2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
26 #include <midgard_pack.h>
34 #include "util/macros.h"
35 #include "util/u_math.h"
37 #include "midgard/disassemble.h"
38 #include "bifrost/disassemble.h"
40 #include "pan_encoder.h"
42 static void pandecode_swizzle(unsigned swizzle
, enum mali_format format
);
44 #define MEMORY_PROP(obj, p) {\
46 char *a = pointer_as_memory_reference(obj->p); \
47 pandecode_prop("%s = %s", #p, a); \
52 #define MEMORY_PROP_DIR(obj, p) {\
54 char *a = pointer_as_memory_reference(obj.p); \
55 pandecode_prop("%s = %s", #p, a); \
60 #define DUMP_CL(title, T, cl, indent) {\
61 fprintf(pandecode_dump_stream, "%s\n", title); \
62 struct MALI_ ## T temp; \
63 MALI_ ## T ## _unpack((const uint8_t *) cl, &temp); \
64 MALI_ ## T ## _print(pandecode_dump_stream, &temp, indent * 2); \
67 #define MAP_ADDR(T, addr, cl) \
68 const uint8_t *cl = 0; \
70 struct pandecode_mapped_memory *mapped_mem = pandecode_find_mapped_gpu_mem_containing(addr); \
71 cl = pandecode_fetch_gpu_mem(mapped_mem, addr, MALI_ ## T ## _LENGTH); \
74 #define DUMP_ADDR(title, T, addr, indent) {\
75 MAP_ADDR(T, addr, cl) \
76 DUMP_CL(title, T, cl, indent); \
79 FILE *pandecode_dump_stream
;
81 /* Semantic logging type.
83 * Raw: for raw messages to be printed as is.
84 * Message: for helpful information to be commented out in replays.
85 * Property: for properties of a struct
87 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
90 enum pandecode_log_type
{
96 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
97 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
98 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
100 unsigned pandecode_indent
= 0;
103 pandecode_make_indent(void)
105 for (unsigned i
= 0; i
< pandecode_indent
; ++i
)
106 fprintf(pandecode_dump_stream
, " ");
110 pandecode_log_typed(enum pandecode_log_type type
, const char *format
, ...)
114 pandecode_make_indent();
116 if (type
== PANDECODE_MESSAGE
)
117 fprintf(pandecode_dump_stream
, "// ");
118 else if (type
== PANDECODE_PROPERTY
)
119 fprintf(pandecode_dump_stream
, ".");
121 va_start(ap
, format
);
122 vfprintf(pandecode_dump_stream
, format
, ap
);
125 if (type
== PANDECODE_PROPERTY
)
126 fprintf(pandecode_dump_stream
, ",\n");
130 pandecode_log_cont(const char *format
, ...)
134 va_start(ap
, format
);
135 vfprintf(pandecode_dump_stream
, format
, ap
);
139 /* To check for memory safety issues, validates that the given pointer in GPU
140 * memory is valid, containing at least sz bytes. The goal is to eliminate
141 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
142 * overruns) by statically validating pointers.
146 pandecode_validate_buffer(mali_ptr addr
, size_t sz
)
149 pandecode_msg("XXX: null pointer deref");
155 struct pandecode_mapped_memory
*bo
=
156 pandecode_find_mapped_gpu_mem_containing(addr
);
159 pandecode_msg("XXX: invalid memory dereference\n");
165 unsigned offset
= addr
- bo
->gpu_va
;
166 unsigned total
= offset
+ sz
;
168 if (total
> bo
->length
) {
169 pandecode_msg("XXX: buffer overrun. "
170 "Chunk of size %zu at offset %d in buffer of size %zu. "
171 "Overrun by %zu bytes. \n",
172 sz
, offset
, bo
->length
, total
- bo
->length
);
177 struct pandecode_flag_info
{
183 pandecode_log_decoded_flags(const struct pandecode_flag_info
*flag_info
,
186 bool decodable_flags_found
= false;
188 for (int i
= 0; flag_info
[i
].name
; i
++) {
189 if ((flags
& flag_info
[i
].flag
) != flag_info
[i
].flag
)
192 if (!decodable_flags_found
) {
193 decodable_flags_found
= true;
195 pandecode_log_cont(" | ");
198 pandecode_log_cont("%s", flag_info
[i
].name
);
200 flags
&= ~flag_info
[i
].flag
;
203 if (decodable_flags_found
) {
205 pandecode_log_cont(" | 0x%" PRIx64
, flags
);
207 pandecode_log_cont("0x%" PRIx64
, flags
);
211 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
212 static const struct pandecode_flag_info gl_enable_flag_info
[] = {
213 FLAG_INFO(OCCLUSION_QUERY
),
214 FLAG_INFO(OCCLUSION_PRECISE
),
215 FLAG_INFO(FRONT_CCW_TOP
),
216 FLAG_INFO(CULL_FACE_FRONT
),
217 FLAG_INFO(CULL_FACE_BACK
),
222 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
223 static const struct pandecode_flag_info clear_flag_info
[] = {
226 FLAG_INFO(SLOW_STENCIL
),
231 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
232 static const struct pandecode_flag_info mfbd_fmt_flag_info
[] = {
238 #define FLAG_INFO(flag) { MALI_AFBC_##flag, "MALI_AFBC_" #flag }
239 static const struct pandecode_flag_info afbc_fmt_flag_info
[] = {
245 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
246 static const struct pandecode_flag_info mfbd_extra_flag_hi_info
[] = {
252 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
253 static const struct pandecode_flag_info mfbd_extra_flag_lo_info
[] = {
259 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
260 static const struct pandecode_flag_info mfbd_flag_info
[] = {
261 FLAG_INFO(DEPTH_WRITE
),
267 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
268 static const struct pandecode_flag_info sfbd_unk1_info
[] = {
275 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
276 static const struct pandecode_flag_info sfbd_unk2_info
[] = {
283 /* Midgard's tiler descriptor is embedded within the
287 pandecode_midgard_tiler_descriptor(
288 const struct midgard_tiler_descriptor
*t
,
294 pandecode_log(".tiler = {\n");
297 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
298 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
300 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
302 /* We know this name from the kernel, but we never see it nonzero */
305 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
307 MEMORY_PROP(t
, polygon_list
);
309 /* The body is offset from the base of the polygon list */
310 //assert(t->polygon_list_body > t->polygon_list);
311 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
313 /* It needs to fit inside the reported size */
314 //assert(t->polygon_list_size >= body_offset);
316 /* Now that we've sanity checked, we'll try to calculate the sizes
317 * ourselves for comparison */
319 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
320 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
322 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
323 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
324 ref_header
, ref_size
);
325 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
326 pandecode_msg("body offset %d\n", body_offset
);
329 /* The tiler heap has a start and end specified -- it should be
330 * identical to what we have in the BO. The exception is if tiling is
333 MEMORY_PROP(t
, heap_start
);
334 assert(t
->heap_end
>= t
->heap_start
);
336 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
338 /* Tiling is enabled with a special flag */
339 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
340 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
342 bool tiling_enabled
= hierarchy_mask
;
344 if (tiling_enabled
) {
345 /* We should also have no other flags */
347 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
349 /* When tiling is disabled, we should have that flag and no others */
351 if (tiler_flags
!= MALI_TILER_DISABLED
) {
352 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
356 /* We should also have an empty heap */
358 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
362 /* Disabled tiling is used only for clear-only jobs, which are
363 * purely FRAGMENT, so we should never see this for
364 * non-FRAGMENT descriptors. */
367 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
370 /* We've never seen weights used in practice, but we know from the
371 * kernel these fields is there */
373 bool nonzero_weights
= false;
375 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
376 nonzero_weights
|= t
->weights
[w
] != 0x0;
379 if (nonzero_weights
) {
380 pandecode_log(".weights = { ");
382 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
383 pandecode_log_cont("%d, ", t
->weights
[w
]);
390 pandecode_log("}\n");
393 /* TODO: The Bifrost tiler is not understood at all yet */
396 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer
*fb
)
398 pandecode_log(".tiler = {\n");
401 MEMORY_PROP(fb
, tiler_meta
);
403 for (int i
= 0; i
< 16; i
++) {
404 if (fb
->zeros
[i
] != 0) {
405 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
410 pandecode_log("},\n");
413 pandecode_log("}\n");
417 /* Information about the framebuffer passed back for
418 * additional analysis */
420 struct pandecode_fbd
{
428 pandecode_sfbd_format(struct mali_sfbd_format format
)
430 pandecode_log(".format = {\n");
433 pandecode_log(".unk1 = ");
434 pandecode_log_decoded_flags(sfbd_unk1_info
, format
.unk1
);
435 pandecode_log_cont(",\n");
437 /* TODO: Map formats so we can check swizzles and print nicely */
438 pandecode_log("swizzle");
439 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
440 pandecode_log_cont(",\n");
442 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
443 (format
.nr_channels
+ 1));
445 pandecode_log(".unk2 = ");
446 pandecode_log_decoded_flags(sfbd_unk2_info
, format
.unk2
);
447 pandecode_log_cont(",\n");
449 pandecode_prop("block = %s", mali_block_format_as_str(format
.block
));
451 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
454 pandecode_log("},\n");
458 pandecode_shared_memory(const struct mali_shared_memory
*desc
, bool is_compute
)
460 pandecode_prop("stack_shift = 0x%x", desc
->stack_shift
);
463 pandecode_prop("unk0 = 0x%x", desc
->unk0
);
465 if (desc
->shared_workgroup_count
!= 0x1F) {
466 pandecode_prop("shared_workgroup_count = %d", desc
->shared_workgroup_count
);
468 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
471 if (desc
->shared_unk1
|| desc
->shared_shift
) {
472 pandecode_prop("shared_unk1 = %X", desc
->shared_unk1
);
473 pandecode_prop("shared_shift = %X", desc
->shared_shift
);
476 pandecode_msg("XXX: shared memory configured in noncompute shader");
479 if (desc
->shared_zero
) {
480 pandecode_msg("XXX: shared memory zero tripped\n");
481 pandecode_prop("shared_zero = 0x%" PRIx32
, desc
->shared_zero
);
484 if (desc
->shared_memory
&& !is_compute
)
485 pandecode_msg("XXX: shared memory used in noncompute shader\n");
487 MEMORY_PROP(desc
, scratchpad
);
488 MEMORY_PROP(desc
, shared_memory
);
489 MEMORY_PROP(desc
, unknown1
);
492 static struct pandecode_fbd
493 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
, unsigned gpu_id
)
495 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
496 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
498 struct pandecode_fbd info
= {
503 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
506 pandecode_log(".shared_memory = {\n");
508 pandecode_shared_memory(&s
->shared_memory
, false);
510 pandecode_log("},\n");
512 pandecode_sfbd_format(s
->format
);
514 info
.width
= s
->width
+ 1;
515 info
.height
= s
->height
+ 1;
517 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", info
.width
);
518 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", info
.height
);
520 MEMORY_PROP(s
, checksum
);
522 if (s
->checksum_stride
)
523 pandecode_prop("checksum_stride = %d", s
->checksum_stride
);
525 MEMORY_PROP(s
, framebuffer
);
526 pandecode_prop("stride = %d", s
->stride
);
528 /* Earlier in the actual commandstream -- right before width -- but we
529 * delay to flow nicer */
531 pandecode_log(".clear_flags = ");
532 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
533 pandecode_log_cont(",\n");
535 if (s
->depth_buffer
) {
536 MEMORY_PROP(s
, depth_buffer
);
537 pandecode_prop("depth_stride = %d", s
->depth_stride
);
540 if (s
->stencil_buffer
) {
541 MEMORY_PROP(s
, stencil_buffer
);
542 pandecode_prop("stencil_stride = %d", s
->stencil_stride
);
545 if (s
->depth_stride_zero
||
546 s
->stencil_stride_zero
||
547 s
->zero7
|| s
->zero8
) {
548 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
549 pandecode_prop("depth_stride_zero = 0x%x",
550 s
->depth_stride_zero
);
551 pandecode_prop("stencil_stride_zero = 0x%x",
552 s
->stencil_stride_zero
);
553 pandecode_prop("zero7 = 0x%" PRIx32
,
555 pandecode_prop("zero8 = 0x%" PRIx32
,
559 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
560 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
561 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
562 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
563 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
566 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
567 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
568 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
569 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
570 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
573 if (s
->clear_stencil
) {
574 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
577 const struct midgard_tiler_descriptor t
= s
->tiler
;
579 bool has_hierarchy
= !(gpu_id
== 0x0720 || gpu_id
== 0x0820 || gpu_id
== 0x0830);
580 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
, has_hierarchy
);
583 pandecode_log("};\n");
585 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
586 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
587 pandecode_prop("zero5 = 0x%" PRIx32
, s
->zero5
);
589 pandecode_log_cont(".zero3 = {");
591 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
592 pandecode_log_cont("%X, ", s
->zero3
[i
]);
594 pandecode_log_cont("},\n");
596 pandecode_log_cont(".zero6 = {");
598 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
599 pandecode_log_cont("%X, ", s
->zero6
[i
]);
601 pandecode_log_cont("},\n");
607 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
609 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
610 const struct mali_shared_memory
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
612 pandecode_log("struct mali_shared_memory shared_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
614 pandecode_shared_memory(s
, true);
616 pandecode_log("},\n");
619 /* Extracts the number of components associated with a Mali format */
622 pandecode_format_component_count(enum mali_format fmt
)
624 /* Mask out the format class */
625 unsigned top
= fmt
& 0b11100000;
628 case MALI_FORMAT_SNORM
:
629 case MALI_FORMAT_UINT
:
630 case MALI_FORMAT_UNORM
:
631 case MALI_FORMAT_SINT
:
632 return ((fmt
>> 3) & 3) + 1;
639 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
642 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
645 assert(MALI_CHANNEL_R
== 0);
647 for (unsigned c
= 0; c
< 4; ++c
) {
648 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
650 if (chan
<= MALI_CHANNEL_A
)
657 /* Validates that a (format, swizzle) pair is valid, in the sense that the
658 * swizzle doesn't access any components that are undefined in the format.
659 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
663 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
665 unsigned nr_comp
= pandecode_format_component_count(fmt
);
666 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
667 unsigned valid_mask
= (1 << nr_comp
) - 1;
668 unsigned invalid_mask
= ~valid_mask
;
670 if (access_mask
& invalid_mask
) {
671 pandecode_msg("XXX: invalid components accessed\n");
675 /* Check for the default non-swizzling swizzle so we can suppress
676 * useless printing for the defaults */
678 unsigned default_swizzles
[4] = {
679 MALI_CHANNEL_R
| (MALI_CHANNEL_0
<< 3) | (MALI_CHANNEL_0
<< 6) | (MALI_CHANNEL_1
<< 9),
680 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_0
<< 6) | (MALI_CHANNEL_1
<< 9),
681 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_B
<< 6) | (MALI_CHANNEL_1
<< 9),
682 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_B
<< 6) | (MALI_CHANNEL_A
<< 9)
685 return (swizzle
== default_swizzles
[nr_comp
- 1]);
689 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
691 /* First, do some validation */
692 bool trivial_swizzle
= pandecode_validate_format_swizzle(
698 /* Next, print the swizzle */
699 pandecode_log_cont(".");
701 static const char components
[] = "rgba01";
703 for (unsigned c
= 0; c
< 4; ++c
) {
704 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
706 if (chan
> MALI_CHANNEL_1
) {
707 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
710 pandecode_log_cont("%c", components
[chan
]);
715 pandecode_rt_format(struct mali_rt_format format
)
717 pandecode_log(".format = {\n");
720 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
721 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
722 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
723 pandecode_prop("unk4 = 0x%" PRIx32
, format
.unk4
);
725 pandecode_prop("block = %s", mali_block_format_as_str(format
.block
));
727 /* TODO: Map formats so we can check swizzles and print nicely */
728 pandecode_log("swizzle");
729 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
730 pandecode_log_cont(",\n");
732 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
733 (format
.nr_channels
+ 1));
735 pandecode_log(".flags = ");
736 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
737 pandecode_log_cont(",\n");
739 pandecode_prop("msaa = %s", mali_msaa_as_str(format
.msaa
));
741 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
742 * which is a faster hardware-based alternative to the wallpaper method
743 * to preserve framebuffer contents across frames. In practice, MFBD
744 * preload is buggy on Midgard, and so this is a chicken bit. If this
745 * bit isn't set, most likely something broke unrelated to preload */
747 if (!format
.no_preload
) {
748 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
749 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
753 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
756 pandecode_log("},\n");
760 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct mali_framebuffer
*fb
)
762 pandecode_log("struct mali_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
765 for (int i
= 0; i
< (fb
->rt_count_1
+ 1); i
++) {
766 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct mali_render_target
);
767 struct pandecode_mapped_memory
*mem
=
768 pandecode_find_mapped_gpu_mem_containing(rt_va
);
769 const struct mali_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
771 pandecode_log("{\n");
774 pandecode_rt_format(rt
->format
);
776 if (rt
->format
.block
== MALI_BLOCK_FORMAT_AFBC
) {
777 pandecode_log(".afbc = {\n");
780 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
781 pandecode_prop("metadata = %s", a
);
784 pandecode_prop("stride = %d", rt
->afbc
.stride
);
786 pandecode_log(".flags = ");
787 pandecode_log_decoded_flags(afbc_fmt_flag_info
, rt
->afbc
.flags
);
788 pandecode_log_cont(",\n");
791 pandecode_log("},\n");
792 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.flags
) {
793 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
799 MEMORY_PROP(rt
, framebuffer
);
800 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
802 if (rt
->layer_stride
)
803 pandecode_prop("layer_stride = %d", rt
->layer_stride
);
805 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
806 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
807 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
808 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
809 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
812 if (rt
->zero1
|| rt
->zero2
) {
813 pandecode_msg("XXX: render target zeros tripped\n");
814 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
815 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
819 pandecode_log("},\n");
823 pandecode_log("};\n");
826 static struct pandecode_fbd
827 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
, bool is_compute
, bool is_bifrost
)
829 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
830 const struct mali_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
832 struct pandecode_fbd info
;
834 if (is_bifrost
&& fb
->msaa
.sample_locations
) {
835 /* The blob stores all possible sample locations in a single buffer
836 * allocated on startup, and just switches the pointer when switching
837 * MSAA state. For now, we just put the data into the cmdstream, but we
838 * should do something like what the blob does with a real driver.
840 * There seem to be 32 slots for sample locations, followed by another
841 * 16. The second 16 is just the center location followed by 15 zeros
842 * in all the cases I've identified (maybe shader vs. depth/color
846 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->msaa
.sample_locations
);
848 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->msaa
.sample_locations
);
850 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
853 for (int i
= 0; i
< 32 + 16; i
++) {
854 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
858 pandecode_log("};\n");
861 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
865 pandecode_log(".msaa = {\n");
868 if (fb
->msaa
.sample_locations
)
869 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
871 pandecode_msg("XXX: sample_locations missing\n");
873 if (fb
->msaa
.zero1
|| fb
->msaa
.zero2
|| fb
->msaa
.zero4
) {
874 pandecode_msg("XXX: multisampling zero tripped\n");
875 pandecode_prop("zero1 = %" PRIx64
, fb
->msaa
.zero1
);
876 pandecode_prop("zero2 = %" PRIx64
, fb
->msaa
.zero2
);
877 pandecode_prop("zero4 = %" PRIx64
, fb
->msaa
.zero4
);
881 pandecode_log("},\n");
883 pandecode_log(".shared_memory = {\n");
885 pandecode_shared_memory(&fb
->shared_memory
, is_compute
);
887 pandecode_log("},\n");
890 info
.width
= fb
->width1
+ 1;
891 info
.height
= fb
->height1
+ 1;
892 info
.rt_count
= fb
->rt_count_1
+ 1;
894 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
895 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
896 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
897 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
899 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
900 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
901 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
902 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
904 pandecode_log(".mfbd_flags = ");
905 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
906 pandecode_log_cont(",\n");
908 if (fb
->clear_stencil
)
909 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
912 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
916 pandecode_bifrost_tiler_descriptor(fb
);
918 const struct midgard_tiler_descriptor t
= fb
->tiler
;
919 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
, true);
922 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
924 if (fb
->zero3
|| fb
->zero4
) {
925 pandecode_msg("XXX: framebuffer zeros tripped\n");
926 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
927 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
931 pandecode_log("};\n");
933 gpu_va
+= sizeof(struct mali_framebuffer
);
935 info
.has_extra
= (fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
;
937 if (info
.has_extra
) {
938 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
939 const struct mali_framebuffer_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
941 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
944 MEMORY_PROP(fbx
, checksum
);
946 if (fbx
->checksum_stride
)
947 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
949 pandecode_log(".flags_hi = ");
950 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info
, fbx
->flags_hi
);
951 pandecode_log_cont(",\n");
953 pandecode_log(".flags_lo = ");
954 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info
, fbx
->flags_lo
);
955 pandecode_log_cont(",\n");
957 pandecode_prop("zs_block = %s", mali_block_format_as_str(fbx
->zs_block
));
958 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx
->zs_samples
+ 1);
960 if (fbx
->zs_block
== MALI_BLOCK_FORMAT_AFBC
) {
961 pandecode_log(".ds_afbc = {\n");
964 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
965 pandecode_prop("depth_stencil_afbc_stride = %d",
966 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
967 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
969 pandecode_log(".flags = ");
970 pandecode_log_decoded_flags(afbc_fmt_flag_info
, fbx
->ds_afbc
.flags
);
971 pandecode_log_cont(",\n");
973 if (fbx
->ds_afbc
.padding
) {
974 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
975 pandecode_prop("padding = 0x%" PRIx64
, fbx
->ds_afbc
.padding
);
979 pandecode_log("},\n");
981 pandecode_log(".ds_linear = {\n");
984 if (fbx
->ds_linear
.depth
) {
985 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
986 pandecode_prop("depth_stride = %d",
987 fbx
->ds_linear
.depth_stride
);
988 pandecode_prop("depth_layer_stride = %d",
989 fbx
->ds_linear
.depth_layer_stride
);
990 } else if (fbx
->ds_linear
.depth_stride
|| fbx
->ds_linear
.depth_layer_stride
) {
991 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx
->ds_linear
.depth_stride
, fbx
->ds_linear
.depth_layer_stride
);
994 if (fbx
->ds_linear
.stencil
) {
995 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
996 pandecode_prop("stencil_stride = %d",
997 fbx
->ds_linear
.stencil_stride
);
998 pandecode_prop("stencil_layer_stride = %d",
999 fbx
->ds_linear
.stencil_layer_stride
);
1000 } else if (fbx
->ds_linear
.stencil_stride
|| fbx
->ds_linear
.stencil_layer_stride
) {
1001 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx
->ds_linear
.stencil_stride
, fbx
->ds_linear
.stencil_layer_stride
);
1004 if (fbx
->ds_linear
.depth_stride_zero
||
1005 fbx
->ds_linear
.stencil_stride_zero
) {
1006 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1007 pandecode_prop("depth_stride_zero = 0x%x",
1008 fbx
->ds_linear
.depth_stride_zero
);
1009 pandecode_prop("stencil_stride_zero = 0x%x",
1010 fbx
->ds_linear
.stencil_stride_zero
);
1014 pandecode_log("},\n");
1017 if (fbx
->clear_color_1
| fbx
->clear_color_2
) {
1018 pandecode_prop("clear_color_1 = 0x%" PRIx32
, fbx
->clear_color_1
);
1019 pandecode_prop("clear_color_2 = 0x%" PRIx32
, fbx
->clear_color_2
);
1023 pandecode_msg("XXX: fb_extra zeros tripped\n");
1024 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1028 pandecode_log("};\n");
1030 gpu_va
+= sizeof(struct mali_framebuffer_extra
);
1034 pandecode_render_target(gpu_va
, job_no
, fb
);
1040 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1041 mali_ptr addr
, int job_no
, char *suffix
,
1042 int count
, bool varying
, enum mali_job_type job_type
)
1044 char *prefix
= varying
? "Varying" : "Attribute";
1048 pandecode_msg("warn: No %s records\n", prefix
);
1052 MAP_ADDR(ATTRIBUTE_BUFFER
, addr
, cl
);
1054 for (int i
= 0; i
< count
; ++i
) {
1055 fprintf(pandecode_dump_stream
, "%s\n", prefix
);
1057 struct MALI_ATTRIBUTE_BUFFER temp
;
1058 MALI_ATTRIBUTE_BUFFER_unpack(cl
+ i
* MALI_ATTRIBUTE_BUFFER_LENGTH
, &temp
);
1059 MALI_ATTRIBUTE_BUFFER_print(pandecode_dump_stream
, &temp
, 2);
1061 if (temp
.type
== MALI_ATTRIBUTE_TYPE_1D_NPOT_DIVISOR
) {
1062 struct MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT temp2
;
1063 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_unpack(cl
+ (i
+ 1) * MALI_ATTRIBUTE_BUFFER_LENGTH
, &temp2
);
1064 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_print(pandecode_dump_stream
, &temp2
, 2);
1070 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1072 /* TODO: Decode flags */
1073 mali_ptr shader_ptr
= ptr
& ~15;
1075 char *a
= pointer_as_memory_reference(shader_ptr
);
1076 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1082 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1085 decode_bifrost_constant(u16 constant
)
1087 float lo
= (float) (constant
& 0xFF);
1088 float hi
= (float) (constant
>> 8);
1090 return (hi
/ 255.0) + (lo
/ 65535.0);
1094 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1096 struct bifrost_blend_rt
*b
=
1097 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1099 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1102 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1103 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1104 b
->constant
, decode_bifrost_constant(b
->constant
));
1106 /* TODO figure out blend shader enable bit */
1107 DUMP_CL("Equation", BLEND_EQUATION
, &b
->equation
, 2);
1109 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1110 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1112 pandecode_log(".format = %s", mali_format_as_str(b
->format
));
1113 pandecode_swizzle(b
->swizzle
, b
->format
);
1114 pandecode_log_cont(",\n");
1116 pandecode_prop("swizzle = 0x%" PRIx32
, b
->swizzle
);
1117 pandecode_prop("format = 0x%" PRIx32
, b
->format
);
1120 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1121 pandecode_prop("zero1 = 0x%" PRIx32
, b
->zero1
);
1124 pandecode_log(".shader_type = ");
1125 switch(b
->shader_type
) {
1126 case BIFROST_BLEND_F16
:
1127 pandecode_log_cont("BIFROST_BLEND_F16");
1129 case BIFROST_BLEND_F32
:
1130 pandecode_log_cont("BIFROST_BLEND_F32");
1132 case BIFROST_BLEND_I32
:
1133 pandecode_log_cont("BIFROST_BLEND_I32");
1135 case BIFROST_BLEND_U32
:
1136 pandecode_log_cont("BIFROST_BLEND_U32");
1138 case BIFROST_BLEND_I16
:
1139 pandecode_log_cont("BIFROST_BLEND_I16");
1141 case BIFROST_BLEND_U16
:
1142 pandecode_log_cont("BIFROST_BLEND_U16");
1145 pandecode_log_cont(",\n");
1148 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1149 pandecode_prop("zero2 = 0x%" PRIx32
, b
->zero2
);
1152 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1155 pandecode_log("},\n");
1161 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1163 /* constant/equation is in a union */
1167 pandecode_log(".blend = {\n");
1171 pandecode_shader_address("shader", blend
->shader
);
1173 DUMP_CL("Equation", BLEND_EQUATION
, &blend
->equation
, 2);
1174 pandecode_prop("constant = %f", blend
->constant
);
1178 pandecode_log("},\n");
1180 /* Return blend shader to disassemble if present */
1181 return is_shader
? (blend
->shader
& ~0xF) : 0;
1185 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1187 struct midgard_blend_rt
*b
=
1188 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1190 /* Flags determine presence of blend shader */
1191 bool is_shader
= b
->flags
.opaque
[0] & 0x2;
1193 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1196 DUMP_CL("Flags", BLEND_FLAGS
, &b
->flags
, 2);
1198 union midgard_blend blend
= b
->blend
;
1199 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1202 pandecode_log("};\n");
1207 /* Attributes and varyings have descriptor records, which contain information
1208 * about their format and ordering with the attribute/varying buffers. We'll
1209 * want to validate that the combinations specified are self-consistent.
1213 pandecode_attribute_meta(int count
, mali_ptr attribute
, bool varying
, char *suffix
)
1215 const char *prefix
= varying
? "Varying" : "Attribute";
1217 for (int i
= 0; i
< count
; ++i
, attribute
+= MALI_ATTRIBUTE_LENGTH
)
1218 DUMP_ADDR(prefix
, ATTRIBUTE
, attribute
, 1);
1223 /* return bits [lo, hi) of word */
1225 bits(u32 word
, u32 lo
, u32 hi
)
1228 return word
; // avoid undefined behavior with the shift
1230 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1234 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool graphics
)
1236 pandecode_log(".prefix = {\n");
1239 /* Decode invocation_count. See the comment before the definition of
1240 * invocation_count for an explanation.
1242 struct MALI_INVOCATION invocation
;
1243 struct mali_invocation_packed invocation_packed
= p
->invocation
;
1244 MALI_INVOCATION_unpack((const uint8_t *) &invocation_packed
, &invocation
);
1246 unsigned size_x
= bits(invocation
.invocations
, 0, invocation
.size_y_shift
) + 1;
1247 unsigned size_y
= bits(invocation
.invocations
, invocation
.size_y_shift
, invocation
.size_z_shift
) + 1;
1248 unsigned size_z
= bits(invocation
.invocations
, invocation
.size_z_shift
, invocation
.workgroups_x_shift
) + 1;
1250 unsigned groups_x
= bits(invocation
.invocations
, invocation
.workgroups_x_shift
, invocation
.workgroups_y_shift
) + 1;
1251 unsigned groups_y
= bits(invocation
.invocations
, invocation
.workgroups_y_shift
, invocation
.workgroups_z_shift
) + 1;
1252 unsigned groups_z
= bits(invocation
.invocations
, invocation
.workgroups_z_shift
, 32) + 1;
1254 /* Even though we have this decoded, we want to ensure that the
1255 * representation is "unique" so we don't lose anything by printing only
1256 * the final result. More specifically, we need to check that we were
1257 * passed something in canonical form, since the definition per the
1258 * hardware is inherently not unique. How? Well, take the resulting
1259 * decode and pack it ourselves! If it is bit exact with what we
1260 * decoded, we're good to go. */
1262 struct mali_invocation_packed ref
;
1263 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, graphics
);
1265 if (memcmp(&ref
, &invocation_packed
, sizeof(ref
))) {
1266 pandecode_msg("XXX: non-canonical workgroups packing\n");
1267 MALI_INVOCATION_print(pandecode_dump_stream
, &invocation
, 1 * 2);
1270 /* Regardless, print the decode */
1271 fprintf(pandecode_dump_stream
,
1272 "Invocation (%d, %d, %d) x (%d, %d, %d)\n",
1273 size_x
, size_y
, size_z
,
1274 groups_x
, groups_y
, groups_z
);
1276 fprintf(pandecode_dump_stream
, "Primitive\n");
1277 struct MALI_PRIMITIVE primitive
;
1278 struct mali_primitive_packed prim_packed
= p
->primitive
;
1279 MALI_PRIMITIVE_unpack((const uint8_t *) &prim_packed
, &primitive
);
1280 MALI_PRIMITIVE_print(pandecode_dump_stream
, &primitive
, 1 * 2);
1282 /* Validate an index buffer is present if we need one. TODO: verify
1283 * relationship between invocation_count and index_count */
1285 if (primitive
.indices
) {
1287 unsigned size
= (primitive
.index_type
== MALI_INDEX_TYPE_UINT32
) ?
1288 sizeof(uint32_t) : primitive
.index_type
;
1290 /* Ensure we got a size, and if so, validate the index buffer
1291 * is large enough to hold a full set of indices of the given
1295 pandecode_msg("XXX: index size missing\n");
1297 pandecode_validate_buffer(primitive
.indices
, primitive
.index_count
* size
);
1298 } else if (primitive
.index_type
)
1299 pandecode_msg("XXX: unexpected index size\n");
1302 pandecode_log("},\n");
1306 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1308 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1309 uint64_t *PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1311 for (int i
= 0; i
< ubufs_count
; i
++) {
1312 unsigned size
= (ubufs
[i
] & ((1 << 10) - 1)) * 16;
1313 mali_ptr addr
= (ubufs
[i
] >> 10) << 2;
1315 pandecode_validate_buffer(addr
, size
);
1317 char *ptr
= pointer_as_memory_reference(addr
);
1318 pandecode_log("ubuf_%d[%u] = %s;\n", i
, size
, ptr
);
1322 pandecode_log("\n");
1326 pandecode_uniforms(mali_ptr uniforms
, unsigned uniform_count
)
1328 pandecode_validate_buffer(uniforms
, uniform_count
* 16);
1330 char *ptr
= pointer_as_memory_reference(uniforms
);
1331 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count
, ptr
);
1336 shader_type_for_job(unsigned type
)
1339 case MALI_JOB_TYPE_VERTEX
: return "VERTEX";
1340 case MALI_JOB_TYPE_TILER
: return "FRAGMENT";
1341 case MALI_JOB_TYPE_COMPUTE
: return "COMPUTE";
1347 static unsigned shader_id
= 0;
1349 static struct midgard_disasm_stats
1350 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1351 bool is_bifrost
, unsigned gpu_id
)
1353 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1354 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1356 /* Compute maximum possible size */
1357 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1359 /* Print some boilerplate to clearly denote the assembly (which doesn't
1360 * obey indentation rules), and actually do the disassembly! */
1362 pandecode_log_cont("\n\n");
1364 struct midgard_disasm_stats stats
;
1367 disassemble_bifrost(pandecode_dump_stream
, code
, sz
, true);
1369 /* TODO: Extend stats to Bifrost */
1370 stats
.texture_count
= -128;
1371 stats
.sampler_count
= -128;
1372 stats
.attribute_count
= -128;
1373 stats
.varying_count
= -128;
1374 stats
.uniform_count
= -128;
1375 stats
.uniform_buffer_count
= -128;
1376 stats
.work_count
= -128;
1378 stats
.instruction_count
= 0;
1379 stats
.bundle_count
= 0;
1380 stats
.quadword_count
= 0;
1381 stats
.helper_invocations
= false;
1383 stats
= disassemble_midgard(pandecode_dump_stream
,
1385 type
== MALI_JOB_TYPE_TILER
?
1386 MESA_SHADER_FRAGMENT
: MESA_SHADER_VERTEX
);
1389 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1390 * driver-internal purposes with the blob and interfere */
1392 bool should_shaderdb
= type
!= MALI_JOB_TYPE_COMPUTE
;
1394 if (should_shaderdb
) {
1395 unsigned nr_threads
=
1396 (stats
.work_count
<= 4) ? 4 :
1397 (stats
.work_count
<= 8) ? 2 :
1400 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
1401 "%u inst, %u bundles, %u quadwords, "
1402 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1404 shader_type_for_job(type
),
1405 stats
.instruction_count
, stats
.bundle_count
, stats
.quadword_count
,
1406 stats
.work_count
, nr_threads
);
1414 pandecode_texture_payload(mali_ptr payload
,
1415 enum mali_texture_dimension dim
,
1416 enum mali_texture_layout layout
,
1420 uint16_t array_size
,
1421 struct pandecode_mapped_memory
*tmem
)
1423 pandecode_log(".payload = {\n");
1426 /* A bunch of bitmap pointers follow.
1427 * We work out the correct number,
1428 * based on the mipmap/cubemap
1429 * properties, but dump extra
1430 * possibilities to futureproof */
1432 int bitmap_count
= levels
+ 1;
1434 /* Miptree for each face */
1435 if (dim
== MALI_TEXTURE_DIMENSION_CUBE
)
1438 /* Array of layers */
1439 bitmap_count
*= depth
;
1441 /* Array of textures */
1442 bitmap_count
*= array_size
;
1444 /* Stride for each element */
1448 mali_ptr
*pointers_and_strides
= pandecode_fetch_gpu_mem(tmem
,
1449 payload
, sizeof(mali_ptr
) * bitmap_count
);
1450 for (int i
= 0; i
< bitmap_count
; ++i
) {
1451 /* How we dump depends if this is a stride or a pointer */
1453 if (manual_stride
&& (i
& 1)) {
1454 /* signed 32-bit snuck in as a 64-bit pointer */
1455 uint64_t stride_set
= pointers_and_strides
[i
];
1456 uint32_t clamped_stride
= stride_set
;
1457 int32_t stride
= clamped_stride
;
1458 assert(stride_set
== clamped_stride
);
1459 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
1461 char *a
= pointer_as_memory_reference(pointers_and_strides
[i
]);
1462 pandecode_log("%s, \n", a
);
1468 pandecode_log("},\n");
1472 pandecode_texture(mali_ptr u
,
1473 struct pandecode_mapped_memory
*tmem
,
1474 unsigned job_no
, unsigned tex
)
1476 struct pandecode_mapped_memory
*mapped_mem
= pandecode_find_mapped_gpu_mem_containing(u
);
1477 const uint8_t *cl
= pandecode_fetch_gpu_mem(mapped_mem
, u
, MALI_MIDGARD_TEXTURE_LENGTH
);
1479 struct MALI_MIDGARD_TEXTURE temp
;
1480 MALI_MIDGARD_TEXTURE_unpack(cl
, &temp
);
1481 MALI_MIDGARD_TEXTURE_print(pandecode_dump_stream
, &temp
, 2);
1483 pandecode_texture_payload(u
+ MALI_MIDGARD_TEXTURE_LENGTH
,
1484 temp
.dimension
, temp
.texel_ordering
, temp
.manual_stride
,
1485 temp
.levels
, temp
.depth
, temp
.array_size
, mapped_mem
);
1489 pandecode_bifrost_texture(
1494 struct MALI_BIFROST_TEXTURE temp
;
1495 MALI_BIFROST_TEXTURE_unpack(cl
, &temp
);
1496 MALI_BIFROST_TEXTURE_print(pandecode_dump_stream
, &temp
, 2);
1498 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(temp
.surfaces
);
1499 pandecode_texture_payload(temp
.surfaces
, temp
.dimension
, temp
.texel_ordering
,
1500 true, temp
.levels
, 1, 1, tmem
);
1503 /* For shader properties like texture_count, we have a claimed property in the shader_meta, and the actual Truth from static analysis (this may just be an upper limit). We validate accordingly */
1506 pandecode_shader_prop(const char *name
, unsigned claim
, signed truth
, bool fuzzy
)
1512 if (fuzzy
&& (truth
< 0))
1513 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name
, claim
, truth
);
1515 if ((truth
>= 0) && !fuzzy
) {
1516 pandecode_msg("%s: expected %s = %d, claimed %u\n",
1517 (truth
< claim
) ? "warn" : "XXX",
1518 name
, truth
, claim
);
1519 } else if ((claim
> -truth
) && !fuzzy
) {
1520 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
1521 name
, -truth
, claim
);
1522 } else if (fuzzy
&& (claim
< truth
))
1523 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
1524 name
, truth
, claim
);
1526 pandecode_log(".%s = %" PRId16
, name
, claim
);
1529 pandecode_log_cont(" /* %u used */", truth
);
1531 pandecode_log_cont(",\n");
1535 pandecode_blend_shader_disassemble(mali_ptr shader
, int job_no
, int job_type
,
1536 bool is_bifrost
, unsigned gpu_id
)
1538 struct midgard_disasm_stats stats
=
1539 pandecode_shader_disassemble(shader
, job_no
, job_type
, is_bifrost
, gpu_id
);
1541 bool has_texture
= (stats
.texture_count
> 0);
1542 bool has_sampler
= (stats
.sampler_count
> 0);
1543 bool has_attribute
= (stats
.attribute_count
> 0);
1544 bool has_varying
= (stats
.varying_count
> 0);
1545 bool has_uniform
= (stats
.uniform_count
> 0);
1546 bool has_ubo
= (stats
.uniform_buffer_count
> 0);
1548 if (has_texture
|| has_sampler
)
1549 pandecode_msg("XXX: blend shader accessing textures\n");
1551 if (has_attribute
|| has_varying
)
1552 pandecode_msg("XXX: blend shader accessing interstage\n");
1554 if (has_uniform
|| has_ubo
)
1555 pandecode_msg("XXX: blend shader accessing uniforms\n");
1559 pandecode_textures(mali_ptr textures
, unsigned texture_count
, int job_no
, bool is_bifrost
)
1561 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(textures
);
1566 pandecode_log("Textures (%"PRIx64
"):\n", textures
);
1569 const void *cl
= pandecode_fetch_gpu_mem(mmem
,
1570 textures
, MALI_BIFROST_TEXTURE_LENGTH
*
1573 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
1574 pandecode_bifrost_texture(cl
+
1575 MALI_BIFROST_TEXTURE_LENGTH
* tex
,
1579 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
);
1581 for (int tex
= 0; tex
< texture_count
; ++tex
) {
1582 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
1583 char *a
= pointer_as_memory_reference(*u
);
1584 pandecode_log("%s,\n", a
);
1588 /* Now, finally, descend down into the texture descriptor */
1589 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
1590 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
1591 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
1593 pandecode_texture(*u
, tmem
, job_no
, tex
);
1599 pandecode_samplers(mali_ptr samplers
, unsigned sampler_count
, int job_no
, bool is_bifrost
)
1601 for (int i
= 0; i
< sampler_count
; ++i
) {
1603 DUMP_ADDR("Sampler", BIFROST_SAMPLER
, samplers
+ (MALI_BIFROST_SAMPLER_LENGTH
* i
), 1);
1605 DUMP_ADDR("Sampler", MIDGARD_SAMPLER
, samplers
+ (MALI_MIDGARD_SAMPLER_LENGTH
* i
), 1);
1611 pandecode_vertex_tiler_postfix_pre(
1612 const struct mali_vertex_tiler_postfix
*p
,
1613 int job_no
, enum mali_job_type job_type
,
1614 char *suffix
, bool is_bifrost
, unsigned gpu_id
)
1616 struct pandecode_mapped_memory
*attr_mem
;
1618 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1619 * are the only things actually needed from the FBD, vertex/tiler jobs
1620 * no longer reference the FBD -- instead, this field points to some
1621 * info about the scratchpad.
1624 struct pandecode_fbd fbd_info
= {
1625 /* Default for Bifrost */
1630 pandecode_log_cont("\t/* %X %/\n", p
->shared_memory
& 1);
1631 pandecode_compute_fbd(p
->shared_memory
& ~1, job_no
);
1632 } else if (p
->shared_memory
& MALI_MFBD
)
1633 fbd_info
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->shared_memory
) & FBD_MASK
, job_no
, false, job_type
== MALI_JOB_TYPE_COMPUTE
, false);
1634 else if (job_type
== MALI_JOB_TYPE_COMPUTE
)
1635 pandecode_compute_fbd((u64
) (uintptr_t) p
->shared_memory
, job_no
);
1637 fbd_info
= pandecode_sfbd((u64
) (uintptr_t) p
->shared_memory
, job_no
, false, gpu_id
);
1639 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
1640 int texture_count
= 0, sampler_count
= 0;
1643 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->shader
);
1644 uint32_t *cl
= pandecode_fetch_gpu_mem(smem
, p
->shader
, MALI_STATE_LENGTH
);
1646 /* Disassemble ahead-of-time to get stats. Initialize with
1647 * stats for the missing-shader case so we get validation
1650 struct midgard_disasm_stats info
= {
1653 .attribute_count
= 0,
1657 .uniform_count
= -128,
1658 .uniform_buffer_count
= 0
1661 struct MALI_STATE state
;
1662 struct MALI_MIDGARD_PROPERTIES midg_props
;
1663 struct MALI_BIFROST_PROPERTIES bi_props
;
1665 MALI_STATE_unpack((const uint8_t *) cl
, &state
);
1667 if (state
.shader
.shader
& ~0xF)
1668 info
= pandecode_shader_disassemble(state
.shader
.shader
& ~0xF, job_no
, job_type
, is_bifrost
, gpu_id
);
1670 fprintf(pandecode_dump_stream
, "State %"PRIx64
"\n", p
->shader
);
1671 MALI_STATE_print(pandecode_dump_stream
, &state
, 1 * 2);
1673 /* Save for dumps */
1674 attribute_count
= state
.shader
.attribute_count
;
1675 varying_count
= state
.shader
.varying_count
;
1676 texture_count
= state
.shader
.texture_count
;
1677 sampler_count
= state
.shader
.sampler_count
;
1679 fprintf(pandecode_dump_stream
, " Properties\n");
1681 MALI_BIFROST_PROPERTIES_unpack((const uint8_t *) &state
.properties
, &bi_props
);
1682 MALI_BIFROST_PROPERTIES_print(pandecode_dump_stream
, &bi_props
, 2 * 2);
1684 uniform_count
= state
.preload
.uniform_count
;
1685 uniform_buffer_count
= bi_props
.uniform_buffer_count
;
1687 MALI_MIDGARD_PROPERTIES_unpack((const uint8_t *) &state
.properties
, &midg_props
);
1688 MALI_MIDGARD_PROPERTIES_print(pandecode_dump_stream
, &midg_props
, 2 * 2);
1690 uniform_count
= midg_props
.uniform_count
;
1691 uniform_buffer_count
= midg_props
.uniform_buffer_count
;
1694 pandecode_shader_prop("texture_count", texture_count
, info
.texture_count
, false);
1695 pandecode_shader_prop("sampler_count", sampler_count
, info
.sampler_count
, false);
1696 pandecode_shader_prop("attribute_count", attribute_count
, info
.attribute_count
, false);
1697 pandecode_shader_prop("varying_count", varying_count
, info
.varying_count
, false);
1700 uint32_t opaque
= state
.preload
.uniform_count
<< 15
1701 | state
.preload
.untyped
;
1704 case MALI_JOB_TYPE_VERTEX
:
1705 DUMP_CL("Preload", PRELOAD_VERTEX
, &opaque
, 2);
1707 case MALI_JOB_TYPE_TILER
:
1708 DUMP_CL("Preload", PRELOAD_FRAGMENT
, &opaque
, 2);
1710 case MALI_JOB_TYPE_COMPUTE
:
1711 DUMP_CL("Preload", PRELOAD_COMPUTE
, &opaque
, 2);
1714 DUMP_CL("Preload", PRELOAD
, &opaque
, 2);
1720 /* TODO: Blend shaders routing/disasm */
1721 union midgard_blend blend
;
1722 memcpy(&blend
, &state
.sfbd_blend
, sizeof(blend
));
1723 mali_ptr shader
= pandecode_midgard_blend(&blend
, state
.multisample_misc
.sfbd_blend_shader
);
1725 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
1728 /* MRT blend fields are used whenever MFBD is used, with
1729 * per-RT descriptors */
1731 if (job_type
== MALI_JOB_TYPE_TILER
&& (is_bifrost
|| p
->shared_memory
& MALI_MFBD
)) {
1732 void* blend_base
= ((void *) cl
) + MALI_STATE_LENGTH
;
1734 for (unsigned i
= 0; i
< fbd_info
.rt_count
; i
++) {
1735 mali_ptr shader
= 0;
1738 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
1740 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
1743 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
1748 pandecode_msg("XXX: missing shader descriptor\n");
1751 DUMP_ADDR("Viewport", VIEWPORT
, p
->viewport
, 1);
1753 unsigned max_attr_index
= 0;
1755 if (p
->attribute_meta
)
1756 max_attr_index
= pandecode_attribute_meta(attribute_count
, p
->attribute_meta
, false, suffix
);
1758 if (p
->attributes
) {
1759 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
1760 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false, job_type
);
1763 /* Varyings are encoded like attributes but not actually sent; we just
1764 * pass a zero buffer with the right stride/size set, (or whatever)
1765 * since the GPU will write to it itself */
1767 if (p
->varying_meta
) {
1768 varying_count
= pandecode_attribute_meta(varying_count
, p
->varying_meta
, true, suffix
);
1772 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
1774 /* Number of descriptors depends on whether there are
1775 * non-internal varyings */
1777 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true, job_type
);
1780 if (p
->uniform_buffers
) {
1781 if (uniform_buffer_count
)
1782 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
1784 pandecode_msg("warn: UBOs specified but not referenced\n");
1785 } else if (uniform_buffer_count
)
1786 pandecode_msg("XXX: UBOs referenced but not specified\n");
1788 /* We don't want to actually dump uniforms, but we do need to validate
1789 * that the counts we were given are sane */
1793 pandecode_uniforms(p
->uniforms
, uniform_count
);
1795 pandecode_msg("warn: Uniforms specified but not referenced\n");
1796 } else if (uniform_count
)
1797 pandecode_msg("XXX: Uniforms referenced but not specified\n");
1800 pandecode_textures(p
->textures
, texture_count
, job_no
, is_bifrost
);
1802 if (p
->sampler_descriptor
)
1803 pandecode_samplers(p
->sampler_descriptor
, sampler_count
, job_no
, is_bifrost
);
1807 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
1809 pandecode_log(".gl_enables = ");
1811 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
1813 pandecode_log_cont(",\n");
1817 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
1819 if (p
->shader
& 0xF)
1820 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p
->shader
& 0xF));
1822 pandecode_log(".postfix = {\n");
1825 pandecode_gl_enables(p
->gl_enables
, MALI_JOB_TYPE_TILER
);
1826 pandecode_prop("instance_shift = 0x%x", p
->instance_shift
);
1827 pandecode_prop("instance_odd = 0x%x", p
->instance_odd
);
1830 pandecode_msg("XXX: vertex only zero tripped");
1831 pandecode_prop("zero4 = 0x%" PRIx32
, p
->zero4
);
1834 pandecode_prop("offset_start = 0x%x", p
->offset_start
);
1837 pandecode_msg("XXX: vertex only zero tripped");
1838 pandecode_prop("zero5 = 0x%" PRIx64
, p
->zero5
);
1841 MEMORY_PROP(p
, position_varying
);
1842 MEMORY_PROP(p
, occlusion_counter
);
1845 pandecode_log("},\n");
1849 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
1851 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1852 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
1854 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1858 pandecode_msg("XXX: tiler heap zero tripped\n");
1859 pandecode_prop("zero = 0x%x", h
->zero
);
1862 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
1863 MEMORY_PROP(h
, tiler_heap_start
);
1864 MEMORY_PROP(h
, tiler_heap_free
);
1866 /* this might point to the beginning of another buffer, when it's
1867 * really the end of the tiler heap buffer, so we have to be careful
1868 * here. but for zero length, we need the same pointer.
1871 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
1872 MEMORY_PROP(h
, tiler_heap_start
);
1874 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
1875 pandecode_prop("tiler_heap_end = %s + 1", a
);
1879 for (int i
= 0; i
< 10; i
++) {
1880 if (h
->zeros
[i
] != 0) {
1881 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
1886 if (h
->unk1
!= 0x1) {
1887 pandecode_msg("XXX: tiler heap unk1 tripped\n");
1888 pandecode_prop("unk1 = 0x%x", h
->unk1
);
1891 if (h
->unk7e007e
!= 0x7e007e) {
1892 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
1893 pandecode_prop("unk7e007e = 0x%x", h
->unk7e007e
);
1897 pandecode_log("};\n");
1901 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
1903 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1904 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
1906 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
1908 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1911 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32
, t
->tiler_heap_next_start
);
1912 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32
, t
->used_hierarchy_mask
);
1914 if (t
->hierarchy_mask
!= 0xa &&
1915 t
->hierarchy_mask
!= 0x14 &&
1916 t
->hierarchy_mask
!= 0x28 &&
1917 t
->hierarchy_mask
!= 0x50 &&
1918 t
->hierarchy_mask
!= 0xa0)
1919 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
1921 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
1923 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
1925 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
1926 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
1929 pandecode_msg("XXX: tiler meta zero tripped\n");
1930 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
1933 for (int i
= 0; i
< 12; i
++) {
1934 if (t
->zeros
[i
] != 0) {
1935 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
1941 pandecode_log("};\n");
1945 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
1947 if (u
.pointer
== 0x0)
1950 pandecode_log(".primitive_size = {\n");
1954 pandecode_prop("constant = %f", u
.constant
);
1956 MEMORY_PROP((&u
), pointer
);
1960 pandecode_log("},\n");
1964 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
1965 const struct pandecode_mapped_memory
*mem
,
1966 mali_ptr payload
, int job_no
, unsigned gpu_id
)
1968 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
1970 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
1972 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
1975 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
1976 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
1979 pandecode_log("};\n");
1985 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
1986 const struct pandecode_mapped_memory
*mem
,
1987 mali_ptr payload
, int job_no
, unsigned gpu_id
)
1989 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
1991 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
1992 pandecode_tiler_meta(t
->tiler_meta
, job_no
);
1994 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
1997 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
1999 /* TODO: gl_PointSize on Bifrost */
2000 pandecode_primitive_size(t
->primitive_size
, true);
2002 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2004 pandecode_msg("XXX: tiler only zero tripped\n");
2005 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2006 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2007 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2008 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2009 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2010 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2013 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2016 pandecode_log("};\n");
2022 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2023 const struct pandecode_mapped_memory
*mem
,
2024 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2026 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2027 bool is_graphics
= (h
->job_type
== MALI_JOB_TYPE_VERTEX
) || (h
->job_type
== MALI_JOB_TYPE_TILER
);
2029 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false, gpu_id
);
2031 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2034 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, is_graphics
);
2035 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2037 struct MALI_PRIMITIVE primitive
;
2038 struct mali_primitive_packed prim_packed
= v
->prefix
.primitive
;
2039 MALI_PRIMITIVE_unpack((const uint8_t *) &prim_packed
, &primitive
);
2041 pandecode_primitive_size(v
->primitive_size
, primitive
.point_size_array
== 0);
2044 pandecode_log("};\n");
2050 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2051 mali_ptr payload
, int job_no
,
2052 bool is_bifrost
, unsigned gpu_id
)
2054 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2056 bool is_mfbd
= s
->framebuffer
& MALI_MFBD
;
2058 if (!is_mfbd
&& is_bifrost
)
2059 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2061 struct pandecode_fbd info
;
2064 info
= pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true, false, is_bifrost
);
2066 info
= pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true, gpu_id
);
2068 /* Compute the tag for the tagged pointer. This contains the type of
2069 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2070 * additional structures follow the MFBD header (an extra payload or
2071 * not, as well as a count of render targets) */
2073 unsigned expected_tag
= is_mfbd
? MALI_MFBD
: 0;
2077 expected_tag
|= MALI_MFBD_TAG_EXTRA
;
2079 expected_tag
|= (MALI_POSITIVE(info
.rt_count
) << 2);
2082 if ((s
->min_tile_coord
| s
->max_tile_coord
) & ~(MALI_X_COORD_MASK
| MALI_Y_COORD_MASK
)) {
2083 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2084 pandecode_prop("min_tile_coord = 0x%X\n", s
->min_tile_coord
);
2085 pandecode_prop("max_tile_coord = 0x%X\n", s
->max_tile_coord
);
2088 /* Extract tile coordinates */
2090 unsigned min_x
= MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2091 unsigned min_y
= MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2093 unsigned max_x
= (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2094 unsigned max_y
= (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2096 /* For the max, we also want the floored (rather than ceiled) version for checking */
2098 unsigned max_x_f
= (MALI_TILE_COORD_X(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2099 unsigned max_y_f
= (MALI_TILE_COORD_Y(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2101 /* Validate the coordinates are well-ordered */
2104 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2105 else if (min_x
> max_x
)
2106 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2109 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2110 else if (min_y
> max_y
)
2111 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2113 /* Validate the coordinates fit inside the framebuffer. We use floor,
2114 * rather than ceil, for the max coordinates, since the tile
2115 * coordinates for something like an 800x600 framebuffer will actually
2116 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2118 if ((min_x
> info
.width
) || (max_x_f
> info
.width
))
2119 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2121 if ((min_y
> info
.height
) || (max_y_f
> info
.height
))
2122 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2124 /* After validation, we print */
2126 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x
, min_y
, max_x
, max_y
);
2128 /* The FBD is a tagged pointer */
2130 unsigned tag
= (s
->framebuffer
& ~FBD_MASK
);
2132 if (tag
!= expected_tag
)
2133 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag
, tag
);
2138 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
2139 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
2140 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
2141 * might wish to combine this with the bifrost parameter) because some details
2142 * are model-specific even within a particular architecture. Minimal traces
2143 * *only* examine the job descriptors, skipping printing entirely if there is
2144 * no faults, and only descends into the payload if there are faults. This is
2145 * useful for looking for faults without the overhead of invasive traces. */
2148 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
, unsigned gpu_id
, bool minimal
)
2150 pandecode_dump_file_open();
2152 struct mali_job_descriptor_header
*h
;
2153 unsigned job_descriptor_number
= 0;
2156 struct pandecode_mapped_memory
*mem
=
2157 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2161 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2163 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
);
2164 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
, 64);
2166 int job_no
= job_descriptor_number
++;
2168 /* If the job is good to go, skip it in minimal mode */
2169 if (minimal
&& (h
->exception_status
== 0x0 || h
->exception_status
== 0x1))
2172 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2175 pandecode_prop("job_type = %s", mali_job_type_as_str(h
->job_type
));
2177 if (h
->job_descriptor_size
)
2178 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2180 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2181 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2182 h
->exception_status
,
2183 (h
->exception_status
>> 16) & 0xFFFF,
2184 mali_exception_access_as_str((h
->exception_status
>> 8) & 0x3),
2185 h
->exception_status
& 0xFF);
2187 if (h
->first_incomplete_task
)
2188 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2190 if (h
->fault_pointer
)
2191 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2194 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2196 pandecode_prop("job_index = %d", h
->job_index
);
2198 if (h
->unknown_flags
)
2199 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2201 if (h
->job_dependency_index_1
)
2202 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2204 if (h
->job_dependency_index_2
)
2205 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2208 pandecode_log("};\n");
2210 switch (h
->job_type
) {
2211 case MALI_JOB_TYPE_WRITE_VALUE
: {
2212 struct mali_payload_write_value
*s
= payload
;
2213 pandecode_log("struct mali_payload_write_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2215 MEMORY_PROP(s
, address
);
2217 if (s
->value_descriptor
!= MALI_WRITE_VALUE_ZERO
) {
2218 pandecode_msg("XXX: unknown value descriptor\n");
2219 pandecode_prop("value_descriptor = 0x%" PRIX32
, s
->value_descriptor
);
2223 pandecode_msg("XXX: set value tripped\n");
2224 pandecode_prop("reserved = 0x%" PRIX32
, s
->reserved
);
2227 pandecode_prop("immediate = 0x%" PRIX64
, s
->immediate
);
2229 pandecode_log("};\n");
2234 case MALI_JOB_TYPE_TILER
:
2235 case MALI_JOB_TYPE_VERTEX
:
2236 case MALI_JOB_TYPE_COMPUTE
:
2238 if (h
->job_type
== MALI_JOB_TYPE_TILER
)
2239 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2241 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2243 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2247 case MALI_JOB_TYPE_FRAGMENT
:
2248 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
, gpu_id
);
2254 } while ((jc_gpu_va
= h
->next_job
));
2256 pandecode_map_read_write();