 attr
->ComponentOverrideW
= true;
2968 attr
->ConstantSource
= CONST_0000
;
2970 if (!(vue_map
->slots_valid
& VARYING_BIT_LAYER
))
2971 attr
->ComponentOverrideY
= true;
2972 if (!(vue_map
->slots_valid
& VARYING_BIT_VIEWPORT
))
2973 attr
->ComponentOverrideZ
= true;
2976 case VARYING_SLOT_PRIMITIVE_ID
:
2977 /* Override if the previous shader stage didn't write gl_PrimitiveID. */
2979 attr
->ComponentOverrideX
= true;
2980 attr
->ComponentOverrideY
= true;
2981 attr
->ComponentOverrideZ
= true;
2982 attr
->ComponentOverrideW
= true;
2983 attr
->ConstantSource
= PRIM_ID
;
2991 if (sprite_coord_enables
& (1 << input_index
))
2994 /* If there was only a back color written but not front, use back
2995 * as the color instead of undefined.
2997 if (slot
== -1 && fs_attr
== VARYING_SLOT_COL0
)
2998 slot
= vue_map
->varying_to_slot
[VARYING_SLOT_BFC0
];
2999 if (slot
== -1 && fs_attr
== VARYING_SLOT_COL1
)
3000 slot
= vue_map
->varying_to_slot
[VARYING_SLOT_BFC1
];
3002 /* Not written by the previous stage - undefined. */
3004 attr
->ComponentOverrideX
= true;
3005 attr
->ComponentOverrideY
= true;
3006 attr
->ComponentOverrideZ
= true;
3007 attr
->ComponentOverrideW
= true;
3008 attr
->ConstantSource
= CONST_0001_FLOAT
;
3012 /* Compute the location of the attribute relative to the read offset,
3013 * which is counted in 256-bit increments (two 128-bit VUE slots).
3015 const int source_attr
= slot
- 2 * urb_read_offset
;
3016 assert(source_attr
>= 0 && source_attr
<= 32);
3017 attr
->SourceAttribute
= source_attr
;
3019 /* If we are doing two-sided color, and the VUE slot following this one
3020 * represents a back-facing color, then we need to instruct the SF unit
3021 * to do back-facing swizzling.
3023 if (cso_rast
->light_twoside
&&
3024 ((vue_map
->slot_to_varying
[slot
] == VARYING_SLOT_COL0
&&
3025 vue_map
->slot_to_varying
[slot
+1] == VARYING_SLOT_BFC0
) ||
3026 (vue_map
->slot_to_varying
[slot
] == VARYING_SLOT_COL1
&&
3027 vue_map
->slot_to_varying
[slot
+1] == VARYING_SLOT_BFC1
)))
3028 attr
->SwizzleSelect
= INPUTATTR_FACING
;
3031 iris_emit_cmd(batch
, GENX(3DSTATE_SBE_SWIZ
), sbes
) {
3032 for (int i
= 0; i
< 16; i
++)
3033 sbes
.Attribute
[i
] = attr_overrides
[i
];
3038 iris_calculate_point_sprite_overrides(const struct brw_wm_prog_data
*prog_data
,
3039 const struct iris_rasterizer_state
*cso
)
3041 unsigned overrides
= 0;
3043 if (prog_data
->urb_setup
[VARYING_SLOT_PNTC
] != -1)
3044 overrides
|= 1 << prog_data
->urb_setup
[VARYING_SLOT_PNTC
];
3046 for (int i
= 0; i
< 8; i
++) {
3047 if ((cso
->sprite_coord_enable
& (1 << i
)) &&
3048 prog_data
->urb_setup
[VARYING_SLOT_TEX0
+ i
] != -1)
3049 overrides
|= 1 << prog_data
->urb_setup
[VARYING_SLOT_TEX0
+ i
];
3056 iris_emit_sbe(struct iris_batch
*batch
, const struct iris_context
*ice
)
3058 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
3059 const struct brw_wm_prog_data
*wm_prog_data
= (void *)
3060 ice
->shaders
.prog
[MESA_SHADER_FRAGMENT
]->prog_data
;
3061 const struct shader_info
*fs_info
=
3062 iris_get_shader_info(ice
, MESA_SHADER_FRAGMENT
);
3064 unsigned urb_read_offset
, urb_read_length
;
3065 iris_compute_sbe_urb_read_interval(fs_info
->inputs_read
,
3066 ice
->shaders
.last_vue_map
,
3067 cso_rast
->light_twoside
,
3068 &urb_read_offset
, &urb_read_length
);
3070 unsigned sprite_coord_overrides
=
3071 iris_calculate_point_sprite_overrides(wm_prog_data
, cso_rast
);
3073 iris_emit_cmd(batch
, GENX(3DSTATE_SBE
), sbe
) {
3074 sbe
.AttributeSwizzleEnable
= true;
3075 sbe
.NumberofSFOutputAttributes
= wm_prog_data
->num_varying_inputs
;
3076 sbe
.PointSpriteTextureCoordinateOrigin
= cso_rast
->sprite_coord_mode
;
3077 sbe
.VertexURBEntryReadOffset
= urb_read_offset
;
3078 sbe
.VertexURBEntryReadLength
= urb_read_length
;
3079 sbe
.ForceVertexURBEntryReadOffset
= true;
3080 sbe
.ForceVertexURBEntryReadLength
= true;
3081 sbe
.ConstantInterpolationEnable
= wm_prog_data
->flat_inputs
;
3082 sbe
.PointSpriteTextureCoordinateEnable
= sprite_coord_overrides
;
3084 for (int i
= 0; i
< 32; i
++) {
3085 sbe
.AttributeActiveComponentFormat
[i
] = ACTIVE_COMPONENT_XYZW
;
3090 iris_emit_sbe_swiz(batch
, ice
, urb_read_offset
, sprite_coord_overrides
);
3093 /* ------------------------------------------------------------------- */
3096 * Populate VS program key fields based on the current state.
3099 iris_populate_vs_key(const struct iris_context
*ice
,
3100 const struct shader_info
*info
,
3101 struct brw_vs_prog_key
*key
)
3103 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
3105 if (info
->clip_distance_array_size
== 0 &&
3106 (info
->outputs_written
& (VARYING_BIT_POS
| VARYING_BIT_CLIP_VERTEX
)))
3107 key
->nr_userclip_plane_consts
= cso_rast
->num_clip_plane_consts
;
3111 * Populate TCS program key fields based on the current state.
3114 iris_populate_tcs_key(const struct iris_context
*ice
,
3115 struct brw_tcs_prog_key
*key
)
3120 * Populate TES program key fields based on the current state.
3123 iris_populate_tes_key(const struct iris_context
*ice
,
3124 struct brw_tes_prog_key
*key
)
3129 * Populate GS program key fields based on the current state.
3132 iris_populate_gs_key(const struct iris_context
*ice
,
3133 struct brw_gs_prog_key
*key
)
3138 * Populate FS program key fields based on the current state.
3141 iris_populate_fs_key(const struct iris_context
*ice
,
3142 struct brw_wm_prog_key
*key
)
3144 const struct pipe_framebuffer_state
*fb
= &ice
->state
.framebuffer
;
3145 const struct iris_depth_stencil_alpha_state
*zsa
= ice
->state
.cso_zsa
;
3146 const struct iris_rasterizer_state
*rast
= ice
->state
.cso_rast
;
3147 const struct iris_blend_state
*blend
= ice
->state
.cso_blend
;
3149 key
->nr_color_regions
= fb
->nr_cbufs
;
3151 key
->clamp_fragment_color
= rast
->clamp_fragment_color
;
3153 key
->replicate_alpha
= fb
->nr_cbufs
> 1 &&
3154 (zsa
->alpha
.enabled
|| blend
->alpha_to_coverage
);
3156 /* XXX: only bother if COL0/1 are read */
3157 key
->flat_shade
= rast
->flatshade
;
3159 key
->persample_interp
= rast
->force_persample_interp
;
3160 key
->multisample_fbo
= rast
->multisample
&& fb
->samples
> 1;
3162 key
->coherent_fb_fetch
= true;
3164 // XXX: key->force_dual_color_blend for unigine
3165 // XXX: respect hint for high_quality_derivatives:1;
3169 iris_populate_cs_key(const struct iris_context
*ice
,
3170 struct brw_cs_prog_key
*key
)
3175 // XXX: these need to go in INIT_THREAD_DISPATCH_FIELDS
3176 pkt
.SamplerCount
= \
3177 DIV_ROUND_UP(CLAMP(stage_state
->sampler_count
, 0, 16), 4); \
3182 KSP(const struct iris_compiled_shader
*shader
)
3184 struct iris_resource
*res
= (void *) shader
->assembly
.res
;
3185 return iris_bo_offset_from_base_address(res
->bo
) + shader
->assembly
.offset
;
3188 // Gen11 workaround table #2056 WABTPPrefetchDisable suggests to disable
3189 // prefetching of binding tables in A0 and B0 steppings. XXX: Revisit
3190 // this WA on C0 stepping.
3192 #define INIT_THREAD_DISPATCH_FIELDS(pkt, prefix, stage) \
3193 pkt.KernelStartPointer = KSP(shader); \
3194 pkt.BindingTableEntryCount = GEN_GEN == 11 ? 0 : \
3195 prog_data->binding_table.size_bytes / 4; \
3196 pkt.FloatingPointMode = prog_data->use_alt_mode; \
3198 pkt.DispatchGRFStartRegisterForURBData = \
3199 prog_data->dispatch_grf_start_reg; \
3200 pkt.prefix##URBEntryReadLength = vue_prog_data->urb_read_length; \
3201 pkt.prefix##URBEntryReadOffset = 0; \
3203 pkt.StatisticsEnable = true; \
3204 pkt.Enable = true; \
3206 if (prog_data->total_scratch) { \
3207 struct iris_bo *bo = \
3208 iris_get_scratch_space(ice, prog_data->total_scratch, stage); \
3209 uint32_t scratch_addr = bo->gtt_offset; \
3210 pkt.PerThreadScratchSpace = ffs(prog_data->total_scratch) - 11; \
3211 pkt.ScratchSpaceBasePointer = rw_bo(NULL, scratch_addr); \
3215 * Encode most of 3DSTATE_VS based on the compiled shader.
3218 iris_store_vs_state(struct iris_context
*ice
,
3219 const struct gen_device_info
*devinfo
,
3220 struct iris_compiled_shader
*shader
)
3222 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3223 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3225 iris_pack_command(GENX(3DSTATE_VS
), shader
->derived_data
, vs
) {
3226 INIT_THREAD_DISPATCH_FIELDS(vs
, Vertex
, MESA_SHADER_VERTEX
);
3227 vs
.MaximumNumberofThreads
= devinfo
->max_vs_threads
- 1;
3228 vs
.SIMD8DispatchEnable
= true;
3229 vs
.UserClipDistanceCullTestEnableBitmask
=
3230 vue_prog_data
->cull_distance_mask
;
3235 * Encode most of 3DSTATE_HS based on the compiled shader.
3238 iris_store_tcs_state(struct iris_context
*ice
,
3239 const struct gen_device_info
*devinfo
,
3240 struct iris_compiled_shader
*shader
)
3242 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3243 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3244 struct brw_tcs_prog_data
*tcs_prog_data
= (void *) prog_data
;
3246 iris_pack_command(GENX(3DSTATE_HS
), shader
->derived_data
, hs
) {
3247 INIT_THREAD_DISPATCH_FIELDS(hs
, Vertex
, MESA_SHADER_TESS_CTRL
);
3249 hs
.InstanceCount
= tcs_prog_data
->instances
- 1;
3250 hs
.MaximumNumberofThreads
= devinfo
->max_tcs_threads
- 1;
3251 hs
.IncludeVertexHandles
= true;
3256 * Encode 3DSTATE_TE and most of 3DSTATE_DS based on the compiled shader.
3259 iris_store_tes_state(struct iris_context
*ice
,
3260 const struct gen_device_info
*devinfo
,
3261 struct iris_compiled_shader
*shader
)
3263 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3264 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3265 struct brw_tes_prog_data
*tes_prog_data
= (void *) prog_data
;
3267 uint32_t *te_state
= (void *) shader
->derived_data
;
3268 uint32_t *ds_state
= te_state
+ GENX(3DSTATE_TE_length
);
3270 iris_pack_command(GENX(3DSTATE_TE
), te_state
, te
) {
3271 te
.Partitioning
= tes_prog_data
->partitioning
;
3272 te
.OutputTopology
= tes_prog_data
->output_topology
;
3273 te
.TEDomain
= tes_prog_data
->domain
;
3275 te
.MaximumTessellationFactorOdd
= 63.0;
3276 te
.MaximumTessellationFactorNotOdd
= 64.0;
3279 iris_pack_command(GENX(3DSTATE_DS
), ds_state
, ds
) {
3280 INIT_THREAD_DISPATCH_FIELDS(ds
, Patch
, MESA_SHADER_TESS_EVAL
);
3282 ds
.DispatchMode
= DISPATCH_MODE_SIMD8_SINGLE_PATCH
;
3283 ds
.MaximumNumberofThreads
= devinfo
->max_tes_threads
- 1;
3284 ds
.ComputeWCoordinateEnable
=
3285 tes_prog_data
->domain
== BRW_TESS_DOMAIN_TRI
;
3287 ds
.UserClipDistanceCullTestEnableBitmask
=
3288 vue_prog_data
->cull_distance_mask
;
3294 * Encode most of 3DSTATE_GS based on the compiled shader.
3297 iris_store_gs_state(struct iris_context
*ice
,
3298 const struct gen_device_info
*devinfo
,
3299 struct iris_compiled_shader
*shader
)
3301 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3302 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3303 struct brw_gs_prog_data
*gs_prog_data
= (void *) prog_data
;
3305 iris_pack_command(GENX(3DSTATE_GS
), shader
->derived_data
, gs
) {
3306 INIT_THREAD_DISPATCH_FIELDS(gs
, Vertex
, MESA_SHADER_GEOMETRY
);
3308 gs
.OutputVertexSize
= gs_prog_data
->output_vertex_size_hwords
* 2 - 1;
3309 gs
.OutputTopology
= gs_prog_data
->output_topology
;
3310 gs
.ControlDataHeaderSize
=
3311 gs_prog_data
->control_data_header_size_hwords
;
3312 gs
.InstanceControl
= gs_prog_data
->invocations
- 1;
3313 gs
.DispatchMode
= DISPATCH_MODE_SIMD8
;
3314 gs
.IncludePrimitiveID
= gs_prog_data
->include_primitive_id
;
3315 gs
.ControlDataFormat
= gs_prog_data
->control_data_format
;
3316 gs
.ReorderMode
= TRAILING
;
3317 gs
.ExpectedVertexCount
= gs_prog_data
->vertices_in
;
3318 gs
.MaximumNumberofThreads
=
3319 GEN_GEN
== 8 ? (devinfo
->max_gs_threads
/ 2 - 1)
3320 : (devinfo
->max_gs_threads
- 1);
3322 if (gs_prog_data
->static_vertex_count
!= -1) {
3323 gs
.StaticOutput
= true;
3324 gs
.StaticOutputVertexCount
= gs_prog_data
->static_vertex_count
;
3326 gs
.IncludeVertexHandles
= vue_prog_data
->include_vue_handles
;
3328 gs
.UserClipDistanceCullTestEnableBitmask
=
3329 vue_prog_data
->cull_distance_mask
;
3331 const int urb_entry_write_offset
= 1;
3332 const uint32_t urb_entry_output_length
=
3333 DIV_ROUND_UP(vue_prog_data
->vue_map
.num_slots
, 2) -
3334 urb_entry_write_offset
;
3336 gs
.VertexURBEntryOutputReadOffset
= urb_entry_write_offset
;
3337 gs
.VertexURBEntryOutputLength
= MAX2(urb_entry_output_length
, 1);
3342 * Encode most of 3DSTATE_PS and 3DSTATE_PS_EXTRA based on the shader.
3345 iris_store_fs_state(struct iris_context
*ice
,
3346 const struct gen_device_info
*devinfo
,
3347 struct iris_compiled_shader
*shader
)
3349 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3350 struct brw_wm_prog_data
*wm_prog_data
= (void *) shader
->prog_data
;
3352 uint32_t *ps_state
= (void *) shader
->derived_data
;
3353 uint32_t *psx_state
= ps_state
+ GENX(3DSTATE_PS_length
);
3355 iris_pack_command(GENX(3DSTATE_PS
), ps_state
, ps
) {
3356 ps
.VectorMaskEnable
= true;
3357 //ps.SamplerCount = ...
3358 // XXX: WABTPPrefetchDisable, see above, drop at C0
3359 ps
.BindingTableEntryCount
= GEN_GEN
== 11 ? 0 :
3360 prog_data
->binding_table
.size_bytes
/ 4;
3361 ps
.FloatingPointMode
= prog_data
->use_alt_mode
;
3362 ps
.MaximumNumberofThreadsPerPSD
= 64 - (GEN_GEN
== 8 ? 2 : 1);
3364 ps
.PushConstantEnable
= prog_data
->ubo_ranges
[0].length
> 0;
3366 /* From the documentation for this packet:
3367 * "If the PS kernel does not need the Position XY Offsets to
3368 * compute a Position Value, then this field should be programmed
3369 * to POSOFFSET_NONE."
3371 * "SW Recommendation: If the PS kernel needs the Position Offsets
3372 * to compute a Position XY value, this field should match Position
3373 * ZW Interpolation Mode to ensure a consistent position.xyzw
3376 * We only require XY sample offsets. So, this recommendation doesn't
3377 * look useful at the moment. We might need this in future.
3379 ps
.PositionXYOffsetSelect
=
3380 wm_prog_data
->uses_pos_offset
? POSOFFSET_SAMPLE
: POSOFFSET_NONE
;
3381 ps
._8PixelDispatchEnable
= wm_prog_data
->dispatch_8
;
3382 ps
._16PixelDispatchEnable
= wm_prog_data
->dispatch_16
;
3383 ps
._32PixelDispatchEnable
= wm_prog_data
->dispatch_32
;
3385 // XXX: Disable SIMD32 with 16x MSAA
3387 ps
.DispatchGRFStartRegisterForConstantSetupData0
=
3388 brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data
, ps
, 0);
3389 ps
.DispatchGRFStartRegisterForConstantSetupData1
=
3390 brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data
, ps
, 1);
3391 ps
.DispatchGRFStartRegisterForConstantSetupData2
=
3392 brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data
, ps
, 2);
3394 ps
.KernelStartPointer0
=
3395 KSP(shader
) + brw_wm_prog_data_prog_offset(wm_prog_data
, ps
, 0);
3396 ps
.KernelStartPointer1
=
3397 KSP(shader
) + brw_wm_prog_data_prog_offset(wm_prog_data
, ps
, 1);
3398 ps
.KernelStartPointer2
=
3399 KSP(shader
) + brw_wm_prog_data_prog_offset(wm_prog_data
, ps
, 2);
3401 if (prog_data
->total_scratch
) {
3402 struct iris_bo
*bo
=
3403 iris_get_scratch_space(ice
, prog_data
->total_scratch
,
3404 MESA_SHADER_FRAGMENT
);
3405 uint32_t scratch_addr
= bo
->gtt_offset
;
3406 ps
.PerThreadScratchSpace
= ffs(prog_data
->total_scratch
) - 11;
3407 ps
.ScratchSpaceBasePointer
= rw_bo(NULL
, scratch_addr
);
3411 iris_pack_command(GENX(3DSTATE_PS_EXTRA
), psx_state
, psx
) {
3412 psx
.PixelShaderValid
= true;
3413 psx
.PixelShaderComputedDepthMode
= wm_prog_data
->computed_depth_mode
;
3414 // XXX: alpha test / alpha to coverage :/
3415 psx
.PixelShaderKillsPixel
= wm_prog_data
->uses_kill
||
3416 wm_prog_data
->uses_omask
;
3417 psx
.AttributeEnable
= wm_prog_data
->num_varying_inputs
!= 0;
3418 psx
.PixelShaderUsesSourceDepth
= wm_prog_data
->uses_src_depth
;
3419 psx
.PixelShaderUsesSourceW
= wm_prog_data
->uses_src_w
;
3420 psx
.PixelShaderIsPerSample
= wm_prog_data
->persample_dispatch
;
3421 psx
.oMaskPresenttoRenderTarget
= wm_prog_data
->uses_omask
;
3424 if (wm_prog_data
->uses_sample_mask
) {
3425 /* TODO: conservative rasterization */
3426 if (wm_prog_data
->post_depth_coverage
)
3427 psx
.InputCoverageMaskState
= ICMS_DEPTH_COVERAGE
;
3429 psx
.InputCoverageMaskState
= ICMS_NORMAL
;
3432 psx
.PixelShaderPullsBary
= wm_prog_data
->pulls_bary
;
3433 psx
.PixelShaderComputesStencil
= wm_prog_data
->computed_stencil
;
3435 psx
.PixelShaderUsesInputCoverageMask
= wm_prog_data
->uses_sample_mask
;
3442 * Compute the size of the derived data (shader command packets).
3444 * This must match the data written by the iris_store_xs_state() functions.
3447 iris_store_cs_state(struct iris_context
*ice
,
3448 const struct gen_device_info
*devinfo
,
3449 struct iris_compiled_shader
*shader
)
3451 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3452 struct brw_cs_prog_data
*cs_prog_data
= (void *) shader
->prog_data
;
3453 void *map
= shader
->derived_data
;
3455 iris_pack_state(GENX(INTERFACE_DESCRIPTOR_DATA
), map
, desc
) {
3456 desc
.KernelStartPointer
= KSP(shader
);
3457 desc
.ConstantURBEntryReadLength
= cs_prog_data
->push
.per_thread
.regs
;
3458 desc
.NumberofThreadsinGPGPUThreadGroup
= cs_prog_data
->threads
;
3459 desc
.SharedLocalMemorySize
=
3460 encode_slm_size(GEN_GEN
, prog_data
->total_shared
);
3461 desc
.BarrierEnable
= cs_prog_data
->uses_barrier
;
3462 desc
.CrossThreadConstantDataReadLength
=
3463 cs_prog_data
->push
.cross_thread
.regs
;
3468 iris_derived_program_state_size(enum iris_program_cache_id cache_id
)
3470 assert(cache_id
<= IRIS_CACHE_BLORP
);
3472 static const unsigned dwords
[] = {
3473 [IRIS_CACHE_VS
] = GENX(3DSTATE_VS_length
),
3474 [IRIS_CACHE_TCS
] = GENX(3DSTATE_HS_length
),
3475 [IRIS_CACHE_TES
] = GENX(3DSTATE_TE_length
) + GENX(3DSTATE_DS_length
),
3476 [IRIS_CACHE_GS
] = GENX(3DSTATE_GS_length
),
3478 GENX(3DSTATE_PS_length
) + GENX(3DSTATE_PS_EXTRA_length
),
3479 [IRIS_CACHE_CS
] = GENX(INTERFACE_DESCRIPTOR_DATA_length
),
3480 [IRIS_CACHE_BLORP
] = 0,
3483 return sizeof(uint32_t) * dwords
[cache_id
];
3487 * Create any state packets corresponding to the given shader stage
3488 * (i.e. 3DSTATE_VS) and save them as "derived data" in the shader variant.
3489 * This means that we can look up a program in the in-memory cache and
3490 * get most of the state packet without having to reconstruct it.
3493 iris_store_derived_program_state(struct iris_context
*ice
,
3494 enum iris_program_cache_id cache_id
,
3495 struct iris_compiled_shader
*shader
)
3497 struct iris_screen
*screen
= (void *) ice
->ctx
.screen
;
3498 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
3502 iris_store_vs_state(ice
, devinfo
, shader
);
3504 case IRIS_CACHE_TCS
:
3505 iris_store_tcs_state(ice
, devinfo
, shader
);
3507 case IRIS_CACHE_TES
:
3508 iris_store_tes_state(ice
, devinfo
, shader
);
3511 iris_store_gs_state(ice
, devinfo
, shader
);
3514 iris_store_fs_state(ice
, devinfo
, shader
);
3517 iris_store_cs_state(ice
, devinfo
, shader
);
3518 case IRIS_CACHE_BLORP
:
3525 /* ------------------------------------------------------------------- */
3528 * Configure the URB.
3530 * XXX: write a real comment.
3533 iris_upload_urb_config(struct iris_context
*ice
, struct iris_batch
*batch
)
3535 const struct gen_device_info
*devinfo
= &batch
->screen
->devinfo
;
3536 const unsigned push_size_kB
= 32;
3537 unsigned entries
[4];
3541 for (int i
= MESA_SHADER_VERTEX
; i
<= MESA_SHADER_GEOMETRY
; i
++) {
3542 if (!ice
->shaders
.prog
[i
]) {
3545 struct brw_vue_prog_data
*vue_prog_data
=
3546 (void *) ice
->shaders
.prog
[i
]->prog_data
;
3547 size
[i
] = vue_prog_data
->urb_entry_size
;
3549 assert(size
[i
] != 0);
3552 gen_get_urb_config(devinfo
, 1024 * push_size_kB
,
3553 1024 * ice
->shaders
.urb_size
,
3554 ice
->shaders
.prog
[MESA_SHADER_TESS_EVAL
] != NULL
,
3555 ice
->shaders
.prog
[MESA_SHADER_GEOMETRY
] != NULL
,
3556 size
, entries
, start
);
3558 for (int i
= MESA_SHADER_VERTEX
; i
<= MESA_SHADER_GEOMETRY
; i
++) {
3559 iris_emit_cmd(batch
, GENX(3DSTATE_URB_VS
), urb
) {
3560 urb
._3DCommandSubOpcode
+= i
;
3561 urb
.VSURBStartingAddress
= start
[i
];
3562 urb
.VSURBEntryAllocationSize
= size
[i
] - 1;
3563 urb
.VSNumberofURBEntries
= entries
[i
];
3568 static const uint32_t push_constant_opcodes
[] = {
3569 [MESA_SHADER_VERTEX
] = 21,
3570 [MESA_SHADER_TESS_CTRL
] = 25, /* HS */
3571 [MESA_SHADER_TESS_EVAL
] = 26, /* DS */
3572 [MESA_SHADER_GEOMETRY
] = 22,
3573 [MESA_SHADER_FRAGMENT
] = 23,
3574 [MESA_SHADER_COMPUTE
] = 0,
3578 use_null_surface(struct iris_batch
*batch
, struct iris_context
*ice
)
3580 struct iris_bo
*state_bo
= iris_resource_bo(ice
->state
.unbound_tex
.res
);
3582 iris_use_pinned_bo(batch
, state_bo
, false);
3584 return ice
->state
.unbound_tex
.offset
;
3588 use_null_fb_surface(struct iris_batch
*batch
, struct iris_context
*ice
)
3590 /* If set_framebuffer_state() was never called, fall back to 1x1x1 */
3591 if (!ice
->state
.null_fb
.res
)
3592 return use_null_surface(batch
, ice
);
3594 struct iris_bo
*state_bo
= iris_resource_bo(ice
->state
.null_fb
.res
);
3596 iris_use_pinned_bo(batch
, state_bo
, false);
3598 return ice
->state
.null_fb
.offset
;
3602 * Add a surface to the validation list, as well as the buffer containing
3603 * the corresponding SURFACE_STATE.
3605 * Returns the binding table entry (offset to SURFACE_STATE).
3608 use_surface(struct iris_batch
*batch
,
3609 struct pipe_surface
*p_surf
,
3612 struct iris_surface
*surf
= (void *) p_surf
;
3614 iris_use_pinned_bo(batch
, iris_resource_bo(p_surf
->texture
), writeable
);
3615 iris_use_pinned_bo(batch
, iris_resource_bo(surf
->surface_state
.res
), false);
3617 return surf
->surface_state
.offset
;
3621 use_sampler_view(struct iris_batch
*batch
, struct iris_sampler_view
*isv
)
3623 iris_use_pinned_bo(batch
, isv
->res
->bo
, false);
3624 iris_use_pinned_bo(batch
, iris_resource_bo(isv
->surface_state
.res
), false);
3626 return isv
->surface_state
.offset
;
3630 use_const_buffer(struct iris_batch
*batch
,
3631 struct iris_context
*ice
,
3632 struct iris_const_buffer
*cbuf
)
3634 if (!cbuf
->surface_state
.res
)
3635 return use_null_surface(batch
, ice
);
3637 iris_use_pinned_bo(batch
, iris_resource_bo(cbuf
->data
.res
), false);
3638 iris_use_pinned_bo(batch
, iris_resource_bo(cbuf
->surface_state
.res
), false);
3640 return cbuf
->surface_state
.offset
;
3644 use_ssbo(struct iris_batch
*batch
, struct iris_context
*ice
,
3645 struct iris_shader_state
*shs
, int i
)
3648 return use_null_surface(batch
, ice
);
3650 struct iris_state_ref
*surf_state
= &shs
->ssbo_surface_state
[i
];
3652 iris_use_pinned_bo(batch
, iris_resource_bo(shs
->ssbo
[i
]), true);
3653 iris_use_pinned_bo(batch
, iris_resource_bo(surf_state
->res
), false);
3655 return surf_state
->offset
;
3659 use_image(struct iris_batch
*batch
, struct iris_context
*ice
,
3660 struct iris_shader_state
*shs
, int i
)
3662 if (!shs
->image
[i
].res
)
3663 return use_null_surface(batch
, ice
);
3665 struct iris_state_ref
*surf_state
= &shs
->image
[i
].surface_state
;
3667 iris_use_pinned_bo(batch
, iris_resource_bo(shs
->image
[i
].res
),
3668 shs
->image
[i
].access
& PIPE_IMAGE_ACCESS_WRITE
);
3669 iris_use_pinned_bo(batch
, iris_resource_bo(surf_state
->res
), false);
3671 return surf_state
->offset
;
3674 #define push_bt_entry(addr) \
3675 assert(addr >= binder_addr); \
3676 assert(s < prog_data->binding_table.size_bytes / sizeof(uint32_t)); \
3677 if (!pin_only) bt_map[s++] = (addr) - binder_addr;
3679 #define bt_assert(section, exists) \
3680 if (!pin_only) assert(prog_data->binding_table.section == \
3681 (exists) ? s : 0xd0d0d0d0)
3684 * Populate the binding table for a given shader stage.
3686 * This fills out the table of pointers to surfaces required by the shader,
3687 * and also adds those buffers to the validation list so the kernel can make
3688 * resident before running our batch.
3691 iris_populate_binding_table(struct iris_context
*ice
,
3692 struct iris_batch
*batch
,
3693 gl_shader_stage stage
,
3696 const struct iris_binder
*binder
= &ice
->state
.binder
;
3697 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3701 UNUSED
struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3702 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
3703 uint32_t binder_addr
= binder
->bo
->gtt_offset
;
3705 //struct brw_stage_prog_data *prog_data = (void *) shader->prog_data;
3706 uint32_t *bt_map
= binder
->map
+ binder
->bt_offset
[stage
];
3709 const struct shader_info
*info
= iris_get_shader_info(ice
, stage
);
3711 /* TCS passthrough doesn't need a binding table. */
3712 assert(stage
== MESA_SHADER_TESS_CTRL
);
3716 if (stage
== MESA_SHADER_COMPUTE
) {
3717 /* surface for gl_NumWorkGroups */
3718 struct iris_state_ref
*grid_data
= &ice
->state
.grid_size
;
3719 struct iris_state_ref
*grid_state
= &ice
->state
.grid_surf_state
;
3720 iris_use_pinned_bo(batch
, iris_resource_bo(grid_data
->res
), false);
3721 iris_use_pinned_bo(batch
, iris_resource_bo(grid_state
->res
), false);
3722 push_bt_entry(grid_state
->offset
);
3725 if (stage
== MESA_SHADER_FRAGMENT
) {
3726 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
3727 /* Note that cso_fb->nr_cbufs == fs_key->nr_color_regions. */
3728 if (cso_fb
->nr_cbufs
) {
3729 for (unsigned i
= 0; i
< cso_fb
->nr_cbufs
; i
++) {
3731 cso_fb
->cbufs
[i
] ? use_surface(batch
, cso_fb
->cbufs
[i
], true)
3732 : use_null_fb_surface(batch
, ice
);
3733 push_bt_entry(addr
);
3736 uint32_t addr
= use_null_fb_surface(batch
, ice
);
3737 push_bt_entry(addr
);
3741 bt_assert(texture_start
, info
->num_textures
> 0);
3743 for (int i
= 0; i
< info
->num_textures
; i
++) {
3744 struct iris_sampler_view
*view
= shs
->textures
[i
];
3745 uint32_t addr
= view
? use_sampler_view(batch
, view
)
3746 : use_null_surface(batch
, ice
);
3747 push_bt_entry(addr
);
3750 bt_assert(image_start
, info
->num_images
> 0);
3752 for (int i
= 0; i
< info
->num_images
; i
++) {
3753 uint32_t addr
= use_image(batch
, ice
, shs
, i
);
3754 push_bt_entry(addr
);
3757 bt_assert(ubo_start
, shader
->num_cbufs
> 0);
3759 for (int i
= 0; i
< shader
->num_cbufs
; i
++) {
3760 uint32_t addr
= use_const_buffer(batch
, ice
, &shs
->constbuf
[i
]);
3761 push_bt_entry(addr
);
3764 bt_assert(ssbo_start
, info
->num_abos
+ info
->num_ssbos
> 0);
3766 /* XXX: st is wasting 16 binding table slots for ABOs. Should add a cap
3767 * for changing nir_lower_atomics_to_ssbos setting and buffer_base offset
3768 * in st_atom_storagebuf.c so it'll compact them into one range, with
3769 * SSBOs starting at info->num_abos. Ideally it'd reset num_abos to 0 too
3771 if (info
->num_abos
+ info
->num_ssbos
> 0) {
3772 for (int i
= 0; i
< IRIS_MAX_ABOS
+ info
->num_ssbos
; i
++) {
3773 uint32_t addr
= use_ssbo(batch
, ice
, shs
, i
);
3774 push_bt_entry(addr
);
3779 // XXX: not implemented yet
3780 bt_assert(plane_start
[1], ...);
3781 bt_assert(plane_start
[2], ...);
3786 iris_use_optional_res(struct iris_batch
*batch
,
3787 struct pipe_resource
*res
,
3791 struct iris_bo
*bo
= iris_resource_bo(res
);
3792 iris_use_pinned_bo(batch
, bo
, writeable
);
3796 /* ------------------------------------------------------------------- */
3799 * Pin any BOs which were installed by a previous batch, and restored
3800 * via the hardware logical context mechanism.
3802 * We don't need to re-emit all state every batch - the hardware context
3803 * mechanism will save and restore it for us. This includes pointers to
3804 * various BOs...which won't exist unless we ask the kernel to pin them
3805 * by adding them to the validation list.
3807 * We can skip buffers if we've re-emitted those packets, as we're
3808 * overwriting those stale pointers with new ones, and don't actually
3809 * refer to the old BOs.
3812 iris_restore_render_saved_bos(struct iris_context
*ice
,
3813 struct iris_batch
*batch
,
3814 const struct pipe_draw_info
*draw
)
3816 struct iris_genx_state
*genx
= ice
->state
.genx
;
3818 const uint64_t clean
= ~ice
->state
.dirty
;
3820 if (clean
& IRIS_DIRTY_CC_VIEWPORT
) {
3821 iris_use_optional_res(batch
, ice
->state
.last_res
.cc_vp
, false);
3824 if (clean
& IRIS_DIRTY_SF_CL_VIEWPORT
) {
3825 iris_use_optional_res(batch
, ice
->state
.last_res
.sf_cl_vp
, false);
3828 if (clean
& IRIS_DIRTY_BLEND_STATE
) {
3829 iris_use_optional_res(batch
, ice
->state
.last_res
.blend
, false);
3832 if (clean
& IRIS_DIRTY_COLOR_CALC_STATE
) {
3833 iris_use_optional_res(batch
, ice
->state
.last_res
.color_calc
, false);
3836 if (clean
& IRIS_DIRTY_SCISSOR_RECT
) {
3837 iris_use_optional_res(batch
, ice
->state
.last_res
.scissor
, false);
3840 if (ice
->state
.streamout_active
&& (clean
& IRIS_DIRTY_SO_BUFFERS
)) {
3841 for (int i
= 0; i
< 4; i
++) {
3842 struct iris_stream_output_target
*tgt
=
3843 (void *) ice
->state
.so_target
[i
];
3845 iris_use_pinned_bo(batch
, iris_resource_bo(tgt
->base
.buffer
),
3847 iris_use_pinned_bo(batch
, iris_resource_bo(tgt
->offset
.res
),
3853 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
3854 if (!(clean
& (IRIS_DIRTY_CONSTANTS_VS
<< stage
)))
3857 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
3858 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3863 struct brw_stage_prog_data
*prog_data
= (void *) shader
->prog_data
;
3865 for (int i
= 0; i
< 4; i
++) {
3866 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[i
];
3868 if (range
->length
== 0)
3871 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[range
->block
];
3872 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
3875 iris_use_pinned_bo(batch
, res
->bo
, false);
3877 iris_use_pinned_bo(batch
, batch
->screen
->workaround_bo
, false);
3881 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
3882 if (clean
& (IRIS_DIRTY_BINDINGS_VS
<< stage
)) {
3883 /* Re-pin any buffers referred to by the binding table. */
3884 iris_populate_binding_table(ice
, batch
, stage
, true);
3888 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
3889 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
3890 struct pipe_resource
*res
= shs
->sampler_table
.res
;
3892 iris_use_pinned_bo(batch
, iris_resource_bo(res
), false);
3895 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
3896 if (clean
& (IRIS_DIRTY_VS
<< stage
)) {
3897 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3900 struct iris_bo
*bo
= iris_resource_bo(shader
->assembly
.res
);
3901 iris_use_pinned_bo(batch
, bo
, false);
3903 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3905 if (prog_data
->total_scratch
> 0) {
3906 struct iris_bo
*bo
=
3907 iris_get_scratch_space(ice
, prog_data
->total_scratch
, stage
);
3908 iris_use_pinned_bo(batch
, bo
, true);
3914 if (clean
& IRIS_DIRTY_DEPTH_BUFFER
) {
3915 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
3917 if (cso_fb
->zsbuf
) {
3918 struct iris_resource
*zres
, *sres
;
3919 iris_get_depth_stencil_resources(cso_fb
->zsbuf
->texture
,
3922 iris_use_pinned_bo(batch
, zres
->bo
,
3923 ice
->state
.depth_writes_enabled
);
3926 iris_use_pinned_bo(batch
, sres
->bo
,
3927 ice
->state
.stencil_writes_enabled
);
3932 if (draw
->index_size
== 0 && ice
->state
.last_res
.index_buffer
) {
3933 /* This draw didn't emit a new index buffer, so we are inheriting the
3934 * older index buffer. This draw didn't need it, but future ones may.
3936 struct iris_bo
*bo
= iris_resource_bo(ice
->state
.last_res
.index_buffer
);
3937 iris_use_pinned_bo(batch
, bo
, false);
3940 if (clean
& IRIS_DIRTY_VERTEX_BUFFERS
) {
3941 uint64_t bound
= ice
->state
.bound_vertex_buffers
;
3943 const int i
= u_bit_scan64(&bound
);
3944 struct pipe_resource
*res
= genx
->vertex_buffers
[i
].resource
;
3945 iris_use_pinned_bo(batch
, iris_resource_bo(res
), false);
3951 iris_restore_compute_saved_bos(struct iris_context
*ice
,
3952 struct iris_batch
*batch
,
3953 const struct pipe_grid_info
*grid
)
3955 const uint64_t clean
= ~ice
->state
.dirty
;
3957 const int stage
= MESA_SHADER_COMPUTE
;
3958 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
3960 if (clean
& IRIS_DIRTY_CONSTANTS_CS
) {
3961 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3964 struct brw_stage_prog_data
*prog_data
= (void *) shader
->prog_data
;
3965 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[0];
3967 if (range
->length
> 0) {
3968 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[range
->block
];
3969 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
3972 iris_use_pinned_bo(batch
, res
->bo
, false);
3974 iris_use_pinned_bo(batch
, batch
->screen
->workaround_bo
, false);
3979 if (clean
& IRIS_DIRTY_BINDINGS_CS
) {
3980 /* Re-pin any buffers referred to by the binding table. */
3981 iris_populate_binding_table(ice
, batch
, stage
, true);
3984 struct pipe_resource
*sampler_res
= shs
->sampler_table
.res
;
3986 iris_use_pinned_bo(batch
, iris_resource_bo(sampler_res
), false);
3988 if (clean
& IRIS_DIRTY_CS
) {
3989 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3992 struct iris_bo
*bo
= iris_resource_bo(shader
->assembly
.res
);
3993 iris_use_pinned_bo(batch
, bo
, false);
3995 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3997 if (prog_data
->total_scratch
> 0) {
3998 struct iris_bo
*bo
=
3999 iris_get_scratch_space(ice
, prog_data
->total_scratch
, stage
);
4000 iris_use_pinned_bo(batch
, bo
, true);
4007 * Possibly emit STATE_BASE_ADDRESS to update Surface State Base Address.
4010 iris_update_surface_base_address(struct iris_batch
*batch
,
4011 struct iris_binder
*binder
)
4013 if (batch
->last_surface_base_address
== binder
->bo
->gtt_offset
)
4016 flush_for_state_base_change(batch
);
4018 iris_emit_cmd(batch
, GENX(STATE_BASE_ADDRESS
), sba
) {
4019 sba
.SurfaceStateMOCS
= MOCS_WB
;
4020 sba
.SurfaceStateBaseAddressModifyEnable
= true;
4021 sba
.SurfaceStateBaseAddress
= ro_bo(binder
->bo
, 0);
4024 batch
->last_surface_base_address
= binder
->bo
->gtt_offset
;
4028 iris_upload_dirty_render_state(struct iris_context
*ice
,
4029 struct iris_batch
*batch
,
4030 const struct pipe_draw_info
*draw
)
4032 const uint64_t dirty
= ice
->state
.dirty
;
4034 if (!(dirty
& IRIS_ALL_DIRTY_FOR_RENDER
))
4037 struct iris_genx_state
*genx
= ice
->state
.genx
;
4038 struct iris_binder
*binder
= &ice
->state
.binder
;
4039 struct brw_wm_prog_data
*wm_prog_data
= (void *)
4040 ice
->shaders
.prog
[MESA_SHADER_FRAGMENT
]->prog_data
;
4042 if (dirty
& IRIS_DIRTY_CC_VIEWPORT
) {
4043 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
4044 uint32_t cc_vp_address
;
4046 /* XXX: could avoid streaming for depth_clip [0,1] case. */
4047 uint32_t *cc_vp_map
=
4048 stream_state(batch
, ice
->state
.dynamic_uploader
,
4049 &ice
->state
.last_res
.cc_vp
,
4050 4 * ice
->state
.num_viewports
*
4051 GENX(CC_VIEWPORT_length
), 32, &cc_vp_address
);
4052 for (int i
= 0; i
< ice
->state
.num_viewports
; i
++) {
4054 util_viewport_zmin_zmax(&ice
->state
.viewports
[i
],
4055 cso_rast
->clip_halfz
, &zmin
, &zmax
);
4056 if (cso_rast
->depth_clip_near
)
4058 if (cso_rast
->depth_clip_far
)
4061 iris_pack_state(GENX(CC_VIEWPORT
), cc_vp_map
, ccv
) {
4062 ccv
.MinimumDepth
= zmin
;
4063 ccv
.MaximumDepth
= zmax
;
4066 cc_vp_map
+= GENX(CC_VIEWPORT_length
);
4069 iris_emit_cmd(batch
, GENX(3DSTATE_VIEWPORT_STATE_POINTERS_CC
), ptr
) {
4070 ptr
.CCViewportPointer
= cc_vp_address
;
4074 if (dirty
& IRIS_DIRTY_SF_CL_VIEWPORT
) {
4075 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4076 uint32_t sf_cl_vp_address
;
4078 stream_state(batch
, ice
->state
.dynamic_uploader
,
4079 &ice
->state
.last_res
.sf_cl_vp
,
4080 4 * ice
->state
.num_viewports
*
4081 GENX(SF_CLIP_VIEWPORT_length
), 64, &sf_cl_vp_address
);
4083 for (unsigned i
= 0; i
< ice
->state
.num_viewports
; i
++) {
4084 const struct pipe_viewport_state
*state
= &ice
->state
.viewports
[i
];
4085 float gb_xmin
, gb_xmax
, gb_ymin
, gb_ymax
;
4087 float vp_xmin
= viewport_extent(state
, 0, -1.0f
);
4088 float vp_xmax
= viewport_extent(state
, 0, 1.0f
);
4089 float vp_ymin
= viewport_extent(state
, 1, -1.0f
);
4090 float vp_ymax
= viewport_extent(state
, 1, 1.0f
);
4092 calculate_guardband_size(cso_fb
->width
, cso_fb
->height
,
4093 state
->scale
[0], state
->scale
[1],
4094 state
->translate
[0], state
->translate
[1],
4095 &gb_xmin
, &gb_xmax
, &gb_ymin
, &gb_ymax
);
4097 iris_pack_state(GENX(SF_CLIP_VIEWPORT
), vp_map
, vp
) {
4098 vp
.ViewportMatrixElementm00
= state
->scale
[0];
4099 vp
.ViewportMatrixElementm11
= state
->scale
[1];
4100 vp
.ViewportMatrixElementm22
= state
->scale
[2];
4101 vp
.ViewportMatrixElementm30
= state
->translate
[0];
4102 vp
.ViewportMatrixElementm31
= state
->translate
[1];
4103 vp
.ViewportMatrixElementm32
= state
->translate
[2];
4104 vp
.XMinClipGuardband
= gb_xmin
;
4105 vp
.XMaxClipGuardband
= gb_xmax
;
4106 vp
.YMinClipGuardband
= gb_ymin
;
4107 vp
.YMaxClipGuardband
= gb_ymax
;
4108 vp
.XMinViewPort
= MAX2(vp_xmin
, 0);
4109 vp
.XMaxViewPort
= MIN2(vp_xmax
, cso_fb
->width
) - 1;
4110 vp
.YMinViewPort
= MAX2(vp_ymin
, 0);
4111 vp
.YMaxViewPort
= MIN2(vp_ymax
, cso_fb
->height
) - 1;
4114 vp_map
+= GENX(SF_CLIP_VIEWPORT_length
);
4117 iris_emit_cmd(batch
, GENX(3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP
), ptr
) {
4118 ptr
.SFClipViewportPointer
= sf_cl_vp_address
;
4122 if (dirty
& IRIS_DIRTY_URB
) {
4123 iris_upload_urb_config(ice
, batch
);
4126 if (dirty
& IRIS_DIRTY_BLEND_STATE
) {
4127 struct iris_blend_state
*cso_blend
= ice
->state
.cso_blend
;
4128 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4129 struct iris_depth_stencil_alpha_state
*cso_zsa
= ice
->state
.cso_zsa
;
4130 const int header_dwords
= GENX(BLEND_STATE_length
);
4131 const int rt_dwords
= cso_fb
->nr_cbufs
* GENX(BLEND_STATE_ENTRY_length
);
4132 uint32_t blend_offset
;
4133 uint32_t *blend_map
=
4134 stream_state(batch
, ice
->state
.dynamic_uploader
,
4135 &ice
->state
.last_res
.blend
,
4136 4 * (header_dwords
+ rt_dwords
), 64, &blend_offset
);
4138 uint32_t blend_state_header
;
4139 iris_pack_state(GENX(BLEND_STATE
), &blend_state_header
, bs
) {
4140 bs
.AlphaTestEnable
= cso_zsa
->alpha
.enabled
;
4141 bs
.AlphaTestFunction
= translate_compare_func(cso_zsa
->alpha
.func
);
4144 blend_map
[0] = blend_state_header
| cso_blend
->blend_state
[0];
4145 memcpy(&blend_map
[1], &cso_blend
->blend_state
[1], 4 * rt_dwords
);
4147 iris_emit_cmd(batch
, GENX(3DSTATE_BLEND_STATE_POINTERS
), ptr
) {
4148 ptr
.BlendStatePointer
= blend_offset
;
4149 ptr
.BlendStatePointerValid
= true;
4153 if (dirty
& IRIS_DIRTY_COLOR_CALC_STATE
) {
4154 struct iris_depth_stencil_alpha_state
*cso
= ice
->state
.cso_zsa
;
4156 struct pipe_stencil_ref
*p_stencil_refs
= &ice
->state
.stencil_ref
;
4160 stream_state(batch
, ice
->state
.dynamic_uploader
,
4161 &ice
->state
.last_res
.color_calc
,
4162 sizeof(uint32_t) * GENX(COLOR_CALC_STATE_length
),
4164 iris_pack_state(GENX(COLOR_CALC_STATE
), cc_map
, cc
) {
4165 cc
.AlphaTestFormat
= ALPHATEST_FLOAT32
;
4166 cc
.AlphaReferenceValueAsFLOAT32
= cso
->alpha
.ref_value
;
4167 cc
.BlendConstantColorRed
= ice
->state
.blend_color
.color
[0];
4168 cc
.BlendConstantColorGreen
= ice
->state
.blend_color
.color
[1];
4169 cc
.BlendConstantColorBlue
= ice
->state
.blend_color
.color
[2];
4170 cc
.BlendConstantColorAlpha
= ice
->state
.blend_color
.color
[3];
4172 cc
.StencilReferenceValue
= p_stencil_refs
->ref_value
[0];
4173 cc
.BackfaceStencilReferenceValue
= p_stencil_refs
->ref_value
[1];
4176 iris_emit_cmd(batch
, GENX(3DSTATE_CC_STATE_POINTERS
), ptr
) {
4177 ptr
.ColorCalcStatePointer
= cc_offset
;
4178 ptr
.ColorCalcStatePointerValid
= true;
4182 /* Upload constants for TCS passthrough. */
4183 if ((dirty
& IRIS_DIRTY_CONSTANTS_TCS
) &&
4184 ice
->shaders
.prog
[MESA_SHADER_TESS_CTRL
] &&
4185 !ice
->shaders
.uncompiled
[MESA_SHADER_TESS_CTRL
]) {
4186 struct iris_compiled_shader
*tes_shader
= ice
->shaders
.prog
[MESA_SHADER_TESS_EVAL
];
4189 /* Passthrough always copies 2 vec4s, so when uploading data we ensure
4190 * it is in the right layout for TES.
4193 struct brw_tes_prog_data
*tes_prog_data
= (void *) tes_shader
->prog_data
;
4194 switch (tes_prog_data
->domain
) {
4195 case BRW_TESS_DOMAIN_QUAD
:
4196 for (int i
= 0; i
< 4; i
++)
4197 hdr
[7 - i
] = ice
->state
.default_outer_level
[i
];
4198 hdr
[3] = ice
->state
.default_inner_level
[0];
4199 hdr
[2] = ice
->state
.default_inner_level
[1];
4201 case BRW_TESS_DOMAIN_TRI
:
4202 for (int i
= 0; i
< 3; i
++)
4203 hdr
[7 - i
] = ice
->state
.default_outer_level
[i
];
4204 hdr
[4] = ice
->state
.default_inner_level
[0];
4206 case BRW_TESS_DOMAIN_ISOLINE
:
4207 hdr
[7] = ice
->state
.default_outer_level
[1];
4208 hdr
[6] = ice
->state
.default_outer_level
[0];
4212 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_TESS_CTRL
];
4213 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[0];
4214 u_upload_data(ice
->ctx
.const_uploader
, 0, sizeof(hdr
), 32,
4215 &hdr
[0], &cbuf
->data
.offset
,
4219 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4220 if (!(dirty
& (IRIS_DIRTY_CONSTANTS_VS
<< stage
)))
4223 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4224 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
4229 if (shs
->cbuf0_needs_upload
)
4230 upload_uniforms(ice
, stage
);
4232 struct brw_stage_prog_data
*prog_data
= (void *) shader
->prog_data
;
4234 iris_emit_cmd(batch
, GENX(3DSTATE_CONSTANT_VS
), pkt
) {
4235 pkt
._3DCommandSubOpcode
= push_constant_opcodes
[stage
];
4237 /* The Skylake PRM contains the following restriction:
4239 * "The driver must ensure The following case does not occur
4240 * without a flush to the 3D engine: 3DSTATE_CONSTANT_* with
4241 * buffer 3 read length equal to zero committed followed by a
4242 * 3DSTATE_CONSTANT_* with buffer 0 read length not equal to
4245 * To avoid this, we program the buffers in the highest slots.
4246 * This way, slot 0 is only used if slot 3 is also used.
4250 for (int i
= 3; i
>= 0; i
--) {
4251 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[i
];
4253 if (range
->length
== 0)
4256 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[range
->block
];
4257 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
4259 assert(cbuf
->data
.offset
% 32 == 0);
4261 pkt
.ConstantBody
.ReadLength
[n
] = range
->length
;
4262 pkt
.ConstantBody
.Buffer
[n
] =
4263 res
? ro_bo(res
->bo
, range
->start
* 32 + cbuf
->data
.offset
)
4264 : ro_bo(batch
->screen
->workaround_bo
, 0);
4271 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4272 if (dirty
& (IRIS_DIRTY_BINDINGS_VS
<< stage
)) {
4273 iris_emit_cmd(batch
, GENX(3DSTATE_BINDING_TABLE_POINTERS_VS
), ptr
) {
4274 ptr
._3DCommandSubOpcode
= 38 + stage
;
4275 ptr
.PointertoVSBindingTable
= binder
->bt_offset
[stage
];
4280 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4281 if (dirty
& (IRIS_DIRTY_BINDINGS_VS
<< stage
)) {
4282 iris_populate_binding_table(ice
, batch
, stage
, false);
4286 if (ice
->state
.need_border_colors
)
4287 iris_use_pinned_bo(batch
, ice
->state
.border_color_pool
.bo
, false);
4289 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4290 if (!(dirty
& (IRIS_DIRTY_SAMPLER_STATES_VS
<< stage
)) ||
4291 !ice
->shaders
.prog
[stage
])
4294 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4295 struct pipe_resource
*res
= shs
->sampler_table
.res
;
4297 iris_use_pinned_bo(batch
, iris_resource_bo(res
), false);
4299 iris_emit_cmd(batch
, GENX(3DSTATE_SAMPLER_STATE_POINTERS_VS
), ptr
) {
4300 ptr
._3DCommandSubOpcode
= 43 + stage
;
4301 ptr
.PointertoVSSamplerState
= shs
->sampler_table
.offset
;
4305 if (dirty
& IRIS_DIRTY_MULTISAMPLE
) {
4306 iris_emit_cmd(batch
, GENX(3DSTATE_MULTISAMPLE
), ms
) {
4308 ice
->state
.cso_rast
->half_pixel_center
? CENTER
: UL_CORNER
;
4309 if (ice
->state
.framebuffer
.samples
> 0)
4310 ms
.NumberofMultisamples
= ffs(ice
->state
.framebuffer
.samples
) - 1;
4314 if (dirty
& IRIS_DIRTY_SAMPLE_MASK
) {
4315 iris_emit_cmd(batch
, GENX(3DSTATE_SAMPLE_MASK
), ms
) {
4316 ms
.SampleMask
= ice
->state
.sample_mask
;
4320 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4321 if (!(dirty
& (IRIS_DIRTY_VS
<< stage
)))
4324 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
4327 struct iris_resource
*cache
= (void *) shader
->assembly
.res
;
4328 iris_use_pinned_bo(batch
, cache
->bo
, false);
4329 iris_batch_emit(batch
, shader
->derived_data
,
4330 iris_derived_program_state_size(stage
));
4332 if (stage
== MESA_SHADER_TESS_EVAL
) {
4333 iris_emit_cmd(batch
, GENX(3DSTATE_HS
), hs
);
4334 iris_emit_cmd(batch
, GENX(3DSTATE_TE
), te
);
4335 iris_emit_cmd(batch
, GENX(3DSTATE_DS
), ds
);
4336 } else if (stage
== MESA_SHADER_GEOMETRY
) {
4337 iris_emit_cmd(batch
, GENX(3DSTATE_GS
), gs
);
4342 if (ice
->state
.streamout_active
) {
4343 if (dirty
& IRIS_DIRTY_SO_BUFFERS
) {
4344 iris_batch_emit(batch
, genx
->so_buffers
,
4345 4 * 4 * GENX(3DSTATE_SO_BUFFER_length
));
4346 for (int i
= 0; i
< 4; i
++) {
4347 struct iris_stream_output_target
*tgt
=
4348 (void *) ice
->state
.so_target
[i
];
4350 iris_use_pinned_bo(batch
, iris_resource_bo(tgt
->base
.buffer
),
4352 iris_use_pinned_bo(batch
, iris_resource_bo(tgt
->offset
.res
),
4358 if ((dirty
& IRIS_DIRTY_SO_DECL_LIST
) && ice
->state
.streamout
) {
4359 uint32_t *decl_list
=
4360 ice
->state
.streamout
+ GENX(3DSTATE_STREAMOUT_length
);
4361 iris_batch_emit(batch
, decl_list
, 4 * ((decl_list
[0] & 0xff) + 2));
4364 if (dirty
& IRIS_DIRTY_STREAMOUT
) {
4365 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
4367 uint32_t dynamic_sol
[GENX(3DSTATE_STREAMOUT_length
)];
4368 iris_pack_command(GENX(3DSTATE_STREAMOUT
), dynamic_sol
, sol
) {
4369 sol
.SOFunctionEnable
= true;
4370 sol
.SOStatisticsEnable
= true;
4372 sol
.RenderingDisable
= cso_rast
->rasterizer_discard
&&
4373 !ice
->state
.prims_generated_query_active
;
4374 sol
.ReorderMode
= cso_rast
->flatshade_first
? LEADING
: TRAILING
;
4377 assert(ice
->state
.streamout
);
4379 iris_emit_merge(batch
, ice
->state
.streamout
, dynamic_sol
,
4380 GENX(3DSTATE_STREAMOUT_length
));
4383 if (dirty
& IRIS_DIRTY_STREAMOUT
) {
4384 iris_emit_cmd(batch
, GENX(3DSTATE_STREAMOUT
), sol
);
4388 if (dirty
& IRIS_DIRTY_CLIP
) {
4389 struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
4390 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4392 uint32_t dynamic_clip
[GENX(3DSTATE_CLIP_length
)];
4393 iris_pack_command(GENX(3DSTATE_CLIP
), &dynamic_clip
, cl
) {
4394 cl
.StatisticsEnable
= ice
->state
.statistics_counters_enabled
;
4395 cl
.ClipMode
= cso_rast
->rasterizer_discard
? CLIPMODE_REJECT_ALL
4397 if (wm_prog_data
->barycentric_interp_modes
&
4398 BRW_BARYCENTRIC_NONPERSPECTIVE_BITS
)
4399 cl
.NonPerspectiveBarycentricEnable
= true;
4401 cl
.ForceZeroRTAIndexEnable
= cso_fb
->layers
== 0;
4402 cl
.MaximumVPIndex
= ice
->state
.num_viewports
- 1;
4404 iris_emit_merge(batch
, cso_rast
->clip
, dynamic_clip
,
4405 ARRAY_SIZE(cso_rast
->clip
));
4408 if (dirty
& IRIS_DIRTY_RASTER
) {
4409 struct iris_rasterizer_state
*cso
= ice
->state
.cso_rast
;
4410 iris_batch_emit(batch
, cso
->raster
, sizeof(cso
->raster
));
4411 iris_batch_emit(batch
, cso
->sf
, sizeof(cso
->sf
));
4415 if (dirty
& IRIS_DIRTY_WM
) {
4416 struct iris_rasterizer_state
*cso
= ice
->state
.cso_rast
;
4417 uint32_t dynamic_wm
[GENX(3DSTATE_WM_length
)];
4419 iris_pack_command(GENX(3DSTATE_WM
), &dynamic_wm
, wm
) {
4420 wm
.StatisticsEnable
= ice
->state
.statistics_counters_enabled
;
4422 wm
.BarycentricInterpolationMode
=
4423 wm_prog_data
->barycentric_interp_modes
;
4425 if (wm_prog_data
->early_fragment_tests
)
4426 wm
.EarlyDepthStencilControl
= EDSC_PREPS
;
4427 else if (wm_prog_data
->has_side_effects
)
4428 wm
.EarlyDepthStencilControl
= EDSC_PSEXEC
;
4430 iris_emit_merge(batch
, cso
->wm
, dynamic_wm
, ARRAY_SIZE(cso
->wm
));
4433 if (dirty
& IRIS_DIRTY_SBE
) {
4434 iris_emit_sbe(batch
, ice
);
4437 if (dirty
& IRIS_DIRTY_PS_BLEND
) {
4438 struct iris_blend_state
*cso_blend
= ice
->state
.cso_blend
;
4439 struct iris_depth_stencil_alpha_state
*cso_zsa
= ice
->state
.cso_zsa
;
4440 uint32_t dynamic_pb
[GENX(3DSTATE_PS_BLEND_length
)];
4441 iris_pack_command(GENX(3DSTATE_PS_BLEND
), &dynamic_pb
, pb
) {
4442 pb
.HasWriteableRT
= true; // XXX: comes from somewhere :(
4443 pb
.AlphaTestEnable
= cso_zsa
->alpha
.enabled
;
4446 iris_emit_merge(batch
, cso_blend
->ps_blend
, dynamic_pb
,
4447 ARRAY_SIZE(cso_blend
->ps_blend
));
4450 if (dirty
& IRIS_DIRTY_WM_DEPTH_STENCIL
) {
4451 struct iris_depth_stencil_alpha_state
*cso
= ice
->state
.cso_zsa
;
4453 struct pipe_stencil_ref
*p_stencil_refs
= &ice
->state
.stencil_ref
;
4454 uint32_t stencil_refs
[GENX(3DSTATE_WM_DEPTH_STENCIL_length
)];
4455 iris_pack_command(GENX(3DSTATE_WM_DEPTH_STENCIL
), &stencil_refs
, wmds
) {
4456 wmds
.StencilReferenceValue
= p_stencil_refs
->ref_value
[0];
4457 wmds
.BackfaceStencilReferenceValue
= p_stencil_refs
->ref_value
[1];
4459 iris_emit_merge(batch
, cso
->wmds
, stencil_refs
, ARRAY_SIZE(cso
->wmds
));
4461 iris_batch_emit(batch
, cso
->wmds
, sizeof(cso
->wmds
));
4465 if (dirty
& IRIS_DIRTY_SCISSOR_RECT
) {
4466 uint32_t scissor_offset
=
4467 emit_state(batch
, ice
->state
.dynamic_uploader
,
4468 &ice
->state
.last_res
.scissor
,
4469 ice
->state
.scissors
,
4470 sizeof(struct pipe_scissor_state
) *
4471 ice
->state
.num_viewports
, 32);
4473 iris_emit_cmd(batch
, GENX(3DSTATE_SCISSOR_STATE_POINTERS
), ptr
) {
4474 ptr
.ScissorRectPointer
= scissor_offset
;
4478 if (dirty
& IRIS_DIRTY_DEPTH_BUFFER
) {
4479 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4480 struct iris_depth_buffer_state
*cso_z
= &ice
->state
.genx
->depth_buffer
;
4482 iris_batch_emit(batch
, cso_z
->packets
, sizeof(cso_z
->packets
));
4484 if (cso_fb
->zsbuf
) {
4485 struct iris_resource
*zres
, *sres
;
4486 iris_get_depth_stencil_resources(cso_fb
->zsbuf
->texture
,
4489 iris_use_pinned_bo(batch
, zres
->bo
,
4490 ice
->state
.depth_writes_enabled
);
4494 iris_use_pinned_bo(batch
, sres
->bo
,
4495 ice
->state
.stencil_writes_enabled
);
4500 if (dirty
& IRIS_DIRTY_POLYGON_STIPPLE
) {
4501 iris_emit_cmd(batch
, GENX(3DSTATE_POLY_STIPPLE_PATTERN
), poly
) {
4502 for (int i
= 0; i
< 32; i
++) {
4503 poly
.PatternRow
[i
] = ice
->state
.poly_stipple
.stipple
[i
];
4508 if (dirty
& IRIS_DIRTY_LINE_STIPPLE
) {
4509 struct iris_rasterizer_state
*cso
= ice
->state
.cso_rast
;
4510 iris_batch_emit(batch
, cso
->line_stipple
, sizeof(cso
->line_stipple
));
4513 if (dirty
& IRIS_DIRTY_VF_TOPOLOGY
) {
4514 iris_emit_cmd(batch
, GENX(3DSTATE_VF_TOPOLOGY
), topo
) {
4515 topo
.PrimitiveTopologyType
=
4516 translate_prim_type(draw
->mode
, draw
->vertices_per_patch
);
4520 if (dirty
& IRIS_DIRTY_VERTEX_BUFFERS
) {
4521 int count
= util_bitcount64(ice
->state
.bound_vertex_buffers
);
4524 /* The VF cache designers cut corners, and made the cache key's
4525 * <VertexBufferIndex, Memory Address> tuple only consider the bottom
4526 * 32 bits of the address. If you have two vertex buffers which get
4527 * placed exactly 4 GiB apart and use them in back-to-back draw calls,
4528 * you can get collisions (even within a single batch).
4530 * So, we need to do a VF cache invalidate if the buffer for a VB
4531 * slot slot changes [48:32] address bits from the previous time.
4533 unsigned flush_flags
= 0;
4535 uint64_t bound
= ice
->state
.bound_vertex_buffers
;
4537 const int i
= u_bit_scan64(&bound
);
4538 uint16_t high_bits
= 0;
4540 struct iris_resource
*res
=
4541 (void *) genx
->vertex_buffers
[i
].resource
;
4543 iris_use_pinned_bo(batch
, res
->bo
, false);
4545 high_bits
= res
->bo
->gtt_offset
>> 32ull;
4546 if (high_bits
!= ice
->state
.last_vbo_high_bits
[i
]) {
4547 flush_flags
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
|
4548 PIPE_CONTROL_CS_STALL
;
4549 ice
->state
.last_vbo_high_bits
[i
] = high_bits
;
4552 /* If the buffer was written to by streamout, we may need
4553 * to stall so those writes land and become visible to the
4556 * TODO: This may stall more than necessary.
4558 if (res
->bind_history
& PIPE_BIND_STREAM_OUTPUT
)
4559 flush_flags
|= PIPE_CONTROL_CS_STALL
;
4564 iris_emit_pipe_control_flush(batch
, flush_flags
);
4566 const unsigned vb_dwords
= GENX(VERTEX_BUFFER_STATE_length
);
4569 iris_get_command_space(batch
, 4 * (1 + vb_dwords
* count
));
4570 _iris_pack_command(batch
, GENX(3DSTATE_VERTEX_BUFFERS
), map
, vb
) {
4571 vb
.DWordLength
= (vb_dwords
* count
+ 1) - 2;
4575 bound
= ice
->state
.bound_vertex_buffers
;
4577 const int i
= u_bit_scan64(&bound
);
4578 memcpy(map
, genx
->vertex_buffers
[i
].state
,
4579 sizeof(uint32_t) * vb_dwords
);
4585 if (dirty
& IRIS_DIRTY_VERTEX_ELEMENTS
) {
4586 struct iris_vertex_element_state
*cso
= ice
->state
.cso_vertex_elements
;
4587 const unsigned entries
= MAX2(cso
->count
, 1);
4588 iris_batch_emit(batch
, cso
->vertex_elements
, sizeof(uint32_t) *
4589 (1 + entries
* GENX(VERTEX_ELEMENT_STATE_length
)));
4590 iris_batch_emit(batch
, cso
->vf_instancing
, sizeof(uint32_t) *
4591 entries
* GENX(3DSTATE_VF_INSTANCING_length
));
4594 if (dirty
& IRIS_DIRTY_VF_SGVS
) {
4595 const struct brw_vs_prog_data
*vs_prog_data
= (void *)
4596 ice
->shaders
.prog
[MESA_SHADER_VERTEX
]->prog_data
;
4597 struct iris_vertex_element_state
*cso
= ice
->state
.cso_vertex_elements
;
4599 iris_emit_cmd(batch
, GENX(3DSTATE_VF_SGVS
), sgv
) {
4600 if (vs_prog_data
->uses_vertexid
) {
4601 sgv
.VertexIDEnable
= true;
4602 sgv
.VertexIDComponentNumber
= 2;
4603 sgv
.VertexIDElementOffset
= cso
->count
;
4606 if (vs_prog_data
->uses_instanceid
) {
4607 sgv
.InstanceIDEnable
= true;
4608 sgv
.InstanceIDComponentNumber
= 3;
4609 sgv
.InstanceIDElementOffset
= cso
->count
;
4614 if (dirty
& IRIS_DIRTY_VF
) {
4615 iris_emit_cmd(batch
, GENX(3DSTATE_VF
), vf
) {
4616 if (draw
->primitive_restart
) {
4617 vf
.IndexedDrawCutIndexEnable
= true;
4618 vf
.CutIndex
= draw
->restart_index
;
4623 // XXX: Gen8 - PMA fix
4627 iris_upload_render_state(struct iris_context
*ice
,
4628 struct iris_batch
*batch
,
4629 const struct pipe_draw_info
*draw
)
4631 /* Always pin the binder. If we're emitting new binding table pointers,
4632 * we need it. If not, we're probably inheriting old tables via the
4633 * context, and need it anyway. Since true zero-bindings cases are
4634 * practically non-existent, just pin it and avoid last_res tracking.
4636 iris_use_pinned_bo(batch
, ice
->state
.binder
.bo
, false);
4638 if (!batch
->contains_draw
) {
4639 iris_restore_render_saved_bos(ice
, batch
, draw
);
4640 batch
->contains_draw
= true;
4643 iris_upload_dirty_render_state(ice
, batch
, draw
);
4645 if (draw
->index_size
> 0) {
4648 if (draw
->has_user_indices
) {
4649 u_upload_data(ice
->ctx
.stream_uploader
, 0,
4650 draw
->count
* draw
->index_size
, 4, draw
->index
.user
,
4651 &offset
, &ice
->state
.last_res
.index_buffer
);
4653 struct iris_resource
*res
= (void *) draw
->index
.resource
;
4654 res
->bind_history
|= PIPE_BIND_INDEX_BUFFER
;
4656 pipe_resource_reference(&ice
->state
.last_res
.index_buffer
,
4657 draw
->index
.resource
);
4661 struct iris_bo
*bo
= iris_resource_bo(ice
->state
.last_res
.index_buffer
);
4663 iris_emit_cmd(batch
, GENX(3DSTATE_INDEX_BUFFER
), ib
) {
4664 ib
.IndexFormat
= draw
->index_size
>> 1;
4666 ib
.BufferSize
= bo
->size
;
4667 ib
.BufferStartingAddress
= ro_bo(bo
, offset
);
4670 /* The VF cache key only uses 32-bits, see vertex buffer comment above */
4671 uint16_t high_bits
= bo
->gtt_offset
>> 32ull;
4672 if (high_bits
!= ice
->state
.last_index_bo_high_bits
) {
4673 iris_emit_pipe_control_flush(batch
, PIPE_CONTROL_VF_CACHE_INVALIDATE
|
4674 PIPE_CONTROL_CS_STALL
);
4675 ice
->state
.last_index_bo_high_bits
= high_bits
;
4679 #define _3DPRIM_END_OFFSET 0x2420
4680 #define _3DPRIM_START_VERTEX 0x2430
4681 #define _3DPRIM_VERTEX_COUNT 0x2434
4682 #define _3DPRIM_INSTANCE_COUNT 0x2438
4683 #define _3DPRIM_START_INSTANCE 0x243C
4684 #define _3DPRIM_BASE_VERTEX 0x2440
4686 if (draw
->indirect
) {
4687 /* We don't support this MultidrawIndirect. */
4688 assert(!draw
->indirect
->indirect_draw_count
);
4690 struct iris_bo
*bo
= iris_resource_bo(draw
->indirect
->buffer
);
4693 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4694 lrm
.RegisterAddress
= _3DPRIM_VERTEX_COUNT
;
4695 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 0);
4697 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4698 lrm
.RegisterAddress
= _3DPRIM_INSTANCE_COUNT
;
4699 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 4);
4701 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4702 lrm
.RegisterAddress
= _3DPRIM_START_VERTEX
;
4703 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 8);
4705 if (draw
->index_size
) {
4706 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4707 lrm
.RegisterAddress
= _3DPRIM_BASE_VERTEX
;
4708 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 12);
4710 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4711 lrm
.RegisterAddress
= _3DPRIM_START_INSTANCE
;
4712 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 16);
4715 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4716 lrm
.RegisterAddress
= _3DPRIM_START_INSTANCE
;
4717 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 12);
4719 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_IMM
), lri
) {
4720 lri
.RegisterOffset
= _3DPRIM_BASE_VERTEX
;
4724 } else if (draw
->count_from_stream_output
) {
4725 struct iris_stream_output_target
*so
=
4726 (void *) draw
->count_from_stream_output
;
4728 // XXX: avoid if possible
4729 iris_emit_pipe_control_flush(batch
, PIPE_CONTROL_CS_STALL
);
4731 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4732 lrm
.RegisterAddress
= CS_GPR(0);
4734 ro_bo(iris_resource_bo(so
->offset
.res
), so
->offset
.offset
);
4736 iris_math_div32_gpr0(ice
, batch
, so
->stride
);
4737 _iris_emit_lrr(batch
, _3DPRIM_VERTEX_COUNT
, CS_GPR(0));
4739 _iris_emit_lri(batch
, _3DPRIM_START_VERTEX
, 0);
4740 _iris_emit_lri(batch
, _3DPRIM_BASE_VERTEX
, 0);
4741 _iris_emit_lri(batch
, _3DPRIM_START_INSTANCE
, 0);
4742 _iris_emit_lri(batch
, _3DPRIM_INSTANCE_COUNT
, draw
->instance_count
);
4745 iris_emit_cmd(batch
, GENX(3DPRIMITIVE
), prim
) {
4746 prim
.VertexAccessType
= draw
->index_size
> 0 ? RANDOM
: SEQUENTIAL
;
4747 prim
.PredicateEnable
=
4748 ice
->state
.predicate
== IRIS_PREDICATE_STATE_USE_BIT
;
4750 if (draw
->indirect
|| draw
->count_from_stream_output
) {
4751 prim
.IndirectParameterEnable
= true;
4753 prim
.StartInstanceLocation
= draw
->start_instance
;
4754 prim
.InstanceCount
= draw
->instance_count
;
4755 prim
.VertexCountPerInstance
= draw
->count
;
4757 // XXX: this is probably bonkers.
4758 prim
.StartVertexLocation
= draw
->start
;
4760 if (draw
->index_size
) {
4761 prim
.BaseVertexLocation
+= draw
->index_bias
;
4763 prim
.StartVertexLocation
+= draw
->index_bias
;
4766 //prim.BaseVertexLocation = ...;
4772 iris_upload_compute_state(struct iris_context
*ice
,
4773 struct iris_batch
*batch
,
4774 const struct pipe_grid_info
*grid
)
4776 const uint64_t dirty
= ice
->state
.dirty
;
4777 struct iris_screen
*screen
= batch
->screen
;
4778 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
4779 struct iris_binder
*binder
= &ice
->state
.binder
;
4780 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_COMPUTE
];
4781 struct iris_compiled_shader
*shader
=
4782 ice
->shaders
.prog
[MESA_SHADER_COMPUTE
];
4783 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
4784 struct brw_cs_prog_data
*cs_prog_data
= (void *) prog_data
;
4786 /* Always pin the binder. If we're emitting new binding table pointers,
4787 * we need it. If not, we're probably inheriting old tables via the
4788 * context, and need it anyway. Since true zero-bindings cases are
4789 * practically non-existent, just pin it and avoid last_res tracking.
4791 iris_use_pinned_bo(batch
, ice
->state
.binder
.bo
, false);
4793 if ((dirty
& IRIS_DIRTY_CONSTANTS_CS
) && shs
->cbuf0_needs_upload
)
4794 upload_uniforms(ice
, MESA_SHADER_COMPUTE
);
4796 if (dirty
& IRIS_DIRTY_BINDINGS_CS
)
4797 iris_populate_binding_table(ice
, batch
, MESA_SHADER_COMPUTE
, false);
4799 iris_use_optional_res(batch
, shs
->sampler_table
.res
, false);
4800 iris_use_pinned_bo(batch
, iris_resource_bo(shader
->assembly
.res
), false);
4802 if (ice
->state
.need_border_colors
)
4803 iris_use_pinned_bo(batch
, ice
->state
.border_color_pool
.bo
, false);
4805 if (dirty
& IRIS_DIRTY_CS
) {
4806 /* The MEDIA_VFE_STATE documentation for Gen8+ says:
4808 * "A stalling PIPE_CONTROL is required before MEDIA_VFE_STATE unless
4809 * the only bits that are changed are scoreboard related: Scoreboard
4810 * Enable, Scoreboard Type, Scoreboard Mask, Scoreboard Delta. For
4811 * these scoreboard related states, a MEDIA_STATE_FLUSH is
4814 iris_emit_pipe_control_flush(batch
, PIPE_CONTROL_CS_STALL
);
4816 iris_emit_cmd(batch
, GENX(MEDIA_VFE_STATE
), vfe
) {
4817 if (prog_data
->total_scratch
) {
4818 struct iris_bo
*bo
=
4819 iris_get_scratch_space(ice
, prog_data
->total_scratch
,
4820 MESA_SHADER_COMPUTE
);
4821 vfe
.PerThreadScratchSpace
= ffs(prog_data
->total_scratch
) - 11;
4822 vfe
.ScratchSpaceBasePointer
= rw_bo(bo
, 0);
4825 vfe
.MaximumNumberofThreads
=
4826 devinfo
->max_cs_threads
* screen
->subslice_total
- 1;
4828 vfe
.ResetGatewayTimer
=
4829 Resettingrelativetimerandlatchingtheglobaltimestamp
;
4832 vfe
.BypassGatewayControl
= true;
4834 vfe
.NumberofURBEntries
= 2;
4835 vfe
.URBEntryAllocationSize
= 2;
4837 // XXX: Use Indirect Payload Storage?
4838 vfe
.CURBEAllocationSize
=
4839 ALIGN(cs_prog_data
->push
.per_thread
.regs
* cs_prog_data
->threads
+
4840 cs_prog_data
->push
.cross_thread
.regs
, 2);
4844 // XXX: hack iris_set_constant_buffers to upload these thread counts
4845 // XXX: along with regular uniforms for compute shaders, somehow.
4847 uint32_t curbe_data_offset
= 0;
4848 // TODO: Move subgroup-id into uniforms ubo so we can push uniforms
4849 assert(cs_prog_data
->push
.cross_thread
.dwords
== 0 &&
4850 cs_prog_data
->push
.per_thread
.dwords
== 1 &&
4851 cs_prog_data
->base
.param
[0] == BRW_PARAM_BUILTIN_SUBGROUP_ID
);
4852 struct pipe_resource
*curbe_data_res
= NULL
;
4853 uint32_t *curbe_data_map
=
4854 stream_state(batch
, ice
->state
.dynamic_uploader
, &curbe_data_res
,
4855 ALIGN(cs_prog_data
->push
.total
.size
, 64), 64,
4856 &curbe_data_offset
);
4857 assert(curbe_data_map
);
4858 memset(curbe_data_map
, 0x5a, ALIGN(cs_prog_data
->push
.total
.size
, 64));
4859 iris_fill_cs_push_const_buffer(cs_prog_data
, curbe_data_map
);
4861 if (dirty
& IRIS_DIRTY_CONSTANTS_CS
) {
4862 iris_emit_cmd(batch
, GENX(MEDIA_CURBE_LOAD
), curbe
) {
4863 curbe
.CURBETotalDataLength
=
4864 ALIGN(cs_prog_data
->push
.total
.size
, 64);
4865 curbe
.CURBEDataStartAddress
= curbe_data_offset
;
4869 if (dirty
& (IRIS_DIRTY_SAMPLER_STATES_CS
|
4870 IRIS_DIRTY_BINDINGS_CS
|
4871 IRIS_DIRTY_CONSTANTS_CS
|
4873 struct pipe_resource
*desc_res
= NULL
;
4874 uint32_t desc
[GENX(INTERFACE_DESCRIPTOR_DATA_length
)];
4876 iris_pack_state(GENX(INTERFACE_DESCRIPTOR_DATA
), desc
, idd
) {
4877 idd
.SamplerStatePointer
= shs
->sampler_table
.offset
;
4878 idd
.BindingTablePointer
= binder
->bt_offset
[MESA_SHADER_COMPUTE
];
4881 for (int i
= 0; i
< GENX(INTERFACE_DESCRIPTOR_DATA_length
); i
++)
4882 desc
[i
] |= ((uint32_t *) shader
->derived_data
)[i
];
4884 iris_emit_cmd(batch
, GENX(MEDIA_INTERFACE_DESCRIPTOR_LOAD
), load
) {
4885 load
.InterfaceDescriptorTotalLength
=
4886 GENX(INTERFACE_DESCRIPTOR_DATA_length
) * sizeof(uint32_t);
4887 load
.InterfaceDescriptorDataStartAddress
=
4888 emit_state(batch
, ice
->state
.dynamic_uploader
,
4889 &desc_res
, desc
, sizeof(desc
), 32);
4892 pipe_resource_reference(&desc_res
, NULL
);
4895 uint32_t group_size
= grid
->block
[0] * grid
->block
[1] * grid
->block
[2];
4896 uint32_t remainder
= group_size
& (cs_prog_data
->simd_size
- 1);
4897 uint32_t right_mask
;
4900 right_mask
= ~0u >> (32 - remainder
);
4902 right_mask
= ~0u >> (32 - cs_prog_data
->simd_size
);
4904 #define GPGPU_DISPATCHDIMX 0x2500
4905 #define GPGPU_DISPATCHDIMY 0x2504
4906 #define GPGPU_DISPATCHDIMZ 0x2508
4908 if (grid
->indirect
) {
4909 struct iris_state_ref
*grid_size
= &ice
->state
.grid_size
;
4910 struct iris_bo
*bo
= iris_resource_bo(grid_size
->res
);
4911 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4912 lrm
.RegisterAddress
= GPGPU_DISPATCHDIMX
;
4913 lrm
.MemoryAddress
= ro_bo(bo
, grid_size
->offset
+ 0);
4915 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4916 lrm
.RegisterAddress
= GPGPU_DISPATCHDIMY
;
4917 lrm
.MemoryAddress
= ro_bo(bo
, grid_size
->offset
+ 4);
4919 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4920 lrm
.RegisterAddress
= GPGPU_DISPATCHDIMZ
;
4921 lrm
.MemoryAddress
= ro_bo(bo
, grid_size
->offset
+ 8);
4925 iris_emit_cmd(batch
, GENX(GPGPU_WALKER
), ggw
) {
4926 ggw
.IndirectParameterEnable
= grid
->indirect
!= NULL
;
4927 ggw
.SIMDSize
= cs_prog_data
->simd_size
/ 16;
4928 ggw
.ThreadDepthCounterMaximum
= 0;
4929 ggw
.ThreadHeightCounterMaximum
= 0;
4930 ggw
.ThreadWidthCounterMaximum
= cs_prog_data
->threads
- 1;
4931 ggw
.ThreadGroupIDXDimension
= grid
->grid
[0];
4932 ggw
.ThreadGroupIDYDimension
= grid
->grid
[1];
4933 ggw
.ThreadGroupIDZDimension
= grid
->grid
[2];
4934 ggw
.RightExecutionMask
= right_mask
;
4935 ggw
.BottomExecutionMask
= 0xffffffff;
4938 iris_emit_cmd(batch
, GENX(MEDIA_STATE_FLUSH
), msf
);
4940 if (!batch
->contains_draw
) {
4941 iris_restore_compute_saved_bos(ice
, batch
, grid
);
4942 batch
->contains_draw
= true;
4947 * State module teardown.
4950 iris_destroy_state(struct iris_context
*ice
)
4952 struct iris_genx_state
*genx
= ice
->state
.genx
;
4954 uint64_t bound_vbs
= ice
->state
.bound_vertex_buffers
;
4956 const int i
= u_bit_scan64(&bound_vbs
);
4957 pipe_resource_reference(&genx
->vertex_buffers
[i
].resource
, NULL
);
4959 free(ice
->state
.genx
);
4961 for (unsigned i
= 0; i
< ice
->state
.framebuffer
.nr_cbufs
; i
++) {
4962 pipe_surface_reference(&ice
->state
.framebuffer
.cbufs
[i
], NULL
);
4964 pipe_surface_reference(&ice
->state
.framebuffer
.zsbuf
, NULL
);
4966 for (int stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
4967 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4968 pipe_resource_reference(&shs
->sampler_table
.res
, NULL
);
4969 for (int i
= 0; i
< PIPE_MAX_CONSTANT_BUFFERS
; i
++) {
4970 pipe_resource_reference(&shs
->constbuf
[i
].data
.res
, NULL
);
4971 pipe_resource_reference(&shs
->constbuf
[i
].surface_state
.res
, NULL
);
4973 for (int i
= 0; i
< PIPE_MAX_SHADER_IMAGES
; i
++) {
4974 pipe_resource_reference(&shs
->image
[i
].res
, NULL
);
4975 pipe_resource_reference(&shs
->image
[i
].surface_state
.res
, NULL
);
4977 for (int i
= 0; i
< PIPE_MAX_SHADER_BUFFERS
; i
++) {
4978 pipe_resource_reference(&shs
->ssbo
[i
], NULL
);
4979 pipe_resource_reference(&shs
->ssbo_surface_state
[i
].res
, NULL
);
4981 for (int i
= 0; i
< IRIS_MAX_TEXTURE_SAMPLERS
; i
++) {
4982 pipe_sampler_view_reference((struct pipe_sampler_view
**)
4983 &shs
->textures
[i
], NULL
);
4987 pipe_resource_reference(&ice
->state
.grid_size
.res
, NULL
);
4988 pipe_resource_reference(&ice
->state
.grid_surf_state
.res
, NULL
);
4990 pipe_resource_reference(&ice
->state
.null_fb
.res
, NULL
);
4991 pipe_resource_reference(&ice
->state
.unbound_tex
.res
, NULL
);
4993 pipe_resource_reference(&ice
->state
.last_res
.cc_vp
, NULL
);
4994 pipe_resource_reference(&ice
->state
.last_res
.sf_cl_vp
, NULL
);
4995 pipe_resource_reference(&ice
->state
.last_res
.color_calc
, NULL
);
4996 pipe_resource_reference(&ice
->state
.last_res
.scissor
, NULL
);
4997 pipe_resource_reference(&ice
->state
.last_res
.blend
, NULL
);
4998 pipe_resource_reference(&ice
->state
.last_res
.index_buffer
, NULL
);
5001 /* ------------------------------------------------------------------- */
5004 iris_load_register_reg32(struct iris_batch
*batch
, uint32_t dst
,
5007 _iris_emit_lrr(batch
, dst
, src
);
5011 iris_load_register_reg64(struct iris_batch
*batch
, uint32_t dst
,
5014 _iris_emit_lrr(batch
, dst
, src
);
5015 _iris_emit_lrr(batch
, dst
+ 4, src
+ 4);
5019 iris_load_register_imm32(struct iris_batch
*batch
, uint32_t reg
,
5022 _iris_emit_lri(batch
, reg
, val
);
5026 iris_load_register_imm64(struct iris_batch
*batch
, uint32_t reg
,
5029 _iris_emit_lri(batch
, reg
+ 0, val
& 0xffffffff);
5030 _iris_emit_lri(batch
, reg
+ 4, val
>> 32);
5034 * Emit MI_LOAD_REGISTER_MEM to load a 32-bit MMIO register from a buffer.
5037 iris_load_register_mem32(struct iris_batch
*batch
, uint32_t reg
,
5038 struct iris_bo
*bo
, uint32_t offset
)
5040 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
5041 lrm
.RegisterAddress
= reg
;
5042 lrm
.MemoryAddress
= ro_bo(bo
, offset
);
5047 * Load a 64-bit value from a buffer into a MMIO register via
5048 * two MI_LOAD_REGISTER_MEM commands.
5051 iris_load_register_mem64(struct iris_batch
*batch
, uint32_t reg
,
5052 struct iris_bo
*bo
, uint32_t offset
)
5054 iris_load_register_mem32(batch
, reg
+ 0, bo
, offset
+ 0);
5055 iris_load_register_mem32(batch
, reg
+ 4, bo
, offset
+ 4);
5059 iris_store_register_mem32(struct iris_batch
*batch
, uint32_t reg
,
5060 struct iris_bo
*bo
, uint32_t offset
,
5063 iris_emit_cmd(batch
, GENX(MI_STORE_REGISTER_MEM
), srm
) {
5064 srm
.RegisterAddress
= reg
;
5065 srm
.MemoryAddress
= rw_bo(bo
, offset
);
5066 srm
.PredicateEnable
= predicated
;
5071 iris_store_register_mem64(struct iris_batch
*batch
, uint32_t reg
,
5072 struct iris_bo
*bo
, uint32_t offset
,
5075 iris_store_register_mem32(batch
, reg
+ 0, bo
, offset
+ 0, predicated
);
5076 iris_store_register_mem32(batch
, reg
+ 4, bo
, offset
+ 4, predicated
);
5080 iris_store_data_imm32(struct iris_batch
*batch
,
5081 struct iris_bo
*bo
, uint32_t offset
,
5084 iris_emit_cmd(batch
, GENX(MI_STORE_DATA_IMM
), sdi
) {
5085 sdi
.Address
= rw_bo(bo
, offset
);
5086 sdi
.ImmediateData
= imm
;
5091 iris_store_data_imm64(struct iris_batch
*batch
,
5092 struct iris_bo
*bo
, uint32_t offset
,
5095 /* Can't use iris_emit_cmd because MI_STORE_DATA_IMM has a length of
5096 * 2 in genxml but it's actually variable length and we need 5 DWords.
5098 void *map
= iris_get_command_space(batch
, 4 * 5);
5099 _iris_pack_command(batch
, GENX(MI_STORE_DATA_IMM
), map
, sdi
) {
5100 sdi
.DWordLength
= 5 - 2;
5101 sdi
.Address
= rw_bo(bo
, offset
);
5102 sdi
.ImmediateData
= imm
;
5107 iris_copy_mem_mem(struct iris_batch
*batch
,
5108 struct iris_bo
*dst_bo
, uint32_t dst_offset
,
5109 struct iris_bo
*src_bo
, uint32_t src_offset
,
5112 /* MI_COPY_MEM_MEM operates on DWords. */
5113 assert(bytes
% 4 == 0);
5114 assert(dst_offset
% 4 == 0);
5115 assert(src_offset
% 4 == 0);
5117 for (unsigned i
= 0; i
< bytes
; i
+= 4) {
5118 iris_emit_cmd(batch
, GENX(MI_COPY_MEM_MEM
), cp
) {
5119 cp
.DestinationMemoryAddress
= rw_bo(dst_bo
, dst_offset
+ i
);
5120 cp
.SourceMemoryAddress
= ro_bo(src_bo
, src_offset
+ i
);
5125 /* ------------------------------------------------------------------- */
5128 flags_to_post_sync_op(uint32_t flags
)
5130 if (flags
& PIPE_CONTROL_WRITE_IMMEDIATE
)
5131 return WriteImmediateData
;
5133 if (flags
& PIPE_CONTROL_WRITE_DEPTH_COUNT
)
5134 return WritePSDepthCount
;
5136 if (flags
& PIPE_CONTROL_WRITE_TIMESTAMP
)
5137 return WriteTimestamp
;
5143 * Do the given flags have a Post Sync or LRI Post Sync operation?
5145 static enum pipe_control_flags
5146 get_post_sync_flags(enum pipe_control_flags flags
)
5148 flags
&= PIPE_CONTROL_WRITE_IMMEDIATE
|
5149 PIPE_CONTROL_WRITE_DEPTH_COUNT
|
5150 PIPE_CONTROL_WRITE_TIMESTAMP
|
5151 PIPE_CONTROL_LRI_POST_SYNC_OP
;
5153 /* Only one "Post Sync Op" is allowed, and it's mutually exclusive with
5154 * "LRI Post Sync Operation". So more than one bit set would be illegal.
5156 assert(util_bitcount(flags
) <= 1);
5161 #define IS_COMPUTE_PIPELINE(batch) (batch->name == IRIS_BATCH_COMPUTE)
5164 * Emit a series of PIPE_CONTROL commands, taking into account any
5165 * workarounds necessary to actually accomplish the caller's request.
5167 * Unless otherwise noted, spec quotations in this function come from:
5169 * Synchronization of the 3D Pipeline > PIPE_CONTROL Command > Programming
5170 * Restrictions for PIPE_CONTROL.
5172 * You should not use this function directly. Use the helpers in
5173 * iris_pipe_control.c instead, which may split the pipe control further.
5176 iris_emit_raw_pipe_control(struct iris_batch
*batch
, uint32_t flags
,
5177 struct iris_bo
*bo
, uint32_t offset
, uint64_t imm
)
5179 UNUSED
const struct gen_device_info
*devinfo
= &batch
->screen
->devinfo
;
5180 enum pipe_control_flags post_sync_flags
= get_post_sync_flags(flags
);
5181 enum pipe_control_flags non_lri_post_sync_flags
=
5182 post_sync_flags
& ~PIPE_CONTROL_LRI_POST_SYNC_OP
;
5184 /* Recursive PIPE_CONTROL workarounds --------------------------------
5185 * (http://knowyourmeme.com/memes/xzibit-yo-dawg)
5187 * We do these first because we want to look at the original operation,
5188 * rather than any workarounds we set.
5190 if (GEN_GEN
== 9 && (flags
& PIPE_CONTROL_VF_CACHE_INVALIDATE
)) {
5191 /* The PIPE_CONTROL "VF Cache Invalidation Enable" bit description
5192 * lists several workarounds:
5194 * "Project: SKL, KBL, BXT
5196 * If the VF Cache Invalidation Enable is set to a 1 in a
5197 * PIPE_CONTROL, a separate Null PIPE_CONTROL, all bitfields
5198 * sets to 0, with the VF Cache Invalidation Enable set to 0
5199 * needs to be sent prior to the PIPE_CONTROL with VF Cache
5200 * Invalidation Enable set to a 1."
5202 iris_emit_raw_pipe_control(batch
, 0, NULL
, 0, 0);
5205 if (GEN_GEN
== 9 && IS_COMPUTE_PIPELINE(batch
) && post_sync_flags
) {
5206 /* Project: SKL / Argument: LRI Post Sync Operation [23]
5208 * "PIPECONTROL command with “Command Streamer Stall Enable” must be
5209 * programmed prior to programming a PIPECONTROL command with "LRI
5210 * Post Sync Operation" in GPGPU mode of operation (i.e when
5211 * PIPELINE_SELECT command is set to GPGPU mode of operation)."
5213 * The same text exists a few rows below for Post Sync Op.
5215 iris_emit_raw_pipe_control(batch
, PIPE_CONTROL_CS_STALL
, bo
, offset
, imm
);
5218 if (GEN_GEN
== 10 && (flags
& PIPE_CONTROL_RENDER_TARGET_FLUSH
)) {
5220 * "Before sending a PIPE_CONTROL command with bit 12 set, SW must issue
5221 * another PIPE_CONTROL with Render Target Cache Flush Enable (bit 12)
5222 * = 0 and Pipe Control Flush Enable (bit 7) = 1"
5224 iris_emit_raw_pipe_control(batch
, PIPE_CONTROL_FLUSH_ENABLE
, bo
,
5228 /* "Flush Types" workarounds ---------------------------------------------
5229 * We do these now because they may add post-sync operations or CS stalls.
5232 if (GEN_GEN
< 11 && flags
& PIPE_CONTROL_VF_CACHE_INVALIDATE
) {
5233 /* Project: BDW, SKL+ (stopping at CNL) / Argument: VF Invalidate
5235 * "'Post Sync Operation' must be enabled to 'Write Immediate Data' or
5236 * 'Write PS Depth Count' or 'Write Timestamp'."
5239 flags
|= PIPE_CONTROL_WRITE_IMMEDIATE
;
5240 post_sync_flags
|= PIPE_CONTROL_WRITE_IMMEDIATE
;
5241 non_lri_post_sync_flags
|= PIPE_CONTROL_WRITE_IMMEDIATE
;
5242 bo
= batch
->screen
->workaround_bo
;
5246 /* #1130 from Gen10 workarounds page:
5248 * "Enable Depth Stall on every Post Sync Op if Render target Cache
5249 * Flush is not enabled in same PIPE CONTROL and Enable Pixel score
5250 * board stall if Render target cache flush is enabled."
5252 * Applicable to CNL B0 and C0 steppings only.
5254 * The wording here is unclear, and this workaround doesn't look anything
5255 * like the internal bug report recommendations, but leave it be for now...
5257 if (GEN_GEN
== 10) {
5258 if (flags
& PIPE_CONTROL_RENDER_TARGET_FLUSH
) {
5259 flags
|= PIPE_CONTROL_STALL_AT_SCOREBOARD
;
5260 } else if (flags
& non_lri_post_sync_flags
) {
5261 flags
|= PIPE_CONTROL_DEPTH_STALL
;
5265 if (flags
& PIPE_CONTROL_DEPTH_STALL
) {
5266 /* From the PIPE_CONTROL instruction table, bit 13 (Depth Stall Enable):
5268 * "This bit must be DISABLED for operations other than writing
5271 * This seems like nonsense. An Ivybridge workaround requires us to
5272 * emit a PIPE_CONTROL with a depth stall and write immediate post-sync
5273 * operation. Gen8+ requires us to emit depth stalls and depth cache
5274 * flushes together. So, it's hard to imagine this means anything other
5275 * than "we originally intended this to be used for PS_DEPTH_COUNT".
5277 * We ignore the supposed restriction and do nothing.
5281 if (flags
& (PIPE_CONTROL_RENDER_TARGET_FLUSH
|
5282 PIPE_CONTROL_STALL_AT_SCOREBOARD
)) {
5283 /* From the PIPE_CONTROL instruction table, bit 12 and bit 1:
5285 * "This bit must be DISABLED for End-of-pipe (Read) fences,
5286 * PS_DEPTH_COUNT or TIMESTAMP queries."
5288 * TODO: Implement end-of-pipe checking.
5290 assert(!(post_sync_flags
& (PIPE_CONTROL_WRITE_DEPTH_COUNT
|
5291 PIPE_CONTROL_WRITE_TIMESTAMP
)));
5294 if (GEN_GEN
< 11 && (flags
& PIPE_CONTROL_STALL_AT_SCOREBOARD
)) {
5295 /* From the PIPE_CONTROL instruction table, bit 1:
5297 * "This bit is ignored if Depth Stall Enable is set.
5298 * Further, the render cache is not flushed even if Write Cache
5299 * Flush Enable bit is set."
5301 * We assert that the caller doesn't do this combination, to try and
5302 * prevent mistakes. It shouldn't hurt the GPU, though.
5304 * We skip this check on Gen11+ as the "Stall at Pixel Scoreboard"
5305 * and "Render Target Flush" combo is explicitly required for BTI
5306 * update workarounds.
5308 assert(!(flags
& (PIPE_CONTROL_DEPTH_STALL
|
5309 PIPE_CONTROL_RENDER_TARGET_FLUSH
)));
5312 /* PIPE_CONTROL page workarounds ------------------------------------- */
5314 if (GEN_GEN
<= 8 && (flags
& PIPE_CONTROL_STATE_CACHE_INVALIDATE
)) {
5315 /* From the PIPE_CONTROL page itself:
5318 * Restriction: Pipe_control with CS-stall bit set must be issued
5319 * before a pipe-control command that has the State Cache
5320 * Invalidate bit set."
5322 flags
|= PIPE_CONTROL_CS_STALL
;
5325 if (flags
& PIPE_CONTROL_FLUSH_LLC
) {
5326 /* From the PIPE_CONTROL instruction table, bit 26 (Flush LLC):
5329 * SW must always program Post-Sync Operation to "Write Immediate
5330 * Data" when Flush LLC is set."
5332 * For now, we just require the caller to do it.
5334 assert(flags
& PIPE_CONTROL_WRITE_IMMEDIATE
);
5337 /* "Post-Sync Operation" workarounds -------------------------------- */
5339 /* Project: All / Argument: Global Snapshot Count Reset [19]
5341 * "This bit must not be exercised on any product.
5342 * Requires stall bit ([20] of DW1) set."
5344 * We don't use this, so we just assert that it isn't used. The
5345 * PIPE_CONTROL instruction page indicates that they intended this
5346 * as a debug feature and don't think it is useful in production,
5347 * but it may actually be usable, should we ever want to.
5349 assert((flags
& PIPE_CONTROL_GLOBAL_SNAPSHOT_COUNT_RESET
) == 0);
5351 if (flags
& (PIPE_CONTROL_MEDIA_STATE_CLEAR
|
5352 PIPE_CONTROL_INDIRECT_STATE_POINTERS_DISABLE
)) {
5353 /* Project: All / Arguments:
5355 * - Generic Media State Clear [16]
5356 * - Indirect State Pointers Disable [16]
5358 * "Requires stall bit ([20] of DW1) set."
5360 * Also, the PIPE_CONTROL instruction table, bit 16 (Generic Media
5361 * State Clear) says:
5363 * "PIPECONTROL command with “Command Streamer Stall Enable” must be
5364 * programmed prior to programming a PIPECONTROL command with "Media
5365 * State Clear" set in GPGPU mode of operation"
5367 * This is a subset of the earlier rule, so there's nothing to do.
5369 flags
|= PIPE_CONTROL_CS_STALL
;
5372 if (flags
& PIPE_CONTROL_STORE_DATA_INDEX
) {
5373 /* Project: All / Argument: Store Data Index
5375 * "Post-Sync Operation ([15:14] of DW1) must be set to something other
5378 * For now, we just assert that the caller does this. We might want to
5379 * automatically add a write to the workaround BO...
5381 assert(non_lri_post_sync_flags
!= 0);
5384 if (flags
& PIPE_CONTROL_SYNC_GFDT
) {
5385 /* Project: All / Argument: Sync GFDT
5387 * "Post-Sync Operation ([15:14] of DW1) must be set to something other
5388 * than '0' or 0x2520[13] must be set."
5390 * For now, we just assert that the caller does this.
5392 assert(non_lri_post_sync_flags
!= 0);
5395 if (flags
& PIPE_CONTROL_TLB_INVALIDATE
) {
5396 /* Project: IVB+ / Argument: TLB inv
5398 * "Requires stall bit ([20] of DW1) set."
5400 * Also, from the PIPE_CONTROL instruction table:
5403 * Post Sync Operation or CS stall must be set to ensure a TLB
5404 * invalidation occurs. Otherwise no cycle will occur to the TLB
5405 * cache to invalidate."
5407 * This is not a subset of the earlier rule, so there's nothing to do.
5409 flags
|= PIPE_CONTROL_CS_STALL
;
5412 if (GEN_GEN
== 9 && devinfo
->gt
== 4) {
5413 /* TODO: The big Skylake GT4 post sync op workaround */
5416 /* "GPGPU specific workarounds" (both post-sync and flush) ------------ */
5418 if (IS_COMPUTE_PIPELINE(batch
)) {
5419 if (GEN_GEN
>= 9 && (flags
& PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
)) {
5420 /* Project: SKL+ / Argument: Tex Invalidate
5421 * "Requires stall bit ([20] of DW) set for all GPGPU Workloads."
5423 flags
|= PIPE_CONTROL_CS_STALL
;
5426 if (GEN_GEN
== 8 && (post_sync_flags
||
5427 (flags
& (PIPE_CONTROL_NOTIFY_ENABLE
|
5428 PIPE_CONTROL_DEPTH_STALL
|
5429 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
5430 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
5431 PIPE_CONTROL_DATA_CACHE_FLUSH
)))) {
5432 /* Project: BDW / Arguments:
5434 * - LRI Post Sync Operation [23]
5435 * - Post Sync Op [15:14]
5437 * - Depth Stall [13]
5438 * - Render Target Cache Flush [12]
5439 * - Depth Cache Flush [0]
5440 * - DC Flush Enable [5]
5442 * "Requires stall bit ([20] of DW) set for all GPGPU and Media
5445 flags
|= PIPE_CONTROL_CS_STALL
;
5447 /* Also, from the PIPE_CONTROL instruction table, bit 20:
5450 * This bit must be always set when PIPE_CONTROL command is
5451 * programmed by GPGPU and MEDIA workloads, except for the cases
5452 * when only Read Only Cache Invalidation bits are set (State
5453 * Cache Invalidation Enable, Instruction cache Invalidation
5454 * Enable, Texture Cache Invalidation Enable, Constant Cache
5455 * Invalidation Enable). This is to WA FFDOP CG issue, this WA
5456 * need not implemented when FF_DOP_CG is disable via "Fixed
5457 * Function DOP Clock Gate Disable" bit in RC_PSMI_CTRL register."
5459 * It sounds like we could avoid CS stalls in some cases, but we
5460 * don't currently bother. This list isn't exactly the list above,
5466 /* "Stall" workarounds ----------------------------------------------
5467 * These have to come after the earlier ones because we may have added
5468 * some additional CS stalls above.
5471 if (GEN_GEN
< 9 && (flags
& PIPE_CONTROL_CS_STALL
)) {
5472 /* Project: PRE-SKL, VLV, CHV
5474 * "[All Stepping][All SKUs]:
5476 * One of the following must also be set:
5478 * - Render Target Cache Flush Enable ([12] of DW1)
5479 * - Depth Cache Flush Enable ([0] of DW1)
5480 * - Stall at Pixel Scoreboard ([1] of DW1)
5481 * - Depth Stall ([13] of DW1)
5482 * - Post-Sync Operation ([13] of DW1)
5483 * - DC Flush Enable ([5] of DW1)"
5485 * If we don't already have one of those bits set, we choose to add
5486 * "Stall at Pixel Scoreboard". Some of the other bits require a
5487 * CS stall as a workaround (see above), which would send us into
5488 * an infinite recursion of PIPE_CONTROLs. "Stall at Pixel Scoreboard"
5489 * appears to be safe, so we choose that.
5491 const uint32_t wa_bits
= PIPE_CONTROL_RENDER_TARGET_FLUSH
|
5492 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
5493 PIPE_CONTROL_WRITE_IMMEDIATE
|
5494 PIPE_CONTROL_WRITE_DEPTH_COUNT
|
5495 PIPE_CONTROL_WRITE_TIMESTAMP
|
5496 PIPE_CONTROL_STALL_AT_SCOREBOARD
|
5497 PIPE_CONTROL_DEPTH_STALL
|
5498 PIPE_CONTROL_DATA_CACHE_FLUSH
;
5499 if (!(flags
& wa_bits
))
5500 flags
|= PIPE_CONTROL_STALL_AT_SCOREBOARD
;
5503 /* Emit --------------------------------------------------------------- */
5505 iris_emit_cmd(batch
, GENX(PIPE_CONTROL
), pc
) {
5506 pc
.LRIPostSyncOperation
= NoLRIOperation
;
5507 pc
.PipeControlFlushEnable
= flags
& PIPE_CONTROL_FLUSH_ENABLE
;
5508 pc
.DCFlushEnable
= flags
& PIPE_CONTROL_DATA_CACHE_FLUSH
;
5509 pc
.StoreDataIndex
= 0;
5510 pc
.CommandStreamerStallEnable
= flags
& PIPE_CONTROL_CS_STALL
;
5511 pc
.GlobalSnapshotCountReset
=
5512 flags
& PIPE_CONTROL_GLOBAL_SNAPSHOT_COUNT_RESET
;
5513 pc
.TLBInvalidate
= flags
& PIPE_CONTROL_TLB_INVALIDATE
;
5514 pc
.GenericMediaStateClear
= flags
& PIPE_CONTROL_MEDIA_STATE_CLEAR
;
5515 pc
.StallAtPixelScoreboard
= flags
& PIPE_CONTROL_STALL_AT_SCOREBOARD
;
5516 pc
.RenderTargetCacheFlushEnable
=
5517 flags
& PIPE_CONTROL_RENDER_TARGET_FLUSH
;
5518 pc
.DepthCacheFlushEnable
= flags
& PIPE_CONTROL_DEPTH_CACHE_FLUSH
;
5519 pc
.StateCacheInvalidationEnable
=
5520 flags
& PIPE_CONTROL_STATE_CACHE_INVALIDATE
;
5521 pc
.VFCacheInvalidationEnable
= flags
& PIPE_CONTROL_VF_CACHE_INVALIDATE
;
5522 pc
.ConstantCacheInvalidationEnable
=
5523 flags
& PIPE_CONTROL_CONST_CACHE_INVALIDATE
;
5524 pc
.PostSyncOperation
= flags_to_post_sync_op(flags
);
5525 pc
.DepthStallEnable
= flags
& PIPE_CONTROL_DEPTH_STALL
;
5526 pc
.InstructionCacheInvalidateEnable
=
5527 flags
& PIPE_CONTROL_INSTRUCTION_INVALIDATE
;
5528 pc
.NotifyEnable
= flags
& PIPE_CONTROL_NOTIFY_ENABLE
;
5529 pc
.IndirectStatePointersDisable
=
5530 flags
& PIPE_CONTROL_INDIRECT_STATE_POINTERS_DISABLE
;
5531 pc
.TextureCacheInvalidationEnable
=
5532 flags
& PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
5533 pc
.Address
= rw_bo(bo
, offset
);
5534 pc
.ImmediateData
= imm
;
5539 genX(init_state
)(struct iris_context
*ice
)
5541 struct pipe_context
*ctx
= &ice
->ctx
;
5542 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
5544 ctx
->create_blend_state
= iris_create_blend_state
;
5545 ctx
->create_depth_stencil_alpha_state
= iris_create_zsa_state
;
5546 ctx
->create_rasterizer_state
= iris_create_rasterizer_state
;
5547 ctx
->create_sampler_state
= iris_create_sampler_state
;
5548 ctx
->create_sampler_view
= iris_create_sampler_view
;
5549 ctx
->create_surface
= iris_create_surface
;
5550 ctx
->create_vertex_elements_state
= iris_create_vertex_elements
;
5551 ctx
->bind_blend_state
= iris_bind_blend_state
;
5552 ctx
->bind_depth_stencil_alpha_state
= iris_bind_zsa_state
;
5553 ctx
->bind_sampler_states
= iris_bind_sampler_states
;
5554 ctx
->bind_rasterizer_state
= iris_bind_rasterizer_state
;
5555 ctx
->bind_vertex_elements_state
= iris_bind_vertex_elements_state
;
5556 ctx
->delete_blend_state
= iris_delete_state
;
5557 ctx
->delete_depth_stencil_alpha_state
= iris_delete_state
;
5558 ctx
->delete_rasterizer_state
= iris_delete_state
;
5559 ctx
->delete_sampler_state
= iris_delete_state
;
5560 ctx
->delete_vertex_elements_state
= iris_delete_state
;
5561 ctx
->set_blend_color
= iris_set_blend_color
;
5562 ctx
->set_clip_state
= iris_set_clip_state
;
5563 ctx
->set_constant_buffer
= iris_set_constant_buffer
;
5564 ctx
->set_shader_buffers
= iris_set_shader_buffers
;
5565 ctx
->set_shader_images
= iris_set_shader_images
;
5566 ctx
->set_sampler_views
= iris_set_sampler_views
;
5567 ctx
->set_tess_state
= iris_set_tess_state
;
5568 ctx
->set_framebuffer_state
= iris_set_framebuffer_state
;
5569 ctx
->set_polygon_stipple
= iris_set_polygon_stipple
;
5570 ctx
->set_sample_mask
= iris_set_sample_mask
;
5571 ctx
->set_scissor_states
= iris_set_scissor_states
;
5572 ctx
->set_stencil_ref
= iris_set_stencil_ref
;
5573 ctx
->set_vertex_buffers
= iris_set_vertex_buffers
;
5574 ctx
->set_viewport_states
= iris_set_viewport_states
;
5575 ctx
->sampler_view_destroy
= iris_sampler_view_destroy
;
5576 ctx
->surface_destroy
= iris_surface_destroy
;
5577 ctx
->draw_vbo
= iris_draw_vbo
;
5578 ctx
->launch_grid
= iris_launch_grid
;
5579 ctx
->create_stream_output_target
= iris_create_stream_output_target
;
5580 ctx
->stream_output_target_destroy
= iris_stream_output_target_destroy
;
5581 ctx
->set_stream_output_targets
= iris_set_stream_output_targets
;
5583 ice
->vtbl
.destroy_state
= iris_destroy_state
;
5584 ice
->vtbl
.init_render_context
= iris_init_render_context
;
5585 ice
->vtbl
.init_compute_context
= iris_init_compute_context
;
5586 ice
->vtbl
.upload_render_state
= iris_upload_render_state
;
5587 ice
->vtbl
.update_surface_base_address
= iris_update_surface_base_address
;
5588 ice
->vtbl
.upload_compute_state
= iris_upload_compute_state
;
5589 ice
->vtbl
.emit_raw_pipe_control
= iris_emit_raw_pipe_control
;
5590 ice
->vtbl
.load_register_reg32
= iris_load_register_reg32
;
5591 ice
->vtbl
.load_register_reg64
= iris_load_register_reg64
;
5592 ice
->vtbl
.load_register_imm32
= iris_load_register_imm32
;
5593 ice
->vtbl
.load_register_imm64
= iris_load_register_imm64
;
5594 ice
->vtbl
.load_register_mem32
= iris_load_register_mem32
;
5595 ice
->vtbl
.load_register_mem64
= iris_load_register_mem64
;
5596 ice
->vtbl
.store_register_mem32
= iris_store_register_mem32
;
5597 ice
->vtbl
.store_register_mem64
= iris_store_register_mem64
;
5598 ice
->vtbl
.store_data_imm32
= iris_store_data_imm32
;
5599 ice
->vtbl
.store_data_imm64
= iris_store_data_imm64
;
5600 ice
->vtbl
.copy_mem_mem
= iris_copy_mem_mem
;
5601 ice
->vtbl
.derived_program_state_size
= iris_derived_program_state_size
;
5602 ice
->vtbl
.store_derived_program_state
= iris_store_derived_program_state
;
5603 ice
->vtbl
.create_so_decl_list
= iris_create_so_decl_list
;
5604 ice
->vtbl
.populate_vs_key
= iris_populate_vs_key
;
5605 ice
->vtbl
.populate_tcs_key
= iris_populate_tcs_key
;
5606 ice
->vtbl
.populate_tes_key
= iris_populate_tes_key
;
5607 ice
->vtbl
.populate_gs_key
= iris_populate_gs_key
;
5608 ice
->vtbl
.populate_fs_key
= iris_populate_fs_key
;
5609 ice
->vtbl
.populate_cs_key
= iris_populate_cs_key
;
5611 ice
->state
.dirty
= ~0ull;
5613 ice
->state
.statistics_counters_enabled
= true;
5615 ice
->state
.sample_mask
= 0xffff;
5616 ice
->state
.num_viewports
= 1;
5617 ice
->state
.genx
= calloc(1, sizeof(struct iris_genx_state
));
5619 /* Make a 1x1x1 null surface for unbound textures */
5620 void *null_surf_map
=
5621 upload_state(ice
->state
.surface_uploader
, &ice
->state
.unbound_tex
,
5622 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
5623 isl_null_fill_state(&screen
->isl_dev
, null_surf_map
, isl_extent3d(1, 1, 1));
5624 ice
->state
.unbound_tex
.offset
+=
5625 iris_bo_offset_from_base_address(iris_resource_bo(ice
->state
.unbound_tex
.res
));
5627 /* Default all scissor rectangles to be empty regions. */
5628 for (int i
= 0; i
< IRIS_MAX_VIEWPORTS
; i
++) {
5629 ice
->state
.scissors
[i
] = (struct pipe_scissor_state
) {
5630 .minx
= 1, .maxx
= 0, .miny
= 1, .maxy
= 0,