1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
4 * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
26 * Rob Clark <robclark@freedesktop.org>
29 #include "pipe/p_state.h"
30 #include "util/u_string.h"
31 #include "util/u_memory.h"
32 #include "util/u_inlines.h"
33 #include "util/u_format.h"
34 #include "tgsi/tgsi_dump.h"
35 #include "tgsi/tgsi_parse.h"
37 #include "freedreno_context.h"
38 #include "freedreno_util.h"
40 #include "ir3_shader.h"
41 #include "ir3_compiler.h"
45 ir3_glsl_type_size(const struct glsl_type
*type
)
47 return glsl_count_attribute_slots(type
, false);
51 delete_variant(struct ir3_shader_variant
*v
)
60 /* for vertex shader, the inputs are loaded into registers before the shader
61 * is executed, so max_regs from the shader instructions might not properly
62 * reflect the # of registers actually used, especially in case passthrough
65 * Likewise, for fragment shader, we can have some regs which are passed
66 * input values but never touched by the resulting shader (ie. as result
67 * of dead code elimination or simply because we don't know how to turn
71 fixup_regfootprint(struct ir3_shader_variant
*v
)
73 if (v
->type
== SHADER_VERTEX
) {
75 for (i
= 0; i
< v
->inputs_count
; i
++) {
76 /* skip frag inputs fetch via bary.f since their reg's are
77 * not written by gpu before shader starts (and in fact the
78 * regid's might not even be valid)
80 if (v
->inputs
[i
].bary
)
83 if (v
->inputs
[i
].compmask
) {
84 int32_t regid
= (v
->inputs
[i
].regid
+ 3) >> 2;
85 v
->info
.max_reg
= MAX2(v
->info
.max_reg
, regid
);
88 for (i
= 0; i
< v
->outputs_count
; i
++) {
89 int32_t regid
= (v
->outputs
[i
].regid
+ 3) >> 2;
90 v
->info
.max_reg
= MAX2(v
->info
.max_reg
, regid
);
92 } else if (v
->type
== SHADER_FRAGMENT
) {
93 /* NOTE: not sure how to turn pos_regid off.. but this could
94 * be, for example, r1.x while max reg used by the shader is
95 * r0.*, in which case we need to fixup the reg footprint:
97 v
->info
.max_reg
= MAX2(v
->info
.max_reg
, v
->pos_regid
>> 2);
99 debug_assert(v
->info
.max_reg
>= 0); /* hard coded r0.x */
101 debug_assert(v
->info
.max_half_reg
>= 0); /* hr0.x */
105 /* wrapper for ir3_assemble() which does some info fixup based on
106 * shader state. Non-static since used by ir3_cmdline too.
108 void * ir3_shader_assemble(struct ir3_shader_variant
*v
, uint32_t gpu_id
)
112 bin
= ir3_assemble(v
->ir
, &v
->info
, gpu_id
);
117 v
->instrlen
= v
->info
.sizedwords
/ (2 * 16);
119 v
->instrlen
= v
->info
.sizedwords
/ (2 * 4);
122 /* NOTE: if relative addressing is used, we set constlen in
123 * the compiler (to worst-case value) since we don't know in
124 * the assembler what the max addr reg value can be:
126 v
->constlen
= MIN2(255, MAX2(v
->constlen
, v
->info
.max_const
+ 1));
128 fixup_regfootprint(v
);
134 assemble_variant(struct ir3_shader_variant
*v
)
136 struct ir3_compiler
*compiler
= v
->shader
->compiler
;
137 uint32_t gpu_id
= compiler
->gpu_id
;
140 bin
= ir3_shader_assemble(v
, gpu_id
);
141 sz
= v
->info
.sizedwords
* 4;
143 v
->bo
= fd_bo_new(compiler
->dev
, sz
,
144 DRM_FREEDRENO_GEM_CACHE_WCOMBINE
|
145 DRM_FREEDRENO_GEM_TYPE_KMEM
);
147 memcpy(fd_bo_map(v
->bo
), bin
, sz
);
149 if (fd_mesa_debug
& FD_DBG_DISASM
) {
150 struct ir3_shader_key key
= v
->key
;
151 DBG("disassemble: type=%d, k={bp=%u,cts=%u,hp=%u}", v
->type
,
152 key
.binning_pass
, key
.color_two_side
, key
.half_precision
);
153 ir3_shader_disasm(v
, bin
);
158 /* no need to keep the ir around beyond this point: */
164 dump_shader_info(struct ir3_shader_variant
*v
, struct pipe_debug_callback
*debug
)
166 if (!unlikely(fd_mesa_debug
& FD_DBG_SHADERDB
))
169 pipe_debug_message(debug
, SHADER_INFO
, "\n"
170 "SHADER-DB: %s prog %d/%d: %u instructions, %u dwords\n"
171 "SHADER-DB: %s prog %d/%d: %u half, %u full\n"
172 "SHADER-DB: %s prog %d/%d: %u const, %u constlen\n",
173 ir3_shader_stage(v
->shader
),
174 v
->shader
->id
, v
->id
,
175 v
->info
.instrs_count
,
177 ir3_shader_stage(v
->shader
),
178 v
->shader
->id
, v
->id
,
179 v
->info
.max_half_reg
+ 1,
181 ir3_shader_stage(v
->shader
),
182 v
->shader
->id
, v
->id
,
183 v
->info
.max_const
+ 1,
187 static struct ir3_shader_variant
*
188 create_variant(struct ir3_shader
*shader
, struct ir3_shader_key key
)
190 struct ir3_shader_variant
*v
= CALLOC_STRUCT(ir3_shader_variant
);
196 v
->id
= ++shader
->variant_count
;
199 v
->type
= shader
->type
;
201 ret
= ir3_compile_shader_nir(shader
->compiler
, v
);
203 debug_error("compile failed!");
209 debug_error("assemble failed!");
220 struct ir3_shader_variant
*
221 ir3_shader_variant(struct ir3_shader
*shader
, struct ir3_shader_key key
,
222 struct pipe_debug_callback
*debug
)
224 struct ir3_shader_variant
*v
;
226 /* some shader key values only apply to vertex or frag shader,
227 * so normalize the key to avoid constructing multiple identical
230 switch (shader
->type
) {
231 case SHADER_FRAGMENT
:
232 key
.binning_pass
= false;
233 if (key
.has_per_samp
) {
241 key
.color_two_side
= false;
242 key
.half_precision
= false;
243 key
.rasterflat
= false;
244 if (key
.has_per_samp
) {
256 for (v
= shader
->variants
; v
; v
= v
->next
)
257 if (ir3_shader_key_equal(&key
, &v
->key
))
260 /* compile new variant if it doesn't exist already: */
261 v
= create_variant(shader
, key
);
263 v
->next
= shader
->variants
;
264 shader
->variants
= v
;
265 dump_shader_info(v
, debug
);
273 ir3_shader_destroy(struct ir3_shader
*shader
)
275 struct ir3_shader_variant
*v
, *t
;
276 for (v
= shader
->variants
; v
; ) {
281 ralloc_free(shader
->nir
);
286 ir3_shader_create(struct ir3_compiler
*compiler
,
287 const struct pipe_shader_state
*cso
, enum shader_t type
,
288 struct pipe_debug_callback
*debug
)
290 struct ir3_shader
*shader
= CALLOC_STRUCT(ir3_shader
);
291 shader
->compiler
= compiler
;
292 shader
->id
= ++shader
->compiler
->shader_count
;
296 if (cso
->type
== PIPE_SHADER_IR_NIR
) {
297 /* we take ownership of the reference: */
300 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, ir3_glsl_type_size
,
301 (nir_lower_io_options
)0);
303 debug_assert(cso
->type
== PIPE_SHADER_IR_TGSI
);
304 if (fd_mesa_debug
& FD_DBG_DISASM
) {
305 DBG("dump tgsi: type=%d", shader
->type
);
306 tgsi_dump(cso
->tokens
, 0);
308 nir
= ir3_tgsi_to_nir(cso
->tokens
);
310 /* do first pass optimization, ignoring the key: */
311 shader
->nir
= ir3_optimize_nir(shader
, nir
, NULL
);
312 if (fd_mesa_debug
& FD_DBG_DISASM
) {
313 DBG("dump nir%d: type=%d", shader
->id
, shader
->type
);
314 nir_print_shader(shader
->nir
, stdout
);
317 shader
->stream_output
= cso
->stream_output
;
318 if (fd_mesa_debug
& FD_DBG_SHADERDB
) {
319 /* if shader-db run, create a standard variant immediately
320 * (as otherwise nothing will trigger the shader to be
323 static struct ir3_shader_key key
;
324 memset(&key
, 0, sizeof(key
));
325 ir3_shader_variant(shader
, key
, debug
);
330 /* a bit annoying that compute-shader and normal shader state objects
331 * aren't a bit more aligned.
334 ir3_shader_create_compute(struct ir3_compiler
*compiler
,
335 const struct pipe_compute_state
*cso
,
336 struct pipe_debug_callback
*debug
)
338 struct ir3_shader
*shader
= CALLOC_STRUCT(ir3_shader
);
340 shader
->compiler
= compiler
;
341 shader
->id
= ++shader
->compiler
->shader_count
;
342 shader
->type
= SHADER_COMPUTE
;
345 if (cso
->ir_type
== PIPE_SHADER_IR_NIR
) {
346 /* we take ownership of the reference: */
347 nir
= (nir_shader
*)cso
->prog
;
349 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, ir3_glsl_type_size
,
350 (nir_lower_io_options
)0);
352 debug_assert(cso
->ir_type
== PIPE_SHADER_IR_TGSI
);
353 if (fd_mesa_debug
& FD_DBG_DISASM
) {
354 DBG("dump tgsi: type=%d", shader
->type
);
355 tgsi_dump(cso
->prog
, 0);
357 nir
= ir3_tgsi_to_nir(cso
->prog
);
360 /* do first pass optimization, ignoring the key: */
361 shader
->nir
= ir3_optimize_nir(shader
, nir
, NULL
);
362 if (fd_mesa_debug
& FD_DBG_DISASM
) {
363 DBG("dump nir%d: type=%d", shader
->id
, shader
->type
);
364 nir_print_shader(shader
->nir
, stdout
);
370 static void dump_reg(const char *name
, uint32_t r
)
372 if (r
!= regid(63,0))
373 debug_printf("; %s: r%d.%c\n", name
, r
>> 2, "xyzw"[r
& 0x3]);
376 static void dump_output(struct ir3_shader_variant
*so
,
377 unsigned slot
, const char *name
)
380 regid
= ir3_find_output_regid(so
, slot
);
381 dump_reg(name
, regid
);
385 ir3_shader_disasm(struct ir3_shader_variant
*so
, uint32_t *bin
)
387 struct ir3
*ir
= so
->ir
;
388 struct ir3_register
*reg
;
389 const char *type
= ir3_shader_stage(so
->shader
);
393 for (i
= 0; i
< ir
->ninputs
; i
++) {
394 if (!ir
->inputs
[i
]) {
395 debug_printf("; in%d unused\n", i
);
398 reg
= ir
->inputs
[i
]->regs
[0];
400 debug_printf("@in(%sr%d.%c)\tin%d\n",
401 (reg
->flags
& IR3_REG_HALF
) ? "h" : "",
402 (regid
>> 2), "xyzw"[regid
& 0x3], i
);
405 for (i
= 0; i
< ir
->noutputs
; i
++) {
406 if (!ir
->outputs
[i
]) {
407 debug_printf("; out%d unused\n", i
);
410 /* kill shows up as a virtual output.. skip it! */
411 if (is_kill(ir
->outputs
[i
]))
413 reg
= ir
->outputs
[i
]->regs
[0];
415 debug_printf("@out(%sr%d.%c)\tout%d\n",
416 (reg
->flags
& IR3_REG_HALF
) ? "h" : "",
417 (regid
>> 2), "xyzw"[regid
& 0x3], i
);
420 for (i
= 0; i
< so
->immediates_count
; i
++) {
421 debug_printf("@const(c%d.x)\t", so
->constbase
.immediate
+ i
);
422 debug_printf("0x%08x, 0x%08x, 0x%08x, 0x%08x\n",
423 so
->immediates
[i
].val
[0],
424 so
->immediates
[i
].val
[1],
425 so
->immediates
[i
].val
[2],
426 so
->immediates
[i
].val
[3]);
429 disasm_a3xx(bin
, so
->info
.sizedwords
, 0, so
->type
);
433 debug_printf("; %s: outputs:", type
);
434 for (i
= 0; i
< so
->outputs_count
; i
++) {
435 uint8_t regid
= so
->outputs
[i
].regid
;
436 debug_printf(" r%d.%c (%s)",
437 (regid
>> 2), "xyzw"[regid
& 0x3],
438 gl_varying_slot_name(so
->outputs
[i
].slot
));
441 debug_printf("; %s: inputs:", type
);
442 for (i
= 0; i
< so
->inputs_count
; i
++) {
443 uint8_t regid
= so
->inputs
[i
].regid
;
444 debug_printf(" r%d.%c (cm=%x,il=%u,b=%u)",
445 (regid
>> 2), "xyzw"[regid
& 0x3],
446 so
->inputs
[i
].compmask
,
452 case SHADER_FRAGMENT
:
453 debug_printf("; %s: outputs:", type
);
454 for (i
= 0; i
< so
->outputs_count
; i
++) {
455 uint8_t regid
= so
->outputs
[i
].regid
;
456 debug_printf(" r%d.%c (%s)",
457 (regid
>> 2), "xyzw"[regid
& 0x3],
458 gl_frag_result_name(so
->outputs
[i
].slot
));
461 debug_printf("; %s: inputs:", type
);
462 for (i
= 0; i
< so
->inputs_count
; i
++) {
463 uint8_t regid
= so
->inputs
[i
].regid
;
464 debug_printf(" r%d.%c (%s,cm=%x,il=%u,b=%u)",
465 (regid
>> 2), "xyzw"[regid
& 0x3],
466 gl_varying_slot_name(so
->inputs
[i
].slot
),
467 so
->inputs
[i
].compmask
,
478 /* print generic shader info: */
479 debug_printf("; %s prog %d/%d: %u instructions, %d half, %d full\n",
480 type
, so
->shader
->id
, so
->id
,
481 so
->info
.instrs_count
,
482 so
->info
.max_half_reg
+ 1,
483 so
->info
.max_reg
+ 1);
485 debug_printf("; %d const, %u constlen\n",
486 so
->info
.max_const
+ 1,
489 /* print shader type specific info: */
492 dump_output(so
, VARYING_SLOT_POS
, "pos");
493 dump_output(so
, VARYING_SLOT_PSIZ
, "psize");
495 case SHADER_FRAGMENT
:
496 dump_reg("pos (bary)", so
->pos_regid
);
497 dump_output(so
, FRAG_RESULT_DEPTH
, "posz");
498 if (so
->color0_mrt
) {
499 dump_output(so
, FRAG_RESULT_COLOR
, "color");
501 dump_output(so
, FRAG_RESULT_DATA0
, "data0");
502 dump_output(so
, FRAG_RESULT_DATA1
, "data1");
503 dump_output(so
, FRAG_RESULT_DATA2
, "data2");
504 dump_output(so
, FRAG_RESULT_DATA3
, "data3");
505 dump_output(so
, FRAG_RESULT_DATA4
, "data4");
506 dump_output(so
, FRAG_RESULT_DATA5
, "data5");
507 dump_output(so
, FRAG_RESULT_DATA6
, "data6");
508 dump_output(so
, FRAG_RESULT_DATA7
, "data7");
510 /* these two are hard-coded since we don't know how to
511 * program them to anything but all 0's...
514 debug_printf("; fragcoord: r0.x\n");
516 debug_printf("; fragface: hr0.x\n");
527 ir3_shader_outputs(const struct ir3_shader
*so
)
529 return so
->nir
->info
.outputs_written
;
532 /* This has to reach into the fd_context a bit more than the rest of
533 * ir3, but it needs to be aligned with the compiler, so both agree
534 * on which const regs hold what. And the logic is identical between
535 * a3xx/a4xx, the only difference is small details in the actual
536 * CP_LOAD_STATE packets (which is handled inside the generation
537 * specific ctx->emit_const(_bo)() fxns)
540 #include "freedreno_resource.h"
543 emit_user_consts(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
,
544 struct fd_ringbuffer
*ring
, struct fd_constbuf_stateobj
*constbuf
)
546 const unsigned index
= 0; /* user consts are index 0 */
547 /* TODO save/restore dirty_mask for binning pass instead: */
548 uint32_t dirty_mask
= constbuf
->enabled_mask
;
550 if (dirty_mask
& (1 << index
)) {
551 struct pipe_constant_buffer
*cb
= &constbuf
->cb
[index
];
552 unsigned size
= align(cb
->buffer_size
, 4) / 4; /* size in dwords */
554 /* in particular, with binning shader we may end up with
555 * unused consts, ie. we could end up w/ constlen that is
556 * smaller than first_driver_param. In that case truncate
557 * the user consts early to avoid HLSQ lockup caused by
558 * writing too many consts
560 uint32_t max_const
= MIN2(v
->num_uniforms
, v
->constlen
);
562 // I expect that size should be a multiple of vec4's:
563 assert(size
== align(size
, 4));
565 /* and even if the start of the const buffer is before
566 * first_immediate, the end may not be:
568 size
= MIN2(size
, 4 * max_const
);
571 fd_wfi(ctx
->batch
, ring
);
572 ctx
->emit_const(ring
, v
->type
, 0,
573 cb
->buffer_offset
, size
,
574 cb
->user_buffer
, cb
->buffer
);
575 constbuf
->dirty_mask
&= ~(1 << index
);
581 emit_ubos(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
,
582 struct fd_ringbuffer
*ring
, struct fd_constbuf_stateobj
*constbuf
)
584 uint32_t offset
= v
->constbase
.ubo
;
585 if (v
->constlen
> offset
) {
586 uint32_t params
= v
->num_ubos
;
587 uint32_t offsets
[params
];
588 struct pipe_resource
*prscs
[params
];
590 for (uint32_t i
= 0; i
< params
; i
++) {
591 const uint32_t index
= i
+ 1; /* UBOs start at index 1 */
592 struct pipe_constant_buffer
*cb
= &constbuf
->cb
[index
];
593 assert(!cb
->user_buffer
);
595 if ((constbuf
->enabled_mask
& (1 << index
)) && cb
->buffer
) {
596 offsets
[i
] = cb
->buffer_offset
;
597 prscs
[i
] = cb
->buffer
;
604 fd_wfi(ctx
->batch
, ring
);
605 ctx
->emit_const_bo(ring
, v
->type
, false, offset
* 4, params
, prscs
, offsets
);
610 emit_ssbo_sizes(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
,
611 struct fd_ringbuffer
*ring
, struct fd_shaderbuf_stateobj
*sb
)
613 uint32_t offset
= v
->constbase
.ssbo_sizes
;
614 if (v
->constlen
> offset
) {
615 uint32_t sizes
[align(v
->const_layout
.ssbo_size
.count
, 4)];
616 unsigned mask
= v
->const_layout
.ssbo_size
.mask
;
619 unsigned index
= u_bit_scan(&mask
);
620 unsigned off
= v
->const_layout
.ssbo_size
.off
[index
];
621 sizes
[off
] = sb
->sb
[index
].buffer_size
;
624 fd_wfi(ctx
->batch
, ring
);
625 ctx
->emit_const(ring
, v
->type
, offset
* 4,
626 0, ARRAY_SIZE(sizes
), sizes
, NULL
);
631 emit_immediates(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
,
632 struct fd_ringbuffer
*ring
)
634 int size
= v
->immediates_count
;
635 uint32_t base
= v
->constbase
.immediate
;
637 /* truncate size to avoid writing constants that shader
640 size
= MIN2(size
+ base
, v
->constlen
) - base
;
642 /* convert out of vec4: */
647 fd_wfi(ctx
->batch
, ring
);
648 ctx
->emit_const(ring
, v
->type
, base
,
649 0, size
, v
->immediates
[0].val
, NULL
);
653 /* emit stream-out buffers: */
655 emit_tfbos(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
,
656 struct fd_ringbuffer
*ring
)
658 /* streamout addresses after driver-params: */
659 uint32_t offset
= v
->constbase
.tfbo
;
660 if (v
->constlen
> offset
) {
661 struct fd_streamout_stateobj
*so
= &ctx
->streamout
;
662 struct pipe_stream_output_info
*info
= &v
->shader
->stream_output
;
664 uint32_t offsets
[params
];
665 struct pipe_resource
*prscs
[params
];
667 for (uint32_t i
= 0; i
< params
; i
++) {
668 struct pipe_stream_output_target
*target
= so
->targets
[i
];
671 offsets
[i
] = (so
->offsets
[i
] * info
->stride
[i
] * 4) +
672 target
->buffer_offset
;
673 prscs
[i
] = target
->buffer
;
680 fd_wfi(ctx
->batch
, ring
);
681 ctx
->emit_const_bo(ring
, v
->type
, true, offset
* 4, params
, prscs
, offsets
);
686 max_tf_vtx(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
)
688 struct fd_streamout_stateobj
*so
= &ctx
->streamout
;
689 struct pipe_stream_output_info
*info
= &v
->shader
->stream_output
;
690 uint32_t maxvtxcnt
= 0x7fffffff;
692 if (ctx
->screen
->gpu_id
>= 500)
694 if (v
->key
.binning_pass
)
696 if (v
->shader
->stream_output
.num_outputs
== 0)
698 if (so
->num_targets
== 0)
701 /* offset to write to is:
703 * total_vtxcnt = vtxcnt + offsets[i]
704 * offset = total_vtxcnt * stride[i]
706 * offset = vtxcnt * stride[i] ; calculated in shader
707 * + offsets[i] * stride[i] ; calculated at emit_tfbos()
709 * assuming for each vtx, each target buffer will have data written
710 * up to 'offset + stride[i]', that leaves maxvtxcnt as:
712 * buffer_size = (maxvtxcnt * stride[i]) + stride[i]
713 * maxvtxcnt = (buffer_size - stride[i]) / stride[i]
715 * but shader is actually doing a less-than (rather than less-than-
716 * equal) check, so we can drop the -stride[i].
718 * TODO is assumption about `offset + stride[i]` legit?
720 for (unsigned i
= 0; i
< so
->num_targets
; i
++) {
721 struct pipe_stream_output_target
*target
= so
->targets
[i
];
722 unsigned stride
= info
->stride
[i
] * 4; /* convert dwords->bytes */
724 uint32_t max
= target
->buffer_size
/ stride
;
725 maxvtxcnt
= MIN2(maxvtxcnt
, max
);
733 emit_common_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
734 struct fd_context
*ctx
, enum pipe_shader_type t
)
736 enum fd_dirty_shader_state dirty
= ctx
->dirty_shader
[t
];
738 if (dirty
& (FD_DIRTY_SHADER_PROG
| FD_DIRTY_SHADER_CONST
)) {
739 struct fd_constbuf_stateobj
*constbuf
;
742 constbuf
= &ctx
->constbuf
[t
];
743 shader_dirty
= !!(dirty
& FD_DIRTY_SHADER_PROG
);
745 emit_user_consts(ctx
, v
, ring
, constbuf
);
746 emit_ubos(ctx
, v
, ring
, constbuf
);
748 emit_immediates(ctx
, v
, ring
);
751 if (dirty
& (FD_DIRTY_SHADER_PROG
| FD_DIRTY_SHADER_SSBO
)) {
752 struct fd_shaderbuf_stateobj
*sb
= &ctx
->shaderbuf
[t
];
753 emit_ssbo_sizes(ctx
, v
, ring
, sb
);
758 ir3_emit_vs_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
759 struct fd_context
*ctx
, const struct pipe_draw_info
*info
)
761 debug_assert(v
->type
== SHADER_VERTEX
);
763 emit_common_consts(v
, ring
, ctx
, PIPE_SHADER_VERTEX
);
765 /* emit driver params every time: */
766 /* TODO skip emit if shader doesn't use driver params to avoid WFI.. */
768 uint32_t offset
= v
->constbase
.driver_param
;
769 if (v
->constlen
> offset
) {
770 uint32_t vertex_params
[IR3_DP_VS_COUNT
] = {
771 [IR3_DP_VTXID_BASE
] = info
->index_size
?
772 info
->index_bias
: info
->start
,
773 [IR3_DP_VTXCNT_MAX
] = max_tf_vtx(ctx
, v
),
775 /* if no user-clip-planes, we don't need to emit the
778 uint32_t vertex_params_size
= 4;
780 if (v
->key
.ucp_enables
) {
781 struct pipe_clip_state
*ucp
= &ctx
->ucp
;
782 unsigned pos
= IR3_DP_UCP0_X
;
783 for (unsigned i
= 0; pos
<= IR3_DP_UCP7_W
; i
++) {
784 for (unsigned j
= 0; j
< 4; j
++) {
785 vertex_params
[pos
] = fui(ucp
->ucp
[i
][j
]);
789 vertex_params_size
= ARRAY_SIZE(vertex_params
);
792 fd_wfi(ctx
->batch
, ring
);
793 ctx
->emit_const(ring
, SHADER_VERTEX
, offset
* 4, 0,
794 vertex_params_size
, vertex_params
, NULL
);
796 /* if needed, emit stream-out buffer addresses: */
797 if (vertex_params
[IR3_DP_VTXCNT_MAX
] > 0) {
798 emit_tfbos(ctx
, v
, ring
);
805 ir3_emit_fs_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
806 struct fd_context
*ctx
)
808 debug_assert(v
->type
== SHADER_FRAGMENT
);
810 emit_common_consts(v
, ring
, ctx
, PIPE_SHADER_FRAGMENT
);
813 /* emit compute-shader consts: */
815 ir3_emit_cs_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
816 struct fd_context
*ctx
, const struct pipe_grid_info
*info
)
818 debug_assert(v
->type
== SHADER_COMPUTE
);
820 emit_common_consts(v
, ring
, ctx
, PIPE_SHADER_COMPUTE
);
822 /* emit compute-shader driver-params: */
823 uint32_t offset
= v
->constbase
.driver_param
;
824 if (v
->constlen
> offset
) {
825 uint32_t compute_params
[IR3_DP_CS_COUNT
] = {
826 [IR3_DP_NUM_WORK_GROUPS_X
] = info
->grid
[0],
827 [IR3_DP_NUM_WORK_GROUPS_Y
] = info
->grid
[1],
828 [IR3_DP_NUM_WORK_GROUPS_Z
] = info
->grid
[2],
829 /* do we need work-group-size? */
832 fd_wfi(ctx
->batch
, ring
);
833 ctx
->emit_const(ring
, SHADER_COMPUTE
, offset
* 4, 0,
834 ARRAY_SIZE(compute_params
), compute_params
, NULL
);