2 * Copyright 2013 Advanced Micro Devices, Inc.
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 #include "radeon/r600_cs.h"
28 /* Recommended maximum sizes for optimal performance.
29 * Fall back to compute or SDMA if the size is greater.
31 #define CP_DMA_COPY_PERF_THRESHOLD (64 * 1024) /* copied from Vulkan */
32 #define CP_DMA_CLEAR_PERF_THRESHOLD (32 * 1024) /* guess (clear is much slower) */
34 /* Set this if you want the ME to wait until CP DMA is done.
35 * It should be set on the last CP DMA packet. */
36 #define CP_DMA_SYNC (1 << 0)
38 /* Set this if the source data was used as a destination in a previous CP DMA
39 * packet. It's for preventing a read-after-write (RAW) hazard between two
41 #define CP_DMA_RAW_WAIT (1 << 1)
42 #define CP_DMA_USE_L2 (1 << 2) /* CIK+ */
43 #define CP_DMA_CLEAR (1 << 3)
45 /* The max number of bytes that can be copied per packet. */
46 static inline unsigned cp_dma_max_byte_count(struct si_context
*sctx
)
48 unsigned max
= sctx
->b
.chip_class
>= GFX9
?
49 S_414_BYTE_COUNT_GFX9(~0u) :
50 S_414_BYTE_COUNT_GFX6(~0u);
52 /* make it aligned for optimal performance */
53 return max
& ~(SI_CPDMA_ALIGNMENT
- 1);
57 /* Emit a CP DMA packet to do a copy from one buffer to another, or to clear
58 * a buffer. The size must fit in bits [20:0]. If CP_DMA_CLEAR is set, src_va is a 32-bit
61 static void si_emit_cp_dma(struct si_context
*sctx
, uint64_t dst_va
,
62 uint64_t src_va
, unsigned size
, unsigned flags
,
63 enum r600_coherency coher
)
65 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
66 uint32_t header
= 0, command
= 0;
69 assert(size
<= cp_dma_max_byte_count(sctx
));
71 if (sctx
->b
.chip_class
>= GFX9
)
72 command
|= S_414_BYTE_COUNT_GFX9(size
);
74 command
|= S_414_BYTE_COUNT_GFX6(size
);
77 if (flags
& CP_DMA_SYNC
)
78 header
|= S_411_CP_SYNC(1);
80 if (sctx
->b
.chip_class
>= GFX9
)
81 command
|= S_414_DISABLE_WR_CONFIRM_GFX9(1);
83 command
|= S_414_DISABLE_WR_CONFIRM_GFX6(1);
86 if (flags
& CP_DMA_RAW_WAIT
)
87 command
|= S_414_RAW_WAIT(1);
89 /* Src and dst flags. */
90 if (sctx
->b
.chip_class
>= GFX9
&& !(flags
& CP_DMA_CLEAR
) &&
92 header
|= S_411_DSL_SEL(V_411_NOWHERE
); /* prefetch only */
93 else if (flags
& CP_DMA_USE_L2
)
94 header
|= S_411_DSL_SEL(V_411_DST_ADDR_TC_L2
);
96 if (flags
& CP_DMA_CLEAR
)
97 header
|= S_411_SRC_SEL(V_411_DATA
);
98 else if (flags
& CP_DMA_USE_L2
)
99 header
|= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2
);
101 if (sctx
->b
.chip_class
>= CIK
) {
102 radeon_emit(cs
, PKT3(PKT3_DMA_DATA
, 5, 0));
103 radeon_emit(cs
, header
);
104 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
105 radeon_emit(cs
, src_va
>> 32); /* SRC_ADDR_HI [31:0] */
106 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
107 radeon_emit(cs
, dst_va
>> 32); /* DST_ADDR_HI [31:0] */
108 radeon_emit(cs
, command
);
110 header
|= S_411_SRC_ADDR_HI(src_va
>> 32);
112 radeon_emit(cs
, PKT3(PKT3_CP_DMA
, 4, 0));
113 radeon_emit(cs
, src_va
); /* SRC_ADDR_LO [31:0] */
114 radeon_emit(cs
, header
); /* SRC_ADDR_HI [15:0] + flags. */
115 radeon_emit(cs
, dst_va
); /* DST_ADDR_LO [31:0] */
116 radeon_emit(cs
, (dst_va
>> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
117 radeon_emit(cs
, command
);
120 /* CP DMA is executed in ME, but index buffers are read by PFP.
121 * This ensures that ME (CP DMA) is idle before PFP starts fetching
122 * indices. If we wanted to execute CP DMA in PFP, this packet
125 if (coher
== R600_COHERENCY_SHADER
&& flags
& CP_DMA_SYNC
) {
126 radeon_emit(cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
131 static unsigned get_flush_flags(struct si_context
*sctx
, enum r600_coherency coher
)
135 case R600_COHERENCY_NONE
:
137 case R600_COHERENCY_SHADER
:
138 return SI_CONTEXT_INV_SMEM_L1
|
139 SI_CONTEXT_INV_VMEM_L1
|
140 (sctx
->b
.chip_class
== SI
? SI_CONTEXT_INV_GLOBAL_L2
: 0);
141 case R600_COHERENCY_CB_META
:
142 return SI_CONTEXT_FLUSH_AND_INV_CB
;
146 static unsigned get_tc_l2_flag(struct si_context
*sctx
, enum r600_coherency coher
)
148 if ((sctx
->b
.chip_class
>= GFX9
&& coher
== R600_COHERENCY_CB_META
) ||
149 (sctx
->b
.chip_class
>= CIK
&& coher
== R600_COHERENCY_SHADER
))
150 return CP_DMA_USE_L2
;
155 static void si_cp_dma_prepare(struct si_context
*sctx
, struct pipe_resource
*dst
,
156 struct pipe_resource
*src
, unsigned byte_count
,
157 uint64_t remaining_size
, unsigned user_flags
,
158 bool *is_first
, unsigned *packet_flags
)
160 /* Fast exit for a CPDMA prefetch. */
161 if ((user_flags
& SI_CPDMA_SKIP_ALL
) == SI_CPDMA_SKIP_ALL
) {
166 if (!(user_flags
& SI_CPDMA_SKIP_BO_LIST_UPDATE
)) {
167 /* Count memory usage in so that need_cs_space can take it into account. */
168 r600_context_add_resource_size(&sctx
->b
.b
, dst
);
170 r600_context_add_resource_size(&sctx
->b
.b
, src
);
173 if (!(user_flags
& SI_CPDMA_SKIP_CHECK_CS_SPACE
))
174 si_need_cs_space(sctx
);
176 /* This must be done after need_cs_space. */
177 if (!(user_flags
& SI_CPDMA_SKIP_BO_LIST_UPDATE
)) {
178 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
179 (struct r600_resource
*)dst
,
180 RADEON_USAGE_WRITE
, RADEON_PRIO_CP_DMA
);
182 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
183 (struct r600_resource
*)src
,
184 RADEON_USAGE_READ
, RADEON_PRIO_CP_DMA
);
187 /* Flush the caches for the first copy only.
188 * Also wait for the previous CP DMA operations.
190 if (!(user_flags
& SI_CPDMA_SKIP_GFX_SYNC
) && sctx
->b
.flags
)
191 si_emit_cache_flush(sctx
);
193 if (!(user_flags
& SI_CPDMA_SKIP_SYNC_BEFORE
) && *is_first
)
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
;
206 static void si_clear_buffer(struct pipe_context
*ctx
, struct pipe_resource
*dst
,
207 uint64_t offset
, uint64_t size
, unsigned value
,
208 enum r600_coherency coher
)
210 struct si_context
*sctx
= (struct si_context
*)ctx
;
211 struct radeon_winsys
*ws
= sctx
->b
.ws
;
212 struct r600_resource
*rdst
= r600_resource(dst
);
213 unsigned tc_l2_flag
= get_tc_l2_flag(sctx
, coher
);
214 unsigned flush_flags
= get_flush_flags(sctx
, coher
);
215 uint64_t dma_clear_size
;
216 bool is_first
= true;
221 dma_clear_size
= size
& ~3ull;
223 /* Mark the buffer range of destination as valid (initialized),
224 * so that transfer_map knows it should wait for the GPU when mapping
226 util_range_add(&rdst
->valid_buffer_range
, offset
,
227 offset
+ dma_clear_size
);
229 /* dma_clear_buffer can use clear_buffer on failure. Make sure that
230 * doesn't happen. We don't want an infinite recursion: */
231 if (sctx
->b
.dma
.cs
&&
232 !(dst
->flags
& PIPE_RESOURCE_FLAG_SPARSE
) &&
234 /* CP DMA is very slow. Always use SDMA for big clears. This
235 * alone improves DeusEx:MD performance by 70%. */
236 (size
> CP_DMA_CLEAR_PERF_THRESHOLD
||
237 /* Buffers not used by the GFX IB yet will be cleared by SDMA.
238 * This happens to move most buffer clears to SDMA, including
239 * DCC and CMASK clears, because pipe->clear clears them before
240 * si_emit_framebuffer_state (in a draw call) adds them.
241 * For example, DeusEx:MD has 21 buffer clears per frame and all
242 * of them are moved to SDMA thanks to this. */
243 !ws
->cs_is_buffer_referenced(sctx
->b
.gfx
.cs
, rdst
->buf
,
244 RADEON_USAGE_READWRITE
))) {
245 sctx
->b
.dma_clear_buffer(ctx
, dst
, offset
, dma_clear_size
, value
);
247 offset
+= dma_clear_size
;
248 size
-= dma_clear_size
;
249 } else if (dma_clear_size
>= 4) {
250 uint64_t va
= rdst
->gpu_address
+ offset
;
252 offset
+= dma_clear_size
;
253 size
-= dma_clear_size
;
255 /* Flush the caches. */
256 sctx
->b
.flags
|= SI_CONTEXT_PS_PARTIAL_FLUSH
|
257 SI_CONTEXT_CS_PARTIAL_FLUSH
| flush_flags
;
259 while (dma_clear_size
) {
260 unsigned byte_count
= MIN2(dma_clear_size
, cp_dma_max_byte_count(sctx
));
261 unsigned dma_flags
= tc_l2_flag
| CP_DMA_CLEAR
;
263 si_cp_dma_prepare(sctx
, dst
, NULL
, byte_count
, dma_clear_size
, 0,
264 &is_first
, &dma_flags
);
266 /* Emit the clear packet. */
267 si_emit_cp_dma(sctx
, va
, value
, byte_count
, dma_flags
, coher
);
269 dma_clear_size
-= byte_count
;
274 rdst
->TC_L2_dirty
= true;
276 /* If it's not a framebuffer fast clear... */
277 if (coher
== R600_COHERENCY_SHADER
)
278 sctx
->b
.num_cp_dma_calls
++;
282 /* Handle non-dword alignment.
284 * This function is called for embedded texture metadata clears,
285 * but those should always be properly aligned. */
286 assert(dst
->target
== PIPE_BUFFER
);
289 pipe_buffer_write(ctx
, dst
, offset
, size
, &value
);
294 * Realign the CP DMA engine. This must be done after a copy with an unaligned
297 * \param size Remaining size to the CP DMA alignment.
299 static void si_cp_dma_realign_engine(struct si_context
*sctx
, unsigned size
,
300 unsigned user_flags
, bool *is_first
)
303 unsigned dma_flags
= 0;
304 unsigned scratch_size
= SI_CPDMA_ALIGNMENT
* 2;
306 assert(size
< SI_CPDMA_ALIGNMENT
);
308 /* Use the scratch buffer as the dummy buffer. The 3D engine should be
309 * idle at this point.
311 if (!sctx
->scratch_buffer
||
312 sctx
->scratch_buffer
->b
.b
.width0
< scratch_size
) {
313 r600_resource_reference(&sctx
->scratch_buffer
, NULL
);
314 sctx
->scratch_buffer
= (struct r600_resource
*)
315 si_aligned_buffer_create(&sctx
->screen
->b
.b
,
316 R600_RESOURCE_FLAG_UNMAPPABLE
,
319 if (!sctx
->scratch_buffer
)
322 si_mark_atom_dirty(sctx
, &sctx
->scratch_state
);
325 si_cp_dma_prepare(sctx
, &sctx
->scratch_buffer
->b
.b
,
326 &sctx
->scratch_buffer
->b
.b
, size
, size
, user_flags
,
327 is_first
, &dma_flags
);
329 va
= sctx
->scratch_buffer
->gpu_address
;
330 si_emit_cp_dma(sctx
, va
, va
+ SI_CPDMA_ALIGNMENT
, size
, dma_flags
,
331 R600_COHERENCY_SHADER
);
335 * Do memcpy between buffers using CP DMA.
337 * \param user_flags bitmask of SI_CPDMA_*
339 void si_copy_buffer(struct si_context
*sctx
,
340 struct pipe_resource
*dst
, struct pipe_resource
*src
,
341 uint64_t dst_offset
, uint64_t src_offset
, unsigned size
,
344 uint64_t main_dst_offset
, main_src_offset
;
345 unsigned skipped_size
= 0;
346 unsigned realign_size
= 0;
347 unsigned tc_l2_flag
= get_tc_l2_flag(sctx
, R600_COHERENCY_SHADER
);
348 unsigned flush_flags
= get_flush_flags(sctx
, R600_COHERENCY_SHADER
);
349 bool is_first
= true;
354 if (dst
!= src
|| dst_offset
!= src_offset
) {
355 /* Mark the buffer range of destination as valid (initialized),
356 * so that transfer_map knows it should wait for the GPU when mapping
358 util_range_add(&r600_resource(dst
)->valid_buffer_range
, dst_offset
,
362 dst_offset
+= r600_resource(dst
)->gpu_address
;
363 src_offset
+= r600_resource(src
)->gpu_address
;
365 /* The workarounds aren't needed on Fiji and beyond. */
366 if (sctx
->b
.family
<= CHIP_CARRIZO
||
367 sctx
->b
.family
== CHIP_STONEY
) {
368 /* If the size is not aligned, we must add a dummy copy at the end
369 * just to align the internal counter. Otherwise, the DMA engine
370 * would slow down by an order of magnitude for following copies.
372 if (size
% SI_CPDMA_ALIGNMENT
)
373 realign_size
= SI_CPDMA_ALIGNMENT
- (size
% SI_CPDMA_ALIGNMENT
);
375 /* If the copy begins unaligned, we must start copying from the next
376 * aligned block and the skipped part should be copied after everything
377 * else has been copied. Only the src alignment matters, not dst.
379 if (src_offset
% SI_CPDMA_ALIGNMENT
) {
380 skipped_size
= SI_CPDMA_ALIGNMENT
- (src_offset
% SI_CPDMA_ALIGNMENT
);
381 /* The main part will be skipped if the size is too small. */
382 skipped_size
= MIN2(skipped_size
, size
);
383 size
-= skipped_size
;
387 /* Flush the caches. */
388 if (!(user_flags
& SI_CPDMA_SKIP_GFX_SYNC
))
389 sctx
->b
.flags
|= SI_CONTEXT_PS_PARTIAL_FLUSH
|
390 SI_CONTEXT_CS_PARTIAL_FLUSH
| flush_flags
;
392 /* This is the main part doing the copying. Src is always aligned. */
393 main_dst_offset
= dst_offset
+ skipped_size
;
394 main_src_offset
= src_offset
+ skipped_size
;
397 unsigned dma_flags
= tc_l2_flag
;
398 unsigned byte_count
= MIN2(size
, cp_dma_max_byte_count(sctx
));
400 si_cp_dma_prepare(sctx
, dst
, src
, byte_count
,
401 size
+ skipped_size
+ realign_size
,
402 user_flags
, &is_first
, &dma_flags
);
404 si_emit_cp_dma(sctx
, main_dst_offset
, main_src_offset
,
405 byte_count
, dma_flags
, R600_COHERENCY_SHADER
);
408 main_src_offset
+= byte_count
;
409 main_dst_offset
+= byte_count
;
412 /* Copy the part we skipped because src wasn't aligned. */
414 unsigned dma_flags
= tc_l2_flag
;
416 si_cp_dma_prepare(sctx
, dst
, src
, skipped_size
,
417 skipped_size
+ realign_size
, user_flags
,
418 &is_first
, &dma_flags
);
420 si_emit_cp_dma(sctx
, dst_offset
, src_offset
, skipped_size
,
421 dma_flags
, R600_COHERENCY_SHADER
);
424 /* Finally, realign the engine if the size wasn't aligned. */
426 si_cp_dma_realign_engine(sctx
, realign_size
, user_flags
,
430 r600_resource(dst
)->TC_L2_dirty
= true;
432 /* If it's not a prefetch... */
433 if (dst_offset
!= src_offset
)
434 sctx
->b
.num_cp_dma_calls
++;
437 void cik_prefetch_TC_L2_async(struct si_context
*sctx
, struct pipe_resource
*buf
,
438 uint64_t offset
, unsigned size
)
440 assert(sctx
->b
.chip_class
>= CIK
);
442 si_copy_buffer(sctx
, buf
, buf
, offset
, offset
, size
, SI_CPDMA_SKIP_ALL
);
445 static void cik_prefetch_shader_async(struct si_context
*sctx
,
446 struct si_pm4_state
*state
)
448 struct pipe_resource
*bo
= &state
->bo
[0]->b
.b
;
449 assert(state
->nbo
== 1);
451 cik_prefetch_TC_L2_async(sctx
, bo
, 0, bo
->width0
);
454 static void cik_prefetch_VBO_descriptors(struct si_context
*sctx
)
456 if (!sctx
->vertex_elements
)
459 cik_prefetch_TC_L2_async(sctx
, &sctx
->vertex_buffers
.buffer
->b
.b
,
460 sctx
->vertex_buffers
.gpu_address
-
461 sctx
->vertex_buffers
.buffer
->gpu_address
,
462 sctx
->vertex_elements
->desc_list_byte_size
);
465 void cik_emit_prefetch_L2(struct si_context
*sctx
)
467 /* Prefetch shaders and VBO descriptors to TC L2. */
468 if (sctx
->b
.chip_class
>= GFX9
) {
469 /* Choose the right spot for the VBO prefetch. */
470 if (sctx
->tes_shader
.cso
) {
471 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_HS
)
472 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.hs
);
473 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_VBO_DESCRIPTORS
)
474 cik_prefetch_VBO_descriptors(sctx
);
475 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_GS
)
476 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.gs
);
477 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_VS
)
478 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.vs
);
479 } else if (sctx
->gs_shader
.cso
) {
480 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_GS
)
481 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.gs
);
482 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_VBO_DESCRIPTORS
)
483 cik_prefetch_VBO_descriptors(sctx
);
484 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_VS
)
485 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.vs
);
487 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_VS
)
488 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.vs
);
489 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_VBO_DESCRIPTORS
)
490 cik_prefetch_VBO_descriptors(sctx
);
494 /* Choose the right spot for the VBO prefetch. */
495 if (sctx
->tes_shader
.cso
) {
496 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_LS
)
497 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.ls
);
498 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_VBO_DESCRIPTORS
)
499 cik_prefetch_VBO_descriptors(sctx
);
500 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_HS
)
501 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.hs
);
502 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_ES
)
503 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.es
);
504 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_GS
)
505 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.gs
);
506 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_VS
)
507 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.vs
);
508 } else if (sctx
->gs_shader
.cso
) {
509 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_ES
)
510 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.es
);
511 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_VBO_DESCRIPTORS
)
512 cik_prefetch_VBO_descriptors(sctx
);
513 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_GS
)
514 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.gs
);
515 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_VS
)
516 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.vs
);
518 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_VS
)
519 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.vs
);
520 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_VBO_DESCRIPTORS
)
521 cik_prefetch_VBO_descriptors(sctx
);
525 if (sctx
->prefetch_L2_mask
& SI_PREFETCH_PS
)
526 cik_prefetch_shader_async(sctx
, sctx
->queued
.named
.ps
);
528 sctx
->prefetch_L2_mask
= 0;
531 void si_init_cp_dma_functions(struct si_context
*sctx
)
533 sctx
->b
.clear_buffer
= si_clear_buffer
;