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.
25 #include "tgsi/tgsi_parse.h"
26 #include "util/u_memory.h"
27 #include "util/u_upload_mgr.h"
28 #include "radeon/r600_pipe_common.h"
29 #include "radeon/radeon_elf_util.h"
31 #include "radeon/r600_cs.h"
33 #include "si_shader.h"
36 #define MAX_GLOBAL_BUFFERS 20
41 unsigned private_size
;
43 struct si_shader shader
;
45 struct pipe_resource
*global_buffers
[MAX_GLOBAL_BUFFERS
];
48 static void *si_create_compute_state(
49 struct pipe_context
*ctx
,
50 const struct pipe_compute_state
*cso
)
52 struct si_context
*sctx
= (struct si_context
*)ctx
;
53 struct si_screen
*sscreen
= (struct si_screen
*)ctx
->screen
;
54 struct si_compute
*program
= CALLOC_STRUCT(si_compute
);
55 struct si_shader
*shader
= &program
->shader
;
58 program
->ir_type
= cso
->ir_type
;
59 program
->local_size
= cso
->req_local_mem
;
60 program
->private_size
= cso
->req_private_mem
;
61 program
->input_size
= cso
->req_input_mem
;
64 if (cso
->ir_type
== PIPE_SHADER_IR_TGSI
) {
65 struct si_shader_selector sel
;
68 memset(&sel
, 0, sizeof(sel
));
70 sel
.tokens
= tgsi_dup_tokens(cso
->prog
);
76 tgsi_scan_shader(cso
->prog
, &sel
.info
);
77 sel
.type
= PIPE_SHADER_COMPUTE
;
78 sel
.local_size
= cso
->req_local_mem
;
80 p_atomic_inc(&sscreen
->b
.num_shaders_created
);
82 program
->shader
.selector
= &sel
;
84 if (si_shader_create(sscreen
, sctx
->tm
, &program
->shader
,
91 scratch_enabled
= shader
->config
.scratch_bytes_per_wave
> 0;
93 shader
->config
.rsrc1
=
94 S_00B848_VGPRS((shader
->config
.num_vgprs
- 1) / 4) |
95 S_00B848_SGPRS((shader
->config
.num_sgprs
- 1) / 8) |
96 S_00B848_DX10_CLAMP(1) |
97 S_00B848_FLOAT_MODE(shader
->config
.float_mode
);
99 shader
->config
.rsrc2
= S_00B84C_USER_SGPR(SI_CS_NUM_USER_SGPR
) |
100 S_00B84C_SCRATCH_EN(scratch_enabled
) |
101 S_00B84C_TGID_X_EN(1) | S_00B84C_TGID_Y_EN(1) |
102 S_00B84C_TGID_Z_EN(1) | S_00B84C_TIDIG_COMP_CNT(2) |
103 S_00B84C_LDS_SIZE(shader
->config
.lds_size
);
107 const struct pipe_llvm_program_header
*header
;
110 code
= cso
->prog
+ sizeof(struct pipe_llvm_program_header
);
112 radeon_elf_read(code
, header
->num_bytes
, &program
->shader
.binary
);
113 si_shader_binary_read_config(&program
->shader
.binary
,
114 &program
->shader
.config
, 0);
115 si_shader_dump(sctx
->screen
, &program
->shader
, &sctx
->b
.debug
,
116 PIPE_SHADER_COMPUTE
, stderr
);
117 si_shader_binary_upload(sctx
->screen
, &program
->shader
);
123 static void si_bind_compute_state(struct pipe_context
*ctx
, void *state
)
125 struct si_context
*sctx
= (struct si_context
*)ctx
;
126 sctx
->cs_shader_state
.program
= (struct si_compute
*)state
;
129 static void si_set_global_binding(
130 struct pipe_context
*ctx
, unsigned first
, unsigned n
,
131 struct pipe_resource
**resources
,
135 struct si_context
*sctx
= (struct si_context
*)ctx
;
136 struct si_compute
*program
= sctx
->cs_shader_state
.program
;
139 for (i
= first
; i
< first
+ n
; i
++) {
140 pipe_resource_reference(&program
->global_buffers
[i
], NULL
);
145 for (i
= first
; i
< first
+ n
; i
++) {
148 pipe_resource_reference(&program
->global_buffers
[i
], resources
[i
]);
149 va
= r600_resource(resources
[i
])->gpu_address
;
150 offset
= util_le32_to_cpu(*handles
[i
]);
152 va
= util_cpu_to_le64(va
);
153 memcpy(handles
[i
], &va
, sizeof(va
));
157 static void si_initialize_compute(struct si_context
*sctx
)
159 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
161 radeon_set_sh_reg_seq(cs
, R_00B810_COMPUTE_START_X
, 3);
166 radeon_set_sh_reg_seq(cs
, R_00B854_COMPUTE_RESOURCE_LIMITS
, 3);
168 /* R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE0 / SE1 */
169 radeon_emit(cs
, S_00B858_SH0_CU_EN(0xffff) | S_00B858_SH1_CU_EN(0xffff));
170 radeon_emit(cs
, S_00B85C_SH0_CU_EN(0xffff) | S_00B85C_SH1_CU_EN(0xffff));
172 if (sctx
->b
.chip_class
>= CIK
) {
173 /* Also set R_00B858_COMPUTE_STATIC_THREAD_MGMT_SE2 / SE3 */
174 radeon_set_sh_reg_seq(cs
,
175 R_00B864_COMPUTE_STATIC_THREAD_MGMT_SE2
, 2);
176 radeon_emit(cs
, S_00B864_SH0_CU_EN(0xffff) |
177 S_00B864_SH1_CU_EN(0xffff));
178 radeon_emit(cs
, S_00B868_SH0_CU_EN(0xffff) |
179 S_00B868_SH1_CU_EN(0xffff));
182 /* This register has been moved to R_00CD20_COMPUTE_MAX_WAVE_ID
183 * and is now per pipe, so it should be handled in the
184 * kernel if we want to use something other than the default value,
185 * which is now 0x22f.
187 if (sctx
->b
.chip_class
<= SI
) {
188 /* XXX: This should be:
189 * (number of compute units) * 4 * (waves per simd) - 1 */
191 radeon_set_sh_reg(cs
, R_00B82C_COMPUTE_MAX_WAVE_ID
,
192 0x190 /* Default value */);
195 sctx
->cs_shader_state
.emitted_program
= NULL
;
196 sctx
->cs_shader_state
.initialized
= true;
199 static bool si_setup_compute_scratch_buffer(struct si_context
*sctx
,
200 struct si_shader
*shader
,
201 struct si_shader_config
*config
)
203 uint64_t scratch_bo_size
, scratch_needed
;
205 scratch_needed
= config
->scratch_bytes_per_wave
* sctx
->scratch_waves
;
206 if (sctx
->compute_scratch_buffer
)
207 scratch_bo_size
= sctx
->compute_scratch_buffer
->b
.b
.width0
;
209 if (scratch_bo_size
< scratch_needed
) {
210 r600_resource_reference(&sctx
->compute_scratch_buffer
, NULL
);
212 sctx
->compute_scratch_buffer
=
213 si_resource_create_custom(&sctx
->screen
->b
.b
,
214 PIPE_USAGE_DEFAULT
, scratch_needed
);
216 if (!sctx
->compute_scratch_buffer
)
220 if (sctx
->compute_scratch_buffer
!= shader
->scratch_bo
&& scratch_needed
) {
221 uint64_t scratch_va
= sctx
->compute_scratch_buffer
->gpu_address
;
223 si_shader_apply_scratch_relocs(sctx
, shader
, config
, scratch_va
);
225 if (si_shader_binary_upload(sctx
->screen
, shader
))
228 r600_resource_reference(&shader
->scratch_bo
,
229 sctx
->compute_scratch_buffer
);
235 static bool si_switch_compute_shader(struct si_context
*sctx
,
236 struct si_compute
*program
,
237 struct si_shader
*shader
, unsigned offset
)
239 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
240 struct si_shader_config inline_config
= {0};
241 struct si_shader_config
*config
;
244 if (sctx
->cs_shader_state
.emitted_program
== program
&&
245 sctx
->cs_shader_state
.offset
== offset
)
248 if (program
->ir_type
== PIPE_SHADER_IR_TGSI
) {
249 config
= &shader
->config
;
253 config
= &inline_config
;
254 si_shader_binary_read_config(&shader
->binary
, config
, offset
);
256 lds_blocks
= config
->lds_size
;
257 /* XXX: We are over allocating LDS. For SI, the shader reports
258 * LDS in blocks of 256 bytes, so if there are 4 bytes lds
259 * allocated in the shader and 4 bytes allocated by the state
260 * tracker, then we will set LDS_SIZE to 512 bytes rather than 256.
262 if (sctx
->b
.chip_class
<= SI
) {
263 lds_blocks
+= align(program
->local_size
, 256) >> 8;
265 lds_blocks
+= align(program
->local_size
, 512) >> 9;
268 assert(lds_blocks
<= 0xFF);
270 config
->rsrc2
&= C_00B84C_LDS_SIZE
;
271 config
->rsrc2
|= S_00B84C_LDS_SIZE(lds_blocks
);
274 if (!si_setup_compute_scratch_buffer(sctx
, shader
, config
))
277 if (shader
->scratch_bo
) {
278 COMPUTE_DBG(sctx
->screen
, "Waves: %u; Scratch per wave: %u bytes; "
279 "Total Scratch: %u bytes\n", sctx
->scratch_waves
,
280 config
->scratch_bytes_per_wave
,
281 config
->scratch_bytes_per_wave
*
282 sctx
->scratch_waves
);
284 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
285 shader
->scratch_bo
, RADEON_USAGE_READWRITE
,
286 RADEON_PRIO_SCRATCH_BUFFER
);
289 shader_va
= shader
->bo
->gpu_address
+ offset
;
291 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
, shader
->bo
,
292 RADEON_USAGE_READ
, RADEON_PRIO_USER_SHADER
);
294 radeon_set_sh_reg_seq(cs
, R_00B830_COMPUTE_PGM_LO
, 2);
295 radeon_emit(cs
, shader_va
>> 8);
296 radeon_emit(cs
, shader_va
>> 40);
298 radeon_set_sh_reg_seq(cs
, R_00B848_COMPUTE_PGM_RSRC1
, 2);
299 radeon_emit(cs
, config
->rsrc1
);
300 radeon_emit(cs
, config
->rsrc2
);
302 radeon_set_sh_reg(cs
, R_00B860_COMPUTE_TMPRING_SIZE
,
303 S_00B860_WAVES(sctx
->scratch_waves
)
304 | S_00B860_WAVESIZE(config
->scratch_bytes_per_wave
>> 10));
306 sctx
->cs_shader_state
.emitted_program
= program
;
307 sctx
->cs_shader_state
.offset
= offset
;
308 sctx
->cs_shader_state
.uses_scratch
=
309 config
->scratch_bytes_per_wave
!= 0;
314 static void si_upload_compute_input(struct si_context
*sctx
,
315 const struct pipe_grid_info
*info
)
317 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
318 struct si_compute
*program
= sctx
->cs_shader_state
.program
;
319 struct r600_resource
*input_buffer
= NULL
;
320 unsigned kernel_args_size
;
321 unsigned num_work_size_bytes
= 36;
322 uint32_t kernel_args_offset
= 0;
323 uint32_t *kernel_args
;
324 void *kernel_args_ptr
;
325 uint64_t kernel_args_va
;
328 /* The extra num_work_size_bytes are for work group / work item size information */
329 kernel_args_size
= program
->input_size
+ num_work_size_bytes
;
331 u_upload_alloc(sctx
->b
.uploader
, 0, kernel_args_size
, 256,
333 (struct pipe_resource
**)&input_buffer
, &kernel_args_ptr
);
335 kernel_args
= (uint32_t*)kernel_args_ptr
;
336 for (i
= 0; i
< 3; i
++) {
337 kernel_args
[i
] = info
->grid
[i
];
338 kernel_args
[i
+ 3] = info
->grid
[i
] * info
->block
[i
];
339 kernel_args
[i
+ 6] = info
->block
[i
];
342 memcpy(kernel_args
+ (num_work_size_bytes
/ 4), info
->input
,
343 program
->input_size
);
346 for (i
= 0; i
< (kernel_args_size
/ 4); i
++) {
347 COMPUTE_DBG(sctx
->screen
, "input %u : %u\n", i
,
351 kernel_args_va
= input_buffer
->gpu_address
+ kernel_args_offset
;
353 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
, input_buffer
,
354 RADEON_USAGE_READ
, RADEON_PRIO_CONST_BUFFER
);
356 radeon_set_sh_reg_seq(cs
, R_00B900_COMPUTE_USER_DATA_0
, 2);
357 radeon_emit(cs
, kernel_args_va
);
358 radeon_emit(cs
, S_008F04_BASE_ADDRESS_HI (kernel_args_va
>> 32) |
361 r600_resource_reference(&input_buffer
, NULL
);
364 static void si_setup_tgsi_grid(struct si_context
*sctx
,
365 const struct pipe_grid_info
*info
)
367 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
368 unsigned grid_size_reg
= R_00B900_COMPUTE_USER_DATA_0
+
369 4 * SI_SGPR_GRID_SIZE
;
371 if (info
->indirect
) {
372 uint64_t base_va
= r600_resource(info
->indirect
)->gpu_address
;
373 uint64_t va
= base_va
+ info
->indirect_offset
;
376 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
377 (struct r600_resource
*)info
->indirect
,
378 RADEON_USAGE_READ
, RADEON_PRIO_DRAW_INDIRECT
);
380 for (i
= 0; i
< 3; ++i
) {
381 radeon_emit(cs
, PKT3(PKT3_COPY_DATA
, 4, 0));
382 radeon_emit(cs
, COPY_DATA_SRC_SEL(COPY_DATA_MEM
) |
383 COPY_DATA_DST_SEL(COPY_DATA_REG
));
384 radeon_emit(cs
, (va
+ 4 * i
));
385 radeon_emit(cs
, (va
+ 4 * i
) >> 32);
386 radeon_emit(cs
, (grid_size_reg
>> 2) + i
);
391 radeon_set_sh_reg_seq(cs
, grid_size_reg
, 3);
392 radeon_emit(cs
, info
->grid
[0]);
393 radeon_emit(cs
, info
->grid
[1]);
394 radeon_emit(cs
, info
->grid
[2]);
398 static void si_emit_dispatch_packets(struct si_context
*sctx
,
399 const struct pipe_grid_info
*info
)
401 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
402 bool render_cond_bit
= sctx
->b
.render_cond
&& !sctx
->b
.render_cond_force_off
;
404 radeon_set_sh_reg_seq(cs
, R_00B81C_COMPUTE_NUM_THREAD_X
, 3);
405 radeon_emit(cs
, S_00B81C_NUM_THREAD_FULL(info
->block
[0]));
406 radeon_emit(cs
, S_00B820_NUM_THREAD_FULL(info
->block
[1]));
407 radeon_emit(cs
, S_00B824_NUM_THREAD_FULL(info
->block
[2]));
409 if (info
->indirect
) {
410 uint64_t base_va
= r600_resource(info
->indirect
)->gpu_address
;
412 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
413 (struct r600_resource
*)info
->indirect
,
414 RADEON_USAGE_READ
, RADEON_PRIO_DRAW_INDIRECT
);
416 radeon_emit(cs
, PKT3(PKT3_SET_BASE
, 2, 0) |
417 PKT3_SHADER_TYPE_S(1));
419 radeon_emit(cs
, base_va
);
420 radeon_emit(cs
, base_va
>> 32);
422 radeon_emit(cs
, PKT3(PKT3_DISPATCH_INDIRECT
, 1, render_cond_bit
) |
423 PKT3_SHADER_TYPE_S(1));
424 radeon_emit(cs
, info
->indirect_offset
);
427 radeon_emit(cs
, PKT3(PKT3_DISPATCH_DIRECT
, 3, render_cond_bit
) |
428 PKT3_SHADER_TYPE_S(1));
429 radeon_emit(cs
, info
->grid
[0]);
430 radeon_emit(cs
, info
->grid
[1]);
431 radeon_emit(cs
, info
->grid
[2]);
437 static void si_launch_grid(
438 struct pipe_context
*ctx
, const struct pipe_grid_info
*info
)
440 struct si_context
*sctx
= (struct si_context
*)ctx
;
441 struct si_compute
*program
= sctx
->cs_shader_state
.program
;
443 /* HW bug workaround when CS threadgroups > 256 threads and async
444 * compute isn't used, i.e. only one compute job can run at a time.
445 * If async compute is possible, the threadgroup size must be limited
446 * to 256 threads on all queues to avoid the bug.
447 * Only SI and certain CIK chips are affected.
449 bool cs_regalloc_hang
=
450 (sctx
->b
.chip_class
== SI
||
451 sctx
->b
.family
== CHIP_BONAIRE
||
452 sctx
->b
.family
== CHIP_KABINI
) &&
453 info
->block
[0] * info
->block
[1] * info
->block
[2] > 256;
455 if (cs_regalloc_hang
)
456 sctx
->b
.flags
|= SI_CONTEXT_PS_PARTIAL_FLUSH
|
457 SI_CONTEXT_CS_PARTIAL_FLUSH
;
459 si_decompress_compute_textures(sctx
);
461 si_need_cs_space(sctx
);
463 if (!sctx
->cs_shader_state
.initialized
)
464 si_initialize_compute(sctx
);
467 si_emit_cache_flush(sctx
, NULL
);
469 if (!si_switch_compute_shader(sctx
, program
, &program
->shader
, info
->pc
))
472 si_upload_compute_shader_descriptors(sctx
);
473 si_emit_compute_shader_userdata(sctx
);
475 if (si_is_atom_dirty(sctx
, sctx
->atoms
.s
.render_cond
)) {
476 sctx
->atoms
.s
.render_cond
->emit(&sctx
->b
,
477 sctx
->atoms
.s
.render_cond
);
478 si_set_atom_dirty(sctx
, sctx
->atoms
.s
.render_cond
, false);
481 if (program
->input_size
|| program
->ir_type
== PIPE_SHADER_IR_NATIVE
)
482 si_upload_compute_input(sctx
, info
);
485 for (i
= 0; i
< MAX_GLOBAL_BUFFERS
; i
++) {
486 struct r600_resource
*buffer
=
487 (struct r600_resource
*)program
->global_buffers
[i
];
491 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
, buffer
,
492 RADEON_USAGE_READWRITE
,
493 RADEON_PRIO_COMPUTE_GLOBAL
);
496 if (program
->ir_type
== PIPE_SHADER_IR_TGSI
)
497 si_setup_tgsi_grid(sctx
, info
);
499 si_ce_pre_draw_synchronization(sctx
);
501 si_emit_dispatch_packets(sctx
, info
);
503 si_ce_post_draw_synchronization(sctx
);
505 sctx
->b
.num_compute_calls
++;
506 if (sctx
->cs_shader_state
.uses_scratch
)
507 sctx
->b
.num_spill_compute_calls
++;
509 if (cs_regalloc_hang
)
510 sctx
->b
.flags
|= SI_CONTEXT_CS_PARTIAL_FLUSH
;
514 static void si_delete_compute_state(struct pipe_context
*ctx
, void* state
){
515 struct si_compute
*program
= (struct si_compute
*)state
;
516 struct si_context
*sctx
= (struct si_context
*)ctx
;
522 if (program
== sctx
->cs_shader_state
.program
)
523 sctx
->cs_shader_state
.program
= NULL
;
525 if (program
== sctx
->cs_shader_state
.emitted_program
)
526 sctx
->cs_shader_state
.emitted_program
= NULL
;
528 si_shader_destroy(&program
->shader
);
532 static void si_set_compute_resources(struct pipe_context
* ctx_
,
533 unsigned start
, unsigned count
,
534 struct pipe_surface
** surfaces
) { }
536 void si_init_compute_functions(struct si_context
*sctx
)
538 sctx
->b
.b
.create_compute_state
= si_create_compute_state
;
539 sctx
->b
.b
.delete_compute_state
= si_delete_compute_state
;
540 sctx
->b
.b
.bind_compute_state
= si_bind_compute_state
;
541 /* ctx->context.create_sampler_view = evergreen_compute_create_sampler_view; */
542 sctx
->b
.b
.set_compute_resources
= si_set_compute_resources
;
543 sctx
->b
.b
.set_global_binding
= si_set_global_binding
;
544 sctx
->b
.b
.launch_grid
= si_launch_grid
;