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/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%s shader: %u inst, %u dwords, "
55 "%u half, %u full, %u constlen, "
56 "%u (ss), %u (sy), %d max_sun, %d loops\n",
57 binning_pass
? "B" : "",
58 ir3_shader_stage(v
->shader
),
61 v
->info
.max_half_reg
+ 1,
64 v
->info
.ss
, v
->info
.sy
,
65 v
->max_sun
, v
->loops
);
68 struct ir3_shader_variant
*
69 ir3_shader_variant(struct ir3_shader
*shader
, struct ir3_shader_key key
,
70 bool binning_pass
, struct pipe_debug_callback
*debug
)
72 struct ir3_shader_variant
*v
;
75 /* some shader key values only apply to vertex or frag shader,
76 * so normalize the key to avoid constructing multiple identical
79 ir3_normalize_key(&key
, shader
->type
);
81 v
= ir3_shader_get_variant(shader
, &key
, binning_pass
, &created
);
84 dump_shader_info(v
, binning_pass
, debug
);
91 copy_stream_out(struct ir3_stream_output_info
*i
,
92 const struct pipe_stream_output_info
*p
)
94 STATIC_ASSERT(ARRAY_SIZE(i
->stride
) == ARRAY_SIZE(p
->stride
));
95 STATIC_ASSERT(ARRAY_SIZE(i
->output
) == ARRAY_SIZE(p
->output
));
97 i
->num_outputs
= p
->num_outputs
;
98 for (int n
= 0; n
< ARRAY_SIZE(i
->stride
); n
++)
99 i
->stride
[n
] = p
->stride
[n
];
101 for (int n
= 0; n
< ARRAY_SIZE(i
->output
); n
++) {
102 i
->output
[n
].register_index
= p
->output
[n
].register_index
;
103 i
->output
[n
].start_component
= p
->output
[n
].start_component
;
104 i
->output
[n
].num_components
= p
->output
[n
].num_components
;
105 i
->output
[n
].output_buffer
= p
->output
[n
].output_buffer
;
106 i
->output
[n
].dst_offset
= p
->output
[n
].dst_offset
;
107 i
->output
[n
].stream
= p
->output
[n
].stream
;
112 ir3_shader_create(struct ir3_compiler
*compiler
,
113 const struct pipe_shader_state
*cso
,
114 struct pipe_debug_callback
*debug
,
115 struct pipe_screen
*screen
)
118 if (cso
->type
== PIPE_SHADER_IR_NIR
) {
119 /* we take ownership of the reference: */
122 debug_assert(cso
->type
== PIPE_SHADER_IR_TGSI
);
123 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
124 tgsi_dump(cso
->tokens
, 0);
126 nir
= tgsi_to_nir(cso
->tokens
, screen
);
129 struct ir3_shader
*shader
= ir3_shader_from_nir(compiler
, nir
);
131 copy_stream_out(&shader
->stream_output
, &cso
->stream_output
);
133 if (fd_mesa_debug
& FD_DBG_SHADERDB
) {
134 /* if shader-db run, create a standard variant immediately
135 * (as otherwise nothing will trigger the shader to be
138 static struct ir3_shader_key key
; /* static is implicitly zeroed */
139 ir3_shader_variant(shader
, key
, false, debug
);
141 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
142 ir3_shader_variant(shader
, key
, true, debug
);
147 /* a bit annoying that compute-shader and normal shader state objects
148 * aren't a bit more aligned.
151 ir3_shader_create_compute(struct ir3_compiler
*compiler
,
152 const struct pipe_compute_state
*cso
,
153 struct pipe_debug_callback
*debug
,
154 struct pipe_screen
*screen
)
157 if (cso
->ir_type
== PIPE_SHADER_IR_NIR
) {
158 /* we take ownership of the reference: */
159 nir
= (nir_shader
*)cso
->prog
;
161 debug_assert(cso
->ir_type
== PIPE_SHADER_IR_TGSI
);
162 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
163 tgsi_dump(cso
->prog
, 0);
165 nir
= tgsi_to_nir(cso
->prog
, screen
);
168 struct ir3_shader
*shader
= ir3_shader_from_nir(compiler
, nir
);
173 /* This has to reach into the fd_context a bit more than the rest of
174 * ir3, but it needs to be aligned with the compiler, so both agree
175 * on which const regs hold what. And the logic is identical between
176 * a3xx/a4xx, the only difference is small details in the actual
177 * CP_LOAD_STATE packets (which is handled inside the generation
178 * specific ctx->emit_const(_bo)() fxns)
181 #include "freedreno_resource.h"
184 is_stateobj(struct fd_ringbuffer
*ring
)
186 /* XXX this is an ugly way to differentiate.. */
187 return !!(ring
->flags
& FD_RINGBUFFER_STREAMING
);
191 ring_wfi(struct fd_batch
*batch
, struct fd_ringbuffer
*ring
)
193 /* when we emit const state via ring (IB2) we need a WFI, but when
194 * it is emit'd via stateobj, we don't
196 if (is_stateobj(ring
))
203 emit_const(struct fd_screen
*screen
, struct fd_ringbuffer
*ring
,
204 const struct ir3_shader_variant
*v
, uint32_t dst_offset
,
205 uint32_t offset
, uint32_t size
,
206 const void *user_buffer
, struct pipe_resource
*buffer
)
208 assert(dst_offset
+ size
<= v
->constlen
* 4);
210 screen
->emit_const(ring
, v
->type
, dst_offset
,
211 offset
, size
, user_buffer
, buffer
);
215 * Indirectly calculates size of cmdstream needed for ir3_emit_user_consts().
216 * Returns number of packets, and total size of all the payload.
218 * The value can be a worst-case, ie. some shader variants may not read all
221 * Returns size in dwords.
224 ir3_user_consts_size(struct ir3_ubo_analysis_state
*state
,
225 unsigned *packets
, unsigned *size
)
227 *packets
= *size
= 0;
229 for (uint32_t i
= 0; i
< ARRAY_SIZE(state
->range
); i
++) {
230 if (state
->range
[i
].start
< state
->range
[i
].end
) {
231 *size
+= state
->range
[i
].end
- state
->range
[i
].start
;
238 ir3_emit_user_consts(struct fd_screen
*screen
, const struct ir3_shader_variant
*v
,
239 struct fd_ringbuffer
*ring
, struct fd_constbuf_stateobj
*constbuf
)
241 struct ir3_ubo_analysis_state
*state
;
242 state
= &v
->shader
->ubo_state
;
244 for (uint32_t i
= 0; i
< ARRAY_SIZE(state
->range
); i
++) {
245 struct pipe_constant_buffer
*cb
= &constbuf
->cb
[i
];
247 if (state
->range
[i
].start
< state
->range
[i
].end
&&
248 constbuf
->enabled_mask
& (1 << i
)) {
250 uint32_t size
= state
->range
[i
].end
- state
->range
[i
].start
;
251 uint32_t offset
= cb
->buffer_offset
+ state
->range
[i
].start
;
253 /* and even if the start of the const buffer is before
254 * first_immediate, the end may not be:
256 size
= MIN2(size
, (16 * v
->constlen
) - state
->range
[i
].offset
);
261 /* things should be aligned to vec4: */
262 debug_assert((state
->range
[i
].offset
% 16) == 0);
263 debug_assert((size
% 16) == 0);
264 debug_assert((offset
% 16) == 0);
266 emit_const(screen
, ring
, v
, state
->range
[i
].offset
/ 4,
267 offset
, size
/ 4, cb
->user_buffer
, cb
->buffer
);
273 ir3_emit_ubos(struct fd_screen
*screen
, const struct ir3_shader_variant
*v
,
274 struct fd_ringbuffer
*ring
, struct fd_constbuf_stateobj
*constbuf
)
276 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
277 uint32_t offset
= const_state
->offsets
.ubo
;
278 if (v
->constlen
> offset
) {
279 uint32_t params
= const_state
->num_ubos
;
280 uint32_t offsets
[params
];
281 struct pipe_resource
*prscs
[params
];
283 for (uint32_t i
= 0; i
< params
; i
++) {
284 const uint32_t index
= i
+ 1; /* UBOs start at index 1 */
285 struct pipe_constant_buffer
*cb
= &constbuf
->cb
[index
];
286 assert(!cb
->user_buffer
);
288 if ((constbuf
->enabled_mask
& (1 << index
)) && cb
->buffer
) {
289 offsets
[i
] = cb
->buffer_offset
;
290 prscs
[i
] = cb
->buffer
;
297 assert(offset
* 4 + params
< v
->constlen
* 4);
299 screen
->emit_const_bo(ring
, v
->type
, false, offset
* 4, params
, prscs
, offsets
);
304 ir3_emit_ssbo_sizes(struct fd_screen
*screen
, const struct ir3_shader_variant
*v
,
305 struct fd_ringbuffer
*ring
, struct fd_shaderbuf_stateobj
*sb
)
307 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
308 uint32_t offset
= const_state
->offsets
.ssbo_sizes
;
309 if (v
->constlen
> offset
) {
310 uint32_t sizes
[align(const_state
->ssbo_size
.count
, 4)];
311 unsigned mask
= const_state
->ssbo_size
.mask
;
314 unsigned index
= u_bit_scan(&mask
);
315 unsigned off
= const_state
->ssbo_size
.off
[index
];
316 sizes
[off
] = sb
->sb
[index
].buffer_size
;
319 emit_const(screen
, ring
, v
, offset
* 4,
320 0, ARRAY_SIZE(sizes
), sizes
, NULL
);
325 ir3_emit_image_dims(struct fd_screen
*screen
, const struct ir3_shader_variant
*v
,
326 struct fd_ringbuffer
*ring
, struct fd_shaderimg_stateobj
*si
)
328 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
329 uint32_t offset
= const_state
->offsets
.image_dims
;
330 if (v
->constlen
> offset
) {
331 uint32_t dims
[align(const_state
->image_dims
.count
, 4)];
332 unsigned mask
= const_state
->image_dims
.mask
;
335 struct pipe_image_view
*img
;
336 struct fd_resource
*rsc
;
337 unsigned index
= u_bit_scan(&mask
);
338 unsigned off
= const_state
->image_dims
.off
[index
];
340 img
= &si
->si
[index
];
341 rsc
= fd_resource(img
->resource
);
343 dims
[off
+ 0] = util_format_get_blocksize(img
->format
);
344 if (img
->resource
->target
!= PIPE_BUFFER
) {
345 unsigned lvl
= img
->u
.tex
.level
;
346 /* note for 2d/cube/etc images, even if re-interpreted
347 * as a different color format, the pixel size should
348 * be the same, so use original dimensions for y and z
351 dims
[off
+ 1] = rsc
->slices
[lvl
].pitch
* rsc
->cpp
;
352 /* see corresponding logic in fd_resource_offset(): */
353 if (rsc
->layer_first
) {
354 dims
[off
+ 2] = rsc
->layer_size
;
356 dims
[off
+ 2] = rsc
->slices
[lvl
].size0
;
359 /* For buffer-backed images, the log2 of the format's
360 * bytes-per-pixel is placed on the 2nd slot. This is useful
361 * when emitting image_size instructions, for which we need
362 * to divide by bpp for image buffers. Since the bpp
363 * can only be power-of-two, the division is implemented
364 * as a SHR, and for that it is handy to have the log2 of
365 * bpp as a constant. (log2 = first-set-bit - 1)
367 dims
[off
+ 1] = ffs(dims
[off
+ 0]) - 1;
370 uint32_t size
= MIN2(ARRAY_SIZE(dims
), v
->constlen
* 4 - offset
* 4);
372 emit_const(screen
, ring
, v
, offset
* 4, 0, size
, dims
, NULL
);
377 ir3_emit_immediates(struct fd_screen
*screen
, const struct ir3_shader_variant
*v
,
378 struct fd_ringbuffer
*ring
)
380 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
381 uint32_t base
= const_state
->offsets
.immediate
;
382 int size
= const_state
->immediates_count
;
384 /* truncate size to avoid writing constants that shader
387 size
= MIN2(size
+ base
, v
->constlen
) - base
;
389 /* convert out of vec4: */
394 emit_const(screen
, ring
, v
, base
,
395 0, size
, const_state
->immediates
[0].val
, NULL
);
399 /* emit stream-out buffers: */
401 emit_tfbos(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
,
402 struct fd_ringbuffer
*ring
)
404 /* streamout addresses after driver-params: */
405 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
406 uint32_t offset
= const_state
->offsets
.tfbo
;
407 if (v
->constlen
> offset
) {
408 struct fd_streamout_stateobj
*so
= &ctx
->streamout
;
409 struct ir3_stream_output_info
*info
= &v
->shader
->stream_output
;
411 uint32_t offsets
[params
];
412 struct pipe_resource
*prscs
[params
];
414 for (uint32_t i
= 0; i
< params
; i
++) {
415 struct pipe_stream_output_target
*target
= so
->targets
[i
];
418 offsets
[i
] = (so
->offsets
[i
] * info
->stride
[i
] * 4) +
419 target
->buffer_offset
;
420 prscs
[i
] = target
->buffer
;
427 assert(offset
* 4 + params
< v
->constlen
* 4);
429 ctx
->screen
->emit_const_bo(ring
, v
->type
, true, offset
* 4, params
, prscs
, offsets
);
434 max_tf_vtx(struct fd_context
*ctx
, const struct ir3_shader_variant
*v
)
436 struct fd_streamout_stateobj
*so
= &ctx
->streamout
;
437 struct ir3_stream_output_info
*info
= &v
->shader
->stream_output
;
438 uint32_t maxvtxcnt
= 0x7fffffff;
440 if (ctx
->screen
->gpu_id
>= 500)
444 if (v
->shader
->stream_output
.num_outputs
== 0)
446 if (so
->num_targets
== 0)
449 /* offset to write to is:
451 * total_vtxcnt = vtxcnt + offsets[i]
452 * offset = total_vtxcnt * stride[i]
454 * offset = vtxcnt * stride[i] ; calculated in shader
455 * + offsets[i] * stride[i] ; calculated at emit_tfbos()
457 * assuming for each vtx, each target buffer will have data written
458 * up to 'offset + stride[i]', that leaves maxvtxcnt as:
460 * buffer_size = (maxvtxcnt * stride[i]) + stride[i]
461 * maxvtxcnt = (buffer_size - stride[i]) / stride[i]
463 * but shader is actually doing a less-than (rather than less-than-
464 * equal) check, so we can drop the -stride[i].
466 * TODO is assumption about `offset + stride[i]` legit?
468 for (unsigned i
= 0; i
< so
->num_targets
; i
++) {
469 struct pipe_stream_output_target
*target
= so
->targets
[i
];
470 unsigned stride
= info
->stride
[i
] * 4; /* convert dwords->bytes */
472 uint32_t max
= target
->buffer_size
/ stride
;
473 maxvtxcnt
= MIN2(maxvtxcnt
, max
);
481 emit_common_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
482 struct fd_context
*ctx
, enum pipe_shader_type t
)
484 enum fd_dirty_shader_state dirty
= ctx
->dirty_shader
[t
];
486 /* When we use CP_SET_DRAW_STATE objects to emit constant state,
487 * if we emit any of it we need to emit all. This is because
488 * we are using the same state-group-id each time for uniform
489 * state, and if previous update is never evaluated (due to no
490 * visible primitives in the current tile) then the new stateobj
491 * completely replaces the old one.
493 * Possibly if we split up different parts of the const state to
494 * different state-objects we could avoid this.
496 if (dirty
&& is_stateobj(ring
))
499 if (dirty
& (FD_DIRTY_SHADER_PROG
| FD_DIRTY_SHADER_CONST
)) {
500 struct fd_constbuf_stateobj
*constbuf
;
503 constbuf
= &ctx
->constbuf
[t
];
504 shader_dirty
= !!(dirty
& FD_DIRTY_SHADER_PROG
);
506 ring_wfi(ctx
->batch
, ring
);
508 ir3_emit_user_consts(ctx
->screen
, v
, ring
, constbuf
);
509 ir3_emit_ubos(ctx
->screen
, v
, ring
, constbuf
);
511 ir3_emit_immediates(ctx
->screen
, v
, ring
);
514 if (dirty
& (FD_DIRTY_SHADER_PROG
| FD_DIRTY_SHADER_SSBO
)) {
515 struct fd_shaderbuf_stateobj
*sb
= &ctx
->shaderbuf
[t
];
516 ring_wfi(ctx
->batch
, ring
);
517 ir3_emit_ssbo_sizes(ctx
->screen
, v
, ring
, sb
);
520 if (dirty
& (FD_DIRTY_SHADER_PROG
| FD_DIRTY_SHADER_IMAGE
)) {
521 struct fd_shaderimg_stateobj
*si
= &ctx
->shaderimg
[t
];
522 ring_wfi(ctx
->batch
, ring
);
523 ir3_emit_image_dims(ctx
->screen
, v
, ring
, si
);
528 ir3_emit_vs_driver_params(const struct ir3_shader_variant
*v
,
529 struct fd_ringbuffer
*ring
, struct fd_context
*ctx
,
530 const struct pipe_draw_info
*info
)
532 debug_assert(ir3_needs_vs_driver_params(v
));
534 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
535 uint32_t offset
= const_state
->offsets
.driver_param
;
536 uint32_t vertex_params
[IR3_DP_VS_COUNT
] = {
537 [IR3_DP_VTXID_BASE
] = info
->index_size
?
538 info
->index_bias
: info
->start
,
539 [IR3_DP_VTXCNT_MAX
] = max_tf_vtx(ctx
, v
),
541 /* if no user-clip-planes, we don't need to emit the
544 uint32_t vertex_params_size
= 4;
546 if (v
->key
.ucp_enables
) {
547 struct pipe_clip_state
*ucp
= &ctx
->ucp
;
548 unsigned pos
= IR3_DP_UCP0_X
;
549 for (unsigned i
= 0; pos
<= IR3_DP_UCP7_W
; i
++) {
550 for (unsigned j
= 0; j
< 4; j
++) {
551 vertex_params
[pos
] = fui(ucp
->ucp
[i
][j
]);
555 vertex_params_size
= ARRAY_SIZE(vertex_params
);
558 vertex_params_size
= MAX2(vertex_params_size
, const_state
->num_driver_params
);
560 bool needs_vtxid_base
=
561 ir3_find_sysval_regid(v
, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
) != regid(63, 0);
563 /* for indirect draw, we need to copy VTXID_BASE from
564 * indirect-draw parameters buffer.. which is annoying
565 * and means we can't easily emit these consts in cmd
566 * stream so need to copy them to bo.
568 if (info
->indirect
&& needs_vtxid_base
) {
569 struct pipe_draw_indirect_info
*indirect
= info
->indirect
;
570 struct pipe_resource
*vertex_params_rsc
=
571 pipe_buffer_create(&ctx
->screen
->base
,
572 PIPE_BIND_CONSTANT_BUFFER
, PIPE_USAGE_STREAM
,
573 vertex_params_size
* 4);
574 unsigned src_off
= info
->indirect
->offset
;;
577 ptr
= fd_bo_map(fd_resource(vertex_params_rsc
)->bo
);
578 memcpy(ptr
, vertex_params
, vertex_params_size
* 4);
580 if (info
->index_size
) {
581 /* indexed draw, index_bias is 4th field: */
584 /* non-indexed draw, start is 3rd field: */
588 /* copy index_bias or start from draw params: */
589 ctx
->screen
->mem_to_mem(ring
, vertex_params_rsc
, 0,
590 indirect
->buffer
, src_off
, 1);
592 emit_const(ctx
->screen
, ring
, v
, offset
* 4, 0,
593 vertex_params_size
, NULL
, vertex_params_rsc
);
595 pipe_resource_reference(&vertex_params_rsc
, NULL
);
597 emit_const(ctx
->screen
, ring
, v
, offset
* 4, 0,
598 vertex_params_size
, vertex_params
, NULL
);
601 /* if needed, emit stream-out buffer addresses: */
602 if (vertex_params
[IR3_DP_VTXCNT_MAX
] > 0) {
603 emit_tfbos(ctx
, v
, ring
);
608 ir3_emit_vs_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
609 struct fd_context
*ctx
, const struct pipe_draw_info
*info
)
611 debug_assert(v
->type
== MESA_SHADER_VERTEX
);
613 emit_common_consts(v
, ring
, ctx
, PIPE_SHADER_VERTEX
);
615 /* emit driver params every time: */
616 if (info
&& ir3_needs_vs_driver_params(v
)) {
617 ring_wfi(ctx
->batch
, ring
);
618 ir3_emit_vs_driver_params(v
, ring
, ctx
, info
);
623 ir3_emit_fs_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
624 struct fd_context
*ctx
)
626 debug_assert(v
->type
== MESA_SHADER_FRAGMENT
);
628 emit_common_consts(v
, ring
, ctx
, PIPE_SHADER_FRAGMENT
);
631 /* emit compute-shader consts: */
633 ir3_emit_cs_consts(const struct ir3_shader_variant
*v
, struct fd_ringbuffer
*ring
,
634 struct fd_context
*ctx
, const struct pipe_grid_info
*info
)
636 debug_assert(gl_shader_stage_is_compute(v
->type
));
638 emit_common_consts(v
, ring
, ctx
, PIPE_SHADER_COMPUTE
);
640 /* emit compute-shader driver-params: */
641 const struct ir3_const_state
*const_state
= &v
->shader
->const_state
;
642 uint32_t offset
= const_state
->offsets
.driver_param
;
643 if (v
->constlen
> offset
) {
644 ring_wfi(ctx
->batch
, ring
);
646 if (info
->indirect
) {
647 struct pipe_resource
*indirect
= NULL
;
648 unsigned indirect_offset
;
650 /* This is a bit awkward, but CP_LOAD_STATE.EXT_SRC_ADDR needs
651 * to be aligned more strongly than 4 bytes. So in this case
652 * we need a temporary buffer to copy NumWorkGroups.xyz to.
654 * TODO if previous compute job is writing to info->indirect,
655 * we might need a WFI.. but since we currently flush for each
656 * compute job, we are probably ok for now.
658 if (info
->indirect_offset
& 0xf) {
659 indirect
= pipe_buffer_create(&ctx
->screen
->base
,
660 PIPE_BIND_COMMAND_ARGS_BUFFER
, PIPE_USAGE_STREAM
,
664 ctx
->screen
->mem_to_mem(ring
, indirect
, 0, info
->indirect
,
665 info
->indirect_offset
, 3);
667 pipe_resource_reference(&indirect
, info
->indirect
);
668 indirect_offset
= info
->indirect_offset
;
671 emit_const(ctx
->screen
, ring
, v
, offset
* 4,
672 indirect_offset
, 4, NULL
, indirect
);
674 pipe_resource_reference(&indirect
, NULL
);
676 uint32_t compute_params
[IR3_DP_CS_COUNT
] = {
677 [IR3_DP_NUM_WORK_GROUPS_X
] = info
->grid
[0],
678 [IR3_DP_NUM_WORK_GROUPS_Y
] = info
->grid
[1],
679 [IR3_DP_NUM_WORK_GROUPS_Z
] = info
->grid
[2],
680 [IR3_DP_LOCAL_GROUP_SIZE_X
] = info
->block
[0],
681 [IR3_DP_LOCAL_GROUP_SIZE_Y
] = info
->block
[1],
682 [IR3_DP_LOCAL_GROUP_SIZE_Z
] = info
->block
[2],
684 uint32_t size
= MIN2(const_state
->num_driver_params
,
685 v
->constlen
* 4 - offset
* 4);
687 emit_const(ctx
->screen
, ring
, v
, offset
* 4, 0, size
,
688 compute_params
, NULL
);
694 ir3_shader_state_create(struct pipe_context
*pctx
, const struct pipe_shader_state
*cso
)
696 struct fd_context
*ctx
= fd_context(pctx
);
697 struct ir3_compiler
*compiler
= ctx
->screen
->compiler
;
698 return ir3_shader_create(compiler
, cso
, &ctx
->debug
, pctx
->screen
);
702 ir3_shader_state_delete(struct pipe_context
*pctx
, void *hwcso
)
704 struct ir3_shader
*so
= hwcso
;
705 ir3_shader_destroy(so
);
709 ir3_prog_init(struct pipe_context
*pctx
)
711 pctx
->create_vs_state
= ir3_shader_state_create
;
712 pctx
->delete_vs_state
= ir3_shader_state_delete
;
714 pctx
->create_tcs_state
= ir3_shader_state_create
;
715 pctx
->delete_tcs_state
= ir3_shader_state_delete
;
717 pctx
->create_tes_state
= ir3_shader_state_create
;
718 pctx
->delete_tes_state
= ir3_shader_state_delete
;
720 pctx
->create_gs_state
= ir3_shader_state_create
;
721 pctx
->delete_gs_state
= ir3_shader_state_delete
;
723 pctx
->create_fs_state
= ir3_shader_state_create
;
724 pctx
->delete_fs_state
= ir3_shader_state_delete
;