2 * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
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 (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
24 * Rob Clark <robclark@freedesktop.org>
27 #include "pipe/p_state.h"
28 #include "pipe/p_screen.h"
29 #include "util/u_string.h"
30 #include "util/u_memory.h"
31 #include "util/u_inlines.h"
32 #include "util/format/u_format.h"
33 #include "tgsi/tgsi_dump.h"
34 #include "tgsi/tgsi_parse.h"
36 #include "nir/tgsi_to_nir.h"
38 #include "freedreno_context.h"
39 #include "freedreno_util.h"
41 #include "ir3/ir3_shader.h"
42 #include "ir3/ir3_gallium.h"
43 #include "ir3/ir3_compiler.h"
44 #include "ir3/ir3_nir.h"
47 dump_shader_info(struct ir3_shader_variant
*v
, bool binning_pass
,
48 struct pipe_debug_callback
*debug
)
50 if (!unlikely(fd_mesa_debug
& FD_DBG_SHADERDB
))
53 pipe_debug_message(debug
, SHADER_INFO
,
54 "%s shader: %u inst, %u nops, %u non-nops, %u dwords, "
55 "%u half, %u full, %u constlen, "
56 "%u (ss), %u (sy), %d max_sun, %d loops\n",
60 v
->info
.instrs_count
- v
->info
.nops_count
,
62 v
->info
.max_half_reg
+ 1,
65 v
->info
.ss
, v
->info
.sy
,
66 v
->max_sun
, v
->loops
);
69 struct ir3_shader_variant
*
70 ir3_shader_variant(struct ir3_shader
*shader
, struct ir3_shader_key key
,
71 bool binning_pass
, struct pipe_debug_callback
*debug
)
73 struct ir3_shader_variant
*v
;
76 /* some shader key values only apply to vertex or frag shader,
77 * so normalize the key to avoid constructing multiple identical
80 ir3_normalize_key(&key
, shader
->type
);
82 v
= ir3_shader_get_variant(shader
, &key
, binning_pass
, &created
);
85 dump_shader_info(v
, binning_pass
, debug
);
92 copy_stream_out(struct ir3_stream_output_info
*i
,
93 const struct pipe_stream_output_info
*p
)
95 STATIC_ASSERT(ARRAY_SIZE(i
->stride
) == ARRAY_SIZE(p
->stride
));
96 STATIC_ASSERT(ARRAY_SIZE(i
->output
) == ARRAY_SIZE(p
->output
));
98 i
->num_outputs
= p
->num_outputs
;
99 for (int n
= 0; n
< ARRAY_SIZE(i
->stride
); n
++)
100 i
->stride
[n
] = p
->stride
[n
];
102 for (int n
= 0; n
< ARRAY_SIZE(i
->output
); n
++) {
103 i
->output
[n
].register_index
= p
->output
[n
].register_index
;
104 i
->output
[n
].start_component
= p
->output
[n
].start_component
;
105 i
->output
[n
].num_components
= p
->output
[n
].num_components
;
106 i
->output
[n
].output_buffer
= p
->output
[n
].output_buffer
;
107 i
->output
[n
].dst_offset
= p
->output
[n
].dst_offset
;
108 i
->output
[n
].stream
= p
->output
[n
].stream
;
113 ir3_shader_create(struct ir3_compiler
*compiler
,
114 const struct pipe_shader_state
*cso
,
115 struct pipe_debug_callback
*debug
,
116 struct pipe_screen
*screen
)
119 if (cso
->type
== PIPE_SHADER_IR_NIR
) {
120 /* we take ownership of the reference: */
123 debug_assert(cso
->type
== PIPE_SHADER_IR_TGSI
);
124 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
125 tgsi_dump(cso
->tokens
, 0);
127 nir
= tgsi_to_nir(cso
->tokens
, screen
);
130 struct ir3_shader
*shader
= ir3_shader_from_nir(compiler
, nir
);
132 copy_stream_out(&shader
->stream_output
, &cso
->stream_output
);
134 if (fd_mesa_debug
& FD_DBG_SHADERDB
) {
135 /* if shader-db run, create a standard variant immediately
136 * (as otherwise nothing will trigger the shader to be
139 static struct ir3_shader_key key
; /* static is implicitly zeroed */
140 ir3_shader_variant(shader
, key
, false, debug
);
142 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
143 ir3_shader_variant(shader
, key
, true, debug
);
148 /* a bit annoying that compute-shader and normal shader state objects
149 * aren't a bit more aligned.
152 ir3_shader_create_compute(struct ir3_compiler
*compiler
,
153 const struct pipe_compute_state
*cso
,
154 struct pipe_debug_callback
*debug
,
155 struct pipe_screen
*screen
)
158 if (cso
->ir_type
== PIPE_SHADER_IR_NIR
) {
159 /* we take ownership of the reference: */
160 nir
= (nir_shader
*)cso
->prog
;
162 debug_assert(cso
->ir_type
== PIPE_SHADER_IR_TGSI
);
163 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
164 tgsi_dump(cso
->prog
, 0);
166 nir
= tgsi_to_nir(cso
->prog
, screen
);
169 struct ir3_shader
*shader
= ir3_shader_from_nir(compiler
, nir
);
174 /* This has to reach into the fd_context a bit more than the rest of
175 * ir3, but it needs to be aligned with the compiler, so both agree
176 * on which const regs hold what. And the logic is identical between
177 * a3xx/a4xx, the only difference is small details in the actual
178 * CP_LOAD_STATE packets (which is handled inside the generation
179 * specific ctx->emit_const(_bo)() fxns)
182 #include "freedreno_resource.h"
185 is_stateobj(struct fd_ringbuffer
*ring
)
187 /* XXX this is an ugly way to differentiate.. */
188 return !!(ring
->flags
& FD_RINGBUFFER_STREAMING
);
192 ring_wfi(struct fd_batch
*batch
, struct fd_ringbuffer
*ring
)
194 /* when we emit const state via ring (IB2) we need a WFI, but when
195 * it is emit'd via stateobj, we don't
197 if (is_stateobj(ring
))
204 emit_const(struct fd_screen
*screen
, struct fd_ringbuffer
*ring
,
205 const struct ir3_shader_variant
*v
, uint32_t dst_offset
,
206 uint32_t offset
, uint32_t size
,
207 const void *user_buffer
, struct pipe_resource
*buffer
)
209 assert(dst_offset
+ size
<= v
->constlen
* 4);
211 screen
->emit_const(ring
, v
->type
, dst_offset
,
212 offset
, size
, user_buffer
, buffer
);
216 * Indirectly calculates size of cmdstream needed for ir3_emit_user_consts().
217 * Returns number of packets, and total size of all the payload.
219 * The value can be a worst-case, ie. some shader variants may not read all
222 * Returns size in dwords.
225 ir3_user_consts_size(struct ir3_ubo_analysis_state
*state
,
226 unsigned *packets
, unsigned *size
)
228 *packets
= *size
= 0;
230 for (uint32_t i
= 0; i
< ARRAY_SIZE(state
->range
); i
++) {
231 if (state
->range
[i
].start
< state
->range
[i
].end
) {
232 *size
+= state
->range
[i
].end
- state
->range
[i
].start
;
239 ir3_emit_user_consts(struct fd_screen
*screen
, const struct ir3_shader_variant
*v
,
240 struct fd_ringbuffer
*ring
, struct fd_constbuf_stateobj
*constbuf
)
242 struct ir3_ubo_analysis_state
*state
;
243 state
= &v
->shader
->ubo_state
;
245 for (uint32_t i
= 0; i
< ARRAY_SIZE(state
->range
); i
++) {
246 struct pipe_constant_buffer
*cb
= &constbuf
->cb
[i
];
248 if (state
->range
[i
].start
< state
->range
[i
].end
&&
249 constbuf
->enabled_mask
& (1 << i
)) {
251 uint32_t size
= state
->range
[i
].end
- state
->range
[i
].start
;
252 uint32_t offset
= cb
->buffer_offset
+ state
->range
[i
].start
;
254 /* and even if the start of the const buffer is before
255 * first_immediate, the end may not be:
257 size
= MIN2(size
, (16 * v
->constlen
) - state
->range
[i
].offset
);
262 /* things should be aligned to vec4: */
263 debug_assert((state
->range
[i
].offset
% 16) == 0);
264 debug_assert((size
% 16) == 0);
265 debug_assert((offset
% 16) == 0);
267 emit_const(screen
, ring
, v
, state
->range
[i
].offset
/ 4,
268 offset
, size
/ 4, cb
->user_buffer
, cb
->buffer
);
274 ir3_emit_ubos(struct fd_screen
*screen
, const struct ir3_shader_variant
*v
,
275 struct fd_ringbuffer
*ring
, struct fd_constbuf_stateobj
*constbuf
)
277 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
278 uint32_t offset
= const_state
->offsets
.ubo
;
279 if (v
->constlen
> offset
) {
280 uint32_t params
= const_state
->num_ubos
;
281 uint32_t offsets
[params
];
282 struct pipe_resource
*prscs
[params
];
284 for (uint32_t i
= 0; i
< params
; i
++) {
285 const uint32_t index
= i
+ 1; /* UBOs start at index 1 */
286 struct pipe_constant_buffer
*cb
= &constbuf
->cb
[index
];
287 assert(!cb
->user_buffer
);
289 if ((constbuf
->enabled_mask
& (1 << index
)) && cb
->buffer
) {
290 offsets
[i
] = cb
->buffer_offset
;
291 prscs
[i
] = cb
->buffer
;
298 assert(offset
* 4 + params
< v
->constlen
* 4);
300 screen
->emit_const_bo(ring
, v
->type
, false, offset
* 4, params
, prscs
, offsets
);
305 ir3_emit_ssbo_sizes(struct fd_screen
*screen
, const struct ir3_shader_variant
*v
,
306 struct fd_ringbuffer
*ring
, struct fd_shaderbuf_stateobj
*sb
)
308 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
309 uint32_t offset
= const_state
->offsets
.ssbo_sizes
;
310 if (v
->constlen
> offset
) {
311 uint32_t sizes
[align(const_state
->ssbo_size
.count
, 4)];
312 unsigned mask
= const_state
->ssbo_size
.mask
;
315 unsigned index
= u_bit_scan(&mask
);
316 unsigned off
= const_state
->ssbo_size
.off
[index
];
317 sizes
[off
] = sb
->sb
[index
].buffer_size
;
320 emit_const(screen
, ring
, v
, offset
* 4,
321 0, ARRAY_SIZE(sizes
), sizes
, NULL
);
326 ir3_emit_image_dims(struct fd_screen
*screen
, const struct ir3_shader_variant
*v
,
327 struct fd_ringbuffer
*ring
, struct fd_shaderimg_stateobj
*si
)
329 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
330 uint32_t offset
= const_state
->offsets
.image_dims
;
331 if (v
->constlen
> offset
) {
332 uint32_t dims
[align(const_state
->image_dims
.count
, 4)];
333 unsigned mask
= const_state
->image_dims
.mask
;
336 struct pipe_image_view
*img
;
337 struct fd_resource
*rsc
;
338 unsigned index
= u_bit_scan(&mask
);
339 unsigned off
= const_state
->image_dims
.off
[index
];
341 img
= &si
->si
[index
];
342 rsc
= fd_resource(img
->resource
);
344 dims
[off
+ 0] = util_format_get_blocksize(img
->format
);
345 if (img
->resource
->target
!= PIPE_BUFFER
) {
346 struct fd_resource_slice
*slice
=
347 fd_resource_slice(rsc
, img
->u
.tex
.level
);
348 /* note for 2d/cube/etc images, even if re-interpreted
349 * as a different color format, the pixel size should
350 * be the same, so use original dimensions for y and z
353 dims
[off
+ 1] = slice
->pitch
* rsc
->cpp
;
354 /* see corresponding logic in fd_resource_offset(): */
355 if (rsc
->layer_first
) {
356 dims
[off
+ 2] = rsc
->layer_size
;
358 dims
[off
+ 2] = slice
->size0
;
361 /* For buffer-backed images, the log2 of the format's
362 * bytes-per-pixel is placed on the 2nd slot. This is useful
363 * when emitting image_size instructions, for which we need
364 * to divide by bpp for image buffers. Since the bpp
365 * can only be power-of-two, the division is implemented
366 * as a SHR, and for that it is handy to have the log2 of
367 * bpp as a constant. (log2 = first-set-bit - 1)
369 dims
[off
+ 1] = ffs(dims
[off
+ 0]) - 1;
372 uint32_t size
= MIN2(ARRAY_SIZE(dims
), v
->constlen
* 4 - offset
* 4);
374 emit_const(screen
, ring
, v
, offset
* 4, 0, size
, dims
, NULL
);
379 ir3_emit_immediates(struct fd_screen
*screen
, const struct ir3_shader_variant
*v
,
380 struct fd_ringbuffer
*ring
)
382 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
383 uint32_t base
= const_state
->offsets
.immediate
;
384 int size
= const_state
->immediates_count
;
386 /* truncate size to avoid writing constants that shader
389 size
= MIN2(size
+ base
, v
->constlen
) - base
;
391 /* convert out of vec4: */
396 emit_const(screen
, ring
, v
, base
,
397 0, size
, const_state
->immediates
[0].val
, NULL
);
402 link_geometry_stages(const struct ir3_shader_variant
*producer
,
403 const struct ir3_shader_variant
*consumer
,
406 uint32_t num_loc
= 0, factor
;
408 switch (consumer
->type
) {
409 case MESA_SHADER_TESS_CTRL
:
410 case MESA_SHADER_GEOMETRY
:
411 /* These stages load with ldlw, which expects byte offsets. */
414 case MESA_SHADER_TESS_EVAL
:
415 /* The tess eval shader uses ldg, which takes dword offsets. */
419 unreachable("bad shader stage");
422 nir_foreach_variable(in_var
, &consumer
->shader
->nir
->inputs
) {
423 nir_foreach_variable(out_var
, &producer
->shader
->nir
->outputs
) {
424 if (in_var
->data
.location
== out_var
->data
.location
) {
425 locs
[in_var
->data
.driver_location
] =
426 producer
->shader
->output_loc
[out_var
->data
.driver_location
] * factor
;
428 debug_assert(num_loc
<= in_var
->data
.driver_location
+ 1);
429 num_loc
= in_var
->data
.driver_location
+ 1;
438 ir3_emit_link_map(struct fd_screen
*screen
,
439 const struct ir3_shader_variant
*producer
,
440 const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
)
442 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
443 uint32_t base
= const_state
->offsets
.primitive_map
;
444 uint32_t patch_locs
[MAX_VARYING
] = { }, num_loc
;
446 num_loc
= link_geometry_stages(producer
, v
, patch_locs
);
448 int size
= DIV_ROUND_UP(num_loc
, 4);
450 /* truncate size to avoid writing constants that shader
453 size
= MIN2(size
+ base
, v
->constlen
) - base
;
455 /* convert out of vec4: */
460 emit_const(screen
, ring
, v
, base
, 0, size
, patch_locs
, NULL
);
463 /* emit stream-out buffers: */
465 emit_tfbos(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
,
466 struct fd_ringbuffer
*ring
)
468 /* streamout addresses after driver-params: */
469 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
470 uint32_t offset
= const_state
->offsets
.tfbo
;
471 if (v
->constlen
> offset
) {
472 struct fd_streamout_stateobj
*so
= &ctx
->streamout
;
473 struct ir3_stream_output_info
*info
= &v
->shader
->stream_output
;
475 uint32_t offsets
[params
];
476 struct pipe_resource
*prscs
[params
];
478 for (uint32_t i
= 0; i
< params
; i
++) {
479 struct pipe_stream_output_target
*target
= so
->targets
[i
];
482 offsets
[i
] = (so
->offsets
[i
] * info
->stride
[i
] * 4) +
483 target
->buffer_offset
;
484 prscs
[i
] = target
->buffer
;
491 assert(offset
* 4 + params
< v
->constlen
* 4);
493 ctx
->screen
->emit_const_bo(ring
, v
->type
, true, offset
* 4, params
, prscs
, offsets
);
498 max_tf_vtx(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
)
500 struct fd_streamout_stateobj
*so
= &ctx
->streamout
;
501 struct ir3_stream_output_info
*info
= &v
->shader
->stream_output
;
502 uint32_t maxvtxcnt
= 0x7fffffff;
504 if (ctx
->screen
->gpu_id
>= 500)
508 if (v
->shader
->stream_output
.num_outputs
== 0)
510 if (so
->num_targets
== 0)
513 /* offset to write to is:
515 * total_vtxcnt = vtxcnt + offsets[i]
516 * offset = total_vtxcnt * stride[i]
518 * offset = vtxcnt * stride[i] ; calculated in shader
519 * + offsets[i] * stride[i] ; calculated at emit_tfbos()
521 * assuming for each vtx, each target buffer will have data written
522 * up to 'offset + stride[i]', that leaves maxvtxcnt as:
524 * buffer_size = (maxvtxcnt * stride[i]) + stride[i]
525 * maxvtxcnt = (buffer_size - stride[i]) / stride[i]
527 * but shader is actually doing a less-than (rather than less-than-
528 * equal) check, so we can drop the -stride[i].
530 * TODO is assumption about `offset + stride[i]` legit?
532 for (unsigned i
= 0; i
< so
->num_targets
; i
++) {
533 struct pipe_stream_output_target
*target
= so
->targets
[i
];
534 unsigned stride
= info
->stride
[i
] * 4; /* convert dwords->bytes */
536 uint32_t max
= target
->buffer_size
/ stride
;
537 maxvtxcnt
= MIN2(maxvtxcnt
, max
);
545 emit_common_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
546 struct fd_context
*ctx
, enum pipe_shader_type t
)
548 enum fd_dirty_shader_state dirty
= ctx
->dirty_shader
[t
];
550 /* When we use CP_SET_DRAW_STATE objects to emit constant state,
551 * if we emit any of it we need to emit all. This is because
552 * we are using the same state-group-id each time for uniform
553 * state, and if previous update is never evaluated (due to no
554 * visible primitives in the current tile) then the new stateobj
555 * completely replaces the old one.
557 * Possibly if we split up different parts of the const state to
558 * different state-objects we could avoid this.
560 if (dirty
&& is_stateobj(ring
))
563 if (dirty
& (FD_DIRTY_SHADER_PROG
| FD_DIRTY_SHADER_CONST
)) {
564 struct fd_constbuf_stateobj
*constbuf
;
567 constbuf
= &ctx
->constbuf
[t
];
568 shader_dirty
= !!(dirty
& FD_DIRTY_SHADER_PROG
);
570 ring_wfi(ctx
->batch
, ring
);
572 ir3_emit_user_consts(ctx
->screen
, v
, ring
, constbuf
);
573 ir3_emit_ubos(ctx
->screen
, v
, ring
, constbuf
);
575 ir3_emit_immediates(ctx
->screen
, v
, ring
);
578 if (dirty
& (FD_DIRTY_SHADER_PROG
| FD_DIRTY_SHADER_SSBO
)) {
579 struct fd_shaderbuf_stateobj
*sb
= &ctx
->shaderbuf
[t
];
580 ring_wfi(ctx
->batch
, ring
);
581 ir3_emit_ssbo_sizes(ctx
->screen
, v
, ring
, sb
);
584 if (dirty
& (FD_DIRTY_SHADER_PROG
| FD_DIRTY_SHADER_IMAGE
)) {
585 struct fd_shaderimg_stateobj
*si
= &ctx
->shaderimg
[t
];
586 ring_wfi(ctx
->batch
, ring
);
587 ir3_emit_image_dims(ctx
->screen
, v
, ring
, si
);
592 ir3_emit_vs_driver_params(const struct ir3_shader_variant
*v
,
593 struct fd_ringbuffer
*ring
, struct fd_context
*ctx
,
594 const struct pipe_draw_info
*info
)
596 debug_assert(ir3_needs_vs_driver_params(v
));
598 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
599 uint32_t offset
= const_state
->offsets
.driver_param
;
600 uint32_t vertex_params
[IR3_DP_VS_COUNT
] = {
601 [IR3_DP_VTXID_BASE
] = info
->index_size
?
602 info
->index_bias
: info
->start
,
603 [IR3_DP_VTXCNT_MAX
] = max_tf_vtx(ctx
, v
),
605 /* if no user-clip-planes, we don't need to emit the
608 uint32_t vertex_params_size
= 4;
610 if (v
->key
.ucp_enables
) {
611 struct pipe_clip_state
*ucp
= &ctx
->ucp
;
612 unsigned pos
= IR3_DP_UCP0_X
;
613 for (unsigned i
= 0; pos
<= IR3_DP_UCP7_W
; i
++) {
614 for (unsigned j
= 0; j
< 4; j
++) {
615 vertex_params
[pos
] = fui(ucp
->ucp
[i
][j
]);
619 vertex_params_size
= ARRAY_SIZE(vertex_params
);
622 vertex_params_size
= MAX2(vertex_params_size
, const_state
->num_driver_params
);
624 bool needs_vtxid_base
=
625 ir3_find_sysval_regid(v
, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
) != regid(63, 0);
627 /* for indirect draw, we need to copy VTXID_BASE from
628 * indirect-draw parameters buffer.. which is annoying
629 * and means we can't easily emit these consts in cmd
630 * stream so need to copy them to bo.
632 if (info
->indirect
&& needs_vtxid_base
) {
633 struct pipe_draw_indirect_info
*indirect
= info
->indirect
;
634 struct pipe_resource
*vertex_params_rsc
=
635 pipe_buffer_create(&ctx
->screen
->base
,
636 PIPE_BIND_CONSTANT_BUFFER
, PIPE_USAGE_STREAM
,
637 vertex_params_size
* 4);
638 unsigned src_off
= info
->indirect
->offset
;;
641 ptr
= fd_bo_map(fd_resource(vertex_params_rsc
)->bo
);
642 memcpy(ptr
, vertex_params
, vertex_params_size
* 4);
644 if (info
->index_size
) {
645 /* indexed draw, index_bias is 4th field: */
648 /* non-indexed draw, start is 3rd field: */
652 /* copy index_bias or start from draw params: */
653 ctx
->screen
->mem_to_mem(ring
, vertex_params_rsc
, 0,
654 indirect
->buffer
, src_off
, 1);
656 emit_const(ctx
->screen
, ring
, v
, offset
* 4, 0,
657 vertex_params_size
, NULL
, vertex_params_rsc
);
659 pipe_resource_reference(&vertex_params_rsc
, NULL
);
661 emit_const(ctx
->screen
, ring
, v
, offset
* 4, 0,
662 vertex_params_size
, vertex_params
, NULL
);
665 /* if needed, emit stream-out buffer addresses: */
666 if (vertex_params
[IR3_DP_VTXCNT_MAX
] > 0) {
667 emit_tfbos(ctx
, v
, ring
);
672 ir3_emit_vs_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
673 struct fd_context
*ctx
, const struct pipe_draw_info
*info
)
675 debug_assert(v
->type
== MESA_SHADER_VERTEX
);
677 emit_common_consts(v
, ring
, ctx
, PIPE_SHADER_VERTEX
);
679 /* emit driver params every time: */
680 if (info
&& ir3_needs_vs_driver_params(v
)) {
681 ring_wfi(ctx
->batch
, ring
);
682 ir3_emit_vs_driver_params(v
, ring
, ctx
, info
);
687 ir3_emit_fs_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
688 struct fd_context
*ctx
)
690 debug_assert(v
->type
== MESA_SHADER_FRAGMENT
);
692 emit_common_consts(v
, ring
, ctx
, PIPE_SHADER_FRAGMENT
);
695 /* emit compute-shader consts: */
697 ir3_emit_cs_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
698 struct fd_context
*ctx
, const struct pipe_grid_info
*info
)
700 debug_assert(gl_shader_stage_is_compute(v
->type
));
702 emit_common_consts(v
, ring
, ctx
, PIPE_SHADER_COMPUTE
);
704 /* emit compute-shader driver-params: */
705 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
706 uint32_t offset
= const_state
->offsets
.driver_param
;
707 if (v
->constlen
> offset
) {
708 ring_wfi(ctx
->batch
, ring
);
710 if (info
->indirect
) {
711 struct pipe_resource
*indirect
= NULL
;
712 unsigned indirect_offset
;
714 /* This is a bit awkward, but CP_LOAD_STATE.EXT_SRC_ADDR needs
715 * to be aligned more strongly than 4 bytes. So in this case
716 * we need a temporary buffer to copy NumWorkGroups.xyz to.
718 * TODO if previous compute job is writing to info->indirect,
719 * we might need a WFI.. but since we currently flush for each
720 * compute job, we are probably ok for now.
722 if (info
->indirect_offset
& 0xf) {
723 indirect
= pipe_buffer_create(&ctx
->screen
->base
,
724 PIPE_BIND_COMMAND_ARGS_BUFFER
, PIPE_USAGE_STREAM
,
728 ctx
->screen
->mem_to_mem(ring
, indirect
, 0, info
->indirect
,
729 info
->indirect_offset
, 3);
731 pipe_resource_reference(&indirect
, info
->indirect
);
732 indirect_offset
= info
->indirect_offset
;
735 emit_const(ctx
->screen
, ring
, v
, offset
* 4,
736 indirect_offset
, 4, NULL
, indirect
);
738 pipe_resource_reference(&indirect
, NULL
);
740 uint32_t compute_params
[IR3_DP_CS_COUNT
] = {
741 [IR3_DP_NUM_WORK_GROUPS_X
] = info
->grid
[0],
742 [IR3_DP_NUM_WORK_GROUPS_Y
] = info
->grid
[1],
743 [IR3_DP_NUM_WORK_GROUPS_Z
] = info
->grid
[2],
744 [IR3_DP_LOCAL_GROUP_SIZE_X
] = info
->block
[0],
745 [IR3_DP_LOCAL_GROUP_SIZE_Y
] = info
->block
[1],
746 [IR3_DP_LOCAL_GROUP_SIZE_Z
] = info
->block
[2],
748 uint32_t size
= MIN2(const_state
->num_driver_params
,
749 v
->constlen
* 4 - offset
* 4);
751 emit_const(ctx
->screen
, ring
, v
, offset
* 4, 0, size
,
752 compute_params
, NULL
);
758 ir3_shader_state_create(struct pipe_context
*pctx
, const struct pipe_shader_state
*cso
)
760 struct fd_context
*ctx
= fd_context(pctx
);
761 struct ir3_compiler
*compiler
= ctx
->screen
->compiler
;
762 return ir3_shader_create(compiler
, cso
, &ctx
->debug
, pctx
->screen
);
766 ir3_shader_state_delete(struct pipe_context
*pctx
, void *hwcso
)
768 struct ir3_shader
*so
= hwcso
;
769 ir3_shader_destroy(so
);
773 ir3_prog_init(struct pipe_context
*pctx
)
775 pctx
->create_vs_state
= ir3_shader_state_create
;
776 pctx
->delete_vs_state
= ir3_shader_state_delete
;
778 pctx
->create_tcs_state
= ir3_shader_state_create
;
779 pctx
->delete_tcs_state
= ir3_shader_state_delete
;
781 pctx
->create_tes_state
= ir3_shader_state_create
;
782 pctx
->delete_tes_state
= ir3_shader_state_delete
;
784 pctx
->create_gs_state
= ir3_shader_state_create
;
785 pctx
->delete_gs_state
= ir3_shader_state_delete
;
787 pctx
->create_fs_state
= ir3_shader_state_create
;
788 pctx
->delete_fs_state
= ir3_shader_state_delete
;