2 * Copyright 2015 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.
26 #include "si_compute.h"
28 #include "sid_tables.h"
29 #include "driver_ddebug/dd_util.h"
30 #include "util/u_dump.h"
31 #include "util/u_log.h"
32 #include "util/u_memory.h"
33 #include "util/u_string.h"
37 static void si_dump_bo_list(struct si_context
*sctx
,
38 const struct radeon_saved_cs
*saved
, FILE *f
);
40 DEBUG_GET_ONCE_OPTION(replace_shaders
, "RADEON_REPLACE_SHADERS", NULL
)
43 * Store a linearized copy of all chunks of \p cs together with the buffer
46 void si_save_cs(struct radeon_winsys
*ws
, struct radeon_cmdbuf
*cs
,
47 struct radeon_saved_cs
*saved
, bool get_buffer_list
)
52 /* Save the IB chunks. */
53 saved
->num_dw
= cs
->prev_dw
+ cs
->current
.cdw
;
54 saved
->ib
= MALLOC(4 * saved
->num_dw
);
59 for (i
= 0; i
< cs
->num_prev
; ++i
) {
60 memcpy(buf
, cs
->prev
[i
].buf
, cs
->prev
[i
].cdw
* 4);
61 buf
+= cs
->prev
[i
].cdw
;
63 memcpy(buf
, cs
->current
.buf
, cs
->current
.cdw
* 4);
68 /* Save the buffer list. */
69 saved
->bo_count
= ws
->cs_get_buffer_list(cs
, NULL
);
70 saved
->bo_list
= CALLOC(saved
->bo_count
,
71 sizeof(saved
->bo_list
[0]));
72 if (!saved
->bo_list
) {
76 ws
->cs_get_buffer_list(cs
, saved
->bo_list
);
81 fprintf(stderr
, "%s: out of memory\n", __func__
);
82 memset(saved
, 0, sizeof(*saved
));
85 void si_clear_saved_cs(struct radeon_saved_cs
*saved
)
90 memset(saved
, 0, sizeof(*saved
));
93 void si_destroy_saved_cs(struct si_saved_cs
*scs
)
95 si_clear_saved_cs(&scs
->gfx
);
96 si_resource_reference(&scs
->trace_buf
, NULL
);
100 static void si_dump_shader(struct si_screen
*sscreen
,
101 enum pipe_shader_type processor
,
102 struct si_shader
*shader
, FILE *f
)
104 if (shader
->shader_log
)
105 fwrite(shader
->shader_log
, shader
->shader_log_size
, 1, f
);
107 si_shader_dump(sscreen
, shader
, NULL
, processor
, f
, false);
109 if (shader
->bo
&& sscreen
->options
.dump_shader_binary
) {
110 unsigned size
= shader
->bo
->b
.b
.width0
;
111 fprintf(f
, "BO: VA=%"PRIx64
" Size=%u\n", shader
->bo
->gpu_address
, size
);
113 const char *mapped
= sscreen
->ws
->buffer_map(shader
->bo
->buf
, NULL
,
114 PIPE_TRANSFER_UNSYNCHRONIZED
|
116 RADEON_TRANSFER_TEMPORARY
);
118 for (unsigned i
= 0; i
< size
; i
+= 4) {
119 fprintf(f
, " %4x: %08x\n", i
, *(uint32_t*)(mapped
+ i
));
122 sscreen
->ws
->buffer_unmap(shader
->bo
->buf
);
128 struct si_log_chunk_shader
{
129 /* The shader destroy code assumes a current context for unlinking of
132 * While we should be able to destroy shaders without a context, doing
133 * so would happen only very rarely and be therefore likely to fail
134 * just when you're trying to debug something. Let's just remember the
135 * current context in the chunk.
137 struct si_context
*ctx
;
138 struct si_shader
*shader
;
139 enum pipe_shader_type processor
;
141 /* For keep-alive reference counts */
142 struct si_shader_selector
*sel
;
143 struct si_compute
*program
;
147 si_log_chunk_shader_destroy(void *data
)
149 struct si_log_chunk_shader
*chunk
= data
;
150 si_shader_selector_reference(chunk
->ctx
, &chunk
->sel
, NULL
);
151 si_compute_reference(&chunk
->program
, NULL
);
156 si_log_chunk_shader_print(void *data
, FILE *f
)
158 struct si_log_chunk_shader
*chunk
= data
;
159 struct si_screen
*sscreen
= chunk
->ctx
->screen
;
160 si_dump_shader(sscreen
, chunk
->processor
,
164 static struct u_log_chunk_type si_log_chunk_type_shader
= {
165 .destroy
= si_log_chunk_shader_destroy
,
166 .print
= si_log_chunk_shader_print
,
169 static void si_dump_gfx_shader(struct si_context
*ctx
,
170 const struct si_shader_ctx_state
*state
,
171 struct u_log_context
*log
)
173 struct si_shader
*current
= state
->current
;
175 if (!state
->cso
|| !current
)
178 struct si_log_chunk_shader
*chunk
= CALLOC_STRUCT(si_log_chunk_shader
);
180 chunk
->processor
= state
->cso
->info
.processor
;
181 chunk
->shader
= current
;
182 si_shader_selector_reference(ctx
, &chunk
->sel
, current
->selector
);
183 u_log_chunk(log
, &si_log_chunk_type_shader
, chunk
);
186 static void si_dump_compute_shader(struct si_context
*ctx
,
187 struct u_log_context
*log
)
189 const struct si_cs_shader_state
*state
= &ctx
->cs_shader_state
;
194 struct si_log_chunk_shader
*chunk
= CALLOC_STRUCT(si_log_chunk_shader
);
196 chunk
->processor
= PIPE_SHADER_COMPUTE
;
197 chunk
->shader
= &state
->program
->shader
;
198 si_compute_reference(&chunk
->program
, state
->program
);
199 u_log_chunk(log
, &si_log_chunk_type_shader
, chunk
);
203 * Shader compiles can be overridden with arbitrary ELF objects by setting
204 * the environment variable RADEON_REPLACE_SHADERS=num1:filename1[;num2:filename2]
206 * TODO: key this off some hash
208 bool si_replace_shader(unsigned num
, struct si_shader_binary
*binary
)
210 const char *p
= debug_get_option_replace_shaders();
211 const char *semicolon
;
214 long filesize
, nread
;
215 bool replaced
= false;
223 i
= strtoul(p
, &endp
, 0);
227 fprintf(stderr
, "RADEON_REPLACE_SHADERS formatted badly.\n");
243 semicolon
= strchr(p
, ';');
245 p
= copy
= strndup(p
, semicolon
- p
);
247 fprintf(stderr
, "out of memory\n");
252 fprintf(stderr
, "radeonsi: replace shader %u by %s\n", num
, p
);
256 perror("radeonsi: failed to open file");
260 if (fseek(f
, 0, SEEK_END
) != 0)
267 if (fseek(f
, 0, SEEK_SET
) != 0)
270 binary
->elf_buffer
= MALLOC(filesize
);
271 if (!binary
->elf_buffer
) {
272 fprintf(stderr
, "out of memory\n");
276 nread
= fread((void*)binary
->elf_buffer
, 1, filesize
, f
);
277 if (nread
!= filesize
) {
278 FREE((void*)binary
->elf_buffer
);
279 binary
->elf_buffer
= NULL
;
283 binary
->elf_size
= nread
;
293 perror("radeonsi: reading shader");
297 /* Parsed IBs are difficult to read without colors. Use "less -R file" to
298 * read them, or use "aha -b -f file" to convert them to html.
300 #define COLOR_RESET "\033[0m"
301 #define COLOR_RED "\033[31m"
302 #define COLOR_GREEN "\033[1;32m"
303 #define COLOR_YELLOW "\033[1;33m"
304 #define COLOR_CYAN "\033[1;36m"
306 static void si_dump_mmapped_reg(struct si_context
*sctx
, FILE *f
,
309 struct radeon_winsys
*ws
= sctx
->ws
;
312 if (ws
->read_registers(ws
, offset
, 1, &value
))
313 ac_dump_reg(f
, sctx
->chip_class
, offset
, value
, ~0);
316 static void si_dump_debug_registers(struct si_context
*sctx
, FILE *f
)
318 if (!sctx
->screen
->info
.has_read_registers_query
)
321 fprintf(f
, "Memory-mapped registers:\n");
322 si_dump_mmapped_reg(sctx
, f
, R_008010_GRBM_STATUS
);
324 /* No other registers can be read on DRM < 3.1.0. */
325 if (!sctx
->screen
->info
.is_amdgpu
||
326 sctx
->screen
->info
.drm_minor
< 1) {
331 si_dump_mmapped_reg(sctx
, f
, R_008008_GRBM_STATUS2
);
332 si_dump_mmapped_reg(sctx
, f
, R_008014_GRBM_STATUS_SE0
);
333 si_dump_mmapped_reg(sctx
, f
, R_008018_GRBM_STATUS_SE1
);
334 si_dump_mmapped_reg(sctx
, f
, R_008038_GRBM_STATUS_SE2
);
335 si_dump_mmapped_reg(sctx
, f
, R_00803C_GRBM_STATUS_SE3
);
336 si_dump_mmapped_reg(sctx
, f
, R_00D034_SDMA0_STATUS_REG
);
337 si_dump_mmapped_reg(sctx
, f
, R_00D834_SDMA1_STATUS_REG
);
338 if (sctx
->chip_class
<= GFX8
) {
339 si_dump_mmapped_reg(sctx
, f
, R_000E50_SRBM_STATUS
);
340 si_dump_mmapped_reg(sctx
, f
, R_000E4C_SRBM_STATUS2
);
341 si_dump_mmapped_reg(sctx
, f
, R_000E54_SRBM_STATUS3
);
343 si_dump_mmapped_reg(sctx
, f
, R_008680_CP_STAT
);
344 si_dump_mmapped_reg(sctx
, f
, R_008674_CP_STALLED_STAT1
);
345 si_dump_mmapped_reg(sctx
, f
, R_008678_CP_STALLED_STAT2
);
346 si_dump_mmapped_reg(sctx
, f
, R_008670_CP_STALLED_STAT3
);
347 si_dump_mmapped_reg(sctx
, f
, R_008210_CP_CPC_STATUS
);
348 si_dump_mmapped_reg(sctx
, f
, R_008214_CP_CPC_BUSY_STAT
);
349 si_dump_mmapped_reg(sctx
, f
, R_008218_CP_CPC_STALLED_STAT1
);
350 si_dump_mmapped_reg(sctx
, f
, R_00821C_CP_CPF_STATUS
);
351 si_dump_mmapped_reg(sctx
, f
, R_008220_CP_CPF_BUSY_STAT
);
352 si_dump_mmapped_reg(sctx
, f
, R_008224_CP_CPF_STALLED_STAT1
);
356 struct si_log_chunk_cs
{
357 struct si_context
*ctx
;
358 struct si_saved_cs
*cs
;
360 unsigned gfx_begin
, gfx_end
;
361 unsigned compute_begin
, compute_end
;
364 static void si_log_chunk_type_cs_destroy(void *data
)
366 struct si_log_chunk_cs
*chunk
= data
;
367 si_saved_cs_reference(&chunk
->cs
, NULL
);
371 static void si_parse_current_ib(FILE *f
, struct radeon_cmdbuf
*cs
,
372 unsigned begin
, unsigned end
,
373 int *last_trace_id
, unsigned trace_id_count
,
374 const char *name
, enum chip_class chip_class
)
376 unsigned orig_end
= end
;
378 assert(begin
<= end
);
380 fprintf(f
, "------------------ %s begin (dw = %u) ------------------\n",
383 for (unsigned prev_idx
= 0; prev_idx
< cs
->num_prev
; ++prev_idx
) {
384 struct radeon_cmdbuf_chunk
*chunk
= &cs
->prev
[prev_idx
];
386 if (begin
< chunk
->cdw
) {
387 ac_parse_ib_chunk(f
, chunk
->buf
+ begin
,
388 MIN2(end
, chunk
->cdw
) - begin
,
389 last_trace_id
, trace_id_count
,
390 chip_class
, NULL
, NULL
);
393 if (end
<= chunk
->cdw
)
396 if (begin
< chunk
->cdw
)
397 fprintf(f
, "\n---------- Next %s Chunk ----------\n\n",
400 begin
-= MIN2(begin
, chunk
->cdw
);
404 assert(end
<= cs
->current
.cdw
);
406 ac_parse_ib_chunk(f
, cs
->current
.buf
+ begin
, end
- begin
, last_trace_id
,
407 trace_id_count
, chip_class
, NULL
, NULL
);
409 fprintf(f
, "------------------- %s end (dw = %u) -------------------\n\n",
413 static void si_log_chunk_type_cs_print(void *data
, FILE *f
)
415 struct si_log_chunk_cs
*chunk
= data
;
416 struct si_context
*ctx
= chunk
->ctx
;
417 struct si_saved_cs
*scs
= chunk
->cs
;
418 int last_trace_id
= -1;
419 int last_compute_trace_id
= -1;
421 /* We are expecting that the ddebug pipe has already
422 * waited for the context, so this buffer should be idle.
423 * If the GPU is hung, there is no point in waiting for it.
425 uint32_t *map
= ctx
->ws
->buffer_map(scs
->trace_buf
->buf
,
427 PIPE_TRANSFER_UNSYNCHRONIZED
|
430 last_trace_id
= map
[0];
431 last_compute_trace_id
= map
[1];
434 if (chunk
->gfx_end
!= chunk
->gfx_begin
) {
435 if (chunk
->gfx_begin
== 0) {
436 if (ctx
->init_config
)
437 ac_parse_ib(f
, ctx
->init_config
->pm4
, ctx
->init_config
->ndw
,
438 NULL
, 0, "IB2: Init config", ctx
->chip_class
,
441 if (ctx
->init_config_gs_rings
)
442 ac_parse_ib(f
, ctx
->init_config_gs_rings
->pm4
,
443 ctx
->init_config_gs_rings
->ndw
,
444 NULL
, 0, "IB2: Init GS rings", ctx
->chip_class
,
449 ac_parse_ib(f
, scs
->gfx
.ib
+ chunk
->gfx_begin
,
450 chunk
->gfx_end
- chunk
->gfx_begin
,
451 &last_trace_id
, map
? 1 : 0, "IB", ctx
->chip_class
,
454 si_parse_current_ib(f
, ctx
->gfx_cs
, chunk
->gfx_begin
,
455 chunk
->gfx_end
, &last_trace_id
, map
? 1 : 0,
456 "IB", ctx
->chip_class
);
460 if (chunk
->compute_end
!= chunk
->compute_begin
) {
461 assert(ctx
->prim_discard_compute_cs
);
464 ac_parse_ib(f
, scs
->compute
.ib
+ chunk
->compute_begin
,
465 chunk
->compute_end
- chunk
->compute_begin
,
466 &last_compute_trace_id
, map
? 1 : 0, "Compute IB", ctx
->chip_class
,
469 si_parse_current_ib(f
, ctx
->prim_discard_compute_cs
, chunk
->compute_begin
,
470 chunk
->compute_end
, &last_compute_trace_id
,
471 map
? 1 : 0, "Compute IB", ctx
->chip_class
);
475 if (chunk
->dump_bo_list
) {
476 fprintf(f
, "Flushing. Time: ");
477 util_dump_ns(f
, scs
->time_flush
);
479 si_dump_bo_list(ctx
, &scs
->gfx
, f
);
483 static const struct u_log_chunk_type si_log_chunk_type_cs
= {
484 .destroy
= si_log_chunk_type_cs_destroy
,
485 .print
= si_log_chunk_type_cs_print
,
488 static void si_log_cs(struct si_context
*ctx
, struct u_log_context
*log
,
491 assert(ctx
->current_saved_cs
);
493 struct si_saved_cs
*scs
= ctx
->current_saved_cs
;
494 unsigned gfx_cur
= ctx
->gfx_cs
->prev_dw
+ ctx
->gfx_cs
->current
.cdw
;
495 unsigned compute_cur
= 0;
497 if (ctx
->prim_discard_compute_cs
)
498 compute_cur
= ctx
->prim_discard_compute_cs
->prev_dw
+ ctx
->prim_discard_compute_cs
->current
.cdw
;
501 gfx_cur
== scs
->gfx_last_dw
&&
502 compute_cur
== scs
->compute_last_dw
)
505 struct si_log_chunk_cs
*chunk
= calloc(1, sizeof(*chunk
));
508 si_saved_cs_reference(&chunk
->cs
, scs
);
509 chunk
->dump_bo_list
= dump_bo_list
;
511 chunk
->gfx_begin
= scs
->gfx_last_dw
;
512 chunk
->gfx_end
= gfx_cur
;
513 scs
->gfx_last_dw
= gfx_cur
;
515 chunk
->compute_begin
= scs
->compute_last_dw
;
516 chunk
->compute_end
= compute_cur
;
517 scs
->compute_last_dw
= compute_cur
;
519 u_log_chunk(log
, &si_log_chunk_type_cs
, chunk
);
522 void si_auto_log_cs(void *data
, struct u_log_context
*log
)
524 struct si_context
*ctx
= (struct si_context
*)data
;
525 si_log_cs(ctx
, log
, false);
528 void si_log_hw_flush(struct si_context
*sctx
)
533 si_log_cs(sctx
, sctx
->log
, true);
535 if (&sctx
->b
== sctx
->screen
->aux_context
) {
536 /* The aux context isn't captured by the ddebug wrapper,
537 * so we dump it on a flush-by-flush basis here.
539 FILE *f
= dd_get_debug_file(false);
541 fprintf(stderr
, "radeonsi: error opening aux context dump file.\n");
543 dd_write_header(f
, &sctx
->screen
->b
, 0);
545 fprintf(f
, "Aux context dump:\n\n");
546 u_log_new_page_print(sctx
->log
, f
);
553 static const char *priority_to_string(enum radeon_bo_priority priority
)
555 #define ITEM(x) [RADEON_PRIO_##x] = #x
556 static const char *table
[64] = {
559 ITEM(SO_FILLED_SIZE
),
569 ITEM(SAMPLER_BUFFER
),
571 ITEM(SHADER_RW_BUFFER
),
572 ITEM(COMPUTE_GLOBAL
),
573 ITEM(SAMPLER_TEXTURE
),
574 ITEM(SHADER_RW_IMAGE
),
575 ITEM(SAMPLER_TEXTURE_MSAA
),
578 ITEM(COLOR_BUFFER_MSAA
),
579 ITEM(DEPTH_BUFFER_MSAA
),
583 ITEM(SCRATCH_BUFFER
),
587 assert(priority
< ARRAY_SIZE(table
));
588 return table
[priority
];
591 static int bo_list_compare_va(const struct radeon_bo_list_item
*a
,
592 const struct radeon_bo_list_item
*b
)
594 return a
->vm_address
< b
->vm_address
? -1 :
595 a
->vm_address
> b
->vm_address
? 1 : 0;
598 static void si_dump_bo_list(struct si_context
*sctx
,
599 const struct radeon_saved_cs
*saved
, FILE *f
)
606 /* Sort the list according to VM adddresses first. */
607 qsort(saved
->bo_list
, saved
->bo_count
,
608 sizeof(saved
->bo_list
[0]), (void*)bo_list_compare_va
);
610 fprintf(f
, "Buffer list (in units of pages = 4kB):\n"
611 COLOR_YELLOW
" Size VM start page "
612 "VM end page Usage" COLOR_RESET
"\n");
614 for (i
= 0; i
< saved
->bo_count
; i
++) {
615 /* Note: Buffer sizes are expected to be aligned to 4k by the winsys. */
616 const unsigned page_size
= sctx
->screen
->info
.gart_page_size
;
617 uint64_t va
= saved
->bo_list
[i
].vm_address
;
618 uint64_t size
= saved
->bo_list
[i
].bo_size
;
621 /* If there's unused virtual memory between 2 buffers, print it. */
623 uint64_t previous_va_end
= saved
->bo_list
[i
-1].vm_address
+
624 saved
->bo_list
[i
-1].bo_size
;
626 if (va
> previous_va_end
) {
627 fprintf(f
, " %10"PRIu64
" -- hole --\n",
628 (va
- previous_va_end
) / page_size
);
632 /* Print the buffer. */
633 fprintf(f
, " %10"PRIu64
" 0x%013"PRIX64
" 0x%013"PRIX64
" ",
634 size
/ page_size
, va
/ page_size
, (va
+ size
) / page_size
);
636 /* Print the usage. */
637 for (j
= 0; j
< 32; j
++) {
638 if (!(saved
->bo_list
[i
].priority_usage
& (1u << j
)))
641 fprintf(f
, "%s%s", !hit
? "" : ", ", priority_to_string(j
));
646 fprintf(f
, "\nNote: The holes represent memory not used by the IB.\n"
647 " Other buffers can still be allocated there.\n\n");
650 static void si_dump_framebuffer(struct si_context
*sctx
, struct u_log_context
*log
)
652 struct pipe_framebuffer_state
*state
= &sctx
->framebuffer
.state
;
653 struct si_texture
*tex
;
656 for (i
= 0; i
< state
->nr_cbufs
; i
++) {
657 if (!state
->cbufs
[i
])
660 tex
= (struct si_texture
*)state
->cbufs
[i
]->texture
;
661 u_log_printf(log
, COLOR_YELLOW
"Color buffer %i:" COLOR_RESET
"\n", i
);
662 si_print_texture_info(sctx
->screen
, tex
, log
);
663 u_log_printf(log
, "\n");
667 tex
= (struct si_texture
*)state
->zsbuf
->texture
;
668 u_log_printf(log
, COLOR_YELLOW
"Depth-stencil buffer:" COLOR_RESET
"\n");
669 si_print_texture_info(sctx
->screen
, tex
, log
);
670 u_log_printf(log
, "\n");
674 typedef unsigned (*slot_remap_func
)(unsigned);
676 struct si_log_chunk_desc_list
{
677 /** Pointer to memory map of buffer where the list is uploader */
679 /** Reference of buffer where the list is uploaded, so that gpu_list
681 struct si_resource
*buf
;
683 const char *shader_name
;
684 const char *elem_name
;
685 slot_remap_func slot_remap
;
686 enum chip_class chip_class
;
687 unsigned element_dw_size
;
688 unsigned num_elements
;
694 si_log_chunk_desc_list_destroy(void *data
)
696 struct si_log_chunk_desc_list
*chunk
= data
;
697 si_resource_reference(&chunk
->buf
, NULL
);
702 si_log_chunk_desc_list_print(void *data
, FILE *f
)
704 struct si_log_chunk_desc_list
*chunk
= data
;
705 unsigned sq_img_rsrc_word0
= chunk
->chip_class
>= GFX10
? R_00A000_SQ_IMG_RSRC_WORD0
706 : R_008F10_SQ_IMG_RSRC_WORD0
;
708 for (unsigned i
= 0; i
< chunk
->num_elements
; i
++) {
709 unsigned cpu_dw_offset
= i
* chunk
->element_dw_size
;
710 unsigned gpu_dw_offset
= chunk
->slot_remap(i
) * chunk
->element_dw_size
;
711 const char *list_note
= chunk
->gpu_list
? "GPU list" : "CPU list";
712 uint32_t *cpu_list
= chunk
->list
+ cpu_dw_offset
;
713 uint32_t *gpu_list
= chunk
->gpu_list
? chunk
->gpu_list
+ gpu_dw_offset
: cpu_list
;
715 fprintf(f
, COLOR_GREEN
"%s%s slot %u (%s):" COLOR_RESET
"\n",
716 chunk
->shader_name
, chunk
->elem_name
, i
, list_note
);
718 switch (chunk
->element_dw_size
) {
720 for (unsigned j
= 0; j
< 4; j
++)
721 ac_dump_reg(f
, chunk
->chip_class
,
722 R_008F00_SQ_BUF_RSRC_WORD0
+ j
*4,
723 gpu_list
[j
], 0xffffffff);
726 for (unsigned j
= 0; j
< 8; j
++)
727 ac_dump_reg(f
, chunk
->chip_class
,
728 sq_img_rsrc_word0
+ j
*4,
729 gpu_list
[j
], 0xffffffff);
731 fprintf(f
, COLOR_CYAN
" Buffer:" COLOR_RESET
"\n");
732 for (unsigned j
= 0; j
< 4; j
++)
733 ac_dump_reg(f
, chunk
->chip_class
,
734 R_008F00_SQ_BUF_RSRC_WORD0
+ j
*4,
735 gpu_list
[4+j
], 0xffffffff);
738 for (unsigned j
= 0; j
< 8; j
++)
739 ac_dump_reg(f
, chunk
->chip_class
,
740 sq_img_rsrc_word0
+ j
*4,
741 gpu_list
[j
], 0xffffffff);
743 fprintf(f
, COLOR_CYAN
" Buffer:" COLOR_RESET
"\n");
744 for (unsigned j
= 0; j
< 4; j
++)
745 ac_dump_reg(f
, chunk
->chip_class
,
746 R_008F00_SQ_BUF_RSRC_WORD0
+ j
*4,
747 gpu_list
[4+j
], 0xffffffff);
749 fprintf(f
, COLOR_CYAN
" FMASK:" COLOR_RESET
"\n");
750 for (unsigned j
= 0; j
< 8; j
++)
751 ac_dump_reg(f
, chunk
->chip_class
,
752 sq_img_rsrc_word0
+ j
*4,
753 gpu_list
[8+j
], 0xffffffff);
755 fprintf(f
, COLOR_CYAN
" Sampler state:" COLOR_RESET
"\n");
756 for (unsigned j
= 0; j
< 4; j
++)
757 ac_dump_reg(f
, chunk
->chip_class
,
758 R_008F30_SQ_IMG_SAMP_WORD0
+ j
*4,
759 gpu_list
[12+j
], 0xffffffff);
763 if (memcmp(gpu_list
, cpu_list
, chunk
->element_dw_size
* 4) != 0) {
764 fprintf(f
, COLOR_RED
"!!!!! This slot was corrupted in GPU memory !!!!!"
773 static const struct u_log_chunk_type si_log_chunk_type_descriptor_list
= {
774 .destroy
= si_log_chunk_desc_list_destroy
,
775 .print
= si_log_chunk_desc_list_print
,
778 static void si_dump_descriptor_list(struct si_screen
*screen
,
779 struct si_descriptors
*desc
,
780 const char *shader_name
,
781 const char *elem_name
,
782 unsigned element_dw_size
,
783 unsigned num_elements
,
784 slot_remap_func slot_remap
,
785 struct u_log_context
*log
)
790 /* In some cases, the caller doesn't know how many elements are really
791 * uploaded. Reduce num_elements to fit in the range of active slots. */
792 unsigned active_range_dw_begin
=
793 desc
->first_active_slot
* desc
->element_dw_size
;
794 unsigned active_range_dw_end
=
795 active_range_dw_begin
+ desc
->num_active_slots
* desc
->element_dw_size
;
797 while (num_elements
> 0) {
798 int i
= slot_remap(num_elements
- 1);
799 unsigned dw_begin
= i
* element_dw_size
;
800 unsigned dw_end
= dw_begin
+ element_dw_size
;
802 if (dw_begin
>= active_range_dw_begin
&& dw_end
<= active_range_dw_end
)
808 struct si_log_chunk_desc_list
*chunk
=
809 CALLOC_VARIANT_LENGTH_STRUCT(si_log_chunk_desc_list
,
810 4 * element_dw_size
* num_elements
);
811 chunk
->shader_name
= shader_name
;
812 chunk
->elem_name
= elem_name
;
813 chunk
->element_dw_size
= element_dw_size
;
814 chunk
->num_elements
= num_elements
;
815 chunk
->slot_remap
= slot_remap
;
816 chunk
->chip_class
= screen
->info
.chip_class
;
818 si_resource_reference(&chunk
->buf
, desc
->buffer
);
819 chunk
->gpu_list
= desc
->gpu_list
;
821 for (unsigned i
= 0; i
< num_elements
; ++i
) {
822 memcpy(&chunk
->list
[i
* element_dw_size
],
823 &desc
->list
[slot_remap(i
) * element_dw_size
],
824 4 * element_dw_size
);
827 u_log_chunk(log
, &si_log_chunk_type_descriptor_list
, chunk
);
830 static unsigned si_identity(unsigned slot
)
835 static void si_dump_descriptors(struct si_context
*sctx
,
836 enum pipe_shader_type processor
,
837 const struct tgsi_shader_info
*info
,
838 struct u_log_context
*log
)
840 struct si_descriptors
*descs
=
841 &sctx
->descriptors
[SI_DESCS_FIRST_SHADER
+
842 processor
* SI_NUM_SHADER_DESCS
];
843 static const char *shader_name
[] = {"VS", "PS", "GS", "TCS", "TES", "CS"};
844 const char *name
= shader_name
[processor
];
845 unsigned enabled_constbuf
, enabled_shaderbuf
, enabled_samplers
;
846 unsigned enabled_images
;
849 enabled_constbuf
= info
->const_buffers_declared
;
850 enabled_shaderbuf
= info
->shader_buffers_declared
;
851 enabled_samplers
= info
->samplers_declared
;
852 enabled_images
= info
->images_declared
;
854 enabled_constbuf
= sctx
->const_and_shader_buffers
[processor
].enabled_mask
>>
855 SI_NUM_SHADER_BUFFERS
;
856 enabled_shaderbuf
= sctx
->const_and_shader_buffers
[processor
].enabled_mask
&
857 u_bit_consecutive(0, SI_NUM_SHADER_BUFFERS
);
858 enabled_shaderbuf
= util_bitreverse(enabled_shaderbuf
) >>
859 (32 - SI_NUM_SHADER_BUFFERS
);
860 enabled_samplers
= sctx
->samplers
[processor
].enabled_mask
;
861 enabled_images
= sctx
->images
[processor
].enabled_mask
;
864 if (processor
== PIPE_SHADER_VERTEX
&&
865 sctx
->vb_descriptors_buffer
&&
866 sctx
->vb_descriptors_gpu_list
&&
867 sctx
->vertex_elements
) {
868 assert(info
); /* only CS may not have an info struct */
869 struct si_descriptors desc
= {};
871 desc
.buffer
= sctx
->vb_descriptors_buffer
;
872 desc
.list
= sctx
->vb_descriptors_gpu_list
;
873 desc
.gpu_list
= sctx
->vb_descriptors_gpu_list
;
874 desc
.element_dw_size
= 4;
875 desc
.num_active_slots
= sctx
->vertex_elements
->desc_list_byte_size
/ 16;
877 si_dump_descriptor_list(sctx
->screen
, &desc
, name
,
878 " - Vertex buffer", 4, info
->num_inputs
,
882 si_dump_descriptor_list(sctx
->screen
,
883 &descs
[SI_SHADER_DESCS_CONST_AND_SHADER_BUFFERS
],
884 name
, " - Constant buffer", 4,
885 util_last_bit(enabled_constbuf
),
886 si_get_constbuf_slot
, log
);
887 si_dump_descriptor_list(sctx
->screen
,
888 &descs
[SI_SHADER_DESCS_CONST_AND_SHADER_BUFFERS
],
889 name
, " - Shader buffer", 4,
890 util_last_bit(enabled_shaderbuf
),
891 si_get_shaderbuf_slot
, log
);
892 si_dump_descriptor_list(sctx
->screen
,
893 &descs
[SI_SHADER_DESCS_SAMPLERS_AND_IMAGES
],
894 name
, " - Sampler", 16,
895 util_last_bit(enabled_samplers
),
896 si_get_sampler_slot
, log
);
897 si_dump_descriptor_list(sctx
->screen
,
898 &descs
[SI_SHADER_DESCS_SAMPLERS_AND_IMAGES
],
900 util_last_bit(enabled_images
),
901 si_get_image_slot
, log
);
904 static void si_dump_gfx_descriptors(struct si_context
*sctx
,
905 const struct si_shader_ctx_state
*state
,
906 struct u_log_context
*log
)
908 if (!state
->cso
|| !state
->current
)
911 si_dump_descriptors(sctx
, state
->cso
->type
, &state
->cso
->info
, log
);
914 static void si_dump_compute_descriptors(struct si_context
*sctx
,
915 struct u_log_context
*log
)
917 if (!sctx
->cs_shader_state
.program
)
920 si_dump_descriptors(sctx
, PIPE_SHADER_COMPUTE
, NULL
, log
);
923 struct si_shader_inst
{
924 const char *text
; /* start of disassembly for this instruction */
926 unsigned size
; /* instruction size = 4 or 8 */
927 uint64_t addr
; /* instruction address */
931 * Open the given \p binary as \p rtld_binary and split the contained
932 * disassembly string into instructions and add them to the array
933 * pointed to by \p instructions, which must be sufficiently large.
935 * Labels are considered to be part of the following instruction.
937 * The caller must keep \p rtld_binary alive as long as \p instructions are
938 * used and then close it afterwards.
940 static void si_add_split_disasm(struct si_screen
*screen
,
941 struct ac_rtld_binary
*rtld_binary
,
942 struct si_shader_binary
*binary
,
945 struct si_shader_inst
*instructions
)
947 if (!ac_rtld_open(rtld_binary
, (struct ac_rtld_open_info
){
948 .info
= &screen
->info
,
950 .elf_ptrs
= &binary
->elf_buffer
,
951 .elf_sizes
= &binary
->elf_size
}))
956 if (!ac_rtld_get_section_by_name(rtld_binary
, ".AMDGPU.disasm",
960 const char *end
= disasm
+ nbytes
;
961 while (disasm
< end
) {
962 const char *semicolon
= memchr(disasm
, ';', end
- disasm
);
966 struct si_shader_inst
*inst
= &instructions
[(*num
)++];
967 const char *inst_end
= memchr(semicolon
+ 1, '\n', end
- semicolon
- 1);
972 inst
->textlen
= inst_end
- disasm
;
975 /* More than 16 chars after ";" means the instruction is 8 bytes long. */
976 inst
->size
= inst_end
- semicolon
> 16 ? 8 : 4;
981 disasm
= inst_end
+ 1;
985 /* If the shader is being executed, print its asm instructions, and annotate
986 * those that are being executed right now with information about waves that
987 * execute them. This is most useful during a GPU hang.
989 static void si_print_annotated_shader(struct si_shader
*shader
,
990 struct ac_wave_info
*waves
,
997 struct si_screen
*screen
= shader
->selector
->screen
;
998 uint64_t start_addr
= shader
->bo
->gpu_address
;
999 uint64_t end_addr
= start_addr
+ shader
->bo
->b
.b
.width0
;
1002 /* See if any wave executes the shader. */
1003 for (i
= 0; i
< num_waves
; i
++) {
1004 if (start_addr
<= waves
[i
].pc
&& waves
[i
].pc
<= end_addr
)
1008 return; /* the shader is not being executed */
1010 /* Remember the first found wave. The waves are sorted according to PC. */
1014 /* Get the list of instructions.
1015 * Buffer size / 4 is the upper bound of the instruction count.
1017 unsigned num_inst
= 0;
1018 uint64_t inst_addr
= start_addr
;
1019 struct ac_rtld_binary rtld_binaries
[5] = {};
1020 struct si_shader_inst
*instructions
=
1021 calloc(shader
->bo
->b
.b
.width0
/ 4, sizeof(struct si_shader_inst
));
1023 if (shader
->prolog
) {
1024 si_add_split_disasm(screen
, &rtld_binaries
[0], &shader
->prolog
->binary
,
1025 &inst_addr
, &num_inst
, instructions
);
1027 if (shader
->previous_stage
) {
1028 si_add_split_disasm(screen
, &rtld_binaries
[1], &shader
->previous_stage
->binary
,
1029 &inst_addr
, &num_inst
, instructions
);
1031 if (shader
->prolog2
) {
1032 si_add_split_disasm(screen
, &rtld_binaries
[2], &shader
->prolog2
->binary
,
1033 &inst_addr
, &num_inst
, instructions
);
1035 si_add_split_disasm(screen
, &rtld_binaries
[3], &shader
->binary
,
1036 &inst_addr
, &num_inst
, instructions
);
1037 if (shader
->epilog
) {
1038 si_add_split_disasm(screen
, &rtld_binaries
[4], &shader
->epilog
->binary
,
1039 &inst_addr
, &num_inst
, instructions
);
1042 fprintf(f
, COLOR_YELLOW
"%s - annotated disassembly:" COLOR_RESET
"\n",
1043 si_get_shader_name(shader
, shader
->selector
->type
));
1045 /* Print instructions with annotations. */
1046 for (i
= 0; i
< num_inst
; i
++) {
1047 struct si_shader_inst
*inst
= &instructions
[i
];
1049 fprintf(f
, "%.*s [PC=0x%"PRIx64
", size=%u]\n",
1050 inst
->textlen
, inst
->text
, inst
->addr
, inst
->size
);
1052 /* Print which waves execute the instruction right now. */
1053 while (num_waves
&& inst
->addr
== waves
->pc
) {
1055 " " COLOR_GREEN
"^ SE%u SH%u CU%u "
1056 "SIMD%u WAVE%u EXEC=%016"PRIx64
" ",
1057 waves
->se
, waves
->sh
, waves
->cu
, waves
->simd
,
1058 waves
->wave
, waves
->exec
);
1060 if (inst
->size
== 4) {
1061 fprintf(f
, "INST32=%08X" COLOR_RESET
"\n",
1064 fprintf(f
, "INST64=%08X %08X" COLOR_RESET
"\n",
1065 waves
->inst_dw0
, waves
->inst_dw1
);
1068 waves
->matched
= true;
1076 for (unsigned i
= 0; i
< ARRAY_SIZE(rtld_binaries
); ++i
)
1077 ac_rtld_close(&rtld_binaries
[i
]);
1080 static void si_dump_annotated_shaders(struct si_context
*sctx
, FILE *f
)
1082 struct ac_wave_info waves
[AC_MAX_WAVES_PER_CHIP
];
1083 unsigned num_waves
= ac_get_wave_info(waves
);
1085 fprintf(f
, COLOR_CYAN
"The number of active waves = %u" COLOR_RESET
1088 si_print_annotated_shader(sctx
->vs_shader
.current
, waves
, num_waves
, f
);
1089 si_print_annotated_shader(sctx
->tcs_shader
.current
, waves
, num_waves
, f
);
1090 si_print_annotated_shader(sctx
->tes_shader
.current
, waves
, num_waves
, f
);
1091 si_print_annotated_shader(sctx
->gs_shader
.current
, waves
, num_waves
, f
);
1092 si_print_annotated_shader(sctx
->ps_shader
.current
, waves
, num_waves
, f
);
1094 /* Print waves executing shaders that are not currently bound. */
1097 for (i
= 0; i
< num_waves
; i
++) {
1098 if (waves
[i
].matched
)
1102 fprintf(f
, COLOR_CYAN
1103 "Waves not executing currently-bound shaders:"
1107 fprintf(f
, " SE%u SH%u CU%u SIMD%u WAVE%u EXEC=%016"PRIx64
1108 " INST=%08X %08X PC=%"PRIx64
"\n",
1109 waves
[i
].se
, waves
[i
].sh
, waves
[i
].cu
, waves
[i
].simd
,
1110 waves
[i
].wave
, waves
[i
].exec
, waves
[i
].inst_dw0
,
1111 waves
[i
].inst_dw1
, waves
[i
].pc
);
1117 static void si_dump_command(const char *title
, const char *command
, FILE *f
)
1121 FILE *p
= popen(command
, "r");
1125 fprintf(f
, COLOR_YELLOW
"%s: " COLOR_RESET
"\n", title
);
1126 while (fgets(line
, sizeof(line
), p
))
1132 static void si_dump_debug_state(struct pipe_context
*ctx
, FILE *f
,
1135 struct si_context
*sctx
= (struct si_context
*)ctx
;
1138 u_log_flush(sctx
->log
);
1140 if (flags
& PIPE_DUMP_DEVICE_STATUS_REGISTERS
) {
1141 si_dump_debug_registers(sctx
, f
);
1143 si_dump_annotated_shaders(sctx
, f
);
1144 si_dump_command("Active waves (raw data)", "umr -O halt_waves -wa | column -t", f
);
1145 si_dump_command("Wave information", "umr -O halt_waves,bits -wa", f
);
1149 void si_log_draw_state(struct si_context
*sctx
, struct u_log_context
*log
)
1151 struct si_shader_ctx_state
*tcs_shader
;
1156 tcs_shader
= &sctx
->tcs_shader
;
1157 if (sctx
->tes_shader
.cso
&& !sctx
->tcs_shader
.cso
)
1158 tcs_shader
= &sctx
->fixed_func_tcs_shader
;
1160 si_dump_framebuffer(sctx
, log
);
1162 si_dump_gfx_shader(sctx
, &sctx
->vs_shader
, log
);
1163 si_dump_gfx_shader(sctx
, tcs_shader
, log
);
1164 si_dump_gfx_shader(sctx
, &sctx
->tes_shader
, log
);
1165 si_dump_gfx_shader(sctx
, &sctx
->gs_shader
, log
);
1166 si_dump_gfx_shader(sctx
, &sctx
->ps_shader
, log
);
1168 si_dump_descriptor_list(sctx
->screen
,
1169 &sctx
->descriptors
[SI_DESCS_RW_BUFFERS
],
1170 "", "RW buffers", 4,
1171 sctx
->descriptors
[SI_DESCS_RW_BUFFERS
].num_active_slots
,
1173 si_dump_gfx_descriptors(sctx
, &sctx
->vs_shader
, log
);
1174 si_dump_gfx_descriptors(sctx
, tcs_shader
, log
);
1175 si_dump_gfx_descriptors(sctx
, &sctx
->tes_shader
, log
);
1176 si_dump_gfx_descriptors(sctx
, &sctx
->gs_shader
, log
);
1177 si_dump_gfx_descriptors(sctx
, &sctx
->ps_shader
, log
);
1180 void si_log_compute_state(struct si_context
*sctx
, struct u_log_context
*log
)
1185 si_dump_compute_shader(sctx
, log
);
1186 si_dump_compute_descriptors(sctx
, log
);
1189 static void si_dump_dma(struct si_context
*sctx
,
1190 struct radeon_saved_cs
*saved
, FILE *f
)
1192 static const char ib_name
[] = "sDMA IB";
1195 si_dump_bo_list(sctx
, saved
, f
);
1197 fprintf(f
, "------------------ %s begin ------------------\n", ib_name
);
1199 for (i
= 0; i
< saved
->num_dw
; ++i
) {
1200 fprintf(f
, " %08x\n", saved
->ib
[i
]);
1203 fprintf(f
, "------------------- %s end -------------------\n", ib_name
);
1206 fprintf(f
, "SDMA Dump Done.\n");
1209 void si_check_vm_faults(struct si_context
*sctx
,
1210 struct radeon_saved_cs
*saved
, enum ring_type ring
)
1212 struct pipe_screen
*screen
= sctx
->b
.screen
;
1215 char cmd_line
[4096];
1217 if (!ac_vm_fault_occured(sctx
->chip_class
,
1218 &sctx
->dmesg_timestamp
, &addr
))
1221 f
= dd_get_debug_file(false);
1225 fprintf(f
, "VM fault report.\n\n");
1226 if (os_get_command_line(cmd_line
, sizeof(cmd_line
)))
1227 fprintf(f
, "Command: %s\n", cmd_line
);
1228 fprintf(f
, "Driver vendor: %s\n", screen
->get_vendor(screen
));
1229 fprintf(f
, "Device vendor: %s\n", screen
->get_device_vendor(screen
));
1230 fprintf(f
, "Device name: %s\n\n", screen
->get_name(screen
));
1231 fprintf(f
, "Failing VM page: 0x%08"PRIx64
"\n\n", addr
);
1233 if (sctx
->apitrace_call_number
)
1234 fprintf(f
, "Last apitrace call: %u\n\n",
1235 sctx
->apitrace_call_number
);
1239 struct u_log_context log
;
1240 u_log_context_init(&log
);
1242 si_log_draw_state(sctx
, &log
);
1243 si_log_compute_state(sctx
, &log
);
1244 si_log_cs(sctx
, &log
, true);
1246 u_log_new_page_print(&log
, f
);
1247 u_log_context_destroy(&log
);
1251 si_dump_dma(sctx
, saved
, f
);
1260 fprintf(stderr
, "Detected a VM fault, exiting...\n");
1264 void si_init_debug_functions(struct si_context
*sctx
)
1266 sctx
->b
.dump_debug_state
= si_dump_debug_state
;
1268 /* Set the initial dmesg timestamp for this context, so that
1269 * only new messages will be checked for VM faults.
1271 if (sctx
->screen
->debug_flags
& DBG(CHECK_VM
))
1272 ac_vm_fault_occured(sctx
->chip_class
,
1273 &sctx
->dmesg_timestamp
, NULL
);