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_pipe.h"
105 #include "iris_resource.h"
107 #define __gen_address_type struct iris_address
108 #define __gen_user_data struct iris_batch
110 #define ARRAY_BYTES(x) (sizeof(uint32_t) * ARRAY_SIZE(x))
113 __gen_combine_address(struct iris_batch
*batch
, void *location
,
114 struct iris_address addr
, uint32_t delta
)
116 uint64_t result
= addr
.offset
+ delta
;
119 iris_use_pinned_bo(batch
, addr
.bo
, addr
.write
);
120 /* Assume this is a general address, not relative to a base. */
121 result
+= addr
.bo
->gtt_offset
;
127 #define __genxml_cmd_length(cmd) cmd ## _length
128 #define __genxml_cmd_length_bias(cmd) cmd ## _length_bias
129 #define __genxml_cmd_header(cmd) cmd ## _header
130 #define __genxml_cmd_pack(cmd) cmd ## _pack
132 #define _iris_pack_command(batch, cmd, dst, name) \
133 for (struct cmd name = { __genxml_cmd_header(cmd) }, \
134 *_dst = (void *)(dst); __builtin_expect(_dst != NULL, 1); \
135 ({ __genxml_cmd_pack(cmd)(batch, (void *)_dst, &name); \
139 #define iris_pack_command(cmd, dst, name) \
140 _iris_pack_command(NULL, cmd, dst, name)
142 #define iris_pack_state(cmd, dst, name) \
143 for (struct cmd name = {}, \
144 *_dst = (void *)(dst); __builtin_expect(_dst != NULL, 1); \
145 __genxml_cmd_pack(cmd)(NULL, (void *)_dst, &name), \
148 #define iris_emit_cmd(batch, cmd, name) \
149 _iris_pack_command(batch, cmd, iris_get_command_space(batch, 4 * __genxml_cmd_length(cmd)), name)
151 #define iris_emit_merge(batch, dwords0, dwords1, num_dwords) \
153 uint32_t *dw = iris_get_command_space(batch, 4 * num_dwords); \
154 for (uint32_t i = 0; i < num_dwords; i++) \
155 dw[i] = (dwords0)[i] | (dwords1)[i]; \
156 VG(VALGRIND_CHECK_MEM_IS_DEFINED(dw, num_dwords)); \
159 #include "genxml/genX_pack.h"
160 #include "genxml/gen_macros.h"
161 #include "genxml/genX_bits.h"
163 #define MOCS_WB (2 << 1)
166 * Statically assert that PIPE_* enums match the hardware packets.
167 * (As long as they match, we don't need to translate them.)
169 UNUSED
static void pipe_asserts()
171 #define PIPE_ASSERT(x) STATIC_ASSERT((int)x)
173 /* pipe_logicop happens to match the hardware. */
174 PIPE_ASSERT(PIPE_LOGICOP_CLEAR
== LOGICOP_CLEAR
);
175 PIPE_ASSERT(PIPE_LOGICOP_NOR
== LOGICOP_NOR
);
176 PIPE_ASSERT(PIPE_LOGICOP_AND_INVERTED
== LOGICOP_AND_INVERTED
);
177 PIPE_ASSERT(PIPE_LOGICOP_COPY_INVERTED
== LOGICOP_COPY_INVERTED
);
178 PIPE_ASSERT(PIPE_LOGICOP_AND_REVERSE
== LOGICOP_AND_REVERSE
);
179 PIPE_ASSERT(PIPE_LOGICOP_INVERT
== LOGICOP_INVERT
);
180 PIPE_ASSERT(PIPE_LOGICOP_XOR
== LOGICOP_XOR
);
181 PIPE_ASSERT(PIPE_LOGICOP_NAND
== LOGICOP_NAND
);
182 PIPE_ASSERT(PIPE_LOGICOP_AND
== LOGICOP_AND
);
183 PIPE_ASSERT(PIPE_LOGICOP_EQUIV
== LOGICOP_EQUIV
);
184 PIPE_ASSERT(PIPE_LOGICOP_NOOP
== LOGICOP_NOOP
);
185 PIPE_ASSERT(PIPE_LOGICOP_OR_INVERTED
== LOGICOP_OR_INVERTED
);
186 PIPE_ASSERT(PIPE_LOGICOP_COPY
== LOGICOP_COPY
);
187 PIPE_ASSERT(PIPE_LOGICOP_OR_REVERSE
== LOGICOP_OR_REVERSE
);
188 PIPE_ASSERT(PIPE_LOGICOP_OR
== LOGICOP_OR
);
189 PIPE_ASSERT(PIPE_LOGICOP_SET
== LOGICOP_SET
);
191 /* pipe_blend_func happens to match the hardware. */
192 PIPE_ASSERT(PIPE_BLENDFACTOR_ONE
== BLENDFACTOR_ONE
);
193 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_COLOR
== BLENDFACTOR_SRC_COLOR
);
194 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_ALPHA
== BLENDFACTOR_SRC_ALPHA
);
195 PIPE_ASSERT(PIPE_BLENDFACTOR_DST_ALPHA
== BLENDFACTOR_DST_ALPHA
);
196 PIPE_ASSERT(PIPE_BLENDFACTOR_DST_COLOR
== BLENDFACTOR_DST_COLOR
);
197 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
== BLENDFACTOR_SRC_ALPHA_SATURATE
);
198 PIPE_ASSERT(PIPE_BLENDFACTOR_CONST_COLOR
== BLENDFACTOR_CONST_COLOR
);
199 PIPE_ASSERT(PIPE_BLENDFACTOR_CONST_ALPHA
== BLENDFACTOR_CONST_ALPHA
);
200 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC1_COLOR
== BLENDFACTOR_SRC1_COLOR
);
201 PIPE_ASSERT(PIPE_BLENDFACTOR_SRC1_ALPHA
== BLENDFACTOR_SRC1_ALPHA
);
202 PIPE_ASSERT(PIPE_BLENDFACTOR_ZERO
== BLENDFACTOR_ZERO
);
203 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC_COLOR
== BLENDFACTOR_INV_SRC_COLOR
);
204 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC_ALPHA
== BLENDFACTOR_INV_SRC_ALPHA
);
205 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_DST_ALPHA
== BLENDFACTOR_INV_DST_ALPHA
);
206 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_DST_COLOR
== BLENDFACTOR_INV_DST_COLOR
);
207 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_CONST_COLOR
== BLENDFACTOR_INV_CONST_COLOR
);
208 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_CONST_ALPHA
== BLENDFACTOR_INV_CONST_ALPHA
);
209 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC1_COLOR
== BLENDFACTOR_INV_SRC1_COLOR
);
210 PIPE_ASSERT(PIPE_BLENDFACTOR_INV_SRC1_ALPHA
== BLENDFACTOR_INV_SRC1_ALPHA
);
212 /* pipe_blend_func happens to match the hardware. */
213 PIPE_ASSERT(PIPE_BLEND_ADD
== BLENDFUNCTION_ADD
);
214 PIPE_ASSERT(PIPE_BLEND_SUBTRACT
== BLENDFUNCTION_SUBTRACT
);
215 PIPE_ASSERT(PIPE_BLEND_REVERSE_SUBTRACT
== BLENDFUNCTION_REVERSE_SUBTRACT
);
216 PIPE_ASSERT(PIPE_BLEND_MIN
== BLENDFUNCTION_MIN
);
217 PIPE_ASSERT(PIPE_BLEND_MAX
== BLENDFUNCTION_MAX
);
219 /* pipe_stencil_op happens to match the hardware. */
220 PIPE_ASSERT(PIPE_STENCIL_OP_KEEP
== STENCILOP_KEEP
);
221 PIPE_ASSERT(PIPE_STENCIL_OP_ZERO
== STENCILOP_ZERO
);
222 PIPE_ASSERT(PIPE_STENCIL_OP_REPLACE
== STENCILOP_REPLACE
);
223 PIPE_ASSERT(PIPE_STENCIL_OP_INCR
== STENCILOP_INCRSAT
);
224 PIPE_ASSERT(PIPE_STENCIL_OP_DECR
== STENCILOP_DECRSAT
);
225 PIPE_ASSERT(PIPE_STENCIL_OP_INCR_WRAP
== STENCILOP_INCR
);
226 PIPE_ASSERT(PIPE_STENCIL_OP_DECR_WRAP
== STENCILOP_DECR
);
227 PIPE_ASSERT(PIPE_STENCIL_OP_INVERT
== STENCILOP_INVERT
);
229 /* pipe_sprite_coord_mode happens to match 3DSTATE_SBE */
230 PIPE_ASSERT(PIPE_SPRITE_COORD_UPPER_LEFT
== UPPERLEFT
);
231 PIPE_ASSERT(PIPE_SPRITE_COORD_LOWER_LEFT
== LOWERLEFT
);
236 translate_prim_type(enum pipe_prim_type prim
, uint8_t verts_per_patch
)
238 static const unsigned map
[] = {
239 [PIPE_PRIM_POINTS
] = _3DPRIM_POINTLIST
,
240 [PIPE_PRIM_LINES
] = _3DPRIM_LINELIST
,
241 [PIPE_PRIM_LINE_LOOP
] = _3DPRIM_LINELOOP
,
242 [PIPE_PRIM_LINE_STRIP
] = _3DPRIM_LINESTRIP
,
243 [PIPE_PRIM_TRIANGLES
] = _3DPRIM_TRILIST
,
244 [PIPE_PRIM_TRIANGLE_STRIP
] = _3DPRIM_TRISTRIP
,
245 [PIPE_PRIM_TRIANGLE_FAN
] = _3DPRIM_TRIFAN
,
246 [PIPE_PRIM_QUADS
] = _3DPRIM_QUADLIST
,
247 [PIPE_PRIM_QUAD_STRIP
] = _3DPRIM_QUADSTRIP
,
248 [PIPE_PRIM_POLYGON
] = _3DPRIM_POLYGON
,
249 [PIPE_PRIM_LINES_ADJACENCY
] = _3DPRIM_LINELIST_ADJ
,
250 [PIPE_PRIM_LINE_STRIP_ADJACENCY
] = _3DPRIM_LINESTRIP_ADJ
,
251 [PIPE_PRIM_TRIANGLES_ADJACENCY
] = _3DPRIM_TRILIST_ADJ
,
252 [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY
] = _3DPRIM_TRISTRIP_ADJ
,
253 [PIPE_PRIM_PATCHES
] = _3DPRIM_PATCHLIST_1
- 1,
256 return map
[prim
] + (prim
== PIPE_PRIM_PATCHES
? verts_per_patch
: 0);
260 translate_compare_func(enum pipe_compare_func pipe_func
)
262 static const unsigned map
[] = {
263 [PIPE_FUNC_NEVER
] = COMPAREFUNCTION_NEVER
,
264 [PIPE_FUNC_LESS
] = COMPAREFUNCTION_LESS
,
265 [PIPE_FUNC_EQUAL
] = COMPAREFUNCTION_EQUAL
,
266 [PIPE_FUNC_LEQUAL
] = COMPAREFUNCTION_LEQUAL
,
267 [PIPE_FUNC_GREATER
] = COMPAREFUNCTION_GREATER
,
268 [PIPE_FUNC_NOTEQUAL
] = COMPAREFUNCTION_NOTEQUAL
,
269 [PIPE_FUNC_GEQUAL
] = COMPAREFUNCTION_GEQUAL
,
270 [PIPE_FUNC_ALWAYS
] = COMPAREFUNCTION_ALWAYS
,
272 return map
[pipe_func
];
276 translate_shadow_func(enum pipe_compare_func pipe_func
)
278 /* Gallium specifies the result of shadow comparisons as:
280 * 1 if ref <op> texel,
285 * 0 if texel <op> ref,
288 * So we need to flip the operator and also negate.
290 static const unsigned map
[] = {
291 [PIPE_FUNC_NEVER
] = PREFILTEROPALWAYS
,
292 [PIPE_FUNC_LESS
] = PREFILTEROPLEQUAL
,
293 [PIPE_FUNC_EQUAL
] = PREFILTEROPNOTEQUAL
,
294 [PIPE_FUNC_LEQUAL
] = PREFILTEROPLESS
,
295 [PIPE_FUNC_GREATER
] = PREFILTEROPGEQUAL
,
296 [PIPE_FUNC_NOTEQUAL
] = PREFILTEROPEQUAL
,
297 [PIPE_FUNC_GEQUAL
] = PREFILTEROPGREATER
,
298 [PIPE_FUNC_ALWAYS
] = PREFILTEROPNEVER
,
300 return map
[pipe_func
];
304 translate_cull_mode(unsigned pipe_face
)
306 static const unsigned map
[4] = {
307 [PIPE_FACE_NONE
] = CULLMODE_NONE
,
308 [PIPE_FACE_FRONT
] = CULLMODE_FRONT
,
309 [PIPE_FACE_BACK
] = CULLMODE_BACK
,
310 [PIPE_FACE_FRONT_AND_BACK
] = CULLMODE_BOTH
,
312 return map
[pipe_face
];
316 translate_fill_mode(unsigned pipe_polymode
)
318 static const unsigned map
[4] = {
319 [PIPE_POLYGON_MODE_FILL
] = FILL_MODE_SOLID
,
320 [PIPE_POLYGON_MODE_LINE
] = FILL_MODE_WIREFRAME
,
321 [PIPE_POLYGON_MODE_POINT
] = FILL_MODE_POINT
,
322 [PIPE_POLYGON_MODE_FILL_RECTANGLE
] = FILL_MODE_SOLID
,
324 return map
[pipe_polymode
];
328 translate_mip_filter(enum pipe_tex_mipfilter pipe_mip
)
330 static const unsigned map
[] = {
331 [PIPE_TEX_MIPFILTER_NEAREST
] = MIPFILTER_NEAREST
,
332 [PIPE_TEX_MIPFILTER_LINEAR
] = MIPFILTER_LINEAR
,
333 [PIPE_TEX_MIPFILTER_NONE
] = MIPFILTER_NONE
,
335 return map
[pipe_mip
];
339 translate_wrap(unsigned pipe_wrap
)
341 static const unsigned map
[] = {
342 [PIPE_TEX_WRAP_REPEAT
] = TCM_WRAP
,
343 [PIPE_TEX_WRAP_CLAMP
] = TCM_HALF_BORDER
,
344 [PIPE_TEX_WRAP_CLAMP_TO_EDGE
] = TCM_CLAMP
,
345 [PIPE_TEX_WRAP_CLAMP_TO_BORDER
] = TCM_CLAMP_BORDER
,
346 [PIPE_TEX_WRAP_MIRROR_REPEAT
] = TCM_MIRROR
,
347 [PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
] = TCM_MIRROR_ONCE
,
349 /* These are unsupported. */
350 [PIPE_TEX_WRAP_MIRROR_CLAMP
] = -1,
351 [PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
] = -1,
353 return map
[pipe_wrap
];
356 static struct iris_address
357 ro_bo(struct iris_bo
*bo
, uint64_t offset
)
359 /* CSOs must pass NULL for bo! Otherwise it will add the BO to the
360 * validation list at CSO creation time, instead of draw time.
362 return (struct iris_address
) { .bo
= bo
, .offset
= offset
};
365 static struct iris_address
366 rw_bo(struct iris_bo
*bo
, uint64_t offset
)
368 /* CSOs must pass NULL for bo! Otherwise it will add the BO to the
369 * validation list at CSO creation time, instead of draw time.
371 return (struct iris_address
) { .bo
= bo
, .offset
= offset
, .write
= true };
375 * Allocate space for some indirect state.
377 * Return a pointer to the map (to fill it out) and a state ref (for
378 * referring to the state in GPU commands).
381 upload_state(struct u_upload_mgr
*uploader
,
382 struct iris_state_ref
*ref
,
387 u_upload_alloc(uploader
, 0, size
, alignment
, &ref
->offset
, &ref
->res
, &p
);
392 * Stream out temporary/short-lived state.
394 * This allocates space, pins the BO, and includes the BO address in the
395 * returned offset (which works because all state lives in 32-bit memory
399 stream_state(struct iris_batch
*batch
,
400 struct u_upload_mgr
*uploader
,
401 struct pipe_resource
**out_res
,
404 uint32_t *out_offset
)
408 u_upload_alloc(uploader
, 0, size
, alignment
, out_offset
, out_res
, &ptr
);
410 struct iris_bo
*bo
= iris_resource_bo(*out_res
);
411 iris_use_pinned_bo(batch
, bo
, false);
413 *out_offset
+= iris_bo_offset_from_base_address(bo
);
419 * stream_state() + memcpy.
422 emit_state(struct iris_batch
*batch
,
423 struct u_upload_mgr
*uploader
,
424 struct pipe_resource
**out_res
,
431 stream_state(batch
, uploader
, out_res
, size
, alignment
, &offset
);
434 memcpy(map
, data
, size
);
440 * Did field 'x' change between 'old_cso' and 'new_cso'?
442 * (If so, we may want to set some dirty flags.)
444 #define cso_changed(x) (!old_cso || (old_cso->x != new_cso->x))
445 #define cso_changed_memcmp(x) \
446 (!old_cso || memcmp(old_cso->x, new_cso->x, sizeof(old_cso->x)) != 0)
449 flush_for_state_base_change(struct iris_batch
*batch
)
451 /* Flush before emitting STATE_BASE_ADDRESS.
453 * This isn't documented anywhere in the PRM. However, it seems to be
454 * necessary prior to changing the surface state base adress. We've
455 * seen issues in Vulkan where we get GPU hangs when using multi-level
456 * command buffers which clear depth, reset state base address, and then
459 * Normally, in GL, we would trust the kernel to do sufficient stalls
460 * and flushes prior to executing our batch. However, it doesn't seem
461 * as if the kernel's flushing is always sufficient and we don't want to
464 * We make this an end-of-pipe sync instead of a normal flush because we
465 * do not know the current status of the GPU. On Haswell at least,
466 * having a fast-clear operation in flight at the same time as a normal
467 * rendering operation can cause hangs. Since the kernel's flushing is
468 * insufficient, we need to ensure that any rendering operations from
469 * other processes are definitely complete before we try to do our own
470 * rendering. It's a bit of a big hammer but it appears to work.
472 iris_emit_end_of_pipe_sync(batch
,
473 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
474 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
475 PIPE_CONTROL_DATA_CACHE_FLUSH
);
479 _iris_emit_lri(struct iris_batch
*batch
, uint32_t reg
, uint32_t val
)
481 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_IMM
), lri
) {
482 lri
.RegisterOffset
= reg
;
486 #define iris_emit_lri(b, r, v) _iris_emit_lri(b, GENX(r##_num), v)
489 emit_pipeline_select(struct iris_batch
*batch
, uint32_t pipeline
)
491 #if GEN_GEN >= 8 && GEN_GEN < 10
492 /* From the Broadwell PRM, Volume 2a: Instructions, PIPELINE_SELECT:
494 * Software must clear the COLOR_CALC_STATE Valid field in
495 * 3DSTATE_CC_STATE_POINTERS command prior to send a PIPELINE_SELECT
496 * with Pipeline Select set to GPGPU.
498 * The internal hardware docs recommend the same workaround for Gen9
501 if (pipeline
== GPGPU
)
502 iris_emit_cmd(batch
, GENX(3DSTATE_CC_STATE_POINTERS
), t
);
506 /* From "BXML » GT » MI » vol1a GPU Overview » [Instruction]
507 * PIPELINE_SELECT [DevBWR+]":
511 * Software must ensure all the write caches are flushed through a
512 * stalling PIPE_CONTROL command followed by another PIPE_CONTROL
513 * command to invalidate read only caches prior to programming
514 * MI_PIPELINE_SELECT command to change the Pipeline Select Mode."
516 iris_emit_pipe_control_flush(batch
,
517 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
518 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
519 PIPE_CONTROL_DATA_CACHE_FLUSH
|
520 PIPE_CONTROL_CS_STALL
);
522 iris_emit_pipe_control_flush(batch
,
523 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
524 PIPE_CONTROL_CONST_CACHE_INVALIDATE
|
525 PIPE_CONTROL_STATE_CACHE_INVALIDATE
|
526 PIPE_CONTROL_INSTRUCTION_INVALIDATE
);
528 iris_emit_cmd(batch
, GENX(PIPELINE_SELECT
), sel
) {
532 sel
.PipelineSelection
= pipeline
;
537 init_glk_barrier_mode(struct iris_batch
*batch
, uint32_t value
)
542 * "This chicken bit works around a hardware issue with barrier
543 * logic encountered when switching between GPGPU and 3D pipelines.
544 * To workaround the issue, this mode bit should be set after a
545 * pipeline is selected."
548 iris_pack_state(GENX(SLICE_COMMON_ECO_CHICKEN1
), ®_val
, reg
) {
549 reg
.GLKBarrierMode
= value
;
550 reg
.GLKBarrierModeMask
= 1;
552 iris_emit_lri(batch
, SLICE_COMMON_ECO_CHICKEN1
, reg_val
);
557 init_state_base_address(struct iris_batch
*batch
)
559 flush_for_state_base_change(batch
);
561 /* We program most base addresses once at context initialization time.
562 * Each base address points at a 4GB memory zone, and never needs to
563 * change. See iris_bufmgr.h for a description of the memory zones.
565 * The one exception is Surface State Base Address, which needs to be
566 * updated occasionally. See iris_binder.c for the details there.
568 iris_emit_cmd(batch
, GENX(STATE_BASE_ADDRESS
), sba
) {
570 // XXX: MOCS is stupid for this.
571 sba
.GeneralStateMemoryObjectControlState
= MOCS_WB
;
572 sba
.StatelessDataPortAccessMemoryObjectControlState
= MOCS_WB
;
573 sba
.DynamicStateMemoryObjectControlState
= MOCS_WB
;
574 sba
.IndirectObjectMemoryObjectControlState
= MOCS_WB
;
575 sba
.InstructionMemoryObjectControlState
= MOCS_WB
;
576 sba
.BindlessSurfaceStateMemoryObjectControlState
= MOCS_WB
;
579 sba
.GeneralStateBaseAddressModifyEnable
= true;
580 sba
.DynamicStateBaseAddressModifyEnable
= true;
581 sba
.IndirectObjectBaseAddressModifyEnable
= true;
582 sba
.InstructionBaseAddressModifyEnable
= true;
583 sba
.GeneralStateBufferSizeModifyEnable
= true;
584 sba
.DynamicStateBufferSizeModifyEnable
= true;
585 sba
.BindlessSurfaceStateBaseAddressModifyEnable
= true;
586 sba
.IndirectObjectBufferSizeModifyEnable
= true;
587 sba
.InstructionBuffersizeModifyEnable
= true;
589 sba
.InstructionBaseAddress
= ro_bo(NULL
, IRIS_MEMZONE_SHADER_START
);
590 sba
.DynamicStateBaseAddress
= ro_bo(NULL
, IRIS_MEMZONE_DYNAMIC_START
);
592 sba
.GeneralStateBufferSize
= 0xfffff;
593 sba
.IndirectObjectBufferSize
= 0xfffff;
594 sba
.InstructionBufferSize
= 0xfffff;
595 sba
.DynamicStateBufferSize
= 0xfffff;
600 * Upload the initial GPU state for a render context.
602 * This sets some invariant state that needs to be programmed a particular
603 * way, but we never actually change.
606 iris_init_render_context(struct iris_screen
*screen
,
607 struct iris_batch
*batch
,
608 struct iris_vtable
*vtbl
,
609 struct pipe_debug_callback
*dbg
)
611 UNUSED
const struct gen_device_info
*devinfo
= &screen
->devinfo
;
614 emit_pipeline_select(batch
, _3D
);
616 init_state_base_address(batch
);
618 // XXX: INSTPM on Gen8
619 iris_pack_state(GENX(CS_DEBUG_MODE2
), ®_val
, reg
) {
620 reg
.CONSTANT_BUFFERAddressOffsetDisable
= true;
621 reg
.CONSTANT_BUFFERAddressOffsetDisableMask
= true;
623 iris_emit_lri(batch
, CS_DEBUG_MODE2
, reg_val
);
626 iris_pack_state(GENX(CACHE_MODE_1
), ®_val
, reg
) {
627 reg
.FloatBlendOptimizationEnable
= true;
628 reg
.FloatBlendOptimizationEnableMask
= true;
629 reg
.PartialResolveDisableInVC
= true;
630 reg
.PartialResolveDisableInVCMask
= true;
632 iris_emit_lri(batch
, CACHE_MODE_1
, reg_val
);
634 if (devinfo
->is_geminilake
)
635 init_glk_barrier_mode(batch
, GLK_BARRIER_MODE_3D_HULL
);
639 iris_pack_state(GENX(SAMPLER_MODE
), ®_val
, reg
) {
640 reg
.HeaderlessMessageforPreemptableContexts
= 1;
641 reg
.HeaderlessMessageforPreemptableContextsMask
= 1;
643 iris_emit_lri(batch
, SAMPLER_MODE
, reg_val
);
648 /* 3DSTATE_DRAWING_RECTANGLE is non-pipelined, so we want to avoid
649 * changing it dynamically. We set it to the maximum size here, and
650 * instead include the render target dimensions in the viewport, so
651 * viewport extents clipping takes care of pruning stray geometry.
653 iris_emit_cmd(batch
, GENX(3DSTATE_DRAWING_RECTANGLE
), rect
) {
654 rect
.ClippedDrawingRectangleXMax
= UINT16_MAX
;
655 rect
.ClippedDrawingRectangleYMax
= UINT16_MAX
;
658 /* Set the initial MSAA sample positions. */
659 iris_emit_cmd(batch
, GENX(3DSTATE_SAMPLE_PATTERN
), pat
) {
660 GEN_SAMPLE_POS_1X(pat
._1xSample
);
661 GEN_SAMPLE_POS_2X(pat
._2xSample
);
662 GEN_SAMPLE_POS_4X(pat
._4xSample
);
663 GEN_SAMPLE_POS_8X(pat
._8xSample
);
664 GEN_SAMPLE_POS_16X(pat
._16xSample
);
667 /* Use the legacy AA line coverage computation. */
668 iris_emit_cmd(batch
, GENX(3DSTATE_AA_LINE_PARAMETERS
), foo
);
670 /* Disable chromakeying (it's for media) */
671 iris_emit_cmd(batch
, GENX(3DSTATE_WM_CHROMAKEY
), foo
);
673 /* We want regular rendering, not special HiZ operations. */
674 iris_emit_cmd(batch
, GENX(3DSTATE_WM_HZ_OP
), foo
);
676 /* No polygon stippling offsets are necessary. */
677 // XXX: may need to set an offset for origin-UL framebuffers
678 iris_emit_cmd(batch
, GENX(3DSTATE_POLY_STIPPLE_OFFSET
), foo
);
680 /* Set a static partitioning of the push constant area. */
681 // XXX: this may be a bad idea...could starve the push ringbuffers...
682 for (int i
= 0; i
<= MESA_SHADER_FRAGMENT
; i
++) {
683 iris_emit_cmd(batch
, GENX(3DSTATE_PUSH_CONSTANT_ALLOC_VS
), alloc
) {
684 alloc
._3DCommandSubOpcode
= 18 + i
;
685 alloc
.ConstantBufferOffset
= 6 * i
;
686 alloc
.ConstantBufferSize
= i
== MESA_SHADER_FRAGMENT
? 8 : 6;
692 iris_init_compute_context(struct iris_screen
*screen
,
693 struct iris_batch
*batch
,
694 struct iris_vtable
*vtbl
,
695 struct pipe_debug_callback
*dbg
)
697 UNUSED
const struct gen_device_info
*devinfo
= &screen
->devinfo
;
699 emit_pipeline_select(batch
, GPGPU
);
701 const bool has_slm
= true;
702 const bool wants_dc_cache
= true;
704 const struct gen_l3_weights w
=
705 gen_get_default_l3_weights(devinfo
, wants_dc_cache
, has_slm
);
706 const struct gen_l3_config
*cfg
= gen_get_l3_config(devinfo
, w
);
709 iris_pack_state(GENX(L3CNTLREG
), ®_val
, reg
) {
710 reg
.SLMEnable
= has_slm
;
712 /* WA_1406697149: Bit 9 "Error Detection Behavior Control" must be set
713 * in L3CNTLREG register. The default setting of the bit is not the
714 * desirable behavior.
716 reg
.ErrorDetectionBehaviorControl
= true;
718 reg
.URBAllocation
= cfg
->n
[GEN_L3P_URB
];
719 reg
.ROAllocation
= cfg
->n
[GEN_L3P_RO
];
720 reg
.DCAllocation
= cfg
->n
[GEN_L3P_DC
];
721 reg
.AllAllocation
= cfg
->n
[GEN_L3P_ALL
];
723 iris_emit_lri(batch
, L3CNTLREG
, reg_val
);
725 init_state_base_address(batch
);
728 if (devinfo
->is_geminilake
)
729 init_glk_barrier_mode(batch
, GLK_BARRIER_MODE_GPGPU
);
733 struct iris_vertex_buffer_state
{
734 /** The 3DSTATE_VERTEX_BUFFERS hardware packet. */
735 uint32_t vertex_buffers
[1 + 33 * GENX(VERTEX_BUFFER_STATE_length
)];
737 /** The resource to source vertex data from. */
738 struct pipe_resource
*resources
[33];
740 /** The number of bound vertex buffers. */
741 unsigned num_buffers
;
744 struct iris_depth_buffer_state
{
745 /* Depth/HiZ/Stencil related hardware packets. */
746 uint32_t packets
[GENX(3DSTATE_DEPTH_BUFFER_length
) +
747 GENX(3DSTATE_STENCIL_BUFFER_length
) +
748 GENX(3DSTATE_HIER_DEPTH_BUFFER_length
) +
749 GENX(3DSTATE_CLEAR_PARAMS_length
)];
753 * Generation-specific context state (ice->state.genx->...).
755 * Most state can go in iris_context directly, but these encode hardware
756 * packets which vary by generation.
758 struct iris_genx_state
{
759 /** SF_CLIP_VIEWPORT */
760 uint32_t sf_cl_vp
[GENX(SF_CLIP_VIEWPORT_length
) * IRIS_MAX_VIEWPORTS
];
762 struct iris_vertex_buffer_state vertex_buffers
;
763 struct iris_depth_buffer_state depth_buffer
;
765 uint32_t so_buffers
[4 * GENX(3DSTATE_SO_BUFFER_length
)];
766 uint32_t streamout
[4 * GENX(3DSTATE_STREAMOUT_length
)];
770 * The pipe->set_blend_color() driver hook.
772 * This corresponds to our COLOR_CALC_STATE.
775 iris_set_blend_color(struct pipe_context
*ctx
,
776 const struct pipe_blend_color
*state
)
778 struct iris_context
*ice
= (struct iris_context
*) ctx
;
780 /* Our COLOR_CALC_STATE is exactly pipe_blend_color, so just memcpy */
781 memcpy(&ice
->state
.blend_color
, state
, sizeof(struct pipe_blend_color
));
782 ice
->state
.dirty
|= IRIS_DIRTY_COLOR_CALC_STATE
;
786 * Gallium CSO for blend state (see pipe_blend_state).
788 struct iris_blend_state
{
789 /** Partial 3DSTATE_PS_BLEND */
790 uint32_t ps_blend
[GENX(3DSTATE_PS_BLEND_length
)];
792 /** Partial BLEND_STATE */
793 uint32_t blend_state
[GENX(BLEND_STATE_length
) +
794 BRW_MAX_DRAW_BUFFERS
* GENX(BLEND_STATE_ENTRY_length
)];
796 bool alpha_to_coverage
; /* for shader key */
800 * The pipe->create_blend_state() driver hook.
802 * Translates a pipe_blend_state into iris_blend_state.
805 iris_create_blend_state(struct pipe_context
*ctx
,
806 const struct pipe_blend_state
*state
)
808 struct iris_blend_state
*cso
= malloc(sizeof(struct iris_blend_state
));
809 uint32_t *blend_entry
= cso
->blend_state
+ GENX(BLEND_STATE_length
);
811 cso
->alpha_to_coverage
= state
->alpha_to_coverage
;
813 bool indep_alpha_blend
= false;
815 for (int i
= 0; i
< BRW_MAX_DRAW_BUFFERS
; i
++) {
816 const struct pipe_rt_blend_state
*rt
=
817 &state
->rt
[state
->independent_blend_enable
? i
: 0];
819 if (rt
->rgb_func
!= rt
->alpha_func
||
820 rt
->rgb_src_factor
!= rt
->alpha_src_factor
||
821 rt
->rgb_dst_factor
!= rt
->alpha_dst_factor
)
822 indep_alpha_blend
= true;
824 iris_pack_state(GENX(BLEND_STATE_ENTRY
), blend_entry
, be
) {
825 be
.LogicOpEnable
= state
->logicop_enable
;
826 be
.LogicOpFunction
= state
->logicop_func
;
828 be
.PreBlendSourceOnlyClampEnable
= false;
829 be
.ColorClampRange
= COLORCLAMP_RTFORMAT
;
830 be
.PreBlendColorClampEnable
= true;
831 be
.PostBlendColorClampEnable
= true;
833 be
.ColorBufferBlendEnable
= rt
->blend_enable
;
835 be
.ColorBlendFunction
= rt
->rgb_func
;
836 be
.AlphaBlendFunction
= rt
->alpha_func
;
837 be
.SourceBlendFactor
= rt
->rgb_src_factor
;
838 be
.SourceAlphaBlendFactor
= rt
->alpha_src_factor
;
839 be
.DestinationBlendFactor
= rt
->rgb_dst_factor
;
840 be
.DestinationAlphaBlendFactor
= rt
->alpha_dst_factor
;
842 be
.WriteDisableRed
= !(rt
->colormask
& PIPE_MASK_R
);
843 be
.WriteDisableGreen
= !(rt
->colormask
& PIPE_MASK_G
);
844 be
.WriteDisableBlue
= !(rt
->colormask
& PIPE_MASK_B
);
845 be
.WriteDisableAlpha
= !(rt
->colormask
& PIPE_MASK_A
);
847 blend_entry
+= GENX(BLEND_STATE_ENTRY_length
);
850 iris_pack_command(GENX(3DSTATE_PS_BLEND
), cso
->ps_blend
, pb
) {
851 /* pb.HasWriteableRT is filled in at draw time. */
852 /* pb.AlphaTestEnable is filled in at draw time. */
853 pb
.AlphaToCoverageEnable
= state
->alpha_to_coverage
;
854 pb
.IndependentAlphaBlendEnable
= indep_alpha_blend
;
856 pb
.ColorBufferBlendEnable
= state
->rt
[0].blend_enable
;
858 pb
.SourceBlendFactor
= state
->rt
[0].rgb_src_factor
;
859 pb
.SourceAlphaBlendFactor
= state
->rt
[0].alpha_src_factor
;
860 pb
.DestinationBlendFactor
= state
->rt
[0].rgb_dst_factor
;
861 pb
.DestinationAlphaBlendFactor
= state
->rt
[0].alpha_dst_factor
;
864 iris_pack_state(GENX(BLEND_STATE
), cso
->blend_state
, bs
) {
865 bs
.AlphaToCoverageEnable
= state
->alpha_to_coverage
;
866 bs
.IndependentAlphaBlendEnable
= indep_alpha_blend
;
867 bs
.AlphaToOneEnable
= state
->alpha_to_one
;
868 bs
.AlphaToCoverageDitherEnable
= state
->alpha_to_coverage
;
869 bs
.ColorDitherEnable
= state
->dither
;
870 /* bl.AlphaTestEnable and bs.AlphaTestFunction are filled in later. */
878 * The pipe->bind_blend_state() driver hook.
880 * Bind a blending CSO and flag related dirty bits.
883 iris_bind_blend_state(struct pipe_context
*ctx
, void *state
)
885 struct iris_context
*ice
= (struct iris_context
*) ctx
;
886 ice
->state
.cso_blend
= state
;
887 ice
->state
.dirty
|= IRIS_DIRTY_PS_BLEND
;
888 ice
->state
.dirty
|= IRIS_DIRTY_BLEND_STATE
;
889 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_BLEND
];
893 * Gallium CSO for depth, stencil, and alpha testing state.
895 struct iris_depth_stencil_alpha_state
{
896 /** Partial 3DSTATE_WM_DEPTH_STENCIL. */
897 uint32_t wmds
[GENX(3DSTATE_WM_DEPTH_STENCIL_length
)];
899 /** Outbound to BLEND_STATE, 3DSTATE_PS_BLEND, COLOR_CALC_STATE. */
900 struct pipe_alpha_state alpha
;
902 /** Outbound to resolve and cache set tracking. */
903 bool depth_writes_enabled
;
904 bool stencil_writes_enabled
;
908 * The pipe->create_depth_stencil_alpha_state() driver hook.
910 * We encode most of 3DSTATE_WM_DEPTH_STENCIL, and just save off the alpha
911 * testing state since we need pieces of it in a variety of places.
914 iris_create_zsa_state(struct pipe_context
*ctx
,
915 const struct pipe_depth_stencil_alpha_state
*state
)
917 struct iris_depth_stencil_alpha_state
*cso
=
918 malloc(sizeof(struct iris_depth_stencil_alpha_state
));
920 bool two_sided_stencil
= state
->stencil
[1].enabled
;
922 cso
->alpha
= state
->alpha
;
923 cso
->depth_writes_enabled
= state
->depth
.writemask
;
924 cso
->stencil_writes_enabled
=
925 state
->stencil
[0].writemask
!= 0 ||
926 (two_sided_stencil
&& state
->stencil
[1].writemask
!= 1);
928 /* The state tracker needs to optimize away EQUAL writes for us. */
929 assert(!(state
->depth
.func
== PIPE_FUNC_EQUAL
&& state
->depth
.writemask
));
931 iris_pack_command(GENX(3DSTATE_WM_DEPTH_STENCIL
), cso
->wmds
, wmds
) {
932 wmds
.StencilFailOp
= state
->stencil
[0].fail_op
;
933 wmds
.StencilPassDepthFailOp
= state
->stencil
[0].zfail_op
;
934 wmds
.StencilPassDepthPassOp
= state
->stencil
[0].zpass_op
;
935 wmds
.StencilTestFunction
=
936 translate_compare_func(state
->stencil
[0].func
);
937 wmds
.BackfaceStencilFailOp
= state
->stencil
[1].fail_op
;
938 wmds
.BackfaceStencilPassDepthFailOp
= state
->stencil
[1].zfail_op
;
939 wmds
.BackfaceStencilPassDepthPassOp
= state
->stencil
[1].zpass_op
;
940 wmds
.BackfaceStencilTestFunction
=
941 translate_compare_func(state
->stencil
[1].func
);
942 wmds
.DepthTestFunction
= translate_compare_func(state
->depth
.func
);
943 wmds
.DoubleSidedStencilEnable
= two_sided_stencil
;
944 wmds
.StencilTestEnable
= state
->stencil
[0].enabled
;
945 wmds
.StencilBufferWriteEnable
=
946 state
->stencil
[0].writemask
!= 0 ||
947 (two_sided_stencil
&& state
->stencil
[1].writemask
!= 0);
948 wmds
.DepthTestEnable
= state
->depth
.enabled
;
949 wmds
.DepthBufferWriteEnable
= state
->depth
.writemask
;
950 wmds
.StencilTestMask
= state
->stencil
[0].valuemask
;
951 wmds
.StencilWriteMask
= state
->stencil
[0].writemask
;
952 wmds
.BackfaceStencilTestMask
= state
->stencil
[1].valuemask
;
953 wmds
.BackfaceStencilWriteMask
= state
->stencil
[1].writemask
;
954 /* wmds.[Backface]StencilReferenceValue are merged later */
961 * The pipe->bind_depth_stencil_alpha_state() driver hook.
963 * Bind a depth/stencil/alpha CSO and flag related dirty bits.
966 iris_bind_zsa_state(struct pipe_context
*ctx
, void *state
)
968 struct iris_context
*ice
= (struct iris_context
*) ctx
;
969 struct iris_depth_stencil_alpha_state
*old_cso
= ice
->state
.cso_zsa
;
970 struct iris_depth_stencil_alpha_state
*new_cso
= state
;
973 if (cso_changed(alpha
.ref_value
))
974 ice
->state
.dirty
|= IRIS_DIRTY_COLOR_CALC_STATE
;
976 if (cso_changed(alpha
.enabled
))
977 ice
->state
.dirty
|= IRIS_DIRTY_PS_BLEND
| IRIS_DIRTY_BLEND_STATE
;
979 if (cso_changed(alpha
.func
))
980 ice
->state
.dirty
|= IRIS_DIRTY_BLEND_STATE
;
982 ice
->state
.depth_writes_enabled
= new_cso
->depth_writes_enabled
;
983 ice
->state
.stencil_writes_enabled
= new_cso
->stencil_writes_enabled
;
986 ice
->state
.cso_zsa
= new_cso
;
987 ice
->state
.dirty
|= IRIS_DIRTY_CC_VIEWPORT
;
988 ice
->state
.dirty
|= IRIS_DIRTY_WM_DEPTH_STENCIL
;
989 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_DEPTH_STENCIL_ALPHA
];
993 * Gallium CSO for rasterizer state.
995 struct iris_rasterizer_state
{
996 uint32_t sf
[GENX(3DSTATE_SF_length
)];
997 uint32_t clip
[GENX(3DSTATE_CLIP_length
)];
998 uint32_t raster
[GENX(3DSTATE_RASTER_length
)];
999 uint32_t wm
[GENX(3DSTATE_WM_length
)];
1000 uint32_t line_stipple
[GENX(3DSTATE_LINE_STIPPLE_length
)];
1002 uint8_t num_clip_plane_consts
;
1003 bool clip_halfz
; /* for CC_VIEWPORT */
1004 bool depth_clip_near
; /* for CC_VIEWPORT */
1005 bool depth_clip_far
; /* for CC_VIEWPORT */
1006 bool flatshade
; /* for shader state */
1007 bool flatshade_first
; /* for stream output */
1008 bool clamp_fragment_color
; /* for shader state */
1009 bool light_twoside
; /* for shader state */
1010 bool rasterizer_discard
; /* for 3DSTATE_STREAMOUT */
1011 bool half_pixel_center
; /* for 3DSTATE_MULTISAMPLE */
1012 bool line_stipple_enable
;
1013 bool poly_stipple_enable
;
1015 bool force_persample_interp
;
1016 enum pipe_sprite_coord_mode sprite_coord_mode
; /* PIPE_SPRITE_* */
1017 uint16_t sprite_coord_enable
;
1021 get_line_width(const struct pipe_rasterizer_state
*state
)
1023 float line_width
= state
->line_width
;
1025 /* From the OpenGL 4.4 spec:
1027 * "The actual width of non-antialiased lines is determined by rounding
1028 * the supplied width to the nearest integer, then clamping it to the
1029 * implementation-dependent maximum non-antialiased line width."
1031 if (!state
->multisample
&& !state
->line_smooth
)
1032 line_width
= roundf(state
->line_width
);
1034 if (!state
->multisample
&& state
->line_smooth
&& line_width
< 1.5f
) {
1035 /* For 1 pixel line thickness or less, the general anti-aliasing
1036 * algorithm gives up, and a garbage line is generated. Setting a
1037 * Line Width of 0.0 specifies the rasterization of the "thinnest"
1038 * (one-pixel-wide), non-antialiased lines.
1040 * Lines rendered with zero Line Width are rasterized using the
1041 * "Grid Intersection Quantization" rules as specified by the
1042 * "Zero-Width (Cosmetic) Line Rasterization" section of the docs.
1051 * The pipe->create_rasterizer_state() driver hook.
1054 iris_create_rasterizer_state(struct pipe_context
*ctx
,
1055 const struct pipe_rasterizer_state
*state
)
1057 struct iris_rasterizer_state
*cso
=
1058 malloc(sizeof(struct iris_rasterizer_state
));
1061 point_quad_rasterization
-> SBE
?
1066 force_persample_interp
- ?
1069 offset_units_unscaled
- cap
not exposed
1073 // XXX: it may make more sense just to store the pipe_rasterizer_state,
1074 // we're copying a lot of booleans here. But we don't need all of them...
1076 cso
->multisample
= state
->multisample
;
1077 cso
->force_persample_interp
= state
->force_persample_interp
;
1078 cso
->clip_halfz
= state
->clip_halfz
;
1079 cso
->depth_clip_near
= state
->depth_clip_near
;
1080 cso
->depth_clip_far
= state
->depth_clip_far
;
1081 cso
->flatshade
= state
->flatshade
;
1082 cso
->flatshade_first
= state
->flatshade_first
;
1083 cso
->clamp_fragment_color
= state
->clamp_fragment_color
;
1084 cso
->light_twoside
= state
->light_twoside
;
1085 cso
->rasterizer_discard
= state
->rasterizer_discard
;
1086 cso
->half_pixel_center
= state
->half_pixel_center
;
1087 cso
->sprite_coord_mode
= state
->sprite_coord_mode
;
1088 cso
->sprite_coord_enable
= state
->sprite_coord_enable
;
1089 cso
->line_stipple_enable
= state
->line_stipple_enable
;
1090 cso
->poly_stipple_enable
= state
->poly_stipple_enable
;
1092 if (state
->clip_plane_enable
!= 0)
1093 cso
->num_clip_plane_consts
= util_logbase2(state
->clip_plane_enable
) + 1;
1095 cso
->num_clip_plane_consts
= 0;
1097 float line_width
= get_line_width(state
);
1099 iris_pack_command(GENX(3DSTATE_SF
), cso
->sf
, sf
) {
1100 sf
.StatisticsEnable
= true;
1101 sf
.ViewportTransformEnable
= true;
1102 sf
.AALineDistanceMode
= AALINEDISTANCE_TRUE
;
1103 sf
.LineEndCapAntialiasingRegionWidth
=
1104 state
->line_smooth
? _10pixels
: _05pixels
;
1105 sf
.LastPixelEnable
= state
->line_last_pixel
;
1106 sf
.LineWidth
= line_width
;
1107 sf
.SmoothPointEnable
= state
->point_smooth
|| state
->multisample
;
1108 sf
.PointWidthSource
= state
->point_size_per_vertex
? Vertex
: State
;
1109 sf
.PointWidth
= state
->point_size
;
1111 if (state
->flatshade_first
) {
1112 sf
.TriangleFanProvokingVertexSelect
= 1;
1114 sf
.TriangleStripListProvokingVertexSelect
= 2;
1115 sf
.TriangleFanProvokingVertexSelect
= 2;
1116 sf
.LineStripListProvokingVertexSelect
= 1;
1120 iris_pack_command(GENX(3DSTATE_RASTER
), cso
->raster
, rr
) {
1121 rr
.FrontWinding
= state
->front_ccw
? CounterClockwise
: Clockwise
;
1122 rr
.CullMode
= translate_cull_mode(state
->cull_face
);
1123 rr
.FrontFaceFillMode
= translate_fill_mode(state
->fill_front
);
1124 rr
.BackFaceFillMode
= translate_fill_mode(state
->fill_back
);
1125 rr
.DXMultisampleRasterizationEnable
= state
->multisample
;
1126 rr
.GlobalDepthOffsetEnableSolid
= state
->offset_tri
;
1127 rr
.GlobalDepthOffsetEnableWireframe
= state
->offset_line
;
1128 rr
.GlobalDepthOffsetEnablePoint
= state
->offset_point
;
1129 rr
.GlobalDepthOffsetConstant
= state
->offset_units
* 2;
1130 rr
.GlobalDepthOffsetScale
= state
->offset_scale
;
1131 rr
.GlobalDepthOffsetClamp
= state
->offset_clamp
;
1132 rr
.SmoothPointEnable
= state
->point_smooth
|| state
->multisample
;
1133 rr
.AntialiasingEnable
= state
->line_smooth
;
1134 rr
.ScissorRectangleEnable
= state
->scissor
;
1135 rr
.ViewportZNearClipTestEnable
= state
->depth_clip_near
;
1136 rr
.ViewportZFarClipTestEnable
= state
->depth_clip_far
;
1137 //rr.ConservativeRasterizationEnable = not yet supported by Gallium...
1140 iris_pack_command(GENX(3DSTATE_CLIP
), cso
->clip
, cl
) {
1141 /* cl.NonPerspectiveBarycentricEnable is filled in at draw time from
1142 * the FS program; cl.ForceZeroRTAIndexEnable is filled in from the FB.
1144 cl
.StatisticsEnable
= true;
1145 cl
.EarlyCullEnable
= true;
1146 cl
.UserClipDistanceClipTestEnableBitmask
= state
->clip_plane_enable
;
1147 cl
.ForceUserClipDistanceClipTestEnableBitmask
= true;
1148 cl
.APIMode
= state
->clip_halfz
? APIMODE_D3D
: APIMODE_OGL
;
1149 cl
.GuardbandClipTestEnable
= true;
1150 cl
.ClipMode
= CLIPMODE_NORMAL
;
1151 cl
.ClipEnable
= true;
1152 cl
.ViewportXYClipTestEnable
= state
->point_tri_clip
;
1153 cl
.MinimumPointWidth
= 0.125;
1154 cl
.MaximumPointWidth
= 255.875;
1156 if (state
->flatshade_first
) {
1157 cl
.TriangleFanProvokingVertexSelect
= 1;
1159 cl
.TriangleStripListProvokingVertexSelect
= 2;
1160 cl
.TriangleFanProvokingVertexSelect
= 2;
1161 cl
.LineStripListProvokingVertexSelect
= 1;
1165 iris_pack_command(GENX(3DSTATE_WM
), cso
->wm
, wm
) {
1166 /* wm.BarycentricInterpolationMode and wm.EarlyDepthStencilControl are
1167 * filled in at draw time from the FS program.
1169 wm
.LineAntialiasingRegionWidth
= _10pixels
;
1170 wm
.LineEndCapAntialiasingRegionWidth
= _05pixels
;
1171 wm
.PointRasterizationRule
= RASTRULE_UPPER_RIGHT
;
1172 wm
.LineStippleEnable
= state
->line_stipple_enable
;
1173 wm
.PolygonStippleEnable
= state
->poly_stipple_enable
;
1176 /* Remap from 0..255 back to 1..256 */
1177 const unsigned line_stipple_factor
= state
->line_stipple_factor
+ 1;
1179 iris_pack_command(GENX(3DSTATE_LINE_STIPPLE
), cso
->line_stipple
, line
) {
1180 line
.LineStipplePattern
= state
->line_stipple_pattern
;
1181 line
.LineStippleInverseRepeatCount
= 1.0f
/ line_stipple_factor
;
1182 line
.LineStippleRepeatCount
= line_stipple_factor
;
1189 * The pipe->bind_rasterizer_state() driver hook.
1191 * Bind a rasterizer CSO and flag related dirty bits.
1194 iris_bind_rasterizer_state(struct pipe_context
*ctx
, void *state
)
1196 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1197 struct iris_rasterizer_state
*old_cso
= ice
->state
.cso_rast
;
1198 struct iris_rasterizer_state
*new_cso
= state
;
1201 /* Try to avoid re-emitting 3DSTATE_LINE_STIPPLE, it's non-pipelined */
1202 if (cso_changed_memcmp(line_stipple
))
1203 ice
->state
.dirty
|= IRIS_DIRTY_LINE_STIPPLE
;
1205 if (cso_changed(half_pixel_center
))
1206 ice
->state
.dirty
|= IRIS_DIRTY_MULTISAMPLE
;
1208 if (cso_changed(line_stipple_enable
) || cso_changed(poly_stipple_enable
))
1209 ice
->state
.dirty
|= IRIS_DIRTY_WM
;
1211 if (cso_changed(rasterizer_discard
) || cso_changed(flatshade_first
))
1212 ice
->state
.dirty
|= IRIS_DIRTY_STREAMOUT
;
1214 if (cso_changed(depth_clip_near
) || cso_changed(depth_clip_far
) ||
1215 cso_changed(clip_halfz
))
1216 ice
->state
.dirty
|= IRIS_DIRTY_CC_VIEWPORT
;
1218 if (cso_changed(sprite_coord_enable
) || cso_changed(light_twoside
))
1219 ice
->state
.dirty
|= IRIS_DIRTY_SBE
;
1222 ice
->state
.cso_rast
= new_cso
;
1223 ice
->state
.dirty
|= IRIS_DIRTY_RASTER
;
1224 ice
->state
.dirty
|= IRIS_DIRTY_CLIP
;
1225 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_RASTERIZER
];
1229 * Return true if the given wrap mode requires the border color to exist.
1231 * (We can skip uploading it if the sampler isn't going to use it.)
1234 wrap_mode_needs_border_color(unsigned wrap_mode
)
1236 return wrap_mode
== TCM_CLAMP_BORDER
|| wrap_mode
== TCM_HALF_BORDER
;
1240 * Gallium CSO for sampler state.
1242 struct iris_sampler_state
{
1243 union pipe_color_union border_color
;
1244 bool needs_border_color
;
1246 uint32_t sampler_state
[GENX(SAMPLER_STATE_length
)];
1250 * The pipe->create_sampler_state() driver hook.
1252 * We fill out SAMPLER_STATE (except for the border color pointer), and
1253 * store that on the CPU. It doesn't make sense to upload it to a GPU
1254 * buffer object yet, because 3DSTATE_SAMPLER_STATE_POINTERS requires
1255 * all bound sampler states to be in contiguous memor.
1258 iris_create_sampler_state(struct pipe_context
*ctx
,
1259 const struct pipe_sampler_state
*state
)
1261 struct iris_sampler_state
*cso
= CALLOC_STRUCT(iris_sampler_state
);
1266 STATIC_ASSERT(PIPE_TEX_FILTER_NEAREST
== MAPFILTER_NEAREST
);
1267 STATIC_ASSERT(PIPE_TEX_FILTER_LINEAR
== MAPFILTER_LINEAR
);
1269 unsigned wrap_s
= translate_wrap(state
->wrap_s
);
1270 unsigned wrap_t
= translate_wrap(state
->wrap_t
);
1271 unsigned wrap_r
= translate_wrap(state
->wrap_r
);
1273 memcpy(&cso
->border_color
, &state
->border_color
, sizeof(cso
->border_color
));
1275 cso
->needs_border_color
= wrap_mode_needs_border_color(wrap_s
) ||
1276 wrap_mode_needs_border_color(wrap_t
) ||
1277 wrap_mode_needs_border_color(wrap_r
);
1279 float min_lod
= state
->min_lod
;
1280 unsigned mag_img_filter
= state
->mag_img_filter
;
1282 // XXX: explain this code ported from ilo...I don't get it at all...
1283 if (state
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
&&
1284 state
->min_lod
> 0.0f
) {
1286 mag_img_filter
= state
->min_img_filter
;
1289 iris_pack_state(GENX(SAMPLER_STATE
), cso
->sampler_state
, samp
) {
1290 samp
.TCXAddressControlMode
= wrap_s
;
1291 samp
.TCYAddressControlMode
= wrap_t
;
1292 samp
.TCZAddressControlMode
= wrap_r
;
1293 samp
.CubeSurfaceControlMode
= state
->seamless_cube_map
;
1294 samp
.NonnormalizedCoordinateEnable
= !state
->normalized_coords
;
1295 samp
.MinModeFilter
= state
->min_img_filter
;
1296 samp
.MagModeFilter
= mag_img_filter
;
1297 samp
.MipModeFilter
= translate_mip_filter(state
->min_mip_filter
);
1298 samp
.MaximumAnisotropy
= RATIO21
;
1300 if (state
->max_anisotropy
>= 2) {
1301 if (state
->min_img_filter
== PIPE_TEX_FILTER_LINEAR
) {
1302 samp
.MinModeFilter
= MAPFILTER_ANISOTROPIC
;
1303 samp
.AnisotropicAlgorithm
= EWAApproximation
;
1306 if (state
->mag_img_filter
== PIPE_TEX_FILTER_LINEAR
)
1307 samp
.MagModeFilter
= MAPFILTER_ANISOTROPIC
;
1309 samp
.MaximumAnisotropy
=
1310 MIN2((state
->max_anisotropy
- 2) / 2, RATIO161
);
1313 /* Set address rounding bits if not using nearest filtering. */
1314 if (state
->min_img_filter
!= PIPE_TEX_FILTER_NEAREST
) {
1315 samp
.UAddressMinFilterRoundingEnable
= true;
1316 samp
.VAddressMinFilterRoundingEnable
= true;
1317 samp
.RAddressMinFilterRoundingEnable
= true;
1320 if (state
->mag_img_filter
!= PIPE_TEX_FILTER_NEAREST
) {
1321 samp
.UAddressMagFilterRoundingEnable
= true;
1322 samp
.VAddressMagFilterRoundingEnable
= true;
1323 samp
.RAddressMagFilterRoundingEnable
= true;
1326 if (state
->compare_mode
== PIPE_TEX_COMPARE_R_TO_TEXTURE
)
1327 samp
.ShadowFunction
= translate_shadow_func(state
->compare_func
);
1329 const float hw_max_lod
= GEN_GEN
>= 7 ? 14 : 13;
1331 samp
.LODPreClampMode
= CLAMP_MODE_OGL
;
1332 samp
.MinLOD
= CLAMP(min_lod
, 0, hw_max_lod
);
1333 samp
.MaxLOD
= CLAMP(state
->max_lod
, 0, hw_max_lod
);
1334 samp
.TextureLODBias
= CLAMP(state
->lod_bias
, -16, 15);
1336 /* .BorderColorPointer is filled in by iris_bind_sampler_states. */
1343 * The pipe->bind_sampler_states() driver hook.
1345 * Now that we know all the sampler states, we upload them all into a
1346 * contiguous area of GPU memory, for 3DSTATE_SAMPLER_STATE_POINTERS_*.
1347 * We also fill out the border color state pointers at this point.
1349 * We could defer this work to draw time, but we assume that binding
1350 * will be less frequent than drawing.
1352 // XXX: this may be a bad idea, need to make sure that st/mesa calls us
1353 // XXX: with the complete set of shaders. If it makes multiple calls to
1354 // XXX: things one at a time, we could waste a lot of time assembling things.
1355 // XXX: it doesn't even BUY us anything to do it here, because we only flag
1356 // XXX: IRIS_DIRTY_SAMPLER_STATE when this is called...
1358 iris_bind_sampler_states(struct pipe_context
*ctx
,
1359 enum pipe_shader_type p_stage
,
1360 unsigned start
, unsigned count
,
1363 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1364 gl_shader_stage stage
= stage_from_pipe(p_stage
);
1365 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
1367 assert(start
+ count
<= IRIS_MAX_TEXTURE_SAMPLERS
);
1368 shs
->num_samplers
= MAX2(shs
->num_samplers
, start
+ count
);
1370 for (int i
= 0; i
< count
; i
++) {
1371 shs
->samplers
[start
+ i
] = states
[i
];
1374 /* Assemble the SAMPLER_STATEs into a contiguous table that lives
1375 * in the dynamic state memory zone, so we can point to it via the
1376 * 3DSTATE_SAMPLER_STATE_POINTERS_* commands.
1379 upload_state(ice
->state
.dynamic_uploader
, &shs
->sampler_table
,
1380 count
* 4 * GENX(SAMPLER_STATE_length
), 32);
1384 struct pipe_resource
*res
= shs
->sampler_table
.res
;
1385 shs
->sampler_table
.offset
+=
1386 iris_bo_offset_from_base_address(iris_resource_bo(res
));
1388 /* Make sure all land in the same BO */
1389 iris_border_color_pool_reserve(ice
, IRIS_MAX_TEXTURE_SAMPLERS
);
1391 for (int i
= 0; i
< count
; i
++) {
1392 struct iris_sampler_state
*state
= shs
->samplers
[i
];
1395 memset(map
, 0, 4 * GENX(SAMPLER_STATE_length
));
1396 } else if (!state
->needs_border_color
) {
1397 memcpy(map
, state
->sampler_state
, 4 * GENX(SAMPLER_STATE_length
));
1399 ice
->state
.need_border_colors
= true;
1401 /* Stream out the border color and merge the pointer. */
1403 iris_upload_border_color(ice
, &state
->border_color
);
1405 uint32_t dynamic
[GENX(SAMPLER_STATE_length
)];
1406 iris_pack_state(GENX(SAMPLER_STATE
), dynamic
, dyns
) {
1407 dyns
.BorderColorPointer
= offset
;
1410 for (uint32_t j
= 0; j
< GENX(SAMPLER_STATE_length
); j
++)
1411 map
[j
] = state
->sampler_state
[j
] | dynamic
[j
];
1414 map
+= GENX(SAMPLER_STATE_length
);
1417 ice
->state
.dirty
|= IRIS_DIRTY_SAMPLER_STATES_VS
<< stage
;
1420 static enum isl_channel_select
1421 fmt_swizzle(const struct iris_format_info
*fmt
, enum pipe_swizzle swz
)
1424 case PIPE_SWIZZLE_X
: return fmt
->swizzle
.r
;
1425 case PIPE_SWIZZLE_Y
: return fmt
->swizzle
.g
;
1426 case PIPE_SWIZZLE_Z
: return fmt
->swizzle
.b
;
1427 case PIPE_SWIZZLE_W
: return fmt
->swizzle
.a
;
1428 case PIPE_SWIZZLE_1
: return SCS_ONE
;
1429 case PIPE_SWIZZLE_0
: return SCS_ZERO
;
1430 default: unreachable("invalid swizzle");
1435 fill_buffer_surface_state(struct isl_device
*isl_dev
,
1438 enum isl_format format
,
1442 const struct isl_format_layout
*fmtl
= isl_format_get_layout(format
);
1443 const unsigned cpp
= fmtl
->bpb
/ 8;
1445 /* The ARB_texture_buffer_specification says:
1447 * "The number of texels in the buffer texture's texel array is given by
1449 * floor(<buffer_size> / (<components> * sizeof(<base_type>)),
1451 * where <buffer_size> is the size of the buffer object, in basic
1452 * machine units and <components> and <base_type> are the element count
1453 * and base data type for elements, as specified in Table X.1. The
1454 * number of texels in the texel array is then clamped to the
1455 * implementation-dependent limit MAX_TEXTURE_BUFFER_SIZE_ARB."
1457 * We need to clamp the size in bytes to MAX_TEXTURE_BUFFER_SIZE * stride,
1458 * so that when ISL divides by stride to obtain the number of texels, that
1459 * texel count is clamped to MAX_TEXTURE_BUFFER_SIZE.
1461 unsigned final_size
=
1462 MIN3(size
, bo
->size
- offset
, IRIS_MAX_TEXTURE_BUFFER_SIZE
* cpp
);
1464 isl_buffer_fill_state(isl_dev
, map
,
1465 .address
= bo
->gtt_offset
+ offset
,
1466 .size_B
= final_size
,
1473 * The pipe->create_sampler_view() driver hook.
1475 static struct pipe_sampler_view
*
1476 iris_create_sampler_view(struct pipe_context
*ctx
,
1477 struct pipe_resource
*tex
,
1478 const struct pipe_sampler_view
*tmpl
)
1480 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1481 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
1482 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1483 struct iris_sampler_view
*isv
= calloc(1, sizeof(struct iris_sampler_view
));
1488 /* initialize base object */
1490 isv
->base
.context
= ctx
;
1491 isv
->base
.texture
= NULL
;
1492 pipe_reference_init(&isv
->base
.reference
, 1);
1493 pipe_resource_reference(&isv
->base
.texture
, tex
);
1495 void *map
= upload_state(ice
->state
.surface_uploader
, &isv
->surface_state
,
1496 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
1500 struct iris_bo
*state_bo
= iris_resource_bo(isv
->surface_state
.res
);
1501 isv
->surface_state
.offset
+= iris_bo_offset_from_base_address(state_bo
);
1503 if (util_format_is_depth_or_stencil(tmpl
->format
)) {
1504 struct iris_resource
*zres
, *sres
;
1505 const struct util_format_description
*desc
=
1506 util_format_description(tmpl
->format
);
1508 iris_get_depth_stencil_resources(tex
, &zres
, &sres
);
1510 tex
= util_format_has_depth(desc
) ? &zres
->base
: &sres
->base
;
1513 isv
->res
= (struct iris_resource
*) tex
;
1515 isl_surf_usage_flags_t usage
=
1516 ISL_SURF_USAGE_TEXTURE_BIT
|
1517 (isv
->res
->surf
.usage
& ISL_SURF_USAGE_CUBE_BIT
);
1519 const struct iris_format_info fmt
=
1520 iris_format_for_usage(devinfo
, tmpl
->format
, usage
);
1522 isv
->view
= (struct isl_view
) {
1524 .swizzle
= (struct isl_swizzle
) {
1525 .r
= fmt_swizzle(&fmt
, tmpl
->swizzle_r
),
1526 .g
= fmt_swizzle(&fmt
, tmpl
->swizzle_g
),
1527 .b
= fmt_swizzle(&fmt
, tmpl
->swizzle_b
),
1528 .a
= fmt_swizzle(&fmt
, tmpl
->swizzle_a
),
1533 /* Fill out SURFACE_STATE for this view. */
1534 if (tmpl
->target
!= PIPE_BUFFER
) {
1535 isv
->view
.base_level
= tmpl
->u
.tex
.first_level
;
1536 isv
->view
.levels
= tmpl
->u
.tex
.last_level
- tmpl
->u
.tex
.first_level
+ 1;
1537 // XXX: do I need to port f9fd0cf4790cb2a530e75d1a2206dbb9d8af7cb2?
1538 isv
->view
.base_array_layer
= tmpl
->u
.tex
.first_layer
;
1539 isv
->view
.array_len
=
1540 tmpl
->u
.tex
.last_layer
- tmpl
->u
.tex
.first_layer
+ 1;
1542 isl_surf_fill_state(&screen
->isl_dev
, map
,
1543 .surf
= &isv
->res
->surf
, .view
= &isv
->view
,
1545 .address
= isv
->res
->bo
->gtt_offset
);
1547 // .clear_color = clear_color,
1549 fill_buffer_surface_state(&screen
->isl_dev
, isv
->res
->bo
, map
,
1550 isv
->view
.format
, tmpl
->u
.buf
.offset
,
1558 iris_sampler_view_destroy(struct pipe_context
*ctx
,
1559 struct pipe_sampler_view
*state
)
1561 struct iris_sampler_view
*isv
= (void *) state
;
1562 pipe_resource_reference(&state
->texture
, NULL
);
1563 pipe_resource_reference(&isv
->surface_state
.res
, NULL
);
1568 * The pipe->create_surface() driver hook.
1570 * In Gallium nomenclature, "surfaces" are a view of a resource that
1571 * can be bound as a render target or depth/stencil buffer.
1573 static struct pipe_surface
*
1574 iris_create_surface(struct pipe_context
*ctx
,
1575 struct pipe_resource
*tex
,
1576 const struct pipe_surface
*tmpl
)
1578 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1579 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
1580 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1581 struct iris_surface
*surf
= calloc(1, sizeof(struct iris_surface
));
1582 struct pipe_surface
*psurf
= &surf
->base
;
1583 struct iris_resource
*res
= (struct iris_resource
*) tex
;
1588 pipe_reference_init(&psurf
->reference
, 1);
1589 pipe_resource_reference(&psurf
->texture
, tex
);
1590 psurf
->context
= ctx
;
1591 psurf
->format
= tmpl
->format
;
1592 psurf
->width
= tex
->width0
;
1593 psurf
->height
= tex
->height0
;
1594 psurf
->texture
= tex
;
1595 psurf
->u
.tex
.first_layer
= tmpl
->u
.tex
.first_layer
;
1596 psurf
->u
.tex
.last_layer
= tmpl
->u
.tex
.last_layer
;
1597 psurf
->u
.tex
.level
= tmpl
->u
.tex
.level
;
1599 isl_surf_usage_flags_t usage
= 0;
1601 usage
= ISL_SURF_USAGE_STORAGE_BIT
;
1602 else if (util_format_is_depth_or_stencil(tmpl
->format
))
1603 usage
= ISL_SURF_USAGE_DEPTH_BIT
;
1605 usage
= ISL_SURF_USAGE_RENDER_TARGET_BIT
;
1607 const struct iris_format_info fmt
=
1608 iris_format_for_usage(devinfo
, psurf
->format
, usage
);
1610 if ((usage
& ISL_SURF_USAGE_RENDER_TARGET_BIT
) &&
1611 !isl_format_supports_rendering(devinfo
, fmt
.fmt
)) {
1612 /* Framebuffer validation will reject this invalid case, but it
1613 * hasn't had the opportunity yet. In the meantime, we need to
1614 * avoid hitting ISL asserts about unsupported formats below.
1620 surf
->view
= (struct isl_view
) {
1622 .base_level
= tmpl
->u
.tex
.level
,
1624 .base_array_layer
= tmpl
->u
.tex
.first_layer
,
1625 .array_len
= tmpl
->u
.tex
.last_layer
- tmpl
->u
.tex
.first_layer
+ 1,
1626 .swizzle
= ISL_SWIZZLE_IDENTITY
,
1630 /* Bail early for depth/stencil - we don't want SURFACE_STATE for them. */
1631 if (res
->surf
.usage
& (ISL_SURF_USAGE_DEPTH_BIT
|
1632 ISL_SURF_USAGE_STENCIL_BIT
))
1636 void *map
= upload_state(ice
->state
.surface_uploader
, &surf
->surface_state
,
1637 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
1641 struct iris_bo
*state_bo
= iris_resource_bo(surf
->surface_state
.res
);
1642 surf
->surface_state
.offset
+= iris_bo_offset_from_base_address(state_bo
);
1644 isl_surf_fill_state(&screen
->isl_dev
, map
,
1645 .surf
= &res
->surf
, .view
= &surf
->view
,
1647 .address
= res
->bo
->gtt_offset
);
1649 // .clear_color = clear_color,
1655 * The pipe->set_shader_images() driver hook.
1658 iris_set_shader_images(struct pipe_context
*ctx
,
1659 enum pipe_shader_type p_stage
,
1660 unsigned start_slot
, unsigned count
,
1661 const struct pipe_image_view
*p_images
)
1663 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1664 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
1665 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1666 gl_shader_stage stage
= stage_from_pipe(p_stage
);
1667 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
1669 for (unsigned i
= 0; i
< count
; i
++) {
1670 if (p_images
&& p_images
[i
].resource
) {
1671 const struct pipe_image_view
*img
= &p_images
[i
];
1672 struct iris_resource
*res
= (void *) img
->resource
;
1673 pipe_resource_reference(&shs
->image
[start_slot
+ i
].res
, &res
->base
);
1675 res
->bind_history
|= PIPE_BIND_SHADER_IMAGE
;
1677 // XXX: these are not retained forever, use a separate uploader?
1679 upload_state(ice
->state
.surface_uploader
,
1680 &shs
->image
[start_slot
+ i
].surface_state
,
1681 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
1682 if (!unlikely(map
)) {
1683 pipe_resource_reference(&shs
->image
[start_slot
+ i
].res
, NULL
);
1687 struct iris_bo
*surf_state_bo
=
1688 iris_resource_bo(shs
->image
[start_slot
+ i
].surface_state
.res
);
1689 shs
->image
[start_slot
+ i
].surface_state
.offset
+=
1690 iris_bo_offset_from_base_address(surf_state_bo
);
1692 isl_surf_usage_flags_t usage
= ISL_SURF_USAGE_STORAGE_BIT
;
1693 enum isl_format isl_format
=
1694 iris_format_for_usage(devinfo
, img
->format
, usage
).fmt
;
1696 if (img
->shader_access
& PIPE_IMAGE_ACCESS_READ
)
1697 isl_format
= isl_lower_storage_image_format(devinfo
, isl_format
);
1699 shs
->image
[start_slot
+ i
].access
= img
->shader_access
;
1701 if (res
->base
.target
!= PIPE_BUFFER
) {
1702 struct isl_view view
= {
1703 .format
= isl_format
,
1704 .base_level
= img
->u
.tex
.level
,
1706 .base_array_layer
= img
->u
.tex
.first_layer
,
1707 .array_len
= img
->u
.tex
.last_layer
- img
->u
.tex
.first_layer
+ 1,
1708 .swizzle
= ISL_SWIZZLE_IDENTITY
,
1712 isl_surf_fill_state(&screen
->isl_dev
, map
,
1713 .surf
= &res
->surf
, .view
= &view
,
1715 .address
= res
->bo
->gtt_offset
);
1717 // .clear_color = clear_color,
1719 fill_buffer_surface_state(&screen
->isl_dev
, res
->bo
, map
,
1720 isl_format
, img
->u
.buf
.offset
,
1724 pipe_resource_reference(&shs
->image
[start_slot
+ i
].res
, NULL
);
1725 pipe_resource_reference(&shs
->image
[start_slot
+ i
].surface_state
.res
,
1730 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_VS
<< stage
;
1735 * The pipe->set_sampler_views() driver hook.
1738 iris_set_sampler_views(struct pipe_context
*ctx
,
1739 enum pipe_shader_type p_stage
,
1740 unsigned start
, unsigned count
,
1741 struct pipe_sampler_view
**views
)
1743 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1744 gl_shader_stage stage
= stage_from_pipe(p_stage
);
1745 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
1748 for (i
= 0; i
< count
; i
++) {
1749 pipe_sampler_view_reference((struct pipe_sampler_view
**)
1750 &shs
->textures
[i
], views
[i
]);
1751 struct iris_sampler_view
*view
= (void *) views
[i
];
1753 view
->res
->bind_history
|= PIPE_BIND_SAMPLER_VIEW
;
1755 for (; i
< shs
->num_textures
; i
++) {
1756 pipe_sampler_view_reference((struct pipe_sampler_view
**)
1757 &shs
->textures
[i
], NULL
);
1760 shs
->num_textures
= count
;
1762 ice
->state
.dirty
|= (IRIS_DIRTY_BINDINGS_VS
<< stage
);
1766 * The pipe->set_tess_state() driver hook.
1769 iris_set_tess_state(struct pipe_context
*ctx
,
1770 const float default_outer_level
[4],
1771 const float default_inner_level
[2])
1773 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1775 memcpy(&ice
->state
.default_outer_level
[0], &default_outer_level
[0], 4 * sizeof(float));
1776 memcpy(&ice
->state
.default_inner_level
[0], &default_inner_level
[0], 2 * sizeof(float));
1778 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_TCS
;
1782 iris_surface_destroy(struct pipe_context
*ctx
, struct pipe_surface
*p_surf
)
1784 struct iris_surface
*surf
= (void *) p_surf
;
1785 pipe_resource_reference(&p_surf
->texture
, NULL
);
1786 pipe_resource_reference(&surf
->surface_state
.res
, NULL
);
1791 iris_set_clip_state(struct pipe_context
*ctx
,
1792 const struct pipe_clip_state
*state
)
1794 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1795 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_VERTEX
];
1797 memcpy(&ice
->state
.clip_planes
, state
, sizeof(*state
));
1799 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_VS
;
1800 shs
->cbuf0_needs_upload
= true;
1804 * The pipe->set_polygon_stipple() driver hook.
1807 iris_set_polygon_stipple(struct pipe_context
*ctx
,
1808 const struct pipe_poly_stipple
*state
)
1810 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1811 memcpy(&ice
->state
.poly_stipple
, state
, sizeof(*state
));
1812 ice
->state
.dirty
|= IRIS_DIRTY_POLYGON_STIPPLE
;
1816 * The pipe->set_sample_mask() driver hook.
1819 iris_set_sample_mask(struct pipe_context
*ctx
, unsigned sample_mask
)
1821 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1823 /* We only support 16x MSAA, so we have 16 bits of sample maks.
1824 * st/mesa may pass us 0xffffffff though, meaning "enable all samples".
1826 ice
->state
.sample_mask
= sample_mask
& 0xffff;
1827 ice
->state
.dirty
|= IRIS_DIRTY_SAMPLE_MASK
;
1831 * The pipe->set_scissor_states() driver hook.
1833 * This corresponds to our SCISSOR_RECT state structures. It's an
1834 * exact match, so we just store them, and memcpy them out later.
1837 iris_set_scissor_states(struct pipe_context
*ctx
,
1838 unsigned start_slot
,
1839 unsigned num_scissors
,
1840 const struct pipe_scissor_state
*rects
)
1842 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1844 for (unsigned i
= 0; i
< num_scissors
; i
++) {
1845 if (rects
[i
].minx
== rects
[i
].maxx
|| rects
[i
].miny
== rects
[i
].maxy
) {
1846 /* If the scissor was out of bounds and got clamped to 0 width/height
1847 * at the bounds, the subtraction of 1 from maximums could produce a
1848 * negative number and thus not clip anything. Instead, just provide
1849 * a min > max scissor inside the bounds, which produces the expected
1852 ice
->state
.scissors
[start_slot
+ i
] = (struct pipe_scissor_state
) {
1853 .minx
= 1, .maxx
= 0, .miny
= 1, .maxy
= 0,
1856 ice
->state
.scissors
[start_slot
+ i
] = (struct pipe_scissor_state
) {
1857 .minx
= rects
[i
].minx
, .miny
= rects
[i
].miny
,
1858 .maxx
= rects
[i
].maxx
- 1, .maxy
= rects
[i
].maxy
- 1,
1863 ice
->state
.dirty
|= IRIS_DIRTY_SCISSOR_RECT
;
1867 * The pipe->set_stencil_ref() driver hook.
1869 * This is added to 3DSTATE_WM_DEPTH_STENCIL dynamically at draw time.
1872 iris_set_stencil_ref(struct pipe_context
*ctx
,
1873 const struct pipe_stencil_ref
*state
)
1875 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1876 memcpy(&ice
->state
.stencil_ref
, state
, sizeof(*state
));
1877 ice
->state
.dirty
|= IRIS_DIRTY_WM_DEPTH_STENCIL
;
1881 viewport_extent(const struct pipe_viewport_state
*state
, int axis
, float sign
)
1883 return copysignf(state
->scale
[axis
], sign
) + state
->translate
[axis
];
1888 calculate_guardband_size(uint32_t fb_width
, uint32_t fb_height
,
1889 float m00
, float m11
, float m30
, float m31
,
1890 float *xmin
, float *xmax
,
1891 float *ymin
, float *ymax
)
1893 /* According to the "Vertex X,Y Clamping and Quantization" section of the
1894 * Strips and Fans documentation:
1896 * "The vertex X and Y screen-space coordinates are also /clamped/ to the
1897 * fixed-point "guardband" range supported by the rasterization hardware"
1901 * "In almost all circumstances, if an object’s vertices are actually
1902 * modified by this clamping (i.e., had X or Y coordinates outside of
1903 * the guardband extent the rendered object will not match the intended
1904 * result. Therefore software should take steps to ensure that this does
1905 * not happen - e.g., by clipping objects such that they do not exceed
1906 * these limits after the Drawing Rectangle is applied."
1908 * I believe the fundamental restriction is that the rasterizer (in
1909 * the SF/WM stages) have a limit on the number of pixels that can be
1910 * rasterized. We need to ensure any coordinates beyond the rasterizer
1911 * limit are handled by the clipper. So effectively that limit becomes
1912 * the clipper's guardband size.
1914 * It goes on to say:
1916 * "In addition, in order to be correctly rendered, objects must have a
1917 * screenspace bounding box not exceeding 8K in the X or Y direction.
1918 * This additional restriction must also be comprehended by software,
1919 * i.e., enforced by use of clipping."
1921 * This makes no sense. Gen7+ hardware supports 16K render targets,
1922 * and you definitely need to be able to draw polygons that fill the
1923 * surface. Our assumption is that the rasterizer was limited to 8K
1924 * on Sandybridge, which only supports 8K surfaces, and it was actually
1925 * increased to 16K on Ivybridge and later.
1927 * So, limit the guardband to 16K on Gen7+ and 8K on Sandybridge.
1929 const float gb_size
= GEN_GEN
>= 7 ? 16384.0f
: 8192.0f
;
1931 if (m00
!= 0 && m11
!= 0) {
1932 /* First, we compute the screen-space render area */
1933 const float ss_ra_xmin
= MIN3( 0, m30
+ m00
, m30
- m00
);
1934 const float ss_ra_xmax
= MAX3( fb_width
, m30
+ m00
, m30
- m00
);
1935 const float ss_ra_ymin
= MIN3( 0, m31
+ m11
, m31
- m11
);
1936 const float ss_ra_ymax
= MAX3(fb_height
, m31
+ m11
, m31
- m11
);
1938 /* We want the guardband to be centered on that */
1939 const float ss_gb_xmin
= (ss_ra_xmin
+ ss_ra_xmax
) / 2 - gb_size
;
1940 const float ss_gb_xmax
= (ss_ra_xmin
+ ss_ra_xmax
) / 2 + gb_size
;
1941 const float ss_gb_ymin
= (ss_ra_ymin
+ ss_ra_ymax
) / 2 - gb_size
;
1942 const float ss_gb_ymax
= (ss_ra_ymin
+ ss_ra_ymax
) / 2 + gb_size
;
1944 /* Now we need it in native device coordinates */
1945 const float ndc_gb_xmin
= (ss_gb_xmin
- m30
) / m00
;
1946 const float ndc_gb_xmax
= (ss_gb_xmax
- m30
) / m00
;
1947 const float ndc_gb_ymin
= (ss_gb_ymin
- m31
) / m11
;
1948 const float ndc_gb_ymax
= (ss_gb_ymax
- m31
) / m11
;
1950 /* Thanks to Y-flipping and ORIGIN_UPPER_LEFT, the Y coordinates may be
1951 * flipped upside-down. X should be fine though.
1953 assert(ndc_gb_xmin
<= ndc_gb_xmax
);
1954 *xmin
= ndc_gb_xmin
;
1955 *xmax
= ndc_gb_xmax
;
1956 *ymin
= MIN2(ndc_gb_ymin
, ndc_gb_ymax
);
1957 *ymax
= MAX2(ndc_gb_ymin
, ndc_gb_ymax
);
1959 /* The viewport scales to 0, so nothing will be rendered. */
1969 * The pipe->set_viewport_states() driver hook.
1971 * This corresponds to our SF_CLIP_VIEWPORT states. We can't calculate
1972 * the guardband yet, as we need the framebuffer dimensions, but we can
1973 * at least fill out the rest.
1976 iris_set_viewport_states(struct pipe_context
*ctx
,
1977 unsigned start_slot
,
1979 const struct pipe_viewport_state
*states
)
1981 struct iris_context
*ice
= (struct iris_context
*) ctx
;
1982 struct iris_genx_state
*genx
= ice
->state
.genx
;
1984 &genx
->sf_cl_vp
[start_slot
* GENX(SF_CLIP_VIEWPORT_length
)];
1986 for (unsigned i
= 0; i
< count
; i
++) {
1987 const struct pipe_viewport_state
*state
= &states
[i
];
1989 memcpy(&ice
->state
.viewports
[start_slot
+ i
], state
, sizeof(*state
));
1991 iris_pack_state(GENX(SF_CLIP_VIEWPORT
), vp_map
, vp
) {
1992 vp
.ViewportMatrixElementm00
= state
->scale
[0];
1993 vp
.ViewportMatrixElementm11
= state
->scale
[1];
1994 vp
.ViewportMatrixElementm22
= state
->scale
[2];
1995 vp
.ViewportMatrixElementm30
= state
->translate
[0];
1996 vp
.ViewportMatrixElementm31
= state
->translate
[1];
1997 vp
.ViewportMatrixElementm32
= state
->translate
[2];
1998 /* XXX: in i965 this is computed based on the drawbuffer size,
1999 * but we don't have that here...
2001 vp
.XMinClipGuardband
= -1.0;
2002 vp
.XMaxClipGuardband
= 1.0;
2003 vp
.YMinClipGuardband
= -1.0;
2004 vp
.YMaxClipGuardband
= 1.0;
2005 vp
.XMinViewPort
= viewport_extent(state
, 0, -1.0f
);
2006 vp
.XMaxViewPort
= viewport_extent(state
, 0, 1.0f
) - 1;
2007 vp
.YMinViewPort
= viewport_extent(state
, 1, -1.0f
);
2008 vp
.YMaxViewPort
= viewport_extent(state
, 1, 1.0f
) - 1;
2011 vp_map
+= GENX(SF_CLIP_VIEWPORT_length
);
2014 ice
->state
.dirty
|= IRIS_DIRTY_SF_CL_VIEWPORT
;
2016 if (ice
->state
.cso_rast
&& (!ice
->state
.cso_rast
->depth_clip_near
||
2017 !ice
->state
.cso_rast
->depth_clip_far
))
2018 ice
->state
.dirty
|= IRIS_DIRTY_CC_VIEWPORT
;
2022 * The pipe->set_framebuffer_state() driver hook.
2024 * Sets the current draw FBO, including color render targets, depth,
2025 * and stencil buffers.
2028 iris_set_framebuffer_state(struct pipe_context
*ctx
,
2029 const struct pipe_framebuffer_state
*state
)
2031 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2032 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
2033 struct isl_device
*isl_dev
= &screen
->isl_dev
;
2034 struct pipe_framebuffer_state
*cso
= &ice
->state
.framebuffer
;
2035 struct iris_resource
*zres
;
2036 struct iris_resource
*stencil_res
;
2038 unsigned samples
= util_framebuffer_get_num_samples(state
);
2040 if (cso
->samples
!= samples
) {
2041 ice
->state
.dirty
|= IRIS_DIRTY_MULTISAMPLE
;
2044 if (cso
->nr_cbufs
!= state
->nr_cbufs
) {
2045 ice
->state
.dirty
|= IRIS_DIRTY_BLEND_STATE
;
2048 if ((cso
->layers
== 0) != (state
->layers
== 0)) {
2049 ice
->state
.dirty
|= IRIS_DIRTY_CLIP
;
2052 util_copy_framebuffer_state(cso
, state
);
2053 cso
->samples
= samples
;
2055 struct iris_depth_buffer_state
*cso_z
= &ice
->state
.genx
->depth_buffer
;
2057 struct isl_view view
= {
2060 .base_array_layer
= 0,
2062 .swizzle
= ISL_SWIZZLE_IDENTITY
,
2065 struct isl_depth_stencil_hiz_emit_info info
= {
2071 iris_get_depth_stencil_resources(cso
->zsbuf
->texture
, &zres
,
2074 view
.base_level
= cso
->zsbuf
->u
.tex
.level
;
2075 view
.base_array_layer
= cso
->zsbuf
->u
.tex
.first_layer
;
2077 cso
->zsbuf
->u
.tex
.last_layer
- cso
->zsbuf
->u
.tex
.first_layer
+ 1;
2080 view
.usage
|= ISL_SURF_USAGE_DEPTH_BIT
;
2082 info
.depth_surf
= &zres
->surf
;
2083 info
.depth_address
= zres
->bo
->gtt_offset
;
2084 info
.hiz_usage
= ISL_AUX_USAGE_NONE
;
2086 view
.format
= zres
->surf
.format
;
2090 view
.usage
|= ISL_SURF_USAGE_STENCIL_BIT
;
2091 info
.stencil_surf
= &stencil_res
->surf
;
2092 info
.stencil_address
= stencil_res
->bo
->gtt_offset
;
2094 view
.format
= stencil_res
->surf
.format
;
2098 isl_emit_depth_stencil_hiz_s(isl_dev
, cso_z
->packets
, &info
);
2100 /* Make a null surface for unbound buffers */
2101 void *null_surf_map
=
2102 upload_state(ice
->state
.surface_uploader
, &ice
->state
.null_fb
,
2103 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
2104 isl_null_fill_state(&screen
->isl_dev
, null_surf_map
,
2105 isl_extent3d(MAX2(cso
->width
, 1),
2106 MAX2(cso
->height
, 1),
2107 cso
->layers
? cso
->layers
: 1));
2108 ice
->state
.null_fb
.offset
+=
2109 iris_bo_offset_from_base_address(iris_resource_bo(ice
->state
.null_fb
.res
));
2111 ice
->state
.dirty
|= IRIS_DIRTY_DEPTH_BUFFER
;
2113 /* Render target change */
2114 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_FS
;
2116 ice
->state
.dirty
|= ice
->state
.dirty_for_nos
[IRIS_NOS_FRAMEBUFFER
];
2119 // XXX: we may want to flag IRIS_DIRTY_MULTISAMPLE (or SAMPLE_MASK?)
2120 // XXX: see commit 979fc1bc9bcc64027ff2cfafd285676f31b930a6
2122 /* The PIPE_CONTROL command description says:
2124 * "Whenever a Binding Table Index (BTI) used by a Render Target Message
2125 * points to a different RENDER_SURFACE_STATE, SW must issue a Render
2126 * Target Cache Flush by enabling this bit. When render target flush
2127 * is set due to new association of BTI, PS Scoreboard Stall bit must
2128 * be set in this packet."
2130 // XXX: does this need to happen at 3DSTATE_BTP_PS time?
2131 iris_emit_pipe_control_flush(&ice
->batches
[IRIS_BATCH_RENDER
],
2132 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
2133 PIPE_CONTROL_STALL_AT_SCOREBOARD
);
2138 upload_ubo_surf_state(struct iris_context
*ice
,
2139 struct iris_const_buffer
*cbuf
,
2140 unsigned buffer_size
)
2142 struct pipe_context
*ctx
= &ice
->ctx
;
2143 struct iris_screen
*screen
= (struct iris_screen
*) ctx
->screen
;
2145 // XXX: these are not retained forever, use a separate uploader?
2147 upload_state(ice
->state
.surface_uploader
, &cbuf
->surface_state
,
2148 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
2149 if (!unlikely(map
)) {
2150 pipe_resource_reference(&cbuf
->data
.res
, NULL
);
2154 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
2155 struct iris_bo
*surf_bo
= iris_resource_bo(cbuf
->surface_state
.res
);
2156 cbuf
->surface_state
.offset
+= iris_bo_offset_from_base_address(surf_bo
);
2158 isl_buffer_fill_state(&screen
->isl_dev
, map
,
2159 .address
= res
->bo
->gtt_offset
+ cbuf
->data
.offset
,
2160 .size_B
= MIN2(buffer_size
,
2161 res
->bo
->size
- cbuf
->data
.offset
),
2162 .format
= ISL_FORMAT_R32G32B32A32_FLOAT
,
2168 * The pipe->set_constant_buffer() driver hook.
2170 * This uploads any constant data in user buffers, and references
2171 * any UBO resources containing constant data.
2174 iris_set_constant_buffer(struct pipe_context
*ctx
,
2175 enum pipe_shader_type p_stage
, unsigned index
,
2176 const struct pipe_constant_buffer
*input
)
2178 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2179 gl_shader_stage stage
= stage_from_pipe(p_stage
);
2180 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
2181 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[index
];
2183 if (input
&& input
->buffer
) {
2186 pipe_resource_reference(&cbuf
->data
.res
, input
->buffer
);
2187 cbuf
->data
.offset
= input
->buffer_offset
;
2189 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
2190 res
->bind_history
|= PIPE_BIND_CONSTANT_BUFFER
;
2192 upload_ubo_surf_state(ice
, cbuf
, input
->buffer_size
);
2194 pipe_resource_reference(&cbuf
->data
.res
, NULL
);
2195 pipe_resource_reference(&cbuf
->surface_state
.res
, NULL
);
2200 memcpy(&shs
->cbuf0
, input
, sizeof(shs
->cbuf0
));
2202 memset(&shs
->cbuf0
, 0, sizeof(shs
->cbuf0
));
2204 shs
->cbuf0_needs_upload
= true;
2207 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_VS
<< stage
;
2208 // XXX: maybe not necessary all the time...?
2209 // XXX: we need 3DS_BTP to commit these changes, and if we fell back to
2210 // XXX: pull model we may need actual new bindings...
2211 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_VS
<< stage
;
2215 upload_uniforms(struct iris_context
*ice
,
2216 gl_shader_stage stage
)
2218 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
2219 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[0];
2220 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
2222 unsigned upload_size
= shader
->num_system_values
* sizeof(uint32_t) +
2223 shs
->cbuf0
.buffer_size
;
2225 if (upload_size
== 0)
2229 upload_state(ice
->ctx
.const_uploader
, &cbuf
->data
, upload_size
, 64);
2231 for (int i
= 0; i
< shader
->num_system_values
; i
++) {
2232 uint32_t sysval
= shader
->system_values
[i
];
2235 if (BRW_PARAM_BUILTIN_IS_CLIP_PLANE(sysval
)) {
2236 int plane
= BRW_PARAM_BUILTIN_CLIP_PLANE_IDX(sysval
);
2237 int comp
= BRW_PARAM_BUILTIN_CLIP_PLANE_COMP(sysval
);
2238 value
= fui(ice
->state
.clip_planes
.ucp
[plane
][comp
]);
2240 assert(!"unhandled system value");
2246 if (shs
->cbuf0
.user_buffer
) {
2247 memcpy(map
, shs
->cbuf0
.user_buffer
, shs
->cbuf0
.buffer_size
);
2250 upload_ubo_surf_state(ice
, cbuf
, upload_size
);
2254 * The pipe->set_shader_buffers() driver hook.
2256 * This binds SSBOs and ABOs. Unfortunately, we need to stream out
2257 * SURFACE_STATE here, as the buffer offset may change each time.
2260 iris_set_shader_buffers(struct pipe_context
*ctx
,
2261 enum pipe_shader_type p_stage
,
2262 unsigned start_slot
, unsigned count
,
2263 const struct pipe_shader_buffer
*buffers
)
2265 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2266 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
2267 gl_shader_stage stage
= stage_from_pipe(p_stage
);
2268 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
2270 for (unsigned i
= 0; i
< count
; i
++) {
2271 if (buffers
&& buffers
[i
].buffer
) {
2272 const struct pipe_shader_buffer
*buffer
= &buffers
[i
];
2273 struct iris_resource
*res
= (void *) buffer
->buffer
;
2274 pipe_resource_reference(&shs
->ssbo
[start_slot
+ i
], &res
->base
);
2276 res
->bind_history
|= PIPE_BIND_SHADER_BUFFER
;
2278 // XXX: these are not retained forever, use a separate uploader?
2280 upload_state(ice
->state
.surface_uploader
,
2281 &shs
->ssbo_surface_state
[start_slot
+ i
],
2282 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
2283 if (!unlikely(map
)) {
2284 pipe_resource_reference(&shs
->ssbo
[start_slot
+ i
], NULL
);
2288 struct iris_bo
*surf_state_bo
=
2289 iris_resource_bo(shs
->ssbo_surface_state
[start_slot
+ i
].res
);
2290 shs
->ssbo_surface_state
[start_slot
+ i
].offset
+=
2291 iris_bo_offset_from_base_address(surf_state_bo
);
2293 isl_buffer_fill_state(&screen
->isl_dev
, map
,
2295 res
->bo
->gtt_offset
+ buffer
->buffer_offset
,
2297 MIN2(buffer
->buffer_size
,
2298 res
->bo
->size
- buffer
->buffer_offset
),
2299 .format
= ISL_FORMAT_RAW
,
2303 pipe_resource_reference(&shs
->ssbo
[start_slot
+ i
], NULL
);
2304 pipe_resource_reference(&shs
->ssbo_surface_state
[start_slot
+ i
].res
,
2309 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_VS
<< stage
;
2313 iris_delete_state(struct pipe_context
*ctx
, void *state
)
2319 iris_free_vertex_buffers(struct iris_vertex_buffer_state
*cso
)
2321 for (unsigned i
= 0; i
< cso
->num_buffers
; i
++)
2322 pipe_resource_reference(&cso
->resources
[i
], NULL
);
2326 * The pipe->set_vertex_buffers() driver hook.
2328 * This translates pipe_vertex_buffer to our 3DSTATE_VERTEX_BUFFERS packet.
2331 iris_set_vertex_buffers(struct pipe_context
*ctx
,
2332 unsigned start_slot
, unsigned count
,
2333 const struct pipe_vertex_buffer
*buffers
)
2335 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2336 struct iris_vertex_buffer_state
*cso
= &ice
->state
.genx
->vertex_buffers
;
2338 iris_free_vertex_buffers(&ice
->state
.genx
->vertex_buffers
);
2343 cso
->num_buffers
= count
;
2345 iris_pack_command(GENX(3DSTATE_VERTEX_BUFFERS
), cso
->vertex_buffers
, vb
) {
2346 vb
.DWordLength
= 4 * MAX2(cso
->num_buffers
, 1) - 1;
2349 uint32_t *vb_pack_dest
= &cso
->vertex_buffers
[1];
2352 iris_pack_state(GENX(VERTEX_BUFFER_STATE
), vb_pack_dest
, vb
) {
2353 vb
.VertexBufferIndex
= start_slot
;
2354 vb
.NullVertexBuffer
= true;
2355 vb
.AddressModifyEnable
= true;
2359 for (unsigned i
= 0; i
< count
; i
++) {
2360 assert(!buffers
[i
].is_user_buffer
);
2362 pipe_resource_reference(&cso
->resources
[i
], buffers
[i
].buffer
.resource
);
2363 struct iris_resource
*res
= (void *) cso
->resources
[i
];
2366 res
->bind_history
|= PIPE_BIND_VERTEX_BUFFER
;
2368 iris_pack_state(GENX(VERTEX_BUFFER_STATE
), vb_pack_dest
, vb
) {
2369 vb
.VertexBufferIndex
= start_slot
+ i
;
2371 vb
.AddressModifyEnable
= true;
2372 vb
.BufferPitch
= buffers
[i
].stride
;
2374 vb
.BufferSize
= res
->bo
->size
;
2375 vb
.BufferStartingAddress
=
2376 ro_bo(NULL
, res
->bo
->gtt_offset
+ buffers
[i
].buffer_offset
);
2378 vb
.NullVertexBuffer
= true;
2382 vb_pack_dest
+= GENX(VERTEX_BUFFER_STATE_length
);
2385 ice
->state
.dirty
|= IRIS_DIRTY_VERTEX_BUFFERS
;
2389 * Gallium CSO for vertex elements.
2391 struct iris_vertex_element_state
{
2392 uint32_t vertex_elements
[1 + 33 * GENX(VERTEX_ELEMENT_STATE_length
)];
2393 uint32_t vf_instancing
[33 * GENX(3DSTATE_VF_INSTANCING_length
)];
2398 * The pipe->create_vertex_elements() driver hook.
2400 * This translates pipe_vertex_element to our 3DSTATE_VERTEX_ELEMENTS
2401 * and 3DSTATE_VF_INSTANCING commands. SGVs are handled at draw time.
2404 iris_create_vertex_elements(struct pipe_context
*ctx
,
2406 const struct pipe_vertex_element
*state
)
2408 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
2409 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
2410 struct iris_vertex_element_state
*cso
=
2411 malloc(sizeof(struct iris_vertex_element_state
));
2416 * - create edge flag one
2418 * - if those are necessary, use count + 1/2/3... OR in the length
2420 iris_pack_command(GENX(3DSTATE_VERTEX_ELEMENTS
), cso
->vertex_elements
, ve
) {
2422 1 + GENX(VERTEX_ELEMENT_STATE_length
) * MAX2(count
, 1) - 2;
2425 uint32_t *ve_pack_dest
= &cso
->vertex_elements
[1];
2426 uint32_t *vfi_pack_dest
= cso
->vf_instancing
;
2429 iris_pack_state(GENX(VERTEX_ELEMENT_STATE
), ve_pack_dest
, ve
) {
2431 ve
.SourceElementFormat
= ISL_FORMAT_R32G32B32A32_FLOAT
;
2432 ve
.Component0Control
= VFCOMP_STORE_0
;
2433 ve
.Component1Control
= VFCOMP_STORE_0
;
2434 ve
.Component2Control
= VFCOMP_STORE_0
;
2435 ve
.Component3Control
= VFCOMP_STORE_1_FP
;
2438 iris_pack_command(GENX(3DSTATE_VF_INSTANCING
), vfi_pack_dest
, vi
) {
2442 for (int i
= 0; i
< count
; i
++) {
2443 const struct iris_format_info fmt
=
2444 iris_format_for_usage(devinfo
, state
[i
].src_format
, 0);
2445 unsigned comp
[4] = { VFCOMP_STORE_SRC
, VFCOMP_STORE_SRC
,
2446 VFCOMP_STORE_SRC
, VFCOMP_STORE_SRC
};
2448 switch (isl_format_get_num_channels(fmt
.fmt
)) {
2449 case 0: comp
[0] = VFCOMP_STORE_0
;
2450 case 1: comp
[1] = VFCOMP_STORE_0
;
2451 case 2: comp
[2] = VFCOMP_STORE_0
;
2453 comp
[3] = isl_format_has_int_channel(fmt
.fmt
) ? VFCOMP_STORE_1_INT
2454 : VFCOMP_STORE_1_FP
;
2457 iris_pack_state(GENX(VERTEX_ELEMENT_STATE
), ve_pack_dest
, ve
) {
2458 ve
.VertexBufferIndex
= state
[i
].vertex_buffer_index
;
2460 ve
.SourceElementOffset
= state
[i
].src_offset
;
2461 ve
.SourceElementFormat
= fmt
.fmt
;
2462 ve
.Component0Control
= comp
[0];
2463 ve
.Component1Control
= comp
[1];
2464 ve
.Component2Control
= comp
[2];
2465 ve
.Component3Control
= comp
[3];
2468 iris_pack_command(GENX(3DSTATE_VF_INSTANCING
), vfi_pack_dest
, vi
) {
2469 vi
.VertexElementIndex
= i
;
2470 vi
.InstancingEnable
= state
[i
].instance_divisor
> 0;
2471 vi
.InstanceDataStepRate
= state
[i
].instance_divisor
;
2474 ve_pack_dest
+= GENX(VERTEX_ELEMENT_STATE_length
);
2475 vfi_pack_dest
+= GENX(3DSTATE_VF_INSTANCING_length
);
2482 * The pipe->bind_vertex_elements_state() driver hook.
2485 iris_bind_vertex_elements_state(struct pipe_context
*ctx
, void *state
)
2487 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2488 struct iris_vertex_element_state
*old_cso
= ice
->state
.cso_vertex_elements
;
2489 struct iris_vertex_element_state
*new_cso
= state
;
2491 /* 3DSTATE_VF_SGVs overrides the last VE, so if the count is changing,
2492 * we need to re-emit it to ensure we're overriding the right one.
2494 if (new_cso
&& cso_changed(count
))
2495 ice
->state
.dirty
|= IRIS_DIRTY_VF_SGVS
;
2497 ice
->state
.cso_vertex_elements
= state
;
2498 ice
->state
.dirty
|= IRIS_DIRTY_VERTEX_ELEMENTS
;
2502 * Gallium CSO for stream output (transform feedback) targets.
2504 struct iris_stream_output_target
{
2505 struct pipe_stream_output_target base
;
2507 uint32_t so_buffer
[GENX(3DSTATE_SO_BUFFER_length
)];
2509 /** Storage holding the offset where we're writing in the buffer */
2510 struct iris_state_ref offset
;
2514 * The pipe->create_stream_output_target() driver hook.
2516 * "Target" here refers to a destination buffer. We translate this into
2517 * a 3DSTATE_SO_BUFFER packet. We can handle most fields, but don't yet
2518 * know which buffer this represents, or whether we ought to zero the
2519 * write-offsets, or append. Those are handled in the set() hook.
2521 static struct pipe_stream_output_target
*
2522 iris_create_stream_output_target(struct pipe_context
*ctx
,
2523 struct pipe_resource
*p_res
,
2524 unsigned buffer_offset
,
2525 unsigned buffer_size
)
2527 struct iris_resource
*res
= (void *) p_res
;
2528 struct iris_stream_output_target
*cso
= calloc(1, sizeof(*cso
));
2532 res
->bind_history
|= PIPE_BIND_STREAM_OUTPUT
;
2534 pipe_reference_init(&cso
->base
.reference
, 1);
2535 pipe_resource_reference(&cso
->base
.buffer
, p_res
);
2536 cso
->base
.buffer_offset
= buffer_offset
;
2537 cso
->base
.buffer_size
= buffer_size
;
2538 cso
->base
.context
= ctx
;
2540 upload_state(ctx
->stream_uploader
, &cso
->offset
, 4 * sizeof(uint32_t), 4);
2542 iris_pack_command(GENX(3DSTATE_SO_BUFFER
), cso
->so_buffer
, sob
) {
2543 sob
.SurfaceBaseAddress
=
2544 rw_bo(NULL
, res
->bo
->gtt_offset
+ buffer_offset
);
2545 sob
.SOBufferEnable
= true;
2546 sob
.StreamOffsetWriteEnable
= true;
2547 sob
.StreamOutputBufferOffsetAddressEnable
= true;
2548 sob
.MOCS
= MOCS_WB
; // XXX: MOCS
2550 sob
.SurfaceSize
= MAX2(buffer_size
/ 4, 1) - 1;
2552 /* .SOBufferIndex, .StreamOffset, and .StreamOutputBufferOffsetAddress
2553 * are filled in later when we have stream IDs.
2561 iris_stream_output_target_destroy(struct pipe_context
*ctx
,
2562 struct pipe_stream_output_target
*state
)
2564 struct iris_stream_output_target
*cso
= (void *) state
;
2566 pipe_resource_reference(&cso
->base
.buffer
, NULL
);
2567 pipe_resource_reference(&cso
->offset
.res
, NULL
);
2573 * The pipe->set_stream_output_targets() driver hook.
2575 * At this point, we know which targets are bound to a particular index,
2576 * and also whether we want to append or start over. We can finish the
2577 * 3DSTATE_SO_BUFFER packets we started earlier.
2580 iris_set_stream_output_targets(struct pipe_context
*ctx
,
2581 unsigned num_targets
,
2582 struct pipe_stream_output_target
**targets
,
2583 const unsigned *offsets
)
2585 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2586 struct iris_genx_state
*genx
= ice
->state
.genx
;
2587 uint32_t *so_buffers
= genx
->so_buffers
;
2589 const bool active
= num_targets
> 0;
2590 if (ice
->state
.streamout_active
!= active
) {
2591 ice
->state
.streamout_active
= active
;
2592 ice
->state
.dirty
|= IRIS_DIRTY_STREAMOUT
;
2594 /* We only emit 3DSTATE_SO_DECL_LIST when streamout is active, because
2595 * it's a non-pipelined command. If we're switching streamout on, we
2596 * may have missed emitting it earlier, so do so now. (We're already
2597 * taking a stall to update 3DSTATE_SO_BUFFERS anyway...)
2600 ice
->state
.dirty
|= IRIS_DIRTY_SO_DECL_LIST
;
2603 for (int i
= 0; i
< 4; i
++) {
2604 pipe_so_target_reference(&ice
->state
.so_target
[i
],
2605 i
< num_targets
? targets
[i
] : NULL
);
2608 /* No need to update 3DSTATE_SO_BUFFER unless SOL is active. */
2612 for (unsigned i
= 0; i
< 4; i
++,
2613 so_buffers
+= GENX(3DSTATE_SO_BUFFER_length
)) {
2615 if (i
>= num_targets
|| !targets
[i
]) {
2616 iris_pack_command(GENX(3DSTATE_SO_BUFFER
), so_buffers
, sob
)
2617 sob
.SOBufferIndex
= i
;
2621 struct iris_stream_output_target
*tgt
= (void *) targets
[i
];
2623 /* Note that offsets[i] will either be 0, causing us to zero
2624 * the value in the buffer, or 0xFFFFFFFF, which happens to mean
2625 * "continue appending at the existing offset."
2627 assert(offsets
[i
] == 0 || offsets
[i
] == 0xFFFFFFFF);
2629 uint32_t dynamic
[GENX(3DSTATE_SO_BUFFER_length
)];
2630 iris_pack_state(GENX(3DSTATE_SO_BUFFER
), dynamic
, dyns
) {
2631 dyns
.SOBufferIndex
= i
;
2632 dyns
.StreamOffset
= offsets
[i
];
2633 dyns
.StreamOutputBufferOffsetAddress
=
2634 rw_bo(NULL
, iris_resource_bo(tgt
->offset
.res
)->gtt_offset
+ tgt
->offset
.offset
+ i
* sizeof(uint32_t));
2637 for (uint32_t j
= 0; j
< GENX(3DSTATE_SO_BUFFER_length
); j
++) {
2638 so_buffers
[j
] = tgt
->so_buffer
[j
] | dynamic
[j
];
2642 ice
->state
.dirty
|= IRIS_DIRTY_SO_BUFFERS
;
2646 * An iris-vtable helper for encoding the 3DSTATE_SO_DECL_LIST and
2647 * 3DSTATE_STREAMOUT packets.
2649 * 3DSTATE_SO_DECL_LIST is a list of shader outputs we want the streamout
2650 * hardware to record. We can create it entirely based on the shader, with
2651 * no dynamic state dependencies.
2653 * 3DSTATE_STREAMOUT is an annoying mix of shader-based information and
2654 * state-based settings. We capture the shader-related ones here, and merge
2655 * the rest in at draw time.
2658 iris_create_so_decl_list(const struct pipe_stream_output_info
*info
,
2659 const struct brw_vue_map
*vue_map
)
2661 struct GENX(SO_DECL
) so_decl
[MAX_VERTEX_STREAMS
][128];
2662 int buffer_mask
[MAX_VERTEX_STREAMS
] = {0, 0, 0, 0};
2663 int next_offset
[MAX_VERTEX_STREAMS
] = {0, 0, 0, 0};
2664 int decls
[MAX_VERTEX_STREAMS
] = {0, 0, 0, 0};
2666 STATIC_ASSERT(ARRAY_SIZE(so_decl
[0]) >= MAX_PROGRAM_OUTPUTS
);
2668 memset(so_decl
, 0, sizeof(so_decl
));
2670 /* Construct the list of SO_DECLs to be emitted. The formatting of the
2671 * command feels strange -- each dword pair contains a SO_DECL per stream.
2673 for (unsigned i
= 0; i
< info
->num_outputs
; i
++) {
2674 const struct pipe_stream_output
*output
= &info
->output
[i
];
2675 const int buffer
= output
->output_buffer
;
2676 const int varying
= output
->register_index
;
2677 const unsigned stream_id
= output
->stream
;
2678 assert(stream_id
< MAX_VERTEX_STREAMS
);
2680 buffer_mask
[stream_id
] |= 1 << buffer
;
2682 assert(vue_map
->varying_to_slot
[varying
] >= 0);
2684 /* Mesa doesn't store entries for gl_SkipComponents in the Outputs[]
2685 * array. Instead, it simply increments DstOffset for the following
2686 * input by the number of components that should be skipped.
2688 * Our hardware is unusual in that it requires us to program SO_DECLs
2689 * for fake "hole" components, rather than simply taking the offset
2690 * for each real varying. Each hole can have size 1, 2, 3, or 4; we
2691 * program as many size = 4 holes as we can, then a final hole to
2692 * accommodate the final 1, 2, or 3 remaining.
2694 int skip_components
= output
->dst_offset
- next_offset
[buffer
];
2696 while (skip_components
> 0) {
2697 so_decl
[stream_id
][decls
[stream_id
]++] = (struct GENX(SO_DECL
)) {
2699 .OutputBufferSlot
= output
->output_buffer
,
2700 .ComponentMask
= (1 << MIN2(skip_components
, 4)) - 1,
2702 skip_components
-= 4;
2705 next_offset
[buffer
] = output
->dst_offset
+ output
->num_components
;
2707 so_decl
[stream_id
][decls
[stream_id
]++] = (struct GENX(SO_DECL
)) {
2708 .OutputBufferSlot
= output
->output_buffer
,
2709 .RegisterIndex
= vue_map
->varying_to_slot
[varying
],
2711 ((1 << output
->num_components
) - 1) << output
->start_component
,
2714 if (decls
[stream_id
] > max_decls
)
2715 max_decls
= decls
[stream_id
];
2718 unsigned dwords
= GENX(3DSTATE_STREAMOUT_length
) + (3 + 2 * max_decls
);
2719 uint32_t *map
= ralloc_size(NULL
, sizeof(uint32_t) * dwords
);
2720 uint32_t *so_decl_map
= map
+ GENX(3DSTATE_STREAMOUT_length
);
2722 iris_pack_command(GENX(3DSTATE_STREAMOUT
), map
, sol
) {
2723 int urb_entry_read_offset
= 0;
2724 int urb_entry_read_length
= (vue_map
->num_slots
+ 1) / 2 -
2725 urb_entry_read_offset
;
2727 /* We always read the whole vertex. This could be reduced at some
2728 * point by reading less and offsetting the register index in the
2731 sol
.Stream0VertexReadOffset
= urb_entry_read_offset
;
2732 sol
.Stream0VertexReadLength
= urb_entry_read_length
- 1;
2733 sol
.Stream1VertexReadOffset
= urb_entry_read_offset
;
2734 sol
.Stream1VertexReadLength
= urb_entry_read_length
- 1;
2735 sol
.Stream2VertexReadOffset
= urb_entry_read_offset
;
2736 sol
.Stream2VertexReadLength
= urb_entry_read_length
- 1;
2737 sol
.Stream3VertexReadOffset
= urb_entry_read_offset
;
2738 sol
.Stream3VertexReadLength
= urb_entry_read_length
- 1;
2740 /* Set buffer pitches; 0 means unbound. */
2741 sol
.Buffer0SurfacePitch
= 4 * info
->stride
[0];
2742 sol
.Buffer1SurfacePitch
= 4 * info
->stride
[1];
2743 sol
.Buffer2SurfacePitch
= 4 * info
->stride
[2];
2744 sol
.Buffer3SurfacePitch
= 4 * info
->stride
[3];
2747 iris_pack_command(GENX(3DSTATE_SO_DECL_LIST
), so_decl_map
, list
) {
2748 list
.DWordLength
= 3 + 2 * max_decls
- 2;
2749 list
.StreamtoBufferSelects0
= buffer_mask
[0];
2750 list
.StreamtoBufferSelects1
= buffer_mask
[1];
2751 list
.StreamtoBufferSelects2
= buffer_mask
[2];
2752 list
.StreamtoBufferSelects3
= buffer_mask
[3];
2753 list
.NumEntries0
= decls
[0];
2754 list
.NumEntries1
= decls
[1];
2755 list
.NumEntries2
= decls
[2];
2756 list
.NumEntries3
= decls
[3];
2759 for (int i
= 0; i
< max_decls
; i
++) {
2760 iris_pack_state(GENX(SO_DECL_ENTRY
), so_decl_map
+ 3 + i
* 2, entry
) {
2761 entry
.Stream0Decl
= so_decl
[0][i
];
2762 entry
.Stream1Decl
= so_decl
[1][i
];
2763 entry
.Stream2Decl
= so_decl
[2][i
];
2764 entry
.Stream3Decl
= so_decl
[3][i
];
2772 iris_compute_sbe_urb_read_interval(uint64_t fs_input_slots
,
2773 const struct brw_vue_map
*last_vue_map
,
2774 bool two_sided_color
,
2775 unsigned *out_offset
,
2776 unsigned *out_length
)
2778 /* The compiler computes the first URB slot without considering COL/BFC
2779 * swizzling (because it doesn't know whether it's enabled), so we need
2780 * to do that here too. This may result in a smaller offset, which
2783 const unsigned first_slot
=
2784 brw_compute_first_urb_slot_required(fs_input_slots
, last_vue_map
);
2786 /* This becomes the URB read offset (counted in pairs of slots). */
2787 assert(first_slot
% 2 == 0);
2788 *out_offset
= first_slot
/ 2;
2790 /* We need to adjust the inputs read to account for front/back color
2791 * swizzling, as it can make the URB length longer.
2793 for (int c
= 0; c
<= 1; c
++) {
2794 if (fs_input_slots
& (VARYING_BIT_COL0
<< c
)) {
2795 /* If two sided color is enabled, the fragment shader's gl_Color
2796 * (COL0) input comes from either the gl_FrontColor (COL0) or
2797 * gl_BackColor (BFC0) input varyings. Mark BFC as used, too.
2799 if (two_sided_color
)
2800 fs_input_slots
|= (VARYING_BIT_BFC0
<< c
);
2802 /* If front color isn't written, we opt to give them back color
2803 * instead of an undefined value. Switch from COL to BFC.
2805 if (last_vue_map
->varying_to_slot
[VARYING_SLOT_COL0
+ c
] == -1) {
2806 fs_input_slots
&= ~(VARYING_BIT_COL0
<< c
);
2807 fs_input_slots
|= (VARYING_BIT_BFC0
<< c
);
2812 /* Compute the minimum URB Read Length necessary for the FS inputs.
2814 * From the Sandy Bridge PRM, Volume 2, Part 1, documentation for
2815 * 3DSTATE_SF DWord 1 bits 15:11, "Vertex URB Entry Read Length":
2817 * "This field should be set to the minimum length required to read the
2818 * maximum source attribute. The maximum source attribute is indicated
2819 * by the maximum value of the enabled Attribute # Source Attribute if
2820 * Attribute Swizzle Enable is set, Number of Output Attributes-1 if
2821 * enable is not set.
2822 * read_length = ceiling((max_source_attr + 1) / 2)
2824 * [errata] Corruption/Hang possible if length programmed larger than
2827 * Similar text exists for Ivy Bridge.
2829 * We find the last URB slot that's actually read by the FS.
2831 unsigned last_read_slot
= last_vue_map
->num_slots
- 1;
2832 while (last_read_slot
> first_slot
&& !(fs_input_slots
&
2833 (1ull << last_vue_map
->slot_to_varying
[last_read_slot
])))
2836 /* The URB read length is the difference of the two, counted in pairs. */
2837 *out_length
= DIV_ROUND_UP(last_read_slot
- first_slot
+ 1, 2);
2841 iris_emit_sbe_swiz(struct iris_batch
*batch
,
2842 const struct iris_context
*ice
,
2843 unsigned urb_read_offset
,
2844 unsigned sprite_coord_enables
)
2846 struct GENX(SF_OUTPUT_ATTRIBUTE_DETAIL
) attr_overrides
[16] = {};
2847 const struct brw_wm_prog_data
*wm_prog_data
= (void *)
2848 ice
->shaders
.prog
[MESA_SHADER_FRAGMENT
]->prog_data
;
2849 const struct brw_vue_map
*vue_map
= ice
->shaders
.last_vue_map
;
2850 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
2852 /* XXX: this should be generated when putting programs in place */
2854 // XXX: raster->sprite_coord_enable
2856 for (int fs_attr
= 0; fs_attr
< VARYING_SLOT_MAX
; fs_attr
++) {
2857 const int input_index
= wm_prog_data
->urb_setup
[fs_attr
];
2858 if (input_index
< 0 || input_index
>= 16)
2861 struct GENX(SF_OUTPUT_ATTRIBUTE_DETAIL
) *attr
=
2862 &attr_overrides
[input_index
];
2863 int slot
= vue_map
->varying_to_slot
[fs_attr
];
2865 /* Viewport and Layer are stored in the VUE header. We need to override
2866 * them to zero if earlier stages didn't write them, as GL requires that
2867 * they read back as zero when not explicitly set.
2870 case VARYING_SLOT_VIEWPORT
:
2871 case VARYING_SLOT_LAYER
:
2872 attr
->ComponentOverrideX
= true;
2873 attr
->ComponentOverrideW
= true;
2874 attr
->ConstantSource
= CONST_0000
;
2876 if (!(vue_map
->slots_valid
& VARYING_BIT_LAYER
))
2877 attr
->ComponentOverrideY
= true;
2878 if (!(vue_map
->slots_valid
& VARYING_BIT_VIEWPORT
))
2879 attr
->ComponentOverrideZ
= true;
2882 case VARYING_SLOT_PRIMITIVE_ID
:
2883 /* Override if the previous shader stage didn't write gl_PrimitiveID. */
2885 attr
->ComponentOverrideX
= true;
2886 attr
->ComponentOverrideY
= true;
2887 attr
->ComponentOverrideZ
= true;
2888 attr
->ComponentOverrideW
= true;
2889 attr
->ConstantSource
= PRIM_ID
;
2897 if (sprite_coord_enables
& (1 << input_index
))
2900 /* If there was only a back color written but not front, use back
2901 * as the color instead of undefined.
2903 if (slot
== -1 && fs_attr
== VARYING_SLOT_COL0
)
2904 slot
= vue_map
->varying_to_slot
[VARYING_SLOT_BFC0
];
2905 if (slot
== -1 && fs_attr
== VARYING_SLOT_COL1
)
2906 slot
= vue_map
->varying_to_slot
[VARYING_SLOT_BFC1
];
2908 /* Not written by the previous stage - undefined. */
2910 attr
->ComponentOverrideX
= true;
2911 attr
->ComponentOverrideY
= true;
2912 attr
->ComponentOverrideZ
= true;
2913 attr
->ComponentOverrideW
= true;
2914 attr
->ConstantSource
= CONST_0001_FLOAT
;
2918 /* Compute the location of the attribute relative to the read offset,
2919 * which is counted in 256-bit increments (two 128-bit VUE slots).
2921 const int source_attr
= slot
- 2 * urb_read_offset
;
2922 assert(source_attr
>= 0 && source_attr
<= 32);
2923 attr
->SourceAttribute
= source_attr
;
2925 /* If we are doing two-sided color, and the VUE slot following this one
2926 * represents a back-facing color, then we need to instruct the SF unit
2927 * to do back-facing swizzling.
2929 if (cso_rast
->light_twoside
&&
2930 ((vue_map
->slot_to_varying
[slot
] == VARYING_SLOT_COL0
&&
2931 vue_map
->slot_to_varying
[slot
+1] == VARYING_SLOT_BFC0
) ||
2932 (vue_map
->slot_to_varying
[slot
] == VARYING_SLOT_COL1
&&
2933 vue_map
->slot_to_varying
[slot
+1] == VARYING_SLOT_BFC1
)))
2934 attr
->SwizzleSelect
= INPUTATTR_FACING
;
2937 iris_emit_cmd(batch
, GENX(3DSTATE_SBE_SWIZ
), sbes
) {
2938 for (int i
= 0; i
< 16; i
++)
2939 sbes
.Attribute
[i
] = attr_overrides
[i
];
2944 iris_calculate_point_sprite_overrides(const struct brw_wm_prog_data
*prog_data
,
2945 const struct iris_rasterizer_state
*cso
)
2947 unsigned overrides
= 0;
2949 if (prog_data
->urb_setup
[VARYING_SLOT_PNTC
] != -1)
2950 overrides
|= 1 << prog_data
->urb_setup
[VARYING_SLOT_PNTC
];
2952 for (int i
= 0; i
< 8; i
++) {
2953 if ((cso
->sprite_coord_enable
& (1 << i
)) &&
2954 prog_data
->urb_setup
[VARYING_SLOT_TEX0
+ i
] != -1)
2955 overrides
|= 1 << prog_data
->urb_setup
[VARYING_SLOT_TEX0
+ i
];
2962 iris_emit_sbe(struct iris_batch
*batch
, const struct iris_context
*ice
)
2964 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
2965 const struct brw_wm_prog_data
*wm_prog_data
= (void *)
2966 ice
->shaders
.prog
[MESA_SHADER_FRAGMENT
]->prog_data
;
2967 const struct shader_info
*fs_info
=
2968 iris_get_shader_info(ice
, MESA_SHADER_FRAGMENT
);
2970 unsigned urb_read_offset
, urb_read_length
;
2971 iris_compute_sbe_urb_read_interval(fs_info
->inputs_read
,
2972 ice
->shaders
.last_vue_map
,
2973 cso_rast
->light_twoside
,
2974 &urb_read_offset
, &urb_read_length
);
2976 unsigned sprite_coord_overrides
=
2977 iris_calculate_point_sprite_overrides(wm_prog_data
, cso_rast
);
2979 iris_emit_cmd(batch
, GENX(3DSTATE_SBE
), sbe
) {
2980 sbe
.AttributeSwizzleEnable
= true;
2981 sbe
.NumberofSFOutputAttributes
= wm_prog_data
->num_varying_inputs
;
2982 sbe
.PointSpriteTextureCoordinateOrigin
= cso_rast
->sprite_coord_mode
;
2983 sbe
.VertexURBEntryReadOffset
= urb_read_offset
;
2984 sbe
.VertexURBEntryReadLength
= urb_read_length
;
2985 sbe
.ForceVertexURBEntryReadOffset
= true;
2986 sbe
.ForceVertexURBEntryReadLength
= true;
2987 sbe
.ConstantInterpolationEnable
= wm_prog_data
->flat_inputs
;
2988 sbe
.PointSpriteTextureCoordinateEnable
= sprite_coord_overrides
;
2990 for (int i
= 0; i
< 32; i
++) {
2991 sbe
.AttributeActiveComponentFormat
[i
] = ACTIVE_COMPONENT_XYZW
;
2995 iris_emit_sbe_swiz(batch
, ice
, urb_read_offset
, sprite_coord_overrides
);
2998 /* ------------------------------------------------------------------- */
3001 * Set sampler-related program key fields based on the current state.
3004 iris_populate_sampler_key(const struct iris_context
*ice
,
3005 struct brw_sampler_prog_key_data
*key
)
3007 for (int i
= 0; i
< MAX_SAMPLERS
; i
++) {
3008 key
->swizzles
[i
] = 0x688; /* XYZW */
3013 * Populate VS program key fields based on the current state.
3016 iris_populate_vs_key(const struct iris_context
*ice
,
3017 const struct shader_info
*info
,
3018 struct brw_vs_prog_key
*key
)
3020 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
3022 iris_populate_sampler_key(ice
, &key
->tex
);
3024 if (info
->clip_distance_array_size
== 0 &&
3025 (info
->outputs_written
& (VARYING_BIT_POS
| VARYING_BIT_CLIP_VERTEX
)))
3026 key
->nr_userclip_plane_consts
= cso_rast
->num_clip_plane_consts
;
3030 * Populate TCS program key fields based on the current state.
3033 iris_populate_tcs_key(const struct iris_context
*ice
,
3034 struct brw_tcs_prog_key
*key
)
3036 iris_populate_sampler_key(ice
, &key
->tex
);
3040 * Populate TES program key fields based on the current state.
3043 iris_populate_tes_key(const struct iris_context
*ice
,
3044 struct brw_tes_prog_key
*key
)
3046 iris_populate_sampler_key(ice
, &key
->tex
);
3050 * Populate GS program key fields based on the current state.
3053 iris_populate_gs_key(const struct iris_context
*ice
,
3054 struct brw_gs_prog_key
*key
)
3056 iris_populate_sampler_key(ice
, &key
->tex
);
3060 * Populate FS program key fields based on the current state.
3063 iris_populate_fs_key(const struct iris_context
*ice
,
3064 struct brw_wm_prog_key
*key
)
3066 iris_populate_sampler_key(ice
, &key
->tex
);
3068 /* XXX: dirty flags? */
3069 const struct pipe_framebuffer_state
*fb
= &ice
->state
.framebuffer
;
3070 const struct iris_depth_stencil_alpha_state
*zsa
= ice
->state
.cso_zsa
;
3071 const struct iris_rasterizer_state
*rast
= ice
->state
.cso_rast
;
3072 const struct iris_blend_state
*blend
= ice
->state
.cso_blend
;
3074 key
->nr_color_regions
= fb
->nr_cbufs
;
3076 key
->clamp_fragment_color
= rast
->clamp_fragment_color
;
3078 key
->replicate_alpha
= fb
->nr_cbufs
> 1 &&
3079 (zsa
->alpha
.enabled
|| blend
->alpha_to_coverage
);
3081 /* XXX: only bother if COL0/1 are read */
3082 key
->flat_shade
= rast
->flatshade
;
3084 key
->persample_interp
= rast
->force_persample_interp
;
3085 key
->multisample_fbo
= rast
->multisample
&& fb
->samples
> 1;
3087 key
->coherent_fb_fetch
= true;
3089 // XXX: uint64_t input_slots_valid; - for >16 inputs
3091 // XXX: key->force_dual_color_blend for unigine
3092 // XXX: respect hint for high_quality_derivatives:1;
3096 iris_populate_cs_key(const struct iris_context
*ice
,
3097 struct brw_cs_prog_key
*key
)
3099 iris_populate_sampler_key(ice
, &key
->tex
);
3103 // XXX: these need to go in INIT_THREAD_DISPATCH_FIELDS
3104 pkt
.SamplerCount
= \
3105 DIV_ROUND_UP(CLAMP(stage_state
->sampler_count
, 0, 16), 4); \
3110 KSP(const struct iris_compiled_shader
*shader
)
3112 struct iris_resource
*res
= (void *) shader
->assembly
.res
;
3113 return iris_bo_offset_from_base_address(res
->bo
) + shader
->assembly
.offset
;
3116 // Gen11 workaround table #2056 WABTPPrefetchDisable suggests to disable
3117 // prefetching of binding tables in A0 and B0 steppings. XXX: Revisit
3118 // this WA on C0 stepping.
3120 #define INIT_THREAD_DISPATCH_FIELDS(pkt, prefix, stage) \
3121 pkt.KernelStartPointer = KSP(shader); \
3122 pkt.BindingTableEntryCount = GEN_GEN == 11 ? 0 : \
3123 prog_data->binding_table.size_bytes / 4; \
3124 pkt.FloatingPointMode = prog_data->use_alt_mode; \
3126 pkt.DispatchGRFStartRegisterForURBData = \
3127 prog_data->dispatch_grf_start_reg; \
3128 pkt.prefix##URBEntryReadLength = vue_prog_data->urb_read_length; \
3129 pkt.prefix##URBEntryReadOffset = 0; \
3131 pkt.StatisticsEnable = true; \
3132 pkt.Enable = true; \
3134 if (prog_data->total_scratch) { \
3135 uint32_t scratch_addr = \
3136 iris_get_scratch_space(ice, prog_data->total_scratch, stage); \
3137 pkt.PerThreadScratchSpace = ffs(prog_data->total_scratch) - 11; \
3138 pkt.ScratchSpaceBasePointer = rw_bo(NULL, scratch_addr); \
3142 * Encode most of 3DSTATE_VS based on the compiled shader.
3145 iris_store_vs_state(struct iris_context
*ice
,
3146 const struct gen_device_info
*devinfo
,
3147 struct iris_compiled_shader
*shader
)
3149 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3150 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3152 iris_pack_command(GENX(3DSTATE_VS
), shader
->derived_data
, vs
) {
3153 INIT_THREAD_DISPATCH_FIELDS(vs
, Vertex
, MESA_SHADER_VERTEX
);
3154 vs
.MaximumNumberofThreads
= devinfo
->max_vs_threads
- 1;
3155 vs
.SIMD8DispatchEnable
= true;
3156 vs
.UserClipDistanceCullTestEnableBitmask
=
3157 vue_prog_data
->cull_distance_mask
;
3162 * Encode most of 3DSTATE_HS based on the compiled shader.
3165 iris_store_tcs_state(struct iris_context
*ice
,
3166 const struct gen_device_info
*devinfo
,
3167 struct iris_compiled_shader
*shader
)
3169 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3170 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3171 struct brw_tcs_prog_data
*tcs_prog_data
= (void *) prog_data
;
3173 iris_pack_command(GENX(3DSTATE_HS
), shader
->derived_data
, hs
) {
3174 INIT_THREAD_DISPATCH_FIELDS(hs
, Vertex
, MESA_SHADER_TESS_CTRL
);
3176 hs
.InstanceCount
= tcs_prog_data
->instances
- 1;
3177 hs
.MaximumNumberofThreads
= devinfo
->max_tcs_threads
- 1;
3178 hs
.IncludeVertexHandles
= true;
3183 * Encode 3DSTATE_TE and most of 3DSTATE_DS based on the compiled shader.
3186 iris_store_tes_state(struct iris_context
*ice
,
3187 const struct gen_device_info
*devinfo
,
3188 struct iris_compiled_shader
*shader
)
3190 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3191 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3192 struct brw_tes_prog_data
*tes_prog_data
= (void *) prog_data
;
3194 uint32_t *te_state
= (void *) shader
->derived_data
;
3195 uint32_t *ds_state
= te_state
+ GENX(3DSTATE_TE_length
);
3197 iris_pack_command(GENX(3DSTATE_TE
), te_state
, te
) {
3198 te
.Partitioning
= tes_prog_data
->partitioning
;
3199 te
.OutputTopology
= tes_prog_data
->output_topology
;
3200 te
.TEDomain
= tes_prog_data
->domain
;
3202 te
.MaximumTessellationFactorOdd
= 63.0;
3203 te
.MaximumTessellationFactorNotOdd
= 64.0;
3206 iris_pack_command(GENX(3DSTATE_DS
), ds_state
, ds
) {
3207 INIT_THREAD_DISPATCH_FIELDS(ds
, Patch
, MESA_SHADER_TESS_EVAL
);
3209 ds
.DispatchMode
= DISPATCH_MODE_SIMD8_SINGLE_PATCH
;
3210 ds
.MaximumNumberofThreads
= devinfo
->max_tes_threads
- 1;
3211 ds
.ComputeWCoordinateEnable
=
3212 tes_prog_data
->domain
== BRW_TESS_DOMAIN_TRI
;
3214 ds
.UserClipDistanceCullTestEnableBitmask
=
3215 vue_prog_data
->cull_distance_mask
;
3221 * Encode most of 3DSTATE_GS based on the compiled shader.
3224 iris_store_gs_state(struct iris_context
*ice
,
3225 const struct gen_device_info
*devinfo
,
3226 struct iris_compiled_shader
*shader
)
3228 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3229 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
3230 struct brw_gs_prog_data
*gs_prog_data
= (void *) prog_data
;
3232 iris_pack_command(GENX(3DSTATE_GS
), shader
->derived_data
, gs
) {
3233 INIT_THREAD_DISPATCH_FIELDS(gs
, Vertex
, MESA_SHADER_GEOMETRY
);
3235 gs
.OutputVertexSize
= gs_prog_data
->output_vertex_size_hwords
* 2 - 1;
3236 gs
.OutputTopology
= gs_prog_data
->output_topology
;
3237 gs
.ControlDataHeaderSize
=
3238 gs_prog_data
->control_data_header_size_hwords
;
3239 gs
.InstanceControl
= gs_prog_data
->invocations
- 1;
3240 gs
.DispatchMode
= DISPATCH_MODE_SIMD8
;
3241 gs
.IncludePrimitiveID
= gs_prog_data
->include_primitive_id
;
3242 gs
.ControlDataFormat
= gs_prog_data
->control_data_format
;
3243 gs
.ReorderMode
= TRAILING
;
3244 gs
.ExpectedVertexCount
= gs_prog_data
->vertices_in
;
3245 gs
.MaximumNumberofThreads
=
3246 GEN_GEN
== 8 ? (devinfo
->max_gs_threads
/ 2 - 1)
3247 : (devinfo
->max_gs_threads
- 1);
3249 if (gs_prog_data
->static_vertex_count
!= -1) {
3250 gs
.StaticOutput
= true;
3251 gs
.StaticOutputVertexCount
= gs_prog_data
->static_vertex_count
;
3253 gs
.IncludeVertexHandles
= vue_prog_data
->include_vue_handles
;
3255 gs
.UserClipDistanceCullTestEnableBitmask
=
3256 vue_prog_data
->cull_distance_mask
;
3258 const int urb_entry_write_offset
= 1;
3259 const uint32_t urb_entry_output_length
=
3260 DIV_ROUND_UP(vue_prog_data
->vue_map
.num_slots
, 2) -
3261 urb_entry_write_offset
;
3263 gs
.VertexURBEntryOutputReadOffset
= urb_entry_write_offset
;
3264 gs
.VertexURBEntryOutputLength
= MAX2(urb_entry_output_length
, 1);
3269 * Encode most of 3DSTATE_PS and 3DSTATE_PS_EXTRA based on the shader.
3272 iris_store_fs_state(struct iris_context
*ice
,
3273 const struct gen_device_info
*devinfo
,
3274 struct iris_compiled_shader
*shader
)
3276 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3277 struct brw_wm_prog_data
*wm_prog_data
= (void *) shader
->prog_data
;
3279 uint32_t *ps_state
= (void *) shader
->derived_data
;
3280 uint32_t *psx_state
= ps_state
+ GENX(3DSTATE_PS_length
);
3282 iris_pack_command(GENX(3DSTATE_PS
), ps_state
, ps
) {
3283 ps
.VectorMaskEnable
= true;
3284 //ps.SamplerCount = ...
3285 // XXX: WABTPPrefetchDisable, see above, drop at C0
3286 ps
.BindingTableEntryCount
= GEN_GEN
== 11 ? 0 :
3287 prog_data
->binding_table
.size_bytes
/ 4;
3288 ps
.FloatingPointMode
= prog_data
->use_alt_mode
;
3289 ps
.MaximumNumberofThreadsPerPSD
= 64 - (GEN_GEN
== 8 ? 2 : 1);
3291 ps
.PushConstantEnable
= shader
->num_system_values
> 0 ||
3292 prog_data
->ubo_ranges
[0].length
> 0;
3294 /* From the documentation for this packet:
3295 * "If the PS kernel does not need the Position XY Offsets to
3296 * compute a Position Value, then this field should be programmed
3297 * to POSOFFSET_NONE."
3299 * "SW Recommendation: If the PS kernel needs the Position Offsets
3300 * to compute a Position XY value, this field should match Position
3301 * ZW Interpolation Mode to ensure a consistent position.xyzw
3304 * We only require XY sample offsets. So, this recommendation doesn't
3305 * look useful at the moment. We might need this in future.
3307 ps
.PositionXYOffsetSelect
=
3308 wm_prog_data
->uses_pos_offset
? POSOFFSET_SAMPLE
: POSOFFSET_NONE
;
3309 ps
._8PixelDispatchEnable
= wm_prog_data
->dispatch_8
;
3310 ps
._16PixelDispatchEnable
= wm_prog_data
->dispatch_16
;
3311 ps
._32PixelDispatchEnable
= wm_prog_data
->dispatch_32
;
3313 // XXX: Disable SIMD32 with 16x MSAA
3315 ps
.DispatchGRFStartRegisterForConstantSetupData0
=
3316 brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data
, ps
, 0);
3317 ps
.DispatchGRFStartRegisterForConstantSetupData1
=
3318 brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data
, ps
, 1);
3319 ps
.DispatchGRFStartRegisterForConstantSetupData2
=
3320 brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data
, ps
, 2);
3322 ps
.KernelStartPointer0
=
3323 KSP(shader
) + brw_wm_prog_data_prog_offset(wm_prog_data
, ps
, 0);
3324 ps
.KernelStartPointer1
=
3325 KSP(shader
) + brw_wm_prog_data_prog_offset(wm_prog_data
, ps
, 1);
3326 ps
.KernelStartPointer2
=
3327 KSP(shader
) + brw_wm_prog_data_prog_offset(wm_prog_data
, ps
, 2);
3329 if (prog_data
->total_scratch
) {
3330 uint32_t scratch_addr
=
3331 iris_get_scratch_space(ice
, prog_data
->total_scratch
,
3332 MESA_SHADER_FRAGMENT
);
3333 ps
.PerThreadScratchSpace
= ffs(prog_data
->total_scratch
) - 11;
3334 ps
.ScratchSpaceBasePointer
= rw_bo(NULL
, scratch_addr
);
3338 iris_pack_command(GENX(3DSTATE_PS_EXTRA
), psx_state
, psx
) {
3339 psx
.PixelShaderValid
= true;
3340 psx
.PixelShaderComputedDepthMode
= wm_prog_data
->computed_depth_mode
;
3341 psx
.PixelShaderKillsPixel
= wm_prog_data
->uses_kill
;
3342 psx
.AttributeEnable
= wm_prog_data
->num_varying_inputs
!= 0;
3343 psx
.PixelShaderUsesSourceDepth
= wm_prog_data
->uses_src_depth
;
3344 psx
.PixelShaderUsesSourceW
= wm_prog_data
->uses_src_w
;
3345 psx
.PixelShaderIsPerSample
= wm_prog_data
->persample_dispatch
;
3347 if (wm_prog_data
->uses_sample_mask
) {
3348 /* TODO: conservative rasterization */
3349 if (wm_prog_data
->post_depth_coverage
)
3350 psx
.InputCoverageMaskState
= ICMS_DEPTH_COVERAGE
;
3352 psx
.InputCoverageMaskState
= ICMS_NORMAL
;
3355 psx
.oMaskPresenttoRenderTarget
= wm_prog_data
->uses_omask
;
3356 psx
.PixelShaderPullsBary
= wm_prog_data
->pulls_bary
;
3357 psx
.PixelShaderComputesStencil
= wm_prog_data
->computed_stencil
;
3364 * Compute the size of the derived data (shader command packets).
3366 * This must match the data written by the iris_store_xs_state() functions.
3369 iris_store_cs_state(struct iris_context
*ice
,
3370 const struct gen_device_info
*devinfo
,
3371 struct iris_compiled_shader
*shader
)
3373 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
3374 struct brw_cs_prog_data
*cs_prog_data
= (void *) shader
->prog_data
;
3375 void *map
= shader
->derived_data
;
3377 iris_pack_state(GENX(INTERFACE_DESCRIPTOR_DATA
), map
, desc
) {
3378 desc
.KernelStartPointer
= KSP(shader
);
3379 desc
.ConstantURBEntryReadLength
= cs_prog_data
->push
.per_thread
.regs
;
3380 desc
.NumberofThreadsinGPGPUThreadGroup
= cs_prog_data
->threads
;
3381 desc
.SharedLocalMemorySize
=
3382 encode_slm_size(GEN_GEN
, prog_data
->total_shared
);
3383 desc
.BarrierEnable
= cs_prog_data
->uses_barrier
;
3384 desc
.CrossThreadConstantDataReadLength
=
3385 cs_prog_data
->push
.cross_thread
.regs
;
3390 iris_derived_program_state_size(enum iris_program_cache_id cache_id
)
3392 assert(cache_id
<= IRIS_CACHE_BLORP
);
3394 static const unsigned dwords
[] = {
3395 [IRIS_CACHE_VS
] = GENX(3DSTATE_VS_length
),
3396 [IRIS_CACHE_TCS
] = GENX(3DSTATE_HS_length
),
3397 [IRIS_CACHE_TES
] = GENX(3DSTATE_TE_length
) + GENX(3DSTATE_DS_length
),
3398 [IRIS_CACHE_GS
] = GENX(3DSTATE_GS_length
),
3400 GENX(3DSTATE_PS_length
) + GENX(3DSTATE_PS_EXTRA_length
),
3401 [IRIS_CACHE_CS
] = GENX(INTERFACE_DESCRIPTOR_DATA_length
),
3402 [IRIS_CACHE_BLORP
] = 0,
3405 return sizeof(uint32_t) * dwords
[cache_id
];
3409 * Create any state packets corresponding to the given shader stage
3410 * (i.e. 3DSTATE_VS) and save them as "derived data" in the shader variant.
3411 * This means that we can look up a program in the in-memory cache and
3412 * get most of the state packet without having to reconstruct it.
3415 iris_store_derived_program_state(struct iris_context
*ice
,
3416 enum iris_program_cache_id cache_id
,
3417 struct iris_compiled_shader
*shader
)
3419 struct iris_screen
*screen
= (void *) ice
->ctx
.screen
;
3420 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
3424 iris_store_vs_state(ice
, devinfo
, shader
);
3426 case IRIS_CACHE_TCS
:
3427 iris_store_tcs_state(ice
, devinfo
, shader
);
3429 case IRIS_CACHE_TES
:
3430 iris_store_tes_state(ice
, devinfo
, shader
);
3433 iris_store_gs_state(ice
, devinfo
, shader
);
3436 iris_store_fs_state(ice
, devinfo
, shader
);
3439 iris_store_cs_state(ice
, devinfo
, shader
);
3440 case IRIS_CACHE_BLORP
:
3447 /* ------------------------------------------------------------------- */
3450 * Configure the URB.
3452 * XXX: write a real comment.
3455 iris_upload_urb_config(struct iris_context
*ice
, struct iris_batch
*batch
)
3457 const struct gen_device_info
*devinfo
= &batch
->screen
->devinfo
;
3458 const unsigned push_size_kB
= 32;
3459 unsigned entries
[4];
3463 for (int i
= MESA_SHADER_VERTEX
; i
<= MESA_SHADER_GEOMETRY
; i
++) {
3464 if (!ice
->shaders
.prog
[i
]) {
3467 struct brw_vue_prog_data
*vue_prog_data
=
3468 (void *) ice
->shaders
.prog
[i
]->prog_data
;
3469 size
[i
] = vue_prog_data
->urb_entry_size
;
3471 assert(size
[i
] != 0);
3474 gen_get_urb_config(devinfo
, 1024 * push_size_kB
,
3475 1024 * ice
->shaders
.urb_size
,
3476 ice
->shaders
.prog
[MESA_SHADER_TESS_EVAL
] != NULL
,
3477 ice
->shaders
.prog
[MESA_SHADER_GEOMETRY
] != NULL
,
3478 size
, entries
, start
);
3480 for (int i
= MESA_SHADER_VERTEX
; i
<= MESA_SHADER_GEOMETRY
; i
++) {
3481 iris_emit_cmd(batch
, GENX(3DSTATE_URB_VS
), urb
) {
3482 urb
._3DCommandSubOpcode
+= i
;
3483 urb
.VSURBStartingAddress
= start
[i
];
3484 urb
.VSURBEntryAllocationSize
= size
[i
] - 1;
3485 urb
.VSNumberofURBEntries
= entries
[i
];
3490 static const uint32_t push_constant_opcodes
[] = {
3491 [MESA_SHADER_VERTEX
] = 21,
3492 [MESA_SHADER_TESS_CTRL
] = 25, /* HS */
3493 [MESA_SHADER_TESS_EVAL
] = 26, /* DS */
3494 [MESA_SHADER_GEOMETRY
] = 22,
3495 [MESA_SHADER_FRAGMENT
] = 23,
3496 [MESA_SHADER_COMPUTE
] = 0,
3500 use_null_surface(struct iris_batch
*batch
, struct iris_context
*ice
)
3502 struct iris_bo
*state_bo
= iris_resource_bo(ice
->state
.unbound_tex
.res
);
3504 iris_use_pinned_bo(batch
, state_bo
, false);
3506 return ice
->state
.unbound_tex
.offset
;
3510 use_null_fb_surface(struct iris_batch
*batch
, struct iris_context
*ice
)
3512 /* If set_framebuffer_state() was never called, fall back to 1x1x1 */
3513 if (!ice
->state
.null_fb
.res
)
3514 return use_null_surface(batch
, ice
);
3516 struct iris_bo
*state_bo
= iris_resource_bo(ice
->state
.null_fb
.res
);
3518 iris_use_pinned_bo(batch
, state_bo
, false);
3520 return ice
->state
.null_fb
.offset
;
3524 * Add a surface to the validation list, as well as the buffer containing
3525 * the corresponding SURFACE_STATE.
3527 * Returns the binding table entry (offset to SURFACE_STATE).
3530 use_surface(struct iris_batch
*batch
,
3531 struct pipe_surface
*p_surf
,
3534 struct iris_surface
*surf
= (void *) p_surf
;
3536 iris_use_pinned_bo(batch
, iris_resource_bo(p_surf
->texture
), writeable
);
3537 iris_use_pinned_bo(batch
, iris_resource_bo(surf
->surface_state
.res
), false);
3539 return surf
->surface_state
.offset
;
3543 use_sampler_view(struct iris_batch
*batch
, struct iris_sampler_view
*isv
)
3545 iris_use_pinned_bo(batch
, isv
->res
->bo
, false);
3546 iris_use_pinned_bo(batch
, iris_resource_bo(isv
->surface_state
.res
), false);
3548 return isv
->surface_state
.offset
;
3552 use_const_buffer(struct iris_batch
*batch
,
3553 struct iris_context
*ice
,
3554 struct iris_const_buffer
*cbuf
)
3556 if (!cbuf
->surface_state
.res
)
3557 return use_null_surface(batch
, ice
);
3559 iris_use_pinned_bo(batch
, iris_resource_bo(cbuf
->data
.res
), false);
3560 iris_use_pinned_bo(batch
, iris_resource_bo(cbuf
->surface_state
.res
), false);
3562 return cbuf
->surface_state
.offset
;
3566 use_ssbo(struct iris_batch
*batch
, struct iris_context
*ice
,
3567 struct iris_shader_state
*shs
, int i
)
3570 return use_null_surface(batch
, ice
);
3572 struct iris_state_ref
*surf_state
= &shs
->ssbo_surface_state
[i
];
3574 iris_use_pinned_bo(batch
, iris_resource_bo(shs
->ssbo
[i
]), true);
3575 iris_use_pinned_bo(batch
, iris_resource_bo(surf_state
->res
), false);
3577 return surf_state
->offset
;
3581 use_image(struct iris_batch
*batch
, struct iris_context
*ice
,
3582 struct iris_shader_state
*shs
, int i
)
3584 if (!shs
->image
[i
].res
)
3585 return use_null_surface(batch
, ice
);
3587 struct iris_state_ref
*surf_state
= &shs
->image
[i
].surface_state
;
3589 iris_use_pinned_bo(batch
, iris_resource_bo(shs
->image
[i
].res
),
3590 shs
->image
[i
].access
& PIPE_IMAGE_ACCESS_WRITE
);
3591 iris_use_pinned_bo(batch
, iris_resource_bo(surf_state
->res
), false);
3593 return surf_state
->offset
;
3596 #define push_bt_entry(addr) \
3597 assert(addr >= binder_addr); \
3598 if (!pin_only) bt_map[s++] = (addr) - binder_addr;
3601 * Populate the binding table for a given shader stage.
3603 * This fills out the table of pointers to surfaces required by the shader,
3604 * and also adds those buffers to the validation list so the kernel can make
3605 * resident before running our batch.
3608 iris_populate_binding_table(struct iris_context
*ice
,
3609 struct iris_batch
*batch
,
3610 gl_shader_stage stage
,
3613 const struct iris_binder
*binder
= &ice
->state
.binder
;
3614 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3618 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
3619 uint32_t binder_addr
= binder
->bo
->gtt_offset
;
3621 //struct brw_stage_prog_data *prog_data = (void *) shader->prog_data;
3622 uint32_t *bt_map
= binder
->map
+ binder
->bt_offset
[stage
];
3625 const struct shader_info
*info
= iris_get_shader_info(ice
, stage
);
3627 /* TCS passthrough doesn't need a binding table. */
3628 assert(stage
== MESA_SHADER_TESS_CTRL
);
3632 if (stage
== MESA_SHADER_COMPUTE
) {
3633 /* surface for gl_NumWorkGroups */
3634 struct iris_state_ref
*grid_data
= &ice
->state
.grid_size
;
3635 struct iris_state_ref
*grid_state
= &ice
->state
.grid_surf_state
;
3636 iris_use_pinned_bo(batch
, iris_resource_bo(grid_data
->res
), false);
3637 iris_use_pinned_bo(batch
, iris_resource_bo(grid_state
->res
), false);
3638 push_bt_entry(grid_state
->offset
);
3641 if (stage
== MESA_SHADER_FRAGMENT
) {
3642 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
3643 /* Note that cso_fb->nr_cbufs == fs_key->nr_color_regions. */
3644 if (cso_fb
->nr_cbufs
) {
3645 for (unsigned i
= 0; i
< cso_fb
->nr_cbufs
; i
++) {
3647 cso_fb
->cbufs
[i
] ? use_surface(batch
, cso_fb
->cbufs
[i
], true)
3648 : use_null_fb_surface(batch
, ice
);
3649 push_bt_entry(addr
);
3652 uint32_t addr
= use_null_fb_surface(batch
, ice
);
3653 push_bt_entry(addr
);
3657 //assert(prog_data->binding_table.texture_start ==
3658 //(ice->state.num_textures[stage] ? s : 0xd0d0d0d0));
3660 for (int i
= 0; i
< shs
->num_textures
; i
++) {
3661 struct iris_sampler_view
*view
= shs
->textures
[i
];
3662 uint32_t addr
= view
? use_sampler_view(batch
, view
)
3663 : use_null_surface(batch
, ice
);
3664 push_bt_entry(addr
);
3667 for (int i
= 0; i
< info
->num_images
; i
++) {
3668 uint32_t addr
= use_image(batch
, ice
, shs
, i
);
3669 push_bt_entry(addr
);
3672 const int num_ubos
= iris_get_shader_num_ubos(ice
, stage
);
3674 for (int i
= 0; i
< num_ubos
; i
++) {
3675 uint32_t addr
= use_const_buffer(batch
, ice
, &shs
->constbuf
[i
]);
3676 push_bt_entry(addr
);
3679 /* XXX: st is wasting 16 binding table slots for ABOs. Should add a cap
3680 * for changing nir_lower_atomics_to_ssbos setting and buffer_base offset
3681 * in st_atom_storagebuf.c so it'll compact them into one range, with
3682 * SSBOs starting at info->num_abos. Ideally it'd reset num_abos to 0 too
3684 if (info
->num_abos
+ info
->num_ssbos
> 0) {
3685 for (int i
= 0; i
< IRIS_MAX_ABOS
+ info
->num_ssbos
; i
++) {
3686 uint32_t addr
= use_ssbo(batch
, ice
, shs
, i
);
3687 push_bt_entry(addr
);
3692 // XXX: not implemented yet
3693 assert(prog_data
->binding_table
.plane_start
[1] == 0xd0d0d0d0);
3694 assert(prog_data
->binding_table
.plane_start
[2] == 0xd0d0d0d0);
3699 iris_use_optional_res(struct iris_batch
*batch
,
3700 struct pipe_resource
*res
,
3704 struct iris_bo
*bo
= iris_resource_bo(res
);
3705 iris_use_pinned_bo(batch
, bo
, writeable
);
3709 /* ------------------------------------------------------------------- */
3712 * Pin any BOs which were installed by a previous batch, and restored
3713 * via the hardware logical context mechanism.
3715 * We don't need to re-emit all state every batch - the hardware context
3716 * mechanism will save and restore it for us. This includes pointers to
3717 * various BOs...which won't exist unless we ask the kernel to pin them
3718 * by adding them to the validation list.
3720 * We can skip buffers if we've re-emitted those packets, as we're
3721 * overwriting those stale pointers with new ones, and don't actually
3722 * refer to the old BOs.
3725 iris_restore_render_saved_bos(struct iris_context
*ice
,
3726 struct iris_batch
*batch
,
3727 const struct pipe_draw_info
*draw
)
3729 // XXX: whack IRIS_SHADER_DIRTY_BINDING_TABLE on new batch
3731 const uint64_t clean
= ~ice
->state
.dirty
;
3733 if (clean
& IRIS_DIRTY_CC_VIEWPORT
) {
3734 iris_use_optional_res(batch
, ice
->state
.last_res
.cc_vp
, false);
3737 if (clean
& IRIS_DIRTY_SF_CL_VIEWPORT
) {
3738 iris_use_optional_res(batch
, ice
->state
.last_res
.sf_cl_vp
, false);
3741 if (clean
& IRIS_DIRTY_BLEND_STATE
) {
3742 iris_use_optional_res(batch
, ice
->state
.last_res
.blend
, false);
3745 if (clean
& IRIS_DIRTY_COLOR_CALC_STATE
) {
3746 iris_use_optional_res(batch
, ice
->state
.last_res
.color_calc
, false);
3749 if (clean
& IRIS_DIRTY_SCISSOR_RECT
) {
3750 iris_use_optional_res(batch
, ice
->state
.last_res
.scissor
, false);
3753 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
3754 if (!(clean
& (IRIS_DIRTY_CONSTANTS_VS
<< stage
)))
3757 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
3758 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3763 struct brw_stage_prog_data
*prog_data
= (void *) shader
->prog_data
;
3765 for (int i
= 0; i
< 4; i
++) {
3766 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[i
];
3768 if (range
->length
== 0)
3771 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[range
->block
];
3772 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
3775 iris_use_pinned_bo(batch
, res
->bo
, false);
3777 iris_use_pinned_bo(batch
, batch
->screen
->workaround_bo
, false);
3781 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
3782 if (clean
& (IRIS_DIRTY_BINDINGS_VS
<< stage
)) {
3783 /* Re-pin any buffers referred to by the binding table. */
3784 iris_populate_binding_table(ice
, batch
, stage
, true);
3788 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
3789 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
3790 struct pipe_resource
*res
= shs
->sampler_table
.res
;
3792 iris_use_pinned_bo(batch
, iris_resource_bo(res
), false);
3795 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
3796 if (clean
& (IRIS_DIRTY_VS
<< stage
)) {
3797 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3799 struct iris_bo
*bo
= iris_resource_bo(shader
->assembly
.res
);
3800 iris_use_pinned_bo(batch
, bo
, false);
3803 // XXX: scratch buffer
3807 if (clean
& IRIS_DIRTY_DEPTH_BUFFER
) {
3808 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
3810 if (cso_fb
->zsbuf
) {
3811 struct iris_resource
*zres
, *sres
;
3812 iris_get_depth_stencil_resources(cso_fb
->zsbuf
->texture
,
3814 // XXX: might not be writable...
3816 iris_use_pinned_bo(batch
, zres
->bo
, true);
3818 iris_use_pinned_bo(batch
, sres
->bo
, true);
3822 if (draw
->index_size
== 0 && ice
->state
.last_res
.index_buffer
) {
3823 /* This draw didn't emit a new index buffer, so we are inheriting the
3824 * older index buffer. This draw didn't need it, but future ones may.
3826 struct iris_bo
*bo
= iris_resource_bo(ice
->state
.last_res
.index_buffer
);
3827 iris_use_pinned_bo(batch
, bo
, false);
3830 if (clean
& IRIS_DIRTY_VERTEX_BUFFERS
) {
3831 struct iris_vertex_buffer_state
*cso
= &ice
->state
.genx
->vertex_buffers
;
3832 for (unsigned i
= 0; i
< cso
->num_buffers
; i
++) {
3833 struct iris_resource
*res
= (void *) cso
->resources
[i
];
3834 iris_use_pinned_bo(batch
, res
->bo
, false);
3840 iris_restore_compute_saved_bos(struct iris_context
*ice
,
3841 struct iris_batch
*batch
,
3842 const struct pipe_grid_info
*grid
)
3844 const uint64_t clean
= ~ice
->state
.dirty
;
3846 const int stage
= MESA_SHADER_COMPUTE
;
3847 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
3849 if (clean
& IRIS_DIRTY_CONSTANTS_CS
) {
3850 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3853 struct brw_stage_prog_data
*prog_data
= (void *) shader
->prog_data
;
3854 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[0];
3856 if (range
->length
> 0) {
3857 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[range
->block
];
3858 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
3861 iris_use_pinned_bo(batch
, res
->bo
, false);
3863 iris_use_pinned_bo(batch
, batch
->screen
->workaround_bo
, false);
3868 if (clean
& IRIS_DIRTY_BINDINGS_CS
) {
3869 /* Re-pin any buffers referred to by the binding table. */
3870 iris_populate_binding_table(ice
, batch
, stage
, true);
3873 struct pipe_resource
*sampler_res
= shs
->sampler_table
.res
;
3875 iris_use_pinned_bo(batch
, iris_resource_bo(sampler_res
), false);
3877 if (clean
& IRIS_DIRTY_CS
) {
3878 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
3880 struct iris_bo
*bo
= iris_resource_bo(shader
->assembly
.res
);
3881 iris_use_pinned_bo(batch
, bo
, false);
3884 // XXX: scratch buffer
3889 * Possibly emit STATE_BASE_ADDRESS to update Surface State Base Address.
3892 iris_update_surface_base_address(struct iris_batch
*batch
,
3893 struct iris_binder
*binder
)
3895 if (batch
->last_surface_base_address
== binder
->bo
->gtt_offset
)
3898 flush_for_state_base_change(batch
);
3900 iris_emit_cmd(batch
, GENX(STATE_BASE_ADDRESS
), sba
) {
3901 // XXX: sba.SurfaceStateMemoryObjectControlState = MOCS_WB;
3902 sba
.SurfaceStateBaseAddressModifyEnable
= true;
3903 sba
.SurfaceStateBaseAddress
= ro_bo(binder
->bo
, 0);
3906 batch
->last_surface_base_address
= binder
->bo
->gtt_offset
;
3910 iris_upload_dirty_render_state(struct iris_context
*ice
,
3911 struct iris_batch
*batch
,
3912 const struct pipe_draw_info
*draw
)
3914 const uint64_t dirty
= ice
->state
.dirty
;
3916 if (!(dirty
& IRIS_ALL_DIRTY_FOR_RENDER
))
3919 struct iris_genx_state
*genx
= ice
->state
.genx
;
3920 struct iris_binder
*binder
= &ice
->state
.binder
;
3921 struct brw_wm_prog_data
*wm_prog_data
= (void *)
3922 ice
->shaders
.prog
[MESA_SHADER_FRAGMENT
]->prog_data
;
3924 if (dirty
& IRIS_DIRTY_CC_VIEWPORT
) {
3925 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
3926 uint32_t cc_vp_address
;
3928 /* XXX: could avoid streaming for depth_clip [0,1] case. */
3929 uint32_t *cc_vp_map
=
3930 stream_state(batch
, ice
->state
.dynamic_uploader
,
3931 &ice
->state
.last_res
.cc_vp
,
3932 4 * ice
->state
.num_viewports
*
3933 GENX(CC_VIEWPORT_length
), 32, &cc_vp_address
);
3934 for (int i
= 0; i
< ice
->state
.num_viewports
; i
++) {
3936 util_viewport_zmin_zmax(&ice
->state
.viewports
[i
],
3937 cso_rast
->clip_halfz
, &zmin
, &zmax
);
3938 if (cso_rast
->depth_clip_near
)
3940 if (cso_rast
->depth_clip_far
)
3943 iris_pack_state(GENX(CC_VIEWPORT
), cc_vp_map
, ccv
) {
3944 ccv
.MinimumDepth
= zmin
;
3945 ccv
.MaximumDepth
= zmax
;
3948 cc_vp_map
+= GENX(CC_VIEWPORT_length
);
3951 iris_emit_cmd(batch
, GENX(3DSTATE_VIEWPORT_STATE_POINTERS_CC
), ptr
) {
3952 ptr
.CCViewportPointer
= cc_vp_address
;
3956 if (dirty
& IRIS_DIRTY_SF_CL_VIEWPORT
) {
3957 iris_emit_cmd(batch
, GENX(3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP
), ptr
) {
3958 ptr
.SFClipViewportPointer
=
3959 emit_state(batch
, ice
->state
.dynamic_uploader
,
3960 &ice
->state
.last_res
.sf_cl_vp
,
3961 genx
->sf_cl_vp
, 4 * GENX(SF_CLIP_VIEWPORT_length
) *
3962 ice
->state
.num_viewports
, 64);
3968 // XXX: this is only flagged at setup, we assume a static configuration
3969 if (dirty
& IRIS_DIRTY_URB
) {
3970 iris_upload_urb_config(ice
, batch
);
3973 if (dirty
& IRIS_DIRTY_BLEND_STATE
) {
3974 struct iris_blend_state
*cso_blend
= ice
->state
.cso_blend
;
3975 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
3976 struct iris_depth_stencil_alpha_state
*cso_zsa
= ice
->state
.cso_zsa
;
3977 const int header_dwords
= GENX(BLEND_STATE_length
);
3978 const int rt_dwords
= cso_fb
->nr_cbufs
* GENX(BLEND_STATE_ENTRY_length
);
3979 uint32_t blend_offset
;
3980 uint32_t *blend_map
=
3981 stream_state(batch
, ice
->state
.dynamic_uploader
,
3982 &ice
->state
.last_res
.blend
,
3983 4 * (header_dwords
+ rt_dwords
), 64, &blend_offset
);
3985 uint32_t blend_state_header
;
3986 iris_pack_state(GENX(BLEND_STATE
), &blend_state_header
, bs
) {
3987 bs
.AlphaTestEnable
= cso_zsa
->alpha
.enabled
;
3988 bs
.AlphaTestFunction
= translate_compare_func(cso_zsa
->alpha
.func
);
3991 blend_map
[0] = blend_state_header
| cso_blend
->blend_state
[0];
3992 memcpy(&blend_map
[1], &cso_blend
->blend_state
[1], 4 * rt_dwords
);
3994 iris_emit_cmd(batch
, GENX(3DSTATE_BLEND_STATE_POINTERS
), ptr
) {
3995 ptr
.BlendStatePointer
= blend_offset
;
3996 ptr
.BlendStatePointerValid
= true;
4000 if (dirty
& IRIS_DIRTY_COLOR_CALC_STATE
) {
4001 struct iris_depth_stencil_alpha_state
*cso
= ice
->state
.cso_zsa
;
4004 stream_state(batch
, ice
->state
.dynamic_uploader
,
4005 &ice
->state
.last_res
.color_calc
,
4006 sizeof(uint32_t) * GENX(COLOR_CALC_STATE_length
),
4008 iris_pack_state(GENX(COLOR_CALC_STATE
), cc_map
, cc
) {
4009 cc
.AlphaTestFormat
= ALPHATEST_FLOAT32
;
4010 cc
.AlphaReferenceValueAsFLOAT32
= cso
->alpha
.ref_value
;
4011 cc
.BlendConstantColorRed
= ice
->state
.blend_color
.color
[0];
4012 cc
.BlendConstantColorGreen
= ice
->state
.blend_color
.color
[1];
4013 cc
.BlendConstantColorBlue
= ice
->state
.blend_color
.color
[2];
4014 cc
.BlendConstantColorAlpha
= ice
->state
.blend_color
.color
[3];
4016 iris_emit_cmd(batch
, GENX(3DSTATE_CC_STATE_POINTERS
), ptr
) {
4017 ptr
.ColorCalcStatePointer
= cc_offset
;
4018 ptr
.ColorCalcStatePointerValid
= true;
4022 /* Upload constants for TCS passthrough. */
4023 if ((dirty
& IRIS_DIRTY_CONSTANTS_TCS
) &&
4024 ice
->shaders
.prog
[MESA_SHADER_TESS_CTRL
] &&
4025 !ice
->shaders
.uncompiled
[MESA_SHADER_TESS_CTRL
]) {
4026 struct iris_compiled_shader
*tes_shader
= ice
->shaders
.prog
[MESA_SHADER_TESS_EVAL
];
4029 /* Passthrough always copies 2 vec4s, so when uploading data we ensure
4030 * it is in the right layout for TES.
4033 struct brw_tes_prog_data
*tes_prog_data
= (void *) tes_shader
->prog_data
;
4034 switch (tes_prog_data
->domain
) {
4035 case BRW_TESS_DOMAIN_QUAD
:
4036 for (int i
= 0; i
< 4; i
++)
4037 hdr
[7 - i
] = ice
->state
.default_outer_level
[i
];
4038 hdr
[3] = ice
->state
.default_inner_level
[0];
4039 hdr
[2] = ice
->state
.default_inner_level
[1];
4041 case BRW_TESS_DOMAIN_TRI
:
4042 for (int i
= 0; i
< 3; i
++)
4043 hdr
[7 - i
] = ice
->state
.default_outer_level
[i
];
4044 hdr
[4] = ice
->state
.default_inner_level
[0];
4046 case BRW_TESS_DOMAIN_ISOLINE
:
4047 hdr
[7] = ice
->state
.default_outer_level
[1];
4048 hdr
[6] = ice
->state
.default_outer_level
[0];
4052 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_TESS_CTRL
];
4053 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[0];
4054 u_upload_data(ice
->ctx
.const_uploader
, 0, sizeof(hdr
), 32,
4055 &hdr
[0], &cbuf
->data
.offset
,
4059 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4060 if (!(dirty
& (IRIS_DIRTY_CONSTANTS_VS
<< stage
)))
4063 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4064 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
4069 if (shs
->cbuf0_needs_upload
)
4070 upload_uniforms(ice
, stage
);
4072 struct brw_stage_prog_data
*prog_data
= (void *) shader
->prog_data
;
4074 iris_emit_cmd(batch
, GENX(3DSTATE_CONSTANT_VS
), pkt
) {
4075 pkt
._3DCommandSubOpcode
= push_constant_opcodes
[stage
];
4077 /* The Skylake PRM contains the following restriction:
4079 * "The driver must ensure The following case does not occur
4080 * without a flush to the 3D engine: 3DSTATE_CONSTANT_* with
4081 * buffer 3 read length equal to zero committed followed by a
4082 * 3DSTATE_CONSTANT_* with buffer 0 read length not equal to
4085 * To avoid this, we program the buffers in the highest slots.
4086 * This way, slot 0 is only used if slot 3 is also used.
4090 for (int i
= 3; i
>= 0; i
--) {
4091 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[i
];
4093 if (range
->length
== 0)
4096 struct iris_const_buffer
*cbuf
= &shs
->constbuf
[range
->block
];
4097 struct iris_resource
*res
= (void *) cbuf
->data
.res
;
4099 assert(cbuf
->data
.offset
% 32 == 0);
4101 pkt
.ConstantBody
.ReadLength
[n
] = range
->length
;
4102 pkt
.ConstantBody
.Buffer
[n
] =
4103 res
? ro_bo(res
->bo
, range
->start
* 32 + cbuf
->data
.offset
)
4104 : ro_bo(batch
->screen
->workaround_bo
, 0);
4111 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4112 if (dirty
& (IRIS_DIRTY_BINDINGS_VS
<< stage
)) {
4113 iris_emit_cmd(batch
, GENX(3DSTATE_BINDING_TABLE_POINTERS_VS
), ptr
) {
4114 ptr
._3DCommandSubOpcode
= 38 + stage
;
4115 ptr
.PointertoVSBindingTable
= binder
->bt_offset
[stage
];
4120 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4121 if (dirty
& (IRIS_DIRTY_BINDINGS_VS
<< stage
)) {
4122 iris_populate_binding_table(ice
, batch
, stage
, false);
4126 if (ice
->state
.need_border_colors
)
4127 iris_use_pinned_bo(batch
, ice
->state
.border_color_pool
.bo
, false);
4129 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4130 if (!(dirty
& (IRIS_DIRTY_SAMPLER_STATES_VS
<< stage
)) ||
4131 !ice
->shaders
.prog
[stage
])
4134 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4135 struct pipe_resource
*res
= shs
->sampler_table
.res
;
4137 iris_use_pinned_bo(batch
, iris_resource_bo(res
), false);
4139 iris_emit_cmd(batch
, GENX(3DSTATE_SAMPLER_STATE_POINTERS_VS
), ptr
) {
4140 ptr
._3DCommandSubOpcode
= 43 + stage
;
4141 ptr
.PointertoVSSamplerState
= shs
->sampler_table
.offset
;
4145 if (dirty
& IRIS_DIRTY_MULTISAMPLE
) {
4146 iris_emit_cmd(batch
, GENX(3DSTATE_MULTISAMPLE
), ms
) {
4148 ice
->state
.cso_rast
->half_pixel_center
? CENTER
: UL_CORNER
;
4149 if (ice
->state
.framebuffer
.samples
> 0)
4150 ms
.NumberofMultisamples
= ffs(ice
->state
.framebuffer
.samples
) - 1;
4154 if (dirty
& IRIS_DIRTY_SAMPLE_MASK
) {
4155 iris_emit_cmd(batch
, GENX(3DSTATE_SAMPLE_MASK
), ms
) {
4156 ms
.SampleMask
= MAX2(ice
->state
.sample_mask
, 1);
4160 for (int stage
= 0; stage
<= MESA_SHADER_FRAGMENT
; stage
++) {
4161 if (!(dirty
& (IRIS_DIRTY_VS
<< stage
)))
4164 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
4167 struct iris_resource
*cache
= (void *) shader
->assembly
.res
;
4168 iris_use_pinned_bo(batch
, cache
->bo
, false);
4169 iris_batch_emit(batch
, shader
->derived_data
,
4170 iris_derived_program_state_size(stage
));
4172 if (stage
== MESA_SHADER_TESS_EVAL
) {
4173 iris_emit_cmd(batch
, GENX(3DSTATE_HS
), hs
);
4174 iris_emit_cmd(batch
, GENX(3DSTATE_TE
), te
);
4175 iris_emit_cmd(batch
, GENX(3DSTATE_DS
), ds
);
4176 } else if (stage
== MESA_SHADER_GEOMETRY
) {
4177 iris_emit_cmd(batch
, GENX(3DSTATE_GS
), gs
);
4182 if (ice
->state
.streamout_active
) {
4183 if (dirty
& IRIS_DIRTY_SO_BUFFERS
) {
4184 iris_batch_emit(batch
, genx
->so_buffers
,
4185 4 * 4 * GENX(3DSTATE_SO_BUFFER_length
));
4186 for (int i
= 0; i
< 4; i
++) {
4187 struct iris_stream_output_target
*tgt
=
4188 (void *) ice
->state
.so_target
[i
];
4190 iris_use_pinned_bo(batch
, iris_resource_bo(tgt
->base
.buffer
),
4192 iris_use_pinned_bo(batch
, iris_resource_bo(tgt
->offset
.res
),
4198 if ((dirty
& IRIS_DIRTY_SO_DECL_LIST
) && ice
->state
.streamout
) {
4199 uint32_t *decl_list
=
4200 ice
->state
.streamout
+ GENX(3DSTATE_STREAMOUT_length
);
4201 iris_batch_emit(batch
, decl_list
, 4 * ((decl_list
[0] & 0xff) + 2));
4204 if (dirty
& IRIS_DIRTY_STREAMOUT
) {
4205 const struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
4207 uint32_t dynamic_sol
[GENX(3DSTATE_STREAMOUT_length
)];
4208 iris_pack_command(GENX(3DSTATE_STREAMOUT
), dynamic_sol
, sol
) {
4209 sol
.SOFunctionEnable
= true;
4210 sol
.SOStatisticsEnable
= true;
4212 sol
.RenderingDisable
= cso_rast
->rasterizer_discard
&&
4213 !ice
->state
.prims_generated_query_active
;
4214 sol
.ReorderMode
= cso_rast
->flatshade_first
? LEADING
: TRAILING
;
4217 assert(ice
->state
.streamout
);
4219 iris_emit_merge(batch
, ice
->state
.streamout
, dynamic_sol
,
4220 GENX(3DSTATE_STREAMOUT_length
));
4223 if (dirty
& IRIS_DIRTY_STREAMOUT
) {
4224 iris_emit_cmd(batch
, GENX(3DSTATE_STREAMOUT
), sol
);
4228 if (dirty
& IRIS_DIRTY_CLIP
) {
4229 struct iris_rasterizer_state
*cso_rast
= ice
->state
.cso_rast
;
4230 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4232 uint32_t dynamic_clip
[GENX(3DSTATE_CLIP_length
)];
4233 iris_pack_command(GENX(3DSTATE_CLIP
), &dynamic_clip
, cl
) {
4234 if (wm_prog_data
->barycentric_interp_modes
&
4235 BRW_BARYCENTRIC_NONPERSPECTIVE_BITS
)
4236 cl
.NonPerspectiveBarycentricEnable
= true;
4238 cl
.ForceZeroRTAIndexEnable
= cso_fb
->layers
== 0;
4239 cl
.MaximumVPIndex
= ice
->state
.num_viewports
- 1;
4241 iris_emit_merge(batch
, cso_rast
->clip
, dynamic_clip
,
4242 ARRAY_SIZE(cso_rast
->clip
));
4245 if (dirty
& IRIS_DIRTY_RASTER
) {
4246 struct iris_rasterizer_state
*cso
= ice
->state
.cso_rast
;
4247 iris_batch_emit(batch
, cso
->raster
, sizeof(cso
->raster
));
4248 iris_batch_emit(batch
, cso
->sf
, sizeof(cso
->sf
));
4252 /* XXX: FS program updates needs to flag IRIS_DIRTY_WM */
4253 if (dirty
& IRIS_DIRTY_WM
) {
4254 struct iris_rasterizer_state
*cso
= ice
->state
.cso_rast
;
4255 uint32_t dynamic_wm
[GENX(3DSTATE_WM_length
)];
4257 iris_pack_command(GENX(3DSTATE_WM
), &dynamic_wm
, wm
) {
4258 wm
.StatisticsEnable
= ice
->state
.statistics_counters_enabled
;
4260 wm
.BarycentricInterpolationMode
=
4261 wm_prog_data
->barycentric_interp_modes
;
4263 if (wm_prog_data
->early_fragment_tests
)
4264 wm
.EarlyDepthStencilControl
= EDSC_PREPS
;
4265 else if (wm_prog_data
->has_side_effects
)
4266 wm
.EarlyDepthStencilControl
= EDSC_PSEXEC
;
4268 iris_emit_merge(batch
, cso
->wm
, dynamic_wm
, ARRAY_SIZE(cso
->wm
));
4271 if (dirty
& IRIS_DIRTY_SBE
) {
4272 iris_emit_sbe(batch
, ice
);
4275 if (dirty
& IRIS_DIRTY_PS_BLEND
) {
4276 struct iris_blend_state
*cso_blend
= ice
->state
.cso_blend
;
4277 struct iris_depth_stencil_alpha_state
*cso_zsa
= ice
->state
.cso_zsa
;
4278 uint32_t dynamic_pb
[GENX(3DSTATE_PS_BLEND_length
)];
4279 iris_pack_command(GENX(3DSTATE_PS_BLEND
), &dynamic_pb
, pb
) {
4280 pb
.HasWriteableRT
= true; // XXX: comes from somewhere :(
4281 pb
.AlphaTestEnable
= cso_zsa
->alpha
.enabled
;
4284 iris_emit_merge(batch
, cso_blend
->ps_blend
, dynamic_pb
,
4285 ARRAY_SIZE(cso_blend
->ps_blend
));
4288 if (dirty
& IRIS_DIRTY_WM_DEPTH_STENCIL
) {
4289 struct iris_depth_stencil_alpha_state
*cso
= ice
->state
.cso_zsa
;
4290 struct pipe_stencil_ref
*p_stencil_refs
= &ice
->state
.stencil_ref
;
4292 uint32_t stencil_refs
[GENX(3DSTATE_WM_DEPTH_STENCIL_length
)];
4293 iris_pack_command(GENX(3DSTATE_WM_DEPTH_STENCIL
), &stencil_refs
, wmds
) {
4294 wmds
.StencilReferenceValue
= p_stencil_refs
->ref_value
[0];
4295 wmds
.BackfaceStencilReferenceValue
= p_stencil_refs
->ref_value
[1];
4297 iris_emit_merge(batch
, cso
->wmds
, stencil_refs
, ARRAY_SIZE(cso
->wmds
));
4300 if (dirty
& IRIS_DIRTY_SCISSOR_RECT
) {
4301 uint32_t scissor_offset
=
4302 emit_state(batch
, ice
->state
.dynamic_uploader
,
4303 &ice
->state
.last_res
.scissor
,
4304 ice
->state
.scissors
,
4305 sizeof(struct pipe_scissor_state
) *
4306 ice
->state
.num_viewports
, 32);
4308 iris_emit_cmd(batch
, GENX(3DSTATE_SCISSOR_STATE_POINTERS
), ptr
) {
4309 ptr
.ScissorRectPointer
= scissor_offset
;
4313 if (dirty
& IRIS_DIRTY_DEPTH_BUFFER
) {
4314 struct pipe_framebuffer_state
*cso_fb
= &ice
->state
.framebuffer
;
4315 struct iris_depth_buffer_state
*cso_z
= &ice
->state
.genx
->depth_buffer
;
4317 iris_batch_emit(batch
, cso_z
->packets
, sizeof(cso_z
->packets
));
4319 if (cso_fb
->zsbuf
) {
4320 struct iris_resource
*zres
, *sres
;
4321 iris_get_depth_stencil_resources(cso_fb
->zsbuf
->texture
,
4323 // XXX: might not be writable...
4325 iris_use_pinned_bo(batch
, zres
->bo
, true);
4327 iris_use_pinned_bo(batch
, sres
->bo
, true);
4331 if (dirty
& IRIS_DIRTY_POLYGON_STIPPLE
) {
4332 iris_emit_cmd(batch
, GENX(3DSTATE_POLY_STIPPLE_PATTERN
), poly
) {
4333 for (int i
= 0; i
< 32; i
++) {
4334 poly
.PatternRow
[i
] = ice
->state
.poly_stipple
.stipple
[i
];
4339 if (dirty
& IRIS_DIRTY_LINE_STIPPLE
) {
4340 struct iris_rasterizer_state
*cso
= ice
->state
.cso_rast
;
4341 iris_batch_emit(batch
, cso
->line_stipple
, sizeof(cso
->line_stipple
));
4344 if (dirty
& IRIS_DIRTY_VF_TOPOLOGY
) {
4345 iris_emit_cmd(batch
, GENX(3DSTATE_VF_TOPOLOGY
), topo
) {
4346 topo
.PrimitiveTopologyType
=
4347 translate_prim_type(draw
->mode
, draw
->vertices_per_patch
);
4351 if (dirty
& IRIS_DIRTY_VERTEX_BUFFERS
) {
4352 struct iris_vertex_buffer_state
*cso
= &ice
->state
.genx
->vertex_buffers
;
4353 const unsigned vb_dwords
= GENX(VERTEX_BUFFER_STATE_length
);
4355 if (cso
->num_buffers
> 0) {
4356 /* The VF cache designers cut corners, and made the cache key's
4357 * <VertexBufferIndex, Memory Address> tuple only consider the bottom
4358 * 32 bits of the address. If you have two vertex buffers which get
4359 * placed exactly 4 GiB apart and use them in back-to-back draw calls,
4360 * you can get collisions (even within a single batch).
4362 * So, we need to do a VF cache invalidate if the buffer for a VB
4363 * slot slot changes [48:32] address bits from the previous time.
4365 bool need_invalidate
= false;
4367 for (unsigned i
= 0; i
< cso
->num_buffers
; i
++) {
4368 uint16_t high_bits
= 0;
4370 struct iris_resource
*res
= (void *) cso
->resources
[i
];
4372 iris_use_pinned_bo(batch
, res
->bo
, false);
4374 high_bits
= res
->bo
->gtt_offset
>> 32ull;
4375 if (high_bits
!= ice
->state
.last_vbo_high_bits
[i
]) {
4376 need_invalidate
= true;
4377 ice
->state
.last_vbo_high_bits
[i
] = high_bits
;
4382 if (need_invalidate
) {
4383 iris_emit_pipe_control_flush(batch
,
4384 PIPE_CONTROL_VF_CACHE_INVALIDATE
);
4387 iris_batch_emit(batch
, cso
->vertex_buffers
, sizeof(uint32_t) *
4388 (1 + vb_dwords
* cso
->num_buffers
));
4392 if (dirty
& IRIS_DIRTY_VERTEX_ELEMENTS
) {
4393 struct iris_vertex_element_state
*cso
= ice
->state
.cso_vertex_elements
;
4394 const unsigned entries
= MAX2(cso
->count
, 1);
4395 iris_batch_emit(batch
, cso
->vertex_elements
, sizeof(uint32_t) *
4396 (1 + entries
* GENX(VERTEX_ELEMENT_STATE_length
)));
4397 iris_batch_emit(batch
, cso
->vf_instancing
, sizeof(uint32_t) *
4398 entries
* GENX(3DSTATE_VF_INSTANCING_length
));
4401 if (dirty
& IRIS_DIRTY_VF_SGVS
) {
4402 const struct brw_vs_prog_data
*vs_prog_data
= (void *)
4403 ice
->shaders
.prog
[MESA_SHADER_VERTEX
]->prog_data
;
4404 struct iris_vertex_element_state
*cso
= ice
->state
.cso_vertex_elements
;
4406 iris_emit_cmd(batch
, GENX(3DSTATE_VF_SGVS
), sgv
) {
4407 if (vs_prog_data
->uses_vertexid
) {
4408 sgv
.VertexIDEnable
= true;
4409 sgv
.VertexIDComponentNumber
= 2;
4410 sgv
.VertexIDElementOffset
= cso
->count
;
4413 if (vs_prog_data
->uses_instanceid
) {
4414 sgv
.InstanceIDEnable
= true;
4415 sgv
.InstanceIDComponentNumber
= 3;
4416 sgv
.InstanceIDElementOffset
= cso
->count
;
4421 if (dirty
& IRIS_DIRTY_VF
) {
4422 iris_emit_cmd(batch
, GENX(3DSTATE_VF
), vf
) {
4423 if (draw
->primitive_restart
) {
4424 vf
.IndexedDrawCutIndexEnable
= true;
4425 vf
.CutIndex
= draw
->restart_index
;
4430 // XXX: Gen8 - PMA fix
4434 iris_upload_render_state(struct iris_context
*ice
,
4435 struct iris_batch
*batch
,
4436 const struct pipe_draw_info
*draw
)
4438 /* Always pin the binder. If we're emitting new binding table pointers,
4439 * we need it. If not, we're probably inheriting old tables via the
4440 * context, and need it anyway. Since true zero-bindings cases are
4441 * practically non-existent, just pin it and avoid last_res tracking.
4443 iris_use_pinned_bo(batch
, ice
->state
.binder
.bo
, false);
4445 if (!batch
->contains_draw
) {
4446 iris_restore_render_saved_bos(ice
, batch
, draw
);
4447 batch
->contains_draw
= true;
4450 iris_upload_dirty_render_state(ice
, batch
, draw
);
4452 if (draw
->index_size
> 0) {
4455 if (draw
->has_user_indices
) {
4456 u_upload_data(ice
->ctx
.stream_uploader
, 0,
4457 draw
->count
* draw
->index_size
, 4, draw
->index
.user
,
4458 &offset
, &ice
->state
.last_res
.index_buffer
);
4460 struct iris_resource
*res
= (void *) draw
->index
.resource
;
4461 res
->bind_history
|= PIPE_BIND_INDEX_BUFFER
;
4463 pipe_resource_reference(&ice
->state
.last_res
.index_buffer
,
4464 draw
->index
.resource
);
4468 struct iris_bo
*bo
= iris_resource_bo(ice
->state
.last_res
.index_buffer
);
4470 iris_emit_cmd(batch
, GENX(3DSTATE_INDEX_BUFFER
), ib
) {
4471 ib
.IndexFormat
= draw
->index_size
>> 1;
4473 ib
.BufferSize
= bo
->size
;
4474 ib
.BufferStartingAddress
= ro_bo(bo
, offset
);
4477 /* The VF cache key only uses 32-bits, see vertex buffer comment above */
4478 uint16_t high_bits
= bo
->gtt_offset
>> 32ull;
4479 if (high_bits
!= ice
->state
.last_index_bo_high_bits
) {
4480 iris_emit_pipe_control_flush(batch
, PIPE_CONTROL_VF_CACHE_INVALIDATE
);
4481 ice
->state
.last_index_bo_high_bits
= high_bits
;
4485 #define _3DPRIM_END_OFFSET 0x2420
4486 #define _3DPRIM_START_VERTEX 0x2430
4487 #define _3DPRIM_VERTEX_COUNT 0x2434
4488 #define _3DPRIM_INSTANCE_COUNT 0x2438
4489 #define _3DPRIM_START_INSTANCE 0x243C
4490 #define _3DPRIM_BASE_VERTEX 0x2440
4492 if (draw
->indirect
) {
4493 /* We don't support this MultidrawIndirect. */
4494 assert(!draw
->indirect
->indirect_draw_count
);
4496 struct iris_bo
*bo
= iris_resource_bo(draw
->indirect
->buffer
);
4499 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4500 lrm
.RegisterAddress
= _3DPRIM_VERTEX_COUNT
;
4501 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 0);
4503 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4504 lrm
.RegisterAddress
= _3DPRIM_INSTANCE_COUNT
;
4505 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 4);
4507 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4508 lrm
.RegisterAddress
= _3DPRIM_START_VERTEX
;
4509 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 8);
4511 if (draw
->index_size
) {
4512 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4513 lrm
.RegisterAddress
= _3DPRIM_BASE_VERTEX
;
4514 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 12);
4516 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4517 lrm
.RegisterAddress
= _3DPRIM_START_INSTANCE
;
4518 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 16);
4521 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4522 lrm
.RegisterAddress
= _3DPRIM_START_INSTANCE
;
4523 lrm
.MemoryAddress
= ro_bo(bo
, draw
->indirect
->offset
+ 12);
4525 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_IMM
), lri
) {
4526 lri
.RegisterOffset
= _3DPRIM_BASE_VERTEX
;
4532 iris_emit_cmd(batch
, GENX(3DPRIMITIVE
), prim
) {
4533 prim
.StartInstanceLocation
= draw
->start_instance
;
4534 prim
.InstanceCount
= draw
->instance_count
;
4535 prim
.VertexCountPerInstance
= draw
->count
;
4536 prim
.VertexAccessType
= draw
->index_size
> 0 ? RANDOM
: SEQUENTIAL
;
4538 // XXX: this is probably bonkers.
4539 prim
.StartVertexLocation
= draw
->start
;
4541 prim
.IndirectParameterEnable
= draw
->indirect
!= NULL
;
4543 if (draw
->index_size
) {
4544 prim
.BaseVertexLocation
+= draw
->index_bias
;
4546 prim
.StartVertexLocation
+= draw
->index_bias
;
4549 //prim.BaseVertexLocation = ...;
4554 iris_upload_compute_state(struct iris_context
*ice
,
4555 struct iris_batch
*batch
,
4556 const struct pipe_grid_info
*grid
)
4558 const uint64_t dirty
= ice
->state
.dirty
;
4559 struct iris_screen
*screen
= batch
->screen
;
4560 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
4561 struct iris_binder
*binder
= &ice
->state
.binder
;
4562 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_COMPUTE
];
4563 struct iris_compiled_shader
*shader
=
4564 ice
->shaders
.prog
[MESA_SHADER_COMPUTE
];
4565 struct brw_stage_prog_data
*prog_data
= shader
->prog_data
;
4566 struct brw_cs_prog_data
*cs_prog_data
= (void *) prog_data
;
4568 if ((dirty
& IRIS_DIRTY_CONSTANTS_CS
) && shs
->cbuf0_needs_upload
)
4569 upload_uniforms(ice
, MESA_SHADER_COMPUTE
);
4571 if (dirty
& IRIS_DIRTY_BINDINGS_CS
)
4572 iris_populate_binding_table(ice
, batch
, MESA_SHADER_COMPUTE
, false);
4574 iris_use_optional_res(batch
, shs
->sampler_table
.res
, false);
4575 iris_use_pinned_bo(batch
, iris_resource_bo(shader
->assembly
.res
), false);
4577 if (ice
->state
.need_border_colors
)
4578 iris_use_pinned_bo(batch
, ice
->state
.border_color_pool
.bo
, false);
4580 if (dirty
& IRIS_DIRTY_CS
) {
4581 /* The MEDIA_VFE_STATE documentation for Gen8+ says:
4583 * "A stalling PIPE_CONTROL is required before MEDIA_VFE_STATE unless
4584 * the only bits that are changed are scoreboard related: Scoreboard
4585 * Enable, Scoreboard Type, Scoreboard Mask, Scoreboard Delta. For
4586 * these scoreboard related states, a MEDIA_STATE_FLUSH is
4589 iris_emit_pipe_control_flush(batch
, PIPE_CONTROL_CS_STALL
);
4591 iris_emit_cmd(batch
, GENX(MEDIA_VFE_STATE
), vfe
) {
4592 if (prog_data
->total_scratch
) {
4593 uint32_t scratch_addr
=
4594 iris_get_scratch_space(ice
, prog_data
->total_scratch
,
4595 MESA_SHADER_COMPUTE
);
4596 vfe
.PerThreadScratchSpace
= ffs(prog_data
->total_scratch
) - 11;
4597 vfe
.ScratchSpaceBasePointer
= rw_bo(NULL
, scratch_addr
);
4600 vfe
.MaximumNumberofThreads
=
4601 devinfo
->max_cs_threads
* screen
->subslice_total
- 1;
4603 vfe
.ResetGatewayTimer
=
4604 Resettingrelativetimerandlatchingtheglobaltimestamp
;
4607 vfe
.NumberofURBEntries
= 2;
4608 vfe
.URBEntryAllocationSize
= 2;
4610 // XXX: Use Indirect Payload Storage?
4611 vfe
.CURBEAllocationSize
=
4612 ALIGN(cs_prog_data
->push
.per_thread
.regs
* cs_prog_data
->threads
+
4613 cs_prog_data
->push
.cross_thread
.regs
, 2);
4617 // XXX: hack iris_set_constant_buffers to upload these thread counts
4618 // XXX: along with regular uniforms for compute shaders, somehow.
4620 uint32_t curbe_data_offset
= 0;
4621 // TODO: Move subgroup-id into uniforms ubo so we can push uniforms
4622 assert(cs_prog_data
->push
.cross_thread
.dwords
== 0 &&
4623 cs_prog_data
->push
.per_thread
.dwords
== 1 &&
4624 cs_prog_data
->base
.param
[0] == BRW_PARAM_BUILTIN_SUBGROUP_ID
);
4625 struct pipe_resource
*curbe_data_res
= NULL
;
4626 uint32_t *curbe_data_map
=
4627 stream_state(batch
, ice
->state
.dynamic_uploader
, &curbe_data_res
,
4628 ALIGN(cs_prog_data
->push
.total
.size
, 64), 64,
4629 &curbe_data_offset
);
4630 assert(curbe_data_map
);
4631 memset(curbe_data_map
, 0x5a, ALIGN(cs_prog_data
->push
.total
.size
, 64));
4632 iris_fill_cs_push_const_buffer(cs_prog_data
, curbe_data_map
);
4634 if (dirty
& IRIS_DIRTY_CONSTANTS_CS
) {
4635 iris_emit_cmd(batch
, GENX(MEDIA_CURBE_LOAD
), curbe
) {
4636 curbe
.CURBETotalDataLength
=
4637 ALIGN(cs_prog_data
->push
.total
.size
, 64);
4638 curbe
.CURBEDataStartAddress
= curbe_data_offset
;
4642 if (dirty
& (IRIS_DIRTY_SAMPLER_STATES_CS
|
4643 IRIS_DIRTY_BINDINGS_CS
|
4644 IRIS_DIRTY_CONSTANTS_CS
|
4646 struct pipe_resource
*desc_res
= NULL
;
4647 uint32_t desc
[GENX(INTERFACE_DESCRIPTOR_DATA_length
)];
4649 iris_pack_state(GENX(INTERFACE_DESCRIPTOR_DATA
), desc
, idd
) {
4650 idd
.SamplerStatePointer
= shs
->sampler_table
.offset
;
4651 idd
.BindingTablePointer
= binder
->bt_offset
[MESA_SHADER_COMPUTE
];
4654 for (int i
= 0; i
< GENX(INTERFACE_DESCRIPTOR_DATA_length
); i
++)
4655 desc
[i
] |= ((uint32_t *) shader
->derived_data
)[i
];
4657 iris_emit_cmd(batch
, GENX(MEDIA_INTERFACE_DESCRIPTOR_LOAD
), load
) {
4658 load
.InterfaceDescriptorTotalLength
=
4659 GENX(INTERFACE_DESCRIPTOR_DATA_length
) * sizeof(uint32_t);
4660 load
.InterfaceDescriptorDataStartAddress
=
4661 emit_state(batch
, ice
->state
.dynamic_uploader
,
4662 &desc_res
, desc
, sizeof(desc
), 32);
4665 pipe_resource_reference(&desc_res
, NULL
);
4668 uint32_t group_size
= grid
->block
[0] * grid
->block
[1] * grid
->block
[2];
4669 uint32_t remainder
= group_size
& (cs_prog_data
->simd_size
- 1);
4670 uint32_t right_mask
;
4673 right_mask
= ~0u >> (32 - remainder
);
4675 right_mask
= ~0u >> (32 - cs_prog_data
->simd_size
);
4677 #define GPGPU_DISPATCHDIMX 0x2500
4678 #define GPGPU_DISPATCHDIMY 0x2504
4679 #define GPGPU_DISPATCHDIMZ 0x2508
4681 if (grid
->indirect
) {
4682 struct iris_state_ref
*grid_size
= &ice
->state
.grid_size
;
4683 struct iris_bo
*bo
= iris_resource_bo(grid_size
->res
);
4684 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4685 lrm
.RegisterAddress
= GPGPU_DISPATCHDIMX
;
4686 lrm
.MemoryAddress
= ro_bo(bo
, grid_size
->offset
+ 0);
4688 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4689 lrm
.RegisterAddress
= GPGPU_DISPATCHDIMY
;
4690 lrm
.MemoryAddress
= ro_bo(bo
, grid_size
->offset
+ 4);
4692 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4693 lrm
.RegisterAddress
= GPGPU_DISPATCHDIMZ
;
4694 lrm
.MemoryAddress
= ro_bo(bo
, grid_size
->offset
+ 8);
4698 iris_emit_cmd(batch
, GENX(GPGPU_WALKER
), ggw
) {
4699 ggw
.IndirectParameterEnable
= grid
->indirect
!= NULL
;
4700 ggw
.SIMDSize
= cs_prog_data
->simd_size
/ 16;
4701 ggw
.ThreadDepthCounterMaximum
= 0;
4702 ggw
.ThreadHeightCounterMaximum
= 0;
4703 ggw
.ThreadWidthCounterMaximum
= cs_prog_data
->threads
- 1;
4704 ggw
.ThreadGroupIDXDimension
= grid
->grid
[0];
4705 ggw
.ThreadGroupIDYDimension
= grid
->grid
[1];
4706 ggw
.ThreadGroupIDZDimension
= grid
->grid
[2];
4707 ggw
.RightExecutionMask
= right_mask
;
4708 ggw
.BottomExecutionMask
= 0xffffffff;
4711 iris_emit_cmd(batch
, GENX(MEDIA_STATE_FLUSH
), msf
);
4713 if (!batch
->contains_draw
) {
4714 iris_restore_compute_saved_bos(ice
, batch
, grid
);
4715 batch
->contains_draw
= true;
4720 * State module teardown.
4723 iris_destroy_state(struct iris_context
*ice
)
4725 iris_free_vertex_buffers(&ice
->state
.genx
->vertex_buffers
);
4727 // XXX: unreference resources/surfaces.
4728 for (unsigned i
= 0; i
< ice
->state
.framebuffer
.nr_cbufs
; i
++) {
4729 pipe_surface_reference(&ice
->state
.framebuffer
.cbufs
[i
], NULL
);
4731 pipe_surface_reference(&ice
->state
.framebuffer
.zsbuf
, NULL
);
4733 for (int stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
4734 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
4735 pipe_resource_reference(&shs
->sampler_table
.res
, NULL
);
4737 free(ice
->state
.genx
);
4739 pipe_resource_reference(&ice
->state
.last_res
.cc_vp
, NULL
);
4740 pipe_resource_reference(&ice
->state
.last_res
.sf_cl_vp
, NULL
);
4741 pipe_resource_reference(&ice
->state
.last_res
.color_calc
, NULL
);
4742 pipe_resource_reference(&ice
->state
.last_res
.scissor
, NULL
);
4743 pipe_resource_reference(&ice
->state
.last_res
.blend
, NULL
);
4744 pipe_resource_reference(&ice
->state
.last_res
.index_buffer
, NULL
);
4747 /* ------------------------------------------------------------------- */
4750 iris_load_register_imm32(struct iris_batch
*batch
, uint32_t reg
,
4753 _iris_emit_lri(batch
, reg
, val
);
4757 iris_load_register_imm64(struct iris_batch
*batch
, uint32_t reg
,
4760 _iris_emit_lri(batch
, reg
+ 0, val
& 0xffffffff);
4761 _iris_emit_lri(batch
, reg
+ 4, val
>> 32);
4765 * Emit MI_LOAD_REGISTER_MEM to load a 32-bit MMIO register from a buffer.
4768 iris_load_register_mem32(struct iris_batch
*batch
, uint32_t reg
,
4769 struct iris_bo
*bo
, uint32_t offset
)
4771 iris_emit_cmd(batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
4772 lrm
.RegisterAddress
= reg
;
4773 lrm
.MemoryAddress
= ro_bo(bo
, offset
);
4778 * Load a 64-bit value from a buffer into a MMIO register via
4779 * two MI_LOAD_REGISTER_MEM commands.
4782 iris_load_register_mem64(struct iris_batch
*batch
, uint32_t reg
,
4783 struct iris_bo
*bo
, uint32_t offset
)
4785 iris_load_register_mem32(batch
, reg
+ 0, bo
, offset
+ 0);
4786 iris_load_register_mem32(batch
, reg
+ 4, bo
, offset
+ 4);
4790 iris_store_register_mem32(struct iris_batch
*batch
, uint32_t reg
,
4791 struct iris_bo
*bo
, uint32_t offset
,
4794 iris_emit_cmd(batch
, GENX(MI_STORE_REGISTER_MEM
), srm
) {
4795 srm
.RegisterAddress
= reg
;
4796 srm
.MemoryAddress
= rw_bo(bo
, offset
);
4797 srm
.PredicateEnable
= predicated
;
4802 iris_store_register_mem64(struct iris_batch
*batch
, uint32_t reg
,
4803 struct iris_bo
*bo
, uint32_t offset
,
4806 iris_store_register_mem32(batch
, reg
+ 0, bo
, offset
+ 0, predicated
);
4807 iris_store_register_mem32(batch
, reg
+ 4, bo
, offset
+ 4, predicated
);
4811 iris_store_data_imm32(struct iris_batch
*batch
,
4812 struct iris_bo
*bo
, uint32_t offset
,
4815 iris_emit_cmd(batch
, GENX(MI_STORE_DATA_IMM
), sdi
) {
4816 sdi
.Address
= rw_bo(bo
, offset
);
4817 sdi
.ImmediateData
= imm
;
4822 iris_store_data_imm64(struct iris_batch
*batch
,
4823 struct iris_bo
*bo
, uint32_t offset
,
4826 /* Can't use iris_emit_cmd because MI_STORE_DATA_IMM has a length of
4827 * 2 in genxml but it's actually variable length and we need 5 DWords.
4829 void *map
= iris_get_command_space(batch
, 4 * 5);
4830 _iris_pack_command(batch
, GENX(MI_STORE_DATA_IMM
), map
, sdi
) {
4831 sdi
.DWordLength
= 5 - 2;
4832 sdi
.Address
= rw_bo(bo
, offset
);
4833 sdi
.ImmediateData
= imm
;
4838 iris_copy_mem_mem(struct iris_batch
*batch
,
4839 struct iris_bo
*dst_bo
, uint32_t dst_offset
,
4840 struct iris_bo
*src_bo
, uint32_t src_offset
,
4843 /* MI_COPY_MEM_MEM operates on DWords. */
4844 assert(bytes
% 4 == 0);
4845 assert(dst_offset
% 4 == 0);
4846 assert(src_offset
% 4 == 0);
4848 for (unsigned i
= 0; i
< bytes
; i
+= 4) {
4849 iris_emit_cmd(batch
, GENX(MI_COPY_MEM_MEM
), cp
) {
4850 cp
.DestinationMemoryAddress
= rw_bo(dst_bo
, dst_offset
+ i
);
4851 cp
.SourceMemoryAddress
= ro_bo(src_bo
, src_offset
+ i
);
4856 /* ------------------------------------------------------------------- */
4859 flags_to_post_sync_op(uint32_t flags
)
4861 if (flags
& PIPE_CONTROL_WRITE_IMMEDIATE
)
4862 return WriteImmediateData
;
4864 if (flags
& PIPE_CONTROL_WRITE_DEPTH_COUNT
)
4865 return WritePSDepthCount
;
4867 if (flags
& PIPE_CONTROL_WRITE_TIMESTAMP
)
4868 return WriteTimestamp
;
4874 * Do the given flags have a Post Sync or LRI Post Sync operation?
4876 static enum pipe_control_flags
4877 get_post_sync_flags(enum pipe_control_flags flags
)
4879 flags
&= PIPE_CONTROL_WRITE_IMMEDIATE
|
4880 PIPE_CONTROL_WRITE_DEPTH_COUNT
|
4881 PIPE_CONTROL_WRITE_TIMESTAMP
|
4882 PIPE_CONTROL_LRI_POST_SYNC_OP
;
4884 /* Only one "Post Sync Op" is allowed, and it's mutually exclusive with
4885 * "LRI Post Sync Operation". So more than one bit set would be illegal.
4887 assert(util_bitcount(flags
) <= 1);
4892 #define IS_COMPUTE_PIPELINE(batch) (batch->name == IRIS_BATCH_COMPUTE)
4895 * Emit a series of PIPE_CONTROL commands, taking into account any
4896 * workarounds necessary to actually accomplish the caller's request.
4898 * Unless otherwise noted, spec quotations in this function come from:
4900 * Synchronization of the 3D Pipeline > PIPE_CONTROL Command > Programming
4901 * Restrictions for PIPE_CONTROL.
4903 * You should not use this function directly. Use the helpers in
4904 * iris_pipe_control.c instead, which may split the pipe control further.
4907 iris_emit_raw_pipe_control(struct iris_batch
*batch
, uint32_t flags
,
4908 struct iris_bo
*bo
, uint32_t offset
, uint64_t imm
)
4910 UNUSED
const struct gen_device_info
*devinfo
= &batch
->screen
->devinfo
;
4911 enum pipe_control_flags post_sync_flags
= get_post_sync_flags(flags
);
4912 enum pipe_control_flags non_lri_post_sync_flags
=
4913 post_sync_flags
& ~PIPE_CONTROL_LRI_POST_SYNC_OP
;
4915 /* Recursive PIPE_CONTROL workarounds --------------------------------
4916 * (http://knowyourmeme.com/memes/xzibit-yo-dawg)
4918 * We do these first because we want to look at the original operation,
4919 * rather than any workarounds we set.
4921 if (GEN_GEN
== 9 && (flags
& PIPE_CONTROL_VF_CACHE_INVALIDATE
)) {
4922 /* The PIPE_CONTROL "VF Cache Invalidation Enable" bit description
4923 * lists several workarounds:
4925 * "Project: SKL, KBL, BXT
4927 * If the VF Cache Invalidation Enable is set to a 1 in a
4928 * PIPE_CONTROL, a separate Null PIPE_CONTROL, all bitfields
4929 * sets to 0, with the VF Cache Invalidation Enable set to 0
4930 * needs to be sent prior to the PIPE_CONTROL with VF Cache
4931 * Invalidation Enable set to a 1."
4933 iris_emit_raw_pipe_control(batch
, 0, NULL
, 0, 0);
4936 if (GEN_GEN
== 9 && IS_COMPUTE_PIPELINE(batch
) && post_sync_flags
) {
4937 /* Project: SKL / Argument: LRI Post Sync Operation [23]
4939 * "PIPECONTROL command with “Command Streamer Stall Enable” must be
4940 * programmed prior to programming a PIPECONTROL command with "LRI
4941 * Post Sync Operation" in GPGPU mode of operation (i.e when
4942 * PIPELINE_SELECT command is set to GPGPU mode of operation)."
4944 * The same text exists a few rows below for Post Sync Op.
4946 iris_emit_raw_pipe_control(batch
, PIPE_CONTROL_CS_STALL
, bo
, offset
, imm
);
4949 if (GEN_GEN
== 10 && (flags
& PIPE_CONTROL_RENDER_TARGET_FLUSH
)) {
4951 * "Before sending a PIPE_CONTROL command with bit 12 set, SW must issue
4952 * another PIPE_CONTROL with Render Target Cache Flush Enable (bit 12)
4953 * = 0 and Pipe Control Flush Enable (bit 7) = 1"
4955 iris_emit_raw_pipe_control(batch
, PIPE_CONTROL_FLUSH_ENABLE
, bo
,
4959 /* "Flush Types" workarounds ---------------------------------------------
4960 * We do these now because they may add post-sync operations or CS stalls.
4963 if (GEN_GEN
< 11 && flags
& PIPE_CONTROL_VF_CACHE_INVALIDATE
) {
4964 /* Project: BDW, SKL+ (stopping at CNL) / Argument: VF Invalidate
4966 * "'Post Sync Operation' must be enabled to 'Write Immediate Data' or
4967 * 'Write PS Depth Count' or 'Write Timestamp'."
4970 flags
|= PIPE_CONTROL_WRITE_IMMEDIATE
;
4971 post_sync_flags
|= PIPE_CONTROL_WRITE_IMMEDIATE
;
4972 non_lri_post_sync_flags
|= PIPE_CONTROL_WRITE_IMMEDIATE
;
4973 bo
= batch
->screen
->workaround_bo
;
4977 /* #1130 from Gen10 workarounds page:
4979 * "Enable Depth Stall on every Post Sync Op if Render target Cache
4980 * Flush is not enabled in same PIPE CONTROL and Enable Pixel score
4981 * board stall if Render target cache flush is enabled."
4983 * Applicable to CNL B0 and C0 steppings only.
4985 * The wording here is unclear, and this workaround doesn't look anything
4986 * like the internal bug report recommendations, but leave it be for now...
4988 if (GEN_GEN
== 10) {
4989 if (flags
& PIPE_CONTROL_RENDER_TARGET_FLUSH
) {
4990 flags
|= PIPE_CONTROL_STALL_AT_SCOREBOARD
;
4991 } else if (flags
& non_lri_post_sync_flags
) {
4992 flags
|= PIPE_CONTROL_DEPTH_STALL
;
4996 if (flags
& PIPE_CONTROL_DEPTH_STALL
) {
4997 /* From the PIPE_CONTROL instruction table, bit 13 (Depth Stall Enable):
4999 * "This bit must be DISABLED for operations other than writing
5002 * This seems like nonsense. An Ivybridge workaround requires us to
5003 * emit a PIPE_CONTROL with a depth stall and write immediate post-sync
5004 * operation. Gen8+ requires us to emit depth stalls and depth cache
5005 * flushes together. So, it's hard to imagine this means anything other
5006 * than "we originally intended this to be used for PS_DEPTH_COUNT".
5008 * We ignore the supposed restriction and do nothing.
5012 if (flags
& (PIPE_CONTROL_RENDER_TARGET_FLUSH
|
5013 PIPE_CONTROL_STALL_AT_SCOREBOARD
)) {
5014 /* From the PIPE_CONTROL instruction table, bit 12 and bit 1:
5016 * "This bit must be DISABLED for End-of-pipe (Read) fences,
5017 * PS_DEPTH_COUNT or TIMESTAMP queries."
5019 * TODO: Implement end-of-pipe checking.
5021 assert(!(post_sync_flags
& (PIPE_CONTROL_WRITE_DEPTH_COUNT
|
5022 PIPE_CONTROL_WRITE_TIMESTAMP
)));
5025 if (GEN_GEN
< 11 && (flags
& PIPE_CONTROL_STALL_AT_SCOREBOARD
)) {
5026 /* From the PIPE_CONTROL instruction table, bit 1:
5028 * "This bit is ignored if Depth Stall Enable is set.
5029 * Further, the render cache is not flushed even if Write Cache
5030 * Flush Enable bit is set."
5032 * We assert that the caller doesn't do this combination, to try and
5033 * prevent mistakes. It shouldn't hurt the GPU, though.
5035 * We skip this check on Gen11+ as the "Stall at Pixel Scoreboard"
5036 * and "Render Target Flush" combo is explicitly required for BTI
5037 * update workarounds.
5039 assert(!(flags
& (PIPE_CONTROL_DEPTH_STALL
|
5040 PIPE_CONTROL_RENDER_TARGET_FLUSH
)));
5043 /* PIPE_CONTROL page workarounds ------------------------------------- */
5045 if (GEN_GEN
<= 8 && (flags
& PIPE_CONTROL_STATE_CACHE_INVALIDATE
)) {
5046 /* From the PIPE_CONTROL page itself:
5049 * Restriction: Pipe_control with CS-stall bit set must be issued
5050 * before a pipe-control command that has the State Cache
5051 * Invalidate bit set."
5053 flags
|= PIPE_CONTROL_CS_STALL
;
5056 if (flags
& PIPE_CONTROL_FLUSH_LLC
) {
5057 /* From the PIPE_CONTROL instruction table, bit 26 (Flush LLC):
5060 * SW must always program Post-Sync Operation to "Write Immediate
5061 * Data" when Flush LLC is set."
5063 * For now, we just require the caller to do it.
5065 assert(flags
& PIPE_CONTROL_WRITE_IMMEDIATE
);
5068 /* "Post-Sync Operation" workarounds -------------------------------- */
5070 /* Project: All / Argument: Global Snapshot Count Reset [19]
5072 * "This bit must not be exercised on any product.
5073 * Requires stall bit ([20] of DW1) set."
5075 * We don't use this, so we just assert that it isn't used. The
5076 * PIPE_CONTROL instruction page indicates that they intended this
5077 * as a debug feature and don't think it is useful in production,
5078 * but it may actually be usable, should we ever want to.
5080 assert((flags
& PIPE_CONTROL_GLOBAL_SNAPSHOT_COUNT_RESET
) == 0);
5082 if (flags
& (PIPE_CONTROL_MEDIA_STATE_CLEAR
|
5083 PIPE_CONTROL_INDIRECT_STATE_POINTERS_DISABLE
)) {
5084 /* Project: All / Arguments:
5086 * - Generic Media State Clear [16]
5087 * - Indirect State Pointers Disable [16]
5089 * "Requires stall bit ([20] of DW1) set."
5091 * Also, the PIPE_CONTROL instruction table, bit 16 (Generic Media
5092 * State Clear) says:
5094 * "PIPECONTROL command with “Command Streamer Stall Enable” must be
5095 * programmed prior to programming a PIPECONTROL command with "Media
5096 * State Clear" set in GPGPU mode of operation"
5098 * This is a subset of the earlier rule, so there's nothing to do.
5100 flags
|= PIPE_CONTROL_CS_STALL
;
5103 if (flags
& PIPE_CONTROL_STORE_DATA_INDEX
) {
5104 /* Project: All / Argument: Store Data Index
5106 * "Post-Sync Operation ([15:14] of DW1) must be set to something other
5109 * For now, we just assert that the caller does this. We might want to
5110 * automatically add a write to the workaround BO...
5112 assert(non_lri_post_sync_flags
!= 0);
5115 if (flags
& PIPE_CONTROL_SYNC_GFDT
) {
5116 /* Project: All / Argument: Sync GFDT
5118 * "Post-Sync Operation ([15:14] of DW1) must be set to something other
5119 * than '0' or 0x2520[13] must be set."
5121 * For now, we just assert that the caller does this.
5123 assert(non_lri_post_sync_flags
!= 0);
5126 if (flags
& PIPE_CONTROL_TLB_INVALIDATE
) {
5127 /* Project: IVB+ / Argument: TLB inv
5129 * "Requires stall bit ([20] of DW1) set."
5131 * Also, from the PIPE_CONTROL instruction table:
5134 * Post Sync Operation or CS stall must be set to ensure a TLB
5135 * invalidation occurs. Otherwise no cycle will occur to the TLB
5136 * cache to invalidate."
5138 * This is not a subset of the earlier rule, so there's nothing to do.
5140 flags
|= PIPE_CONTROL_CS_STALL
;
5143 if (GEN_GEN
== 9 && devinfo
->gt
== 4) {
5144 /* TODO: The big Skylake GT4 post sync op workaround */
5147 /* "GPGPU specific workarounds" (both post-sync and flush) ------------ */
5149 if (IS_COMPUTE_PIPELINE(batch
)) {
5150 if (GEN_GEN
>= 9 && (flags
& PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
)) {
5151 /* Project: SKL+ / Argument: Tex Invalidate
5152 * "Requires stall bit ([20] of DW) set for all GPGPU Workloads."
5154 flags
|= PIPE_CONTROL_CS_STALL
;
5157 if (GEN_GEN
== 8 && (post_sync_flags
||
5158 (flags
& (PIPE_CONTROL_NOTIFY_ENABLE
|
5159 PIPE_CONTROL_DEPTH_STALL
|
5160 PIPE_CONTROL_RENDER_TARGET_FLUSH
|
5161 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
5162 PIPE_CONTROL_DATA_CACHE_FLUSH
)))) {
5163 /* Project: BDW / Arguments:
5165 * - LRI Post Sync Operation [23]
5166 * - Post Sync Op [15:14]
5168 * - Depth Stall [13]
5169 * - Render Target Cache Flush [12]
5170 * - Depth Cache Flush [0]
5171 * - DC Flush Enable [5]
5173 * "Requires stall bit ([20] of DW) set for all GPGPU and Media
5176 flags
|= PIPE_CONTROL_CS_STALL
;
5178 /* Also, from the PIPE_CONTROL instruction table, bit 20:
5181 * This bit must be always set when PIPE_CONTROL command is
5182 * programmed by GPGPU and MEDIA workloads, except for the cases
5183 * when only Read Only Cache Invalidation bits are set (State
5184 * Cache Invalidation Enable, Instruction cache Invalidation
5185 * Enable, Texture Cache Invalidation Enable, Constant Cache
5186 * Invalidation Enable). This is to WA FFDOP CG issue, this WA
5187 * need not implemented when FF_DOP_CG is disable via "Fixed
5188 * Function DOP Clock Gate Disable" bit in RC_PSMI_CTRL register."
5190 * It sounds like we could avoid CS stalls in some cases, but we
5191 * don't currently bother. This list isn't exactly the list above,
5197 /* "Stall" workarounds ----------------------------------------------
5198 * These have to come after the earlier ones because we may have added
5199 * some additional CS stalls above.
5202 if (GEN_GEN
< 9 && (flags
& PIPE_CONTROL_CS_STALL
)) {
5203 /* Project: PRE-SKL, VLV, CHV
5205 * "[All Stepping][All SKUs]:
5207 * One of the following must also be set:
5209 * - Render Target Cache Flush Enable ([12] of DW1)
5210 * - Depth Cache Flush Enable ([0] of DW1)
5211 * - Stall at Pixel Scoreboard ([1] of DW1)
5212 * - Depth Stall ([13] of DW1)
5213 * - Post-Sync Operation ([13] of DW1)
5214 * - DC Flush Enable ([5] of DW1)"
5216 * If we don't already have one of those bits set, we choose to add
5217 * "Stall at Pixel Scoreboard". Some of the other bits require a
5218 * CS stall as a workaround (see above), which would send us into
5219 * an infinite recursion of PIPE_CONTROLs. "Stall at Pixel Scoreboard"
5220 * appears to be safe, so we choose that.
5222 const uint32_t wa_bits
= PIPE_CONTROL_RENDER_TARGET_FLUSH
|
5223 PIPE_CONTROL_DEPTH_CACHE_FLUSH
|
5224 PIPE_CONTROL_WRITE_IMMEDIATE
|
5225 PIPE_CONTROL_WRITE_DEPTH_COUNT
|
5226 PIPE_CONTROL_WRITE_TIMESTAMP
|
5227 PIPE_CONTROL_STALL_AT_SCOREBOARD
|
5228 PIPE_CONTROL_DEPTH_STALL
|
5229 PIPE_CONTROL_DATA_CACHE_FLUSH
;
5230 if (!(flags
& wa_bits
))
5231 flags
|= PIPE_CONTROL_STALL_AT_SCOREBOARD
;
5234 /* Emit --------------------------------------------------------------- */
5236 iris_emit_cmd(batch
, GENX(PIPE_CONTROL
), pc
) {
5237 pc
.LRIPostSyncOperation
= NoLRIOperation
;
5238 pc
.PipeControlFlushEnable
= flags
& PIPE_CONTROL_FLUSH_ENABLE
;
5239 pc
.DCFlushEnable
= flags
& PIPE_CONTROL_DATA_CACHE_FLUSH
;
5240 pc
.StoreDataIndex
= 0;
5241 pc
.CommandStreamerStallEnable
= flags
& PIPE_CONTROL_CS_STALL
;
5242 pc
.GlobalSnapshotCountReset
=
5243 flags
& PIPE_CONTROL_GLOBAL_SNAPSHOT_COUNT_RESET
;
5244 pc
.TLBInvalidate
= flags
& PIPE_CONTROL_TLB_INVALIDATE
;
5245 pc
.GenericMediaStateClear
= flags
& PIPE_CONTROL_MEDIA_STATE_CLEAR
;
5246 pc
.StallAtPixelScoreboard
= flags
& PIPE_CONTROL_STALL_AT_SCOREBOARD
;
5247 pc
.RenderTargetCacheFlushEnable
=
5248 flags
& PIPE_CONTROL_RENDER_TARGET_FLUSH
;
5249 pc
.DepthCacheFlushEnable
= flags
& PIPE_CONTROL_DEPTH_CACHE_FLUSH
;
5250 pc
.StateCacheInvalidationEnable
=
5251 flags
& PIPE_CONTROL_STATE_CACHE_INVALIDATE
;
5252 pc
.VFCacheInvalidationEnable
= flags
& PIPE_CONTROL_VF_CACHE_INVALIDATE
;
5253 pc
.ConstantCacheInvalidationEnable
=
5254 flags
& PIPE_CONTROL_CONST_CACHE_INVALIDATE
;
5255 pc
.PostSyncOperation
= flags_to_post_sync_op(flags
);
5256 pc
.DepthStallEnable
= flags
& PIPE_CONTROL_DEPTH_STALL
;
5257 pc
.InstructionCacheInvalidateEnable
=
5258 flags
& PIPE_CONTROL_INSTRUCTION_INVALIDATE
;
5259 pc
.NotifyEnable
= flags
& PIPE_CONTROL_NOTIFY_ENABLE
;
5260 pc
.IndirectStatePointersDisable
=
5261 flags
& PIPE_CONTROL_INDIRECT_STATE_POINTERS_DISABLE
;
5262 pc
.TextureCacheInvalidationEnable
=
5263 flags
& PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
5264 pc
.Address
= rw_bo(bo
, offset
);
5265 pc
.ImmediateData
= imm
;
5270 genX(init_state
)(struct iris_context
*ice
)
5272 struct pipe_context
*ctx
= &ice
->ctx
;
5273 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
5275 ctx
->create_blend_state
= iris_create_blend_state
;
5276 ctx
->create_depth_stencil_alpha_state
= iris_create_zsa_state
;
5277 ctx
->create_rasterizer_state
= iris_create_rasterizer_state
;
5278 ctx
->create_sampler_state
= iris_create_sampler_state
;
5279 ctx
->create_sampler_view
= iris_create_sampler_view
;
5280 ctx
->create_surface
= iris_create_surface
;
5281 ctx
->create_vertex_elements_state
= iris_create_vertex_elements
;
5282 ctx
->bind_blend_state
= iris_bind_blend_state
;
5283 ctx
->bind_depth_stencil_alpha_state
= iris_bind_zsa_state
;
5284 ctx
->bind_sampler_states
= iris_bind_sampler_states
;
5285 ctx
->bind_rasterizer_state
= iris_bind_rasterizer_state
;
5286 ctx
->bind_vertex_elements_state
= iris_bind_vertex_elements_state
;
5287 ctx
->delete_blend_state
= iris_delete_state
;
5288 ctx
->delete_depth_stencil_alpha_state
= iris_delete_state
;
5289 ctx
->delete_fs_state
= iris_delete_state
;
5290 ctx
->delete_rasterizer_state
= iris_delete_state
;
5291 ctx
->delete_sampler_state
= iris_delete_state
;
5292 ctx
->delete_vertex_elements_state
= iris_delete_state
;
5293 ctx
->delete_tcs_state
= iris_delete_state
;
5294 ctx
->delete_tes_state
= iris_delete_state
;
5295 ctx
->delete_gs_state
= iris_delete_state
;
5296 ctx
->delete_vs_state
= iris_delete_state
;
5297 ctx
->set_blend_color
= iris_set_blend_color
;
5298 ctx
->set_clip_state
= iris_set_clip_state
;
5299 ctx
->set_constant_buffer
= iris_set_constant_buffer
;
5300 ctx
->set_shader_buffers
= iris_set_shader_buffers
;
5301 ctx
->set_shader_images
= iris_set_shader_images
;
5302 ctx
->set_sampler_views
= iris_set_sampler_views
;
5303 ctx
->set_tess_state
= iris_set_tess_state
;
5304 ctx
->set_framebuffer_state
= iris_set_framebuffer_state
;
5305 ctx
->set_polygon_stipple
= iris_set_polygon_stipple
;
5306 ctx
->set_sample_mask
= iris_set_sample_mask
;
5307 ctx
->set_scissor_states
= iris_set_scissor_states
;
5308 ctx
->set_stencil_ref
= iris_set_stencil_ref
;
5309 ctx
->set_vertex_buffers
= iris_set_vertex_buffers
;
5310 ctx
->set_viewport_states
= iris_set_viewport_states
;
5311 ctx
->sampler_view_destroy
= iris_sampler_view_destroy
;
5312 ctx
->surface_destroy
= iris_surface_destroy
;
5313 ctx
->draw_vbo
= iris_draw_vbo
;
5314 ctx
->launch_grid
= iris_launch_grid
;
5315 ctx
->create_stream_output_target
= iris_create_stream_output_target
;
5316 ctx
->stream_output_target_destroy
= iris_stream_output_target_destroy
;
5317 ctx
->set_stream_output_targets
= iris_set_stream_output_targets
;
5319 ice
->vtbl
.destroy_state
= iris_destroy_state
;
5320 ice
->vtbl
.init_render_context
= iris_init_render_context
;
5321 ice
->vtbl
.init_compute_context
= iris_init_compute_context
;
5322 ice
->vtbl
.upload_render_state
= iris_upload_render_state
;
5323 ice
->vtbl
.update_surface_base_address
= iris_update_surface_base_address
;
5324 ice
->vtbl
.upload_compute_state
= iris_upload_compute_state
;
5325 ice
->vtbl
.emit_raw_pipe_control
= iris_emit_raw_pipe_control
;
5326 ice
->vtbl
.load_register_imm32
= iris_load_register_imm32
;
5327 ice
->vtbl
.load_register_imm64
= iris_load_register_imm64
;
5328 ice
->vtbl
.load_register_mem32
= iris_load_register_mem32
;
5329 ice
->vtbl
.load_register_mem64
= iris_load_register_mem64
;
5330 ice
->vtbl
.store_register_mem32
= iris_store_register_mem32
;
5331 ice
->vtbl
.store_register_mem64
= iris_store_register_mem64
;
5332 ice
->vtbl
.store_data_imm32
= iris_store_data_imm32
;
5333 ice
->vtbl
.store_data_imm64
= iris_store_data_imm64
;
5334 ice
->vtbl
.copy_mem_mem
= iris_copy_mem_mem
;
5335 ice
->vtbl
.derived_program_state_size
= iris_derived_program_state_size
;
5336 ice
->vtbl
.store_derived_program_state
= iris_store_derived_program_state
;
5337 ice
->vtbl
.create_so_decl_list
= iris_create_so_decl_list
;
5338 ice
->vtbl
.populate_vs_key
= iris_populate_vs_key
;
5339 ice
->vtbl
.populate_tcs_key
= iris_populate_tcs_key
;
5340 ice
->vtbl
.populate_tes_key
= iris_populate_tes_key
;
5341 ice
->vtbl
.populate_gs_key
= iris_populate_gs_key
;
5342 ice
->vtbl
.populate_fs_key
= iris_populate_fs_key
;
5343 ice
->vtbl
.populate_cs_key
= iris_populate_cs_key
;
5345 ice
->state
.dirty
= ~0ull;
5347 ice
->state
.statistics_counters_enabled
= true;
5349 ice
->state
.sample_mask
= 0xffff;
5350 ice
->state
.num_viewports
= 1;
5351 ice
->state
.genx
= calloc(1, sizeof(struct iris_genx_state
));
5353 /* Make a 1x1x1 null surface for unbound textures */
5354 void *null_surf_map
=
5355 upload_state(ice
->state
.surface_uploader
, &ice
->state
.unbound_tex
,
5356 4 * GENX(RENDER_SURFACE_STATE_length
), 64);
5357 isl_null_fill_state(&screen
->isl_dev
, null_surf_map
, isl_extent3d(1, 1, 1));
5358 ice
->state
.unbound_tex
.offset
+=
5359 iris_bo_offset_from_base_address(iris_resource_bo(ice
->state
.unbound_tex
.res
));
5361 /* Default all scissor rectangles to be empty regions. */
5362 for (int i
= 0; i
< IRIS_MAX_VIEWPORTS
; i
++) {
5363 ice
->state
.scissors
[i
] = (struct pipe_scissor_state
) {
5364 .minx
= 1, .maxx
= 0, .miny
= 1, .maxy
= 0,