2 * Copyright 2013 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
28 /* Set this if you want the ME to wait until CP DMA is done.
29 * It should be set on the last CP DMA packet. */
30 #define CP_DMA_SYNC (1 << 0)
32 /* Set this if the source data was used as a destination in a previous CP DMA
33 * packet. It's for preventing a read-after-write (RAW) hazard between two
35 #define CP_DMA_RAW_WAIT (1 << 1)
36 #define CP_DMA_DST_IS_GDS (1 << 2)
37 #define CP_DMA_CLEAR (1 << 3)
38 #define CP_DMA_PFP_SYNC_ME (1 << 4)
39 #define CP_DMA_SRC_IS_GDS (1 << 5)
41 /* The max number of bytes that can be copied per packet. */
42 static inline unsigned cp_dma_max_byte_count(struct si_context
*sctx
)
44 unsigned max
= sctx
->chip_class
>= GFX9
?
45 S_414_BYTE_COUNT_GFX9(~0u) :
46 S_414_BYTE_COUNT_GFX6(~0u);
48 /* make it aligned for optimal performance */
49 return max
& ~(SI_CPDMA_ALIGNMENT
- 1);
53 /* Emit a CP DMA packet to do a copy from one buffer to another, or to clear
54 * a buffer. The size must fit in bits [20:0]. If CP_DMA_CLEAR is set, src_va is a 32-bit
57 static void si_emit_cp_dma(struct si_context
*sctx
, struct radeon_cmdbuf
*cs
,
58 uint64_t dst_va
, uint64_t src_va
, unsigned size
,
59 unsigned flags
, enum si_cache_policy cache_policy
)
61 uint32_t header
= 0, command
= 0;
63 assert(size
<= cp_dma_max_byte_count(sctx
));
64 assert(sctx
->chip_class
!= GFX6
|| cache_policy
== L2_BYPASS
);
66 if (sctx
->chip_class
>= GFX9
)
67 command
|= S_414_BYTE_COUNT_GFX9(size
);
69 command
|= S_414_BYTE_COUNT_GFX6(size
);
72 if (flags
& CP_DMA_SYNC
)
73 header
|= S_411_CP_SYNC(1);
75 if (sctx
->chip_class
>= GFX9
)
76 command
|= S_414_DISABLE_WR_CONFIRM_GFX9(1);
78 command
|= S_414_DISABLE_WR_CONFIRM_GFX6(1);
81 if (flags
& CP_DMA_RAW_WAIT
)
82 command
|= S_414_RAW_WAIT(1);
84 /* Src and dst flags. */
85 if (sctx
->chip_class
>= GFX9
&& !(flags
& CP_DMA_CLEAR
) &&
87 header
|= S_411_DST_SEL(V_411_NOWHERE
); /* prefetch only */
88 } else if (flags
& CP_DMA_DST_IS_GDS
) {
89 header
|= S_411_DST_SEL(V_411_GDS
);
90 /* GDS increments the address, not CP. */
91 command
|= S_414_DAS(V_414_REGISTER
) |
92 S_414_DAIC(V_414_NO_INCREMENT
);
93 } else if (sctx
->chip_class
>= GFX7
&& cache_policy
!= L2_BYPASS
) {
94 header
|= S_411_DST_SEL(V_411_DST_ADDR_TC_L2
) |
95 S_500_DST_CACHE_POLICY(cache_policy
== L2_STREAM
);
98 if (flags
& CP_DMA_CLEAR
) {
99 header
|= S_411_SRC_SEL(V_411_DATA
);
100 } else if (flags
& CP_DMA_SRC_IS_GDS
) {
101 header
|= S_411_SRC_SEL(V_411_GDS
);
102 /* Both of these are required for GDS. It does increment the address. */
103 command
|= S_414_SAS(V_414_REGISTER
) |
104 S_414_SAIC(V_414_NO_INCREMENT
);
105 } else if (sctx
->chip_class
>= GFX7
&& cache_policy
!= L2_BYPASS
) {
106 header
|= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2
) |
107 S_500_SRC_CACHE_POLICY(cache_policy
== L2_STREAM
);
110 if (sctx
->chip_class
>= GFX7
) {
111 radeon_emit(cs
, PKT3(PKT3_DMA_DATA
, 5, 0));
112 radeon_emit(cs
, header
);
113 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
114 radeon_emit(cs
, src_va
>> 32); /* SRC_ADDR_HI [31:0] */
115 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
116 radeon_emit(cs
, dst_va
>> 32); /* DST_ADDR_HI [31:0] */
117 radeon_emit(cs
, command
);
119 header
|= S_411_SRC_ADDR_HI(src_va
>> 32);
121 radeon_emit(cs
, PKT3(PKT3_CP_DMA
, 4, 0));
122 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
123 radeon_emit(cs
, header
); /* SRC_ADDR_HI [15:0] + flags. */
124 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
125 radeon_emit(cs
, (dst_va
>> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
126 radeon_emit(cs
, command
);
129 /* CP DMA is executed in ME, but index buffers are read by PFP.
130 * This ensures that ME (CP DMA) is idle before PFP starts fetching
131 * indices. If we wanted to execute CP DMA in PFP, this packet
134 if (sctx
->has_graphics
&& flags
& CP_DMA_PFP_SYNC_ME
) {
135 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
140 void si_cp_dma_wait_for_idle(struct si_context
*sctx
)
142 /* Issue a dummy DMA that copies zero bytes.
144 * The DMA engine will see that there's no work to do and skip this
145 * DMA request, however, the CP will see the sync flag and still wait
146 * for all DMAs to complete.
148 si_emit_cp_dma(sctx
, sctx
->gfx_cs
, 0, 0, 0, CP_DMA_SYNC
, L2_BYPASS
);
151 static void si_cp_dma_prepare(struct si_context
*sctx
, struct pipe_resource
*dst
,
152 struct pipe_resource
*src
, unsigned byte_count
,
153 uint64_t remaining_size
, unsigned user_flags
,
154 enum si_coherency coher
, bool *is_first
,
155 unsigned *packet_flags
)
157 /* Fast exit for a CPDMA prefetch. */
158 if ((user_flags
& SI_CPDMA_SKIP_ALL
) == SI_CPDMA_SKIP_ALL
) {
163 if (!(user_flags
& SI_CPDMA_SKIP_BO_LIST_UPDATE
)) {
164 /* Count memory usage in so that need_cs_space can take it into account. */
166 si_context_add_resource_size(sctx
, dst
);
168 si_context_add_resource_size(sctx
, src
);
171 if (!(user_flags
& SI_CPDMA_SKIP_CHECK_CS_SPACE
))
172 si_need_gfx_cs_space(sctx
);
174 /* This must be done after need_cs_space. */
175 if (!(user_flags
& SI_CPDMA_SKIP_BO_LIST_UPDATE
)) {
177 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
179 RADEON_USAGE_WRITE
, RADEON_PRIO_CP_DMA
);
181 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
183 RADEON_USAGE_READ
, RADEON_PRIO_CP_DMA
);
186 /* Flush the caches for the first copy only.
187 * Also wait for the previous CP DMA operations.
189 if (!(user_flags
& SI_CPDMA_SKIP_GFX_SYNC
) && sctx
->flags
)
190 sctx
->emit_cache_flush(sctx
);
192 if (!(user_flags
& SI_CPDMA_SKIP_SYNC_BEFORE
) && *is_first
&&
193 !(*packet_flags
& CP_DMA_CLEAR
))
194 *packet_flags
|= CP_DMA_RAW_WAIT
;
198 /* Do the synchronization after the last dma, so that all data
199 * is written to memory.
201 if (!(user_flags
& SI_CPDMA_SKIP_SYNC_AFTER
) &&
202 byte_count
== remaining_size
) {
203 *packet_flags
|= CP_DMA_SYNC
;
205 if (coher
== SI_COHERENCY_SHADER
)
206 *packet_flags
|= CP_DMA_PFP_SYNC_ME
;
210 void si_cp_dma_clear_buffer(struct si_context
*sctx
, struct radeon_cmdbuf
*cs
,
211 struct pipe_resource
*dst
, uint64_t offset
,
212 uint64_t size
, unsigned value
, unsigned user_flags
,
213 enum si_coherency coher
, enum si_cache_policy cache_policy
)
215 struct si_resource
*sdst
= si_resource(dst
);
216 uint64_t va
= (sdst
? sdst
->gpu_address
: 0) + offset
;
217 bool is_first
= true;
219 assert(size
&& size
% 4 == 0);
221 /* Mark the buffer range of destination as valid (initialized),
222 * so that transfer_map knows it should wait for the GPU when mapping
225 util_range_add(&sdst
->valid_buffer_range
, offset
, offset
+ size
);
227 /* Flush the caches. */
228 if (sdst
&& !(user_flags
& SI_CPDMA_SKIP_GFX_SYNC
)) {
229 sctx
->flags
|= SI_CONTEXT_PS_PARTIAL_FLUSH
|
230 SI_CONTEXT_CS_PARTIAL_FLUSH
|
231 si_get_flush_flags(sctx
, coher
, cache_policy
);
235 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(sctx
));
236 unsigned dma_flags
= CP_DMA_CLEAR
| (sdst
? 0 : CP_DMA_DST_IS_GDS
);
238 si_cp_dma_prepare(sctx
, dst
, NULL
, byte_count
, size
, user_flags
,
239 coher
, &is_first
, &dma_flags
);
241 /* Emit the clear packet. */
242 si_emit_cp_dma(sctx
, cs
, va
, value
, byte_count
, dma_flags
, cache_policy
);
248 if (sdst
&& cache_policy
!= L2_BYPASS
)
249 sdst
->TC_L2_dirty
= true;
251 /* If it's not a framebuffer fast clear... */
252 if (coher
== SI_COHERENCY_SHADER
) {
253 sctx
->num_cp_dma_calls
++;
254 si_prim_discard_signal_next_compute_ib_start(sctx
);
259 * Realign the CP DMA engine. This must be done after a copy with an unaligned
262 * \param size Remaining size to the CP DMA alignment.
264 static void si_cp_dma_realign_engine(struct si_context
*sctx
, unsigned size
,
265 unsigned user_flags
, enum si_coherency coher
,
266 enum si_cache_policy cache_policy
,
270 unsigned dma_flags
= 0;
271 unsigned scratch_size
= SI_CPDMA_ALIGNMENT
* 2;
273 assert(size
< SI_CPDMA_ALIGNMENT
);
275 /* Use the scratch buffer as the dummy buffer. The 3D engine should be
276 * idle at this point.
278 if (!sctx
->scratch_buffer
||
279 sctx
->scratch_buffer
->b
.b
.width0
< scratch_size
) {
280 si_resource_reference(&sctx
->scratch_buffer
, NULL
);
281 sctx
->scratch_buffer
=
282 si_aligned_buffer_create(&sctx
->screen
->b
,
283 SI_RESOURCE_FLAG_UNMAPPABLE
,
286 if (!sctx
->scratch_buffer
)
289 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.scratch_state
);
292 si_cp_dma_prepare(sctx
, &sctx
->scratch_buffer
->b
.b
,
293 &sctx
->scratch_buffer
->b
.b
, size
, size
, user_flags
,
294 coher
, is_first
, &dma_flags
);
296 va
= sctx
->scratch_buffer
->gpu_address
;
297 si_emit_cp_dma(sctx
, sctx
->gfx_cs
, va
, va
+ SI_CPDMA_ALIGNMENT
, size
, dma_flags
,
302 * Do memcpy between buffers using CP DMA.
303 * If src or dst is NULL, it means read or write GDS, respectively.
305 * \param user_flags bitmask of SI_CPDMA_*
307 void si_cp_dma_copy_buffer(struct si_context
*sctx
,
308 struct pipe_resource
*dst
, struct pipe_resource
*src
,
309 uint64_t dst_offset
, uint64_t src_offset
, unsigned size
,
310 unsigned user_flags
, enum si_coherency coher
,
311 enum si_cache_policy cache_policy
)
313 uint64_t main_dst_offset
, main_src_offset
;
314 unsigned skipped_size
= 0;
315 unsigned realign_size
= 0;
316 unsigned gds_flags
= (dst
? 0 : CP_DMA_DST_IS_GDS
) |
317 (src
? 0 : CP_DMA_SRC_IS_GDS
);
318 bool is_first
= true;
323 /* Skip this for the L2 prefetch. */
324 if (dst
!= src
|| dst_offset
!= src_offset
) {
325 /* Mark the buffer range of destination as valid (initialized),
326 * so that transfer_map knows it should wait for the GPU when mapping
328 util_range_add(&si_resource(dst
)->valid_buffer_range
, dst_offset
,
332 dst_offset
+= si_resource(dst
)->gpu_address
;
335 src_offset
+= si_resource(src
)->gpu_address
;
337 /* The workarounds aren't needed on Fiji and beyond. */
338 if (sctx
->family
<= CHIP_CARRIZO
||
339 sctx
->family
== CHIP_STONEY
) {
340 /* If the size is not aligned, we must add a dummy copy at the end
341 * just to align the internal counter. Otherwise, the DMA engine
342 * would slow down by an order of magnitude for following copies.
344 if (size
% SI_CPDMA_ALIGNMENT
)
345 realign_size
= SI_CPDMA_ALIGNMENT
- (size
% SI_CPDMA_ALIGNMENT
);
347 /* If the copy begins unaligned, we must start copying from the next
348 * aligned block and the skipped part should be copied after everything
349 * else has been copied. Only the src alignment matters, not dst.
351 * GDS doesn't need the source address to be aligned.
353 if (src
&& src_offset
% SI_CPDMA_ALIGNMENT
) {
354 skipped_size
= SI_CPDMA_ALIGNMENT
- (src_offset
% SI_CPDMA_ALIGNMENT
);
355 /* The main part will be skipped if the size is too small. */
356 skipped_size
= MIN2(skipped_size
, size
);
357 size
-= skipped_size
;
361 /* Flush the caches. */
362 if ((dst
|| src
) && !(user_flags
& SI_CPDMA_SKIP_GFX_SYNC
)) {
363 sctx
->flags
|= SI_CONTEXT_PS_PARTIAL_FLUSH
|
364 SI_CONTEXT_CS_PARTIAL_FLUSH
|
365 si_get_flush_flags(sctx
, coher
, cache_policy
);
368 /* This is the main part doing the copying. Src is always aligned. */
369 main_dst_offset
= dst_offset
+ skipped_size
;
370 main_src_offset
= src_offset
+ skipped_size
;
373 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(sctx
));
374 unsigned dma_flags
= gds_flags
;
376 si_cp_dma_prepare(sctx
, dst
, src
, byte_count
,
377 size
+ skipped_size
+ realign_size
,
378 user_flags
, coher
, &is_first
, &dma_flags
);
380 si_emit_cp_dma(sctx
, sctx
->gfx_cs
, main_dst_offset
, main_src_offset
,
381 byte_count
, dma_flags
, cache_policy
);
384 main_src_offset
+= byte_count
;
385 main_dst_offset
+= byte_count
;
388 /* Copy the part we skipped because src wasn't aligned. */
390 unsigned dma_flags
= gds_flags
;
392 si_cp_dma_prepare(sctx
, dst
, src
, skipped_size
,
393 skipped_size
+ realign_size
, user_flags
,
394 coher
, &is_first
, &dma_flags
);
396 si_emit_cp_dma(sctx
, sctx
->gfx_cs
, dst_offset
, src_offset
, skipped_size
,
397 dma_flags
, cache_policy
);
400 /* Finally, realign the engine if the size wasn't aligned. */
402 si_cp_dma_realign_engine(sctx
, realign_size
, user_flags
, coher
,
403 cache_policy
, &is_first
);
406 if (dst
&& cache_policy
!= L2_BYPASS
)
407 si_resource(dst
)->TC_L2_dirty
= true;
409 /* If it's not a prefetch or GDS copy... */
410 if (dst
&& src
&& (dst
!= src
|| dst_offset
!= src_offset
)) {
411 sctx
->num_cp_dma_calls
++;
412 si_prim_discard_signal_next_compute_ib_start(sctx
);
416 void cik_prefetch_TC_L2_async(struct si_context
*sctx
, struct pipe_resource
*buf
,
417 uint64_t offset
, unsigned size
)
419 assert(sctx
->chip_class
>= GFX7
);
421 si_cp_dma_copy_buffer(sctx
, buf
, buf
, offset
, offset
, size
,
422 SI_CPDMA_SKIP_ALL
, SI_COHERENCY_SHADER
, L2_LRU
);
425 static void cik_prefetch_shader_async(struct si_context
*sctx
,
426 struct si_pm4_state
*state
)
428 struct pipe_resource
*bo
= &state
->bo
[0]->b
.b
;
429 assert(state
->nbo
== 1);
431 cik_prefetch_TC_L2_async(sctx
, bo
, 0, bo
->width0
);
434 static void cik_prefetch_VBO_descriptors(struct si_context
*sctx
)
436 if (!sctx
->vertex_elements
|| !sctx
->vertex_elements
->desc_list_byte_size
)
439 cik_prefetch_TC_L2_async(sctx
, &sctx
->vb_descriptors_buffer
->b
.b
,
440 sctx
->vb_descriptors_offset
,
441 sctx
->vertex_elements
->desc_list_byte_size
);
445 * Prefetch shaders and VBO descriptors.
447 * \param vertex_stage_only Whether only the the API VS and VBO descriptors
448 * should be prefetched.
450 void cik_emit_prefetch_L2(struct si_context
*sctx
, bool vertex_stage_only
)
452 unsigned mask
= sctx
->prefetch_L2_mask
;
455 /* Prefetch shaders and VBO descriptors to TC L2. */
456 if (sctx
->chip_class
>= GFX9
) {
457 /* Choose the right spot for the VBO prefetch. */
458 if (sctx
->queued
.named
.hs
) {
459 if (mask
& SI_PREFETCH_HS
)
460 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.hs
);
461 if (mask
& SI_PREFETCH_VBO_DESCRIPTORS
)
462 cik_prefetch_VBO_descriptors(sctx
);
463 if (vertex_stage_only
) {
464 sctx
->prefetch_L2_mask
&= ~(SI_PREFETCH_HS
|
465 SI_PREFETCH_VBO_DESCRIPTORS
);
469 if (mask
& SI_PREFETCH_GS
)
470 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.gs
);
471 if (mask
& SI_PREFETCH_VS
)
472 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.vs
);
473 } else if (sctx
->queued
.named
.gs
) {
474 if (mask
& SI_PREFETCH_GS
)
475 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.gs
);
476 if (mask
& SI_PREFETCH_VBO_DESCRIPTORS
)
477 cik_prefetch_VBO_descriptors(sctx
);
478 if (vertex_stage_only
) {
479 sctx
->prefetch_L2_mask
&= ~(SI_PREFETCH_GS
|
480 SI_PREFETCH_VBO_DESCRIPTORS
);
484 if (mask
& SI_PREFETCH_VS
)
485 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.vs
);
487 if (mask
& SI_PREFETCH_VS
)
488 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.vs
);
489 if (mask
& SI_PREFETCH_VBO_DESCRIPTORS
)
490 cik_prefetch_VBO_descriptors(sctx
);
491 if (vertex_stage_only
) {
492 sctx
->prefetch_L2_mask
&= ~(SI_PREFETCH_VS
|
493 SI_PREFETCH_VBO_DESCRIPTORS
);
499 /* Choose the right spot for the VBO prefetch. */
500 if (sctx
->tes_shader
.cso
) {
501 if (mask
& SI_PREFETCH_LS
)
502 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.ls
);
503 if (mask
& SI_PREFETCH_VBO_DESCRIPTORS
)
504 cik_prefetch_VBO_descriptors(sctx
);
505 if (vertex_stage_only
) {
506 sctx
->prefetch_L2_mask
&= ~(SI_PREFETCH_LS
|
507 SI_PREFETCH_VBO_DESCRIPTORS
);
511 if (mask
& SI_PREFETCH_HS
)
512 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.hs
);
513 if (mask
& SI_PREFETCH_ES
)
514 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.es
);
515 if (mask
& SI_PREFETCH_GS
)
516 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.gs
);
517 if (mask
& SI_PREFETCH_VS
)
518 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.vs
);
519 } else if (sctx
->gs_shader
.cso
) {
520 if (mask
& SI_PREFETCH_ES
)
521 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.es
);
522 if (mask
& SI_PREFETCH_VBO_DESCRIPTORS
)
523 cik_prefetch_VBO_descriptors(sctx
);
524 if (vertex_stage_only
) {
525 sctx
->prefetch_L2_mask
&= ~(SI_PREFETCH_ES
|
526 SI_PREFETCH_VBO_DESCRIPTORS
);
530 if (mask
& SI_PREFETCH_GS
)
531 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.gs
);
532 if (mask
& SI_PREFETCH_VS
)
533 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.vs
);
535 if (mask
& SI_PREFETCH_VS
)
536 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.vs
);
537 if (mask
& SI_PREFETCH_VBO_DESCRIPTORS
)
538 cik_prefetch_VBO_descriptors(sctx
);
539 if (vertex_stage_only
) {
540 sctx
->prefetch_L2_mask
&= ~(SI_PREFETCH_VS
|
541 SI_PREFETCH_VBO_DESCRIPTORS
);
547 if (mask
& SI_PREFETCH_PS
)
548 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.ps
);
550 sctx
->prefetch_L2_mask
= 0;
553 void si_test_gds(struct si_context
*sctx
)
555 struct pipe_context
*ctx
= &sctx
->b
;
556 struct pipe_resource
*src
, *dst
;
558 unsigned offset
= debug_get_num_option("OFFSET", 16);
560 src
= pipe_buffer_create(ctx
->screen
, 0, PIPE_USAGE_DEFAULT
, 16);
561 dst
= pipe_buffer_create(ctx
->screen
, 0, PIPE_USAGE_DEFAULT
, 16);
562 si_cp_dma_clear_buffer(sctx
, sctx
->gfx_cs
, src
, 0, 4, 0xabcdef01, 0, SI_COHERENCY_SHADER
, L2_BYPASS
);
563 si_cp_dma_clear_buffer(sctx
, sctx
->gfx_cs
, src
, 4, 4, 0x23456789, 0, SI_COHERENCY_SHADER
, L2_BYPASS
);
564 si_cp_dma_clear_buffer(sctx
, sctx
->gfx_cs
, src
, 8, 4, 0x87654321, 0, SI_COHERENCY_SHADER
, L2_BYPASS
);
565 si_cp_dma_clear_buffer(sctx
, sctx
->gfx_cs
, src
, 12, 4, 0xfedcba98, 0, SI_COHERENCY_SHADER
, L2_BYPASS
);
566 si_cp_dma_clear_buffer(sctx
, sctx
->gfx_cs
, dst
, 0, 16, 0xdeadbeef, 0, SI_COHERENCY_SHADER
, L2_BYPASS
);
568 si_cp_dma_copy_buffer(sctx
, NULL
, src
, offset
, 0, 16, 0, SI_COHERENCY_NONE
, L2_BYPASS
);
569 si_cp_dma_copy_buffer(sctx
, dst
, NULL
, 0, offset
, 16, 0, SI_COHERENCY_NONE
, L2_BYPASS
);
571 pipe_buffer_read(ctx
, dst
, 0, sizeof(r
), r
);
572 printf("GDS copy = %08x %08x %08x %08x -> %s\n", r
[0], r
[1], r
[2], r
[3],
573 r
[0] == 0xabcdef01 && r
[1] == 0x23456789 &&
574 r
[2] == 0x87654321 && r
[3] == 0xfedcba98 ? "pass" : "fail");
576 si_cp_dma_clear_buffer(sctx
, sctx
->gfx_cs
, NULL
, offset
, 16, 0xc1ea4146, 0, SI_COHERENCY_NONE
, L2_BYPASS
);
577 si_cp_dma_copy_buffer(sctx
, dst
, NULL
, 0, offset
, 16, 0, SI_COHERENCY_NONE
, L2_BYPASS
);
579 pipe_buffer_read(ctx
, dst
, 0, sizeof(r
), r
);
580 printf("GDS clear = %08x %08x %08x %08x -> %s\n", r
[0], r
[1], r
[2], r
[3],
581 r
[0] == 0xc1ea4146 && r
[1] == 0xc1ea4146 &&
582 r
[2] == 0xc1ea4146 && r
[3] == 0xc1ea4146 ? "pass" : "fail");
584 pipe_resource_reference(&src
, NULL
);
585 pipe_resource_reference(&dst
, NULL
);
589 void si_cp_write_data(struct si_context
*sctx
, struct si_resource
*buf
,
590 unsigned offset
, unsigned size
, unsigned dst_sel
,
591 unsigned engine
, const void *data
)
593 struct radeon_cmdbuf
*cs
= sctx
->gfx_cs
;
595 assert(offset
% 4 == 0);
596 assert(size
% 4 == 0);
598 if (sctx
->chip_class
== GFX6
&& dst_sel
== V_370_MEM
)
599 dst_sel
= V_370_MEM_GRBM
;
601 radeon_add_to_buffer_list(sctx
, cs
, buf
,
602 RADEON_USAGE_WRITE
, RADEON_PRIO_CP_DMA
);
603 uint64_t va
= buf
->gpu_address
+ offset
;
605 radeon_emit(cs
, PKT3(PKT3_WRITE_DATA
, 2 + size
/4, 0));
606 radeon_emit(cs
, S_370_DST_SEL(dst_sel
) |
607 S_370_WR_CONFIRM(1) |
608 S_370_ENGINE_SEL(engine
));
610 radeon_emit(cs
, va
>> 32);
611 radeon_emit_array(cs
, (const uint32_t*)data
, size
/4);
614 void si_cp_copy_data(struct si_context
*sctx
, struct radeon_cmdbuf
*cs
,
615 unsigned dst_sel
, struct si_resource
*dst
, unsigned dst_offset
,
616 unsigned src_sel
, struct si_resource
*src
, unsigned src_offset
)
618 /* cs can point to the compute IB, which has the buffer list in gfx_cs. */
620 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
, dst
,
621 RADEON_USAGE_WRITE
, RADEON_PRIO_CP_DMA
);
624 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
, src
,
625 RADEON_USAGE_READ
, RADEON_PRIO_CP_DMA
);
628 uint64_t dst_va
= (dst
? dst
->gpu_address
: 0ull) + dst_offset
;
629 uint64_t src_va
= (src
? src
->gpu_address
: 0ull) + src_offset
;
631 radeon_emit(cs
, PKT3(PKT3_COPY_DATA
, 4, 0));
632 radeon_emit(cs
, COPY_DATA_SRC_SEL(src_sel
) |
633 COPY_DATA_DST_SEL(dst_sel
) |
634 COPY_DATA_WR_CONFIRM
);
635 radeon_emit(cs
, src_va
);
636 radeon_emit(cs
, src_va
>> 32);
637 radeon_emit(cs
, dst_va
);
638 radeon_emit(cs
, dst_va
>> 32);