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
)
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 /* ignore high regs that are global to all threads in a warp
84 * (they exist by default) (a5xx+)
86 if (v
->inputs
[i
].regid
>= regid(48,0))
89 if (v
->inputs
[i
].compmask
) {
90 unsigned n
= util_last_bit(v
->inputs
[i
].compmask
) - 1;
91 int32_t regid
= (v
->inputs
[i
].regid
+ n
) >> 2;
92 v
->info
.max_reg
= MAX2(v
->info
.max_reg
, regid
);
96 for (i
= 0; i
< v
->outputs_count
; i
++) {
97 int32_t regid
= (v
->outputs
[i
].regid
+ 3) >> 2;
98 v
->info
.max_reg
= MAX2(v
->info
.max_reg
, regid
);
102 /* wrapper for ir3_assemble() which does some info fixup based on
103 * shader state. Non-static since used by ir3_cmdline too.
105 void * ir3_shader_assemble(struct ir3_shader_variant
*v
, uint32_t gpu_id
)
109 bin
= ir3_assemble(v
->ir
, &v
->info
, gpu_id
);
114 v
->instrlen
= v
->info
.sizedwords
/ (2 * 16);
116 v
->instrlen
= v
->info
.sizedwords
/ (2 * 4);
119 /* NOTE: if relative addressing is used, we set constlen in
120 * the compiler (to worst-case value) since we don't know in
121 * the assembler what the max addr reg value can be:
123 v
->constlen
= MIN2(255, MAX2(v
->constlen
, v
->info
.max_const
+ 1));
125 fixup_regfootprint(v
);
131 assemble_variant(struct ir3_shader_variant
*v
)
133 struct ir3_compiler
*compiler
= v
->shader
->compiler
;
134 uint32_t gpu_id
= compiler
->gpu_id
;
137 bin
= ir3_shader_assemble(v
, gpu_id
);
138 sz
= v
->info
.sizedwords
* 4;
140 v
->bo
= fd_bo_new(compiler
->dev
, sz
,
141 DRM_FREEDRENO_GEM_CACHE_WCOMBINE
|
142 DRM_FREEDRENO_GEM_TYPE_KMEM
);
144 memcpy(fd_bo_map(v
->bo
), bin
, sz
);
146 if (fd_mesa_debug
& FD_DBG_DISASM
) {
147 struct ir3_shader_key key
= v
->key
;
148 printf("disassemble: type=%d, k={bp=%u,cts=%u,hp=%u}", v
->type
,
149 key
.binning_pass
, key
.color_two_side
, key
.half_precision
);
150 ir3_shader_disasm(v
, bin
, stdout
);
153 if (shader_debug_enabled(v
->shader
->type
)) {
154 fprintf(stderr
, "Native code for unnamed %s shader %s:\n",
155 shader_stage_name(v
->shader
->type
), v
->shader
->nir
->info
.name
);
156 if (v
->shader
->type
== SHADER_FRAGMENT
)
157 fprintf(stderr
, "SIMD0\n");
158 ir3_shader_disasm(v
, bin
, stderr
);
163 /* no need to keep the ir around beyond this point: */
169 dump_shader_info(struct ir3_shader_variant
*v
, struct pipe_debug_callback
*debug
)
171 if (!unlikely(fd_mesa_debug
& FD_DBG_SHADERDB
))
174 pipe_debug_message(debug
, SHADER_INFO
, "\n"
175 "SHADER-DB: %s prog %d/%d: %u instructions, %u dwords\n"
176 "SHADER-DB: %s prog %d/%d: %u half, %u full\n"
177 "SHADER-DB: %s prog %d/%d: %u const, %u constlen\n"
178 "SHADER-DB: %s prog %d/%d: %u (ss), %u (sy)\n",
179 ir3_shader_stage(v
->shader
),
180 v
->shader
->id
, v
->id
,
181 v
->info
.instrs_count
,
183 ir3_shader_stage(v
->shader
),
184 v
->shader
->id
, v
->id
,
185 v
->info
.max_half_reg
+ 1,
187 ir3_shader_stage(v
->shader
),
188 v
->shader
->id
, v
->id
,
189 v
->info
.max_const
+ 1,
191 ir3_shader_stage(v
->shader
),
192 v
->shader
->id
, v
->id
,
193 v
->info
.ss
, v
->info
.sy
);
196 static struct ir3_shader_variant
*
197 create_variant(struct ir3_shader
*shader
, struct ir3_shader_key key
)
199 struct ir3_shader_variant
*v
= CALLOC_STRUCT(ir3_shader_variant
);
205 v
->id
= ++shader
->variant_count
;
208 v
->type
= shader
->type
;
210 ret
= ir3_compile_shader_nir(shader
->compiler
, v
);
212 debug_error("compile failed!");
218 debug_error("assemble failed!");
229 struct ir3_shader_variant
*
230 ir3_shader_variant(struct ir3_shader
*shader
, struct ir3_shader_key key
,
231 struct pipe_debug_callback
*debug
)
233 struct ir3_shader_variant
*v
;
235 /* some shader key values only apply to vertex or frag shader,
236 * so normalize the key to avoid constructing multiple identical
239 switch (shader
->type
) {
240 case SHADER_FRAGMENT
:
241 key
.binning_pass
= false;
242 if (key
.has_per_samp
) {
251 key
.color_two_side
= false;
252 key
.half_precision
= false;
253 key
.rasterflat
= false;
254 if (key
.has_per_samp
) {
267 for (v
= shader
->variants
; v
; v
= v
->next
)
268 if (ir3_shader_key_equal(&key
, &v
->key
))
271 /* compile new variant if it doesn't exist already: */
272 v
= create_variant(shader
, key
);
274 v
->next
= shader
->variants
;
275 shader
->variants
= v
;
276 dump_shader_info(v
, debug
);
284 ir3_shader_destroy(struct ir3_shader
*shader
)
286 struct ir3_shader_variant
*v
, *t
;
287 for (v
= shader
->variants
; v
; ) {
292 ralloc_free(shader
->nir
);
297 ir3_shader_create(struct ir3_compiler
*compiler
,
298 const struct pipe_shader_state
*cso
, enum shader_t type
,
299 struct pipe_debug_callback
*debug
)
301 struct ir3_shader
*shader
= CALLOC_STRUCT(ir3_shader
);
302 shader
->compiler
= compiler
;
303 shader
->id
= ++shader
->compiler
->shader_count
;
307 if (cso
->type
== PIPE_SHADER_IR_NIR
) {
308 /* we take ownership of the reference: */
311 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, ir3_glsl_type_size
,
312 (nir_lower_io_options
)0);
314 debug_assert(cso
->type
== PIPE_SHADER_IR_TGSI
);
315 if (fd_mesa_debug
& FD_DBG_DISASM
) {
316 DBG("dump tgsi: type=%d", shader
->type
);
317 tgsi_dump(cso
->tokens
, 0);
319 nir
= ir3_tgsi_to_nir(cso
->tokens
);
321 /* do first pass optimization, ignoring the key: */
322 shader
->nir
= ir3_optimize_nir(shader
, nir
, NULL
);
323 if (fd_mesa_debug
& FD_DBG_DISASM
) {
324 DBG("dump nir%d: type=%d", shader
->id
, shader
->type
);
325 nir_print_shader(shader
->nir
, stdout
);
328 shader
->stream_output
= cso
->stream_output
;
329 if (fd_mesa_debug
& FD_DBG_SHADERDB
) {
330 /* if shader-db run, create a standard variant immediately
331 * (as otherwise nothing will trigger the shader to be
334 static struct ir3_shader_key key
;
335 memset(&key
, 0, sizeof(key
));
336 ir3_shader_variant(shader
, key
, debug
);
341 /* a bit annoying that compute-shader and normal shader state objects
342 * aren't a bit more aligned.
345 ir3_shader_create_compute(struct ir3_compiler
*compiler
,
346 const struct pipe_compute_state
*cso
,
347 struct pipe_debug_callback
*debug
)
349 struct ir3_shader
*shader
= CALLOC_STRUCT(ir3_shader
);
351 shader
->compiler
= compiler
;
352 shader
->id
= ++shader
->compiler
->shader_count
;
353 shader
->type
= SHADER_COMPUTE
;
356 if (cso
->ir_type
== PIPE_SHADER_IR_NIR
) {
357 /* we take ownership of the reference: */
358 nir
= (nir_shader
*)cso
->prog
;
360 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, ir3_glsl_type_size
,
361 (nir_lower_io_options
)0);
363 debug_assert(cso
->ir_type
== PIPE_SHADER_IR_TGSI
);
364 if (fd_mesa_debug
& FD_DBG_DISASM
) {
365 DBG("dump tgsi: type=%d", shader
->type
);
366 tgsi_dump(cso
->prog
, 0);
368 nir
= ir3_tgsi_to_nir(cso
->prog
);
371 /* do first pass optimization, ignoring the key: */
372 shader
->nir
= ir3_optimize_nir(shader
, nir
, NULL
);
373 if (fd_mesa_debug
& FD_DBG_DISASM
) {
374 printf("dump nir%d: type=%d\n", shader
->id
, shader
->type
);
375 nir_print_shader(shader
->nir
, stdout
);
381 static void dump_reg(FILE *out
, const char *name
, uint32_t r
)
383 if (r
!= regid(63,0))
384 fprintf(out
, "; %s: r%d.%c\n", name
, r
>> 2, "xyzw"[r
& 0x3]);
387 static void dump_output(FILE *out
, struct ir3_shader_variant
*so
,
388 unsigned slot
, const char *name
)
391 regid
= ir3_find_output_regid(so
, slot
);
392 dump_reg(out
, name
, regid
);
396 ir3_shader_disasm(struct ir3_shader_variant
*so
, uint32_t *bin
, FILE *out
)
398 struct ir3
*ir
= so
->ir
;
399 struct ir3_register
*reg
;
400 const char *type
= ir3_shader_stage(so
->shader
);
404 for (i
= 0; i
< ir
->ninputs
; i
++) {
405 if (!ir
->inputs
[i
]) {
406 fprintf(out
, "; in%d unused\n", i
);
409 reg
= ir
->inputs
[i
]->regs
[0];
411 fprintf(out
, "@in(%sr%d.%c)\tin%d\n",
412 (reg
->flags
& IR3_REG_HALF
) ? "h" : "",
413 (regid
>> 2), "xyzw"[regid
& 0x3], i
);
416 for (i
= 0; i
< ir
->noutputs
; i
++) {
417 if (!ir
->outputs
[i
]) {
418 fprintf(out
, "; out%d unused\n", i
);
421 /* kill shows up as a virtual output.. skip it! */
422 if (is_kill(ir
->outputs
[i
]))
424 reg
= ir
->outputs
[i
]->regs
[0];
426 fprintf(out
, "@out(%sr%d.%c)\tout%d\n",
427 (reg
->flags
& IR3_REG_HALF
) ? "h" : "",
428 (regid
>> 2), "xyzw"[regid
& 0x3], i
);
431 for (i
= 0; i
< so
->immediates_count
; i
++) {
432 fprintf(out
, "@const(c%d.x)\t", so
->constbase
.immediate
+ i
);
433 fprintf(out
, "0x%08x, 0x%08x, 0x%08x, 0x%08x\n",
434 so
->immediates
[i
].val
[0],
435 so
->immediates
[i
].val
[1],
436 so
->immediates
[i
].val
[2],
437 so
->immediates
[i
].val
[3]);
440 disasm_a3xx(bin
, so
->info
.sizedwords
, 0, out
);
444 fprintf(out
, "; %s: outputs:", type
);
445 for (i
= 0; i
< so
->outputs_count
; i
++) {
446 uint8_t regid
= so
->outputs
[i
].regid
;
447 fprintf(out
, " r%d.%c (%s)",
448 (regid
>> 2), "xyzw"[regid
& 0x3],
449 gl_varying_slot_name(so
->outputs
[i
].slot
));
452 fprintf(out
, "; %s: inputs:", type
);
453 for (i
= 0; i
< so
->inputs_count
; i
++) {
454 uint8_t regid
= so
->inputs
[i
].regid
;
455 fprintf(out
, " r%d.%c (cm=%x,il=%u,b=%u)",
456 (regid
>> 2), "xyzw"[regid
& 0x3],
457 so
->inputs
[i
].compmask
,
463 case SHADER_FRAGMENT
:
464 fprintf(out
, "; %s: outputs:", type
);
465 for (i
= 0; i
< so
->outputs_count
; i
++) {
466 uint8_t regid
= so
->outputs
[i
].regid
;
467 fprintf(out
, " r%d.%c (%s)",
468 (regid
>> 2), "xyzw"[regid
& 0x3],
469 gl_frag_result_name(so
->outputs
[i
].slot
));
472 fprintf(out
, "; %s: inputs:", type
);
473 for (i
= 0; i
< so
->inputs_count
; i
++) {
474 uint8_t regid
= so
->inputs
[i
].regid
;
475 fprintf(out
, " r%d.%c (%s,cm=%x,il=%u,b=%u)",
476 (regid
>> 2), "xyzw"[regid
& 0x3],
477 gl_varying_slot_name(so
->inputs
[i
].slot
),
478 so
->inputs
[i
].compmask
,
489 /* print generic shader info: */
490 fprintf(out
, "; %s prog %d/%d: %u instructions, %d half, %d full\n",
491 type
, so
->shader
->id
, so
->id
,
492 so
->info
.instrs_count
,
493 so
->info
.max_half_reg
+ 1,
494 so
->info
.max_reg
+ 1);
496 fprintf(out
, "; %d const, %u constlen\n",
497 so
->info
.max_const
+ 1,
500 fprintf(out
, "; %u (ss), %u (sy)\n", so
->info
.ss
, so
->info
.sy
);
502 /* print shader type specific info: */
505 dump_output(out
, so
, VARYING_SLOT_POS
, "pos");
506 dump_output(out
, so
, VARYING_SLOT_PSIZ
, "psize");
508 case SHADER_FRAGMENT
:
509 dump_reg(out
, "pos (bary)",
510 ir3_find_sysval_regid(so
, SYSTEM_VALUE_VARYING_COORD
));
511 dump_output(out
, so
, FRAG_RESULT_DEPTH
, "posz");
512 if (so
->color0_mrt
) {
513 dump_output(out
, so
, FRAG_RESULT_COLOR
, "color");
515 dump_output(out
, so
, FRAG_RESULT_DATA0
, "data0");
516 dump_output(out
, so
, FRAG_RESULT_DATA1
, "data1");
517 dump_output(out
, so
, FRAG_RESULT_DATA2
, "data2");
518 dump_output(out
, so
, FRAG_RESULT_DATA3
, "data3");
519 dump_output(out
, so
, FRAG_RESULT_DATA4
, "data4");
520 dump_output(out
, so
, FRAG_RESULT_DATA5
, "data5");
521 dump_output(out
, so
, FRAG_RESULT_DATA6
, "data6");
522 dump_output(out
, so
, FRAG_RESULT_DATA7
, "data7");
524 /* these two are hard-coded since we don't know how to
525 * program them to anything but all 0's...
528 fprintf(out
, "; fragcoord: r0.x\n");
530 fprintf(out
, "; fragface: hr0.x\n");
541 ir3_shader_outputs(const struct ir3_shader
*so
)
543 return so
->nir
->info
.outputs_written
;
546 /* This has to reach into the fd_context a bit more than the rest of
547 * ir3, but it needs to be aligned with the compiler, so both agree
548 * on which const regs hold what. And the logic is identical between
549 * a3xx/a4xx, the only difference is small details in the actual
550 * CP_LOAD_STATE packets (which is handled inside the generation
551 * specific ctx->emit_const(_bo)() fxns)
554 #include "freedreno_resource.h"
557 emit_user_consts(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
,
558 struct fd_ringbuffer
*ring
, struct fd_constbuf_stateobj
*constbuf
)
560 const unsigned index
= 0; /* user consts are index 0 */
562 if (constbuf
->enabled_mask
& (1 << index
)) {
563 struct pipe_constant_buffer
*cb
= &constbuf
->cb
[index
];
564 unsigned size
= align(cb
->buffer_size
, 4) / 4; /* size in dwords */
566 /* in particular, with binning shader we may end up with
567 * unused consts, ie. we could end up w/ constlen that is
568 * smaller than first_driver_param. In that case truncate
569 * the user consts early to avoid HLSQ lockup caused by
570 * writing too many consts
572 uint32_t max_const
= MIN2(v
->num_uniforms
, v
->constlen
);
574 // I expect that size should be a multiple of vec4's:
575 assert(size
== align(size
, 4));
577 /* and even if the start of the const buffer is before
578 * first_immediate, the end may not be:
580 size
= MIN2(size
, 4 * max_const
);
583 fd_wfi(ctx
->batch
, ring
);
584 ctx
->emit_const(ring
, v
->type
, 0,
585 cb
->buffer_offset
, size
,
586 cb
->user_buffer
, cb
->buffer
);
592 emit_ubos(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
,
593 struct fd_ringbuffer
*ring
, struct fd_constbuf_stateobj
*constbuf
)
595 uint32_t offset
= v
->constbase
.ubo
;
596 if (v
->constlen
> offset
) {
597 uint32_t params
= v
->num_ubos
;
598 uint32_t offsets
[params
];
599 struct pipe_resource
*prscs
[params
];
601 for (uint32_t i
= 0; i
< params
; i
++) {
602 const uint32_t index
= i
+ 1; /* UBOs start at index 1 */
603 struct pipe_constant_buffer
*cb
= &constbuf
->cb
[index
];
604 assert(!cb
->user_buffer
);
606 if ((constbuf
->enabled_mask
& (1 << index
)) && cb
->buffer
) {
607 offsets
[i
] = cb
->buffer_offset
;
608 prscs
[i
] = cb
->buffer
;
615 fd_wfi(ctx
->batch
, ring
);
616 ctx
->emit_const_bo(ring
, v
->type
, false, offset
* 4, params
, prscs
, offsets
);
621 emit_ssbo_sizes(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
,
622 struct fd_ringbuffer
*ring
, struct fd_shaderbuf_stateobj
*sb
)
624 uint32_t offset
= v
->constbase
.ssbo_sizes
;
625 if (v
->constlen
> offset
) {
626 uint32_t sizes
[align(v
->const_layout
.ssbo_size
.count
, 4)];
627 unsigned mask
= v
->const_layout
.ssbo_size
.mask
;
630 unsigned index
= u_bit_scan(&mask
);
631 unsigned off
= v
->const_layout
.ssbo_size
.off
[index
];
632 sizes
[off
] = sb
->sb
[index
].buffer_size
;
635 fd_wfi(ctx
->batch
, ring
);
636 ctx
->emit_const(ring
, v
->type
, offset
* 4,
637 0, ARRAY_SIZE(sizes
), sizes
, NULL
);
642 emit_image_dims(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
,
643 struct fd_ringbuffer
*ring
, struct fd_shaderimg_stateobj
*si
)
645 uint32_t offset
= v
->constbase
.image_dims
;
646 if (v
->constlen
> offset
) {
647 uint32_t dims
[align(v
->const_layout
.image_dims
.count
, 4)];
648 unsigned mask
= v
->const_layout
.image_dims
.mask
;
651 struct pipe_image_view
*img
;
652 struct fd_resource
*rsc
;
653 unsigned index
= u_bit_scan(&mask
);
654 unsigned off
= v
->const_layout
.image_dims
.off
[index
];
656 img
= &si
->si
[index
];
657 rsc
= fd_resource(img
->resource
);
659 dims
[off
+ 0] = util_format_get_blocksize(img
->format
);
660 if (img
->resource
->target
!= PIPE_BUFFER
) {
661 unsigned lvl
= img
->u
.tex
.level
;
662 /* note for 2d/cube/etc images, even if re-interpreted
663 * as a different color format, the pixel size should
664 * be the same, so use original dimensions for y and z
667 dims
[off
+ 1] = rsc
->slices
[lvl
].pitch
* rsc
->cpp
;
668 /* see corresponding logic in fd_resource_offset(): */
669 if (rsc
->layer_first
) {
670 dims
[off
+ 2] = rsc
->layer_size
;
672 dims
[off
+ 2] = rsc
->slices
[lvl
].size0
;
677 fd_wfi(ctx
->batch
, ring
);
678 ctx
->emit_const(ring
, v
->type
, offset
* 4,
679 0, ARRAY_SIZE(dims
), dims
, NULL
);
684 emit_immediates(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
,
685 struct fd_ringbuffer
*ring
)
687 int size
= v
->immediates_count
;
688 uint32_t base
= v
->constbase
.immediate
;
690 /* truncate size to avoid writing constants that shader
693 size
= MIN2(size
+ base
, v
->constlen
) - base
;
695 /* convert out of vec4: */
700 fd_wfi(ctx
->batch
, ring
);
701 ctx
->emit_const(ring
, v
->type
, base
,
702 0, size
, v
->immediates
[0].val
, NULL
);
706 /* emit stream-out buffers: */
708 emit_tfbos(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
,
709 struct fd_ringbuffer
*ring
)
711 /* streamout addresses after driver-params: */
712 uint32_t offset
= v
->constbase
.tfbo
;
713 if (v
->constlen
> offset
) {
714 struct fd_streamout_stateobj
*so
= &ctx
->streamout
;
715 struct pipe_stream_output_info
*info
= &v
->shader
->stream_output
;
717 uint32_t offsets
[params
];
718 struct pipe_resource
*prscs
[params
];
720 for (uint32_t i
= 0; i
< params
; i
++) {
721 struct pipe_stream_output_target
*target
= so
->targets
[i
];
724 offsets
[i
] = (so
->offsets
[i
] * info
->stride
[i
] * 4) +
725 target
->buffer_offset
;
726 prscs
[i
] = target
->buffer
;
733 fd_wfi(ctx
->batch
, ring
);
734 ctx
->emit_const_bo(ring
, v
->type
, true, offset
* 4, params
, prscs
, offsets
);
739 max_tf_vtx(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
)
741 struct fd_streamout_stateobj
*so
= &ctx
->streamout
;
742 struct pipe_stream_output_info
*info
= &v
->shader
->stream_output
;
743 uint32_t maxvtxcnt
= 0x7fffffff;
745 if (ctx
->screen
->gpu_id
>= 500)
747 if (v
->key
.binning_pass
)
749 if (v
->shader
->stream_output
.num_outputs
== 0)
751 if (so
->num_targets
== 0)
754 /* offset to write to is:
756 * total_vtxcnt = vtxcnt + offsets[i]
757 * offset = total_vtxcnt * stride[i]
759 * offset = vtxcnt * stride[i] ; calculated in shader
760 * + offsets[i] * stride[i] ; calculated at emit_tfbos()
762 * assuming for each vtx, each target buffer will have data written
763 * up to 'offset + stride[i]', that leaves maxvtxcnt as:
765 * buffer_size = (maxvtxcnt * stride[i]) + stride[i]
766 * maxvtxcnt = (buffer_size - stride[i]) / stride[i]
768 * but shader is actually doing a less-than (rather than less-than-
769 * equal) check, so we can drop the -stride[i].
771 * TODO is assumption about `offset + stride[i]` legit?
773 for (unsigned i
= 0; i
< so
->num_targets
; i
++) {
774 struct pipe_stream_output_target
*target
= so
->targets
[i
];
775 unsigned stride
= info
->stride
[i
] * 4; /* convert dwords->bytes */
777 uint32_t max
= target
->buffer_size
/ stride
;
778 maxvtxcnt
= MIN2(maxvtxcnt
, max
);
786 emit_common_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
787 struct fd_context
*ctx
, enum pipe_shader_type t
)
789 enum fd_dirty_shader_state dirty
= ctx
->dirty_shader
[t
];
791 if (dirty
& (FD_DIRTY_SHADER_PROG
| FD_DIRTY_SHADER_CONST
)) {
792 struct fd_constbuf_stateobj
*constbuf
;
795 constbuf
= &ctx
->constbuf
[t
];
796 shader_dirty
= !!(dirty
& FD_DIRTY_SHADER_PROG
);
798 emit_user_consts(ctx
, v
, ring
, constbuf
);
799 emit_ubos(ctx
, v
, ring
, constbuf
);
801 emit_immediates(ctx
, v
, ring
);
804 if (dirty
& (FD_DIRTY_SHADER_PROG
| FD_DIRTY_SHADER_SSBO
)) {
805 struct fd_shaderbuf_stateobj
*sb
= &ctx
->shaderbuf
[t
];
806 emit_ssbo_sizes(ctx
, v
, ring
, sb
);
809 if (dirty
& (FD_DIRTY_SHADER_PROG
| FD_DIRTY_SHADER_IMAGE
)) {
810 struct fd_shaderimg_stateobj
*si
= &ctx
->shaderimg
[t
];
811 emit_image_dims(ctx
, v
, ring
, si
);
816 ir3_emit_vs_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
817 struct fd_context
*ctx
, const struct pipe_draw_info
*info
)
819 debug_assert(v
->type
== SHADER_VERTEX
);
821 emit_common_consts(v
, ring
, ctx
, PIPE_SHADER_VERTEX
);
823 /* emit driver params every time: */
824 /* TODO skip emit if shader doesn't use driver params to avoid WFI.. */
826 uint32_t offset
= v
->constbase
.driver_param
;
827 if (v
->constlen
> offset
) {
828 uint32_t vertex_params
[IR3_DP_VS_COUNT
] = {
829 [IR3_DP_VTXID_BASE
] = info
->index_size
?
830 info
->index_bias
: info
->start
,
831 [IR3_DP_VTXCNT_MAX
] = max_tf_vtx(ctx
, v
),
833 /* if no user-clip-planes, we don't need to emit the
836 uint32_t vertex_params_size
= 4;
838 if (v
->key
.ucp_enables
) {
839 struct pipe_clip_state
*ucp
= &ctx
->ucp
;
840 unsigned pos
= IR3_DP_UCP0_X
;
841 for (unsigned i
= 0; pos
<= IR3_DP_UCP7_W
; i
++) {
842 for (unsigned j
= 0; j
< 4; j
++) {
843 vertex_params
[pos
] = fui(ucp
->ucp
[i
][j
]);
847 vertex_params_size
= ARRAY_SIZE(vertex_params
);
850 fd_wfi(ctx
->batch
, ring
);
852 bool needs_vtxid_base
=
853 ir3_find_sysval_regid(v
, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
) != regid(63, 0);
855 /* for indirect draw, we need to copy VTXID_BASE from
856 * indirect-draw parameters buffer.. which is annoying
857 * and means we can't easily emit these consts in cmd
858 * stream so need to copy them to bo.
860 if (info
->indirect
&& needs_vtxid_base
) {
861 struct pipe_draw_indirect_info
*indirect
= info
->indirect
;
862 struct pipe_resource
*vertex_params_rsc
=
863 pipe_buffer_create(&ctx
->screen
->base
,
864 PIPE_BIND_CONSTANT_BUFFER
, PIPE_USAGE_STREAM
,
865 vertex_params_size
* 4);
866 unsigned src_off
= info
->indirect
->offset
;;
869 ptr
= fd_bo_map(fd_resource(vertex_params_rsc
)->bo
);
870 memcpy(ptr
, vertex_params
, vertex_params_size
* 4);
872 if (info
->index_size
) {
873 /* indexed draw, index_bias is 4th field: */
876 /* non-indexed draw, start is 3rd field: */
880 /* copy index_bias or start from draw params: */
881 ctx
->mem_to_mem(ring
, vertex_params_rsc
, 0,
882 indirect
->buffer
, src_off
, 1);
884 ctx
->emit_const(ring
, SHADER_VERTEX
, offset
* 4, 0,
885 vertex_params_size
, NULL
, vertex_params_rsc
);
887 pipe_resource_reference(&vertex_params_rsc
, NULL
);
889 ctx
->emit_const(ring
, SHADER_VERTEX
, offset
* 4, 0,
890 vertex_params_size
, vertex_params
, NULL
);
893 /* if needed, emit stream-out buffer addresses: */
894 if (vertex_params
[IR3_DP_VTXCNT_MAX
] > 0) {
895 emit_tfbos(ctx
, v
, ring
);
902 ir3_emit_fs_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
903 struct fd_context
*ctx
)
905 debug_assert(v
->type
== SHADER_FRAGMENT
);
907 emit_common_consts(v
, ring
, ctx
, PIPE_SHADER_FRAGMENT
);
910 /* emit compute-shader consts: */
912 ir3_emit_cs_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
913 struct fd_context
*ctx
, const struct pipe_grid_info
*info
)
915 debug_assert(v
->type
== SHADER_COMPUTE
);
917 emit_common_consts(v
, ring
, ctx
, PIPE_SHADER_COMPUTE
);
919 /* emit compute-shader driver-params: */
920 uint32_t offset
= v
->constbase
.driver_param
;
921 if (v
->constlen
> offset
) {
922 fd_wfi(ctx
->batch
, ring
);
924 if (info
->indirect
) {
925 struct pipe_resource
*indirect
= NULL
;
926 unsigned indirect_offset
;
928 /* This is a bit awkward, but CP_LOAD_STATE.EXT_SRC_ADDR needs
929 * to be aligned more strongly than 4 bytes. So in this case
930 * we need a temporary buffer to copy NumWorkGroups.xyz to.
932 * TODO if previous compute job is writing to info->indirect,
933 * we might need a WFI.. but since we currently flush for each
934 * compute job, we are probably ok for now.
936 if (info
->indirect_offset
& 0xf) {
937 indirect
= pipe_buffer_create(&ctx
->screen
->base
,
938 PIPE_BIND_COMMAND_ARGS_BUFFER
, PIPE_USAGE_STREAM
,
942 ctx
->mem_to_mem(ring
, indirect
, 0, info
->indirect
,
943 info
->indirect_offset
, 3);
945 pipe_resource_reference(&indirect
, info
->indirect
);
946 indirect_offset
= info
->indirect_offset
;
949 ctx
->emit_const(ring
, SHADER_COMPUTE
, offset
* 4,
950 indirect_offset
, 4, NULL
, indirect
);
952 pipe_resource_reference(&indirect
, NULL
);
954 uint32_t compute_params
[IR3_DP_CS_COUNT
] = {
955 [IR3_DP_NUM_WORK_GROUPS_X
] = info
->grid
[0],
956 [IR3_DP_NUM_WORK_GROUPS_Y
] = info
->grid
[1],
957 [IR3_DP_NUM_WORK_GROUPS_Z
] = info
->grid
[2],
958 [IR3_DP_LOCAL_GROUP_SIZE_X
] = info
->block
[0],
959 [IR3_DP_LOCAL_GROUP_SIZE_Y
] = info
->block
[1],
960 [IR3_DP_LOCAL_GROUP_SIZE_Z
] = info
->block
[2],
963 ctx
->emit_const(ring
, SHADER_COMPUTE
, offset
* 4, 0,
964 ARRAY_SIZE(compute_params
), compute_params
, NULL
);