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_MASK_##flag, "MALI_MASK_" #flag }
232 static const struct pandecode_flag_info mask_flag_info
[] = {
241 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
242 static const struct pandecode_flag_info u3_flag_info
[] = {
244 FLAG_INFO(PER_SAMPLE
),
245 FLAG_INFO(CAN_DISCARD
),
246 FLAG_INFO(HAS_BLEND_SHADER
),
247 FLAG_INFO(DEPTH_WRITEMASK
),
248 FLAG_INFO(DEPTH_CLIP_NEAR
),
249 FLAG_INFO(DEPTH_CLIP_FAR
),
253 static const struct pandecode_flag_info u4_flag_info
[] = {
255 FLAG_INFO(NO_DITHER
),
256 FLAG_INFO(DEPTH_RANGE_A
),
257 FLAG_INFO(DEPTH_RANGE_B
),
258 FLAG_INFO(STENCIL_TEST
),
259 FLAG_INFO(ALPHA_TO_COVERAGE
),
264 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
265 static const struct pandecode_flag_info mfbd_fmt_flag_info
[] = {
271 #define FLAG_INFO(flag) { MALI_AFBC_##flag, "MALI_AFBC_" #flag }
272 static const struct pandecode_flag_info afbc_fmt_flag_info
[] = {
278 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
279 static const struct pandecode_flag_info mfbd_extra_flag_hi_info
[] = {
285 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
286 static const struct pandecode_flag_info mfbd_extra_flag_lo_info
[] = {
292 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
293 static const struct pandecode_flag_info shader_midgard1_flag_lo_info
[] = {
296 FLAG_INFO(READS_TILEBUFFER
),
297 FLAG_INFO(WRITES_GLOBAL
),
302 static const struct pandecode_flag_info shader_midgard1_flag_hi_info
[] = {
304 FLAG_INFO(SUPPRESS_INF_NAN
),
309 #define FLAG_INFO(flag) { MALI_BIFROST_##flag, "MALI_BIFROST_" #flag }
310 static const struct pandecode_flag_info shader_bifrost_info
[] = {
311 FLAG_INFO(FULL_THREAD
),
313 FLAG_INFO(FIRST_ATEST
),
319 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
320 static const struct pandecode_flag_info mfbd_flag_info
[] = {
321 FLAG_INFO(DEPTH_WRITE
),
327 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
328 static const struct pandecode_flag_info sfbd_unk1_info
[] = {
335 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
336 static const struct pandecode_flag_info sfbd_unk2_info
[] = {
343 /* Midgard's tiler descriptor is embedded within the
347 pandecode_midgard_tiler_descriptor(
348 const struct midgard_tiler_descriptor
*t
,
354 pandecode_log(".tiler = {\n");
357 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
358 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
360 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
362 /* We know this name from the kernel, but we never see it nonzero */
365 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
367 MEMORY_PROP(t
, polygon_list
);
369 /* The body is offset from the base of the polygon list */
370 //assert(t->polygon_list_body > t->polygon_list);
371 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
373 /* It needs to fit inside the reported size */
374 //assert(t->polygon_list_size >= body_offset);
376 /* Now that we've sanity checked, we'll try to calculate the sizes
377 * ourselves for comparison */
379 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
380 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
382 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
383 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
384 ref_header
, ref_size
);
385 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
386 pandecode_msg("body offset %d\n", body_offset
);
389 /* The tiler heap has a start and end specified -- it should be
390 * identical to what we have in the BO. The exception is if tiling is
393 MEMORY_PROP(t
, heap_start
);
394 assert(t
->heap_end
>= t
->heap_start
);
396 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
398 /* Tiling is enabled with a special flag */
399 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
400 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
402 bool tiling_enabled
= hierarchy_mask
;
404 if (tiling_enabled
) {
405 /* We should also have no other flags */
407 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
409 /* When tiling is disabled, we should have that flag and no others */
411 if (tiler_flags
!= MALI_TILER_DISABLED
) {
412 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
416 /* We should also have an empty heap */
418 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
422 /* Disabled tiling is used only for clear-only jobs, which are
423 * purely FRAGMENT, so we should never see this for
424 * non-FRAGMENT descriptors. */
427 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
430 /* We've never seen weights used in practice, but we know from the
431 * kernel these fields is there */
433 bool nonzero_weights
= false;
435 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
436 nonzero_weights
|= t
->weights
[w
] != 0x0;
439 if (nonzero_weights
) {
440 pandecode_log(".weights = { ");
442 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
443 pandecode_log_cont("%d, ", t
->weights
[w
]);
450 pandecode_log("}\n");
453 /* TODO: The Bifrost tiler is not understood at all yet */
456 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer
*fb
)
458 pandecode_log(".tiler = {\n");
461 MEMORY_PROP(fb
, tiler_meta
);
463 for (int i
= 0; i
< 16; i
++) {
464 if (fb
->zeros
[i
] != 0) {
465 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
470 pandecode_log("},\n");
473 pandecode_log("}\n");
477 /* Information about the framebuffer passed back for
478 * additional analysis */
480 struct pandecode_fbd
{
488 pandecode_sfbd_format(struct mali_sfbd_format format
)
490 pandecode_log(".format = {\n");
493 pandecode_log(".unk1 = ");
494 pandecode_log_decoded_flags(sfbd_unk1_info
, format
.unk1
);
495 pandecode_log_cont(",\n");
497 /* TODO: Map formats so we can check swizzles and print nicely */
498 pandecode_log("swizzle");
499 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
500 pandecode_log_cont(",\n");
502 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
503 (format
.nr_channels
+ 1));
505 pandecode_log(".unk2 = ");
506 pandecode_log_decoded_flags(sfbd_unk2_info
, format
.unk2
);
507 pandecode_log_cont(",\n");
509 pandecode_prop("block = %s", mali_block_format_as_str(format
.block
));
511 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
514 pandecode_log("},\n");
518 pandecode_shared_memory(const struct mali_shared_memory
*desc
, bool is_compute
)
520 pandecode_prop("stack_shift = 0x%x", desc
->stack_shift
);
523 pandecode_prop("unk0 = 0x%x", desc
->unk0
);
525 if (desc
->shared_workgroup_count
!= 0x1F) {
526 pandecode_prop("shared_workgroup_count = %d", desc
->shared_workgroup_count
);
528 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
531 if (desc
->shared_unk1
|| desc
->shared_shift
) {
532 pandecode_prop("shared_unk1 = %X", desc
->shared_unk1
);
533 pandecode_prop("shared_shift = %X", desc
->shared_shift
);
536 pandecode_msg("XXX: shared memory configured in noncompute shader");
539 if (desc
->shared_zero
) {
540 pandecode_msg("XXX: shared memory zero tripped\n");
541 pandecode_prop("shared_zero = 0x%" PRIx32
, desc
->shared_zero
);
544 if (desc
->shared_memory
&& !is_compute
)
545 pandecode_msg("XXX: shared memory used in noncompute shader\n");
547 MEMORY_PROP(desc
, scratchpad
);
548 MEMORY_PROP(desc
, shared_memory
);
549 MEMORY_PROP(desc
, unknown1
);
551 if (desc
->scratchpad
) {
552 struct pandecode_mapped_memory
*smem
=
553 pandecode_find_mapped_gpu_mem_containing(desc
->scratchpad
);
555 pandecode_msg("scratchpad size %u\n", smem
->length
);
560 static struct pandecode_fbd
561 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
, unsigned gpu_id
)
563 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
564 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
566 struct pandecode_fbd info
= {
571 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
574 pandecode_log(".shared_memory = {\n");
576 pandecode_shared_memory(&s
->shared_memory
, false);
578 pandecode_log("},\n");
580 pandecode_sfbd_format(s
->format
);
582 info
.width
= s
->width
+ 1;
583 info
.height
= s
->height
+ 1;
585 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", info
.width
);
586 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", info
.height
);
588 MEMORY_PROP(s
, checksum
);
590 if (s
->checksum_stride
)
591 pandecode_prop("checksum_stride = %d", s
->checksum_stride
);
593 MEMORY_PROP(s
, framebuffer
);
594 pandecode_prop("stride = %d", s
->stride
);
596 /* Earlier in the actual commandstream -- right before width -- but we
597 * delay to flow nicer */
599 pandecode_log(".clear_flags = ");
600 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
601 pandecode_log_cont(",\n");
603 if (s
->depth_buffer
) {
604 MEMORY_PROP(s
, depth_buffer
);
605 pandecode_prop("depth_stride = %d", s
->depth_stride
);
608 if (s
->stencil_buffer
) {
609 MEMORY_PROP(s
, stencil_buffer
);
610 pandecode_prop("stencil_stride = %d", s
->stencil_stride
);
613 if (s
->depth_stride_zero
||
614 s
->stencil_stride_zero
||
615 s
->zero7
|| s
->zero8
) {
616 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
617 pandecode_prop("depth_stride_zero = 0x%x",
618 s
->depth_stride_zero
);
619 pandecode_prop("stencil_stride_zero = 0x%x",
620 s
->stencil_stride_zero
);
621 pandecode_prop("zero7 = 0x%" PRIx32
,
623 pandecode_prop("zero8 = 0x%" PRIx32
,
627 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
628 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
629 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
630 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
631 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
634 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
635 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
636 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
637 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
638 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
641 if (s
->clear_stencil
) {
642 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
645 const struct midgard_tiler_descriptor t
= s
->tiler
;
647 bool has_hierarchy
= !(gpu_id
== 0x0720 || gpu_id
== 0x0820 || gpu_id
== 0x0830);
648 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
, has_hierarchy
);
651 pandecode_log("};\n");
653 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
654 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
655 pandecode_prop("zero5 = 0x%" PRIx32
, s
->zero5
);
657 pandecode_log_cont(".zero3 = {");
659 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
660 pandecode_log_cont("%X, ", s
->zero3
[i
]);
662 pandecode_log_cont("},\n");
664 pandecode_log_cont(".zero6 = {");
666 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
667 pandecode_log_cont("%X, ", s
->zero6
[i
]);
669 pandecode_log_cont("},\n");
675 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
677 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
678 const struct mali_shared_memory
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
680 pandecode_log("struct mali_shared_memory shared_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
682 pandecode_shared_memory(s
, true);
684 pandecode_log("},\n");
687 /* Extracts the number of components associated with a Mali format */
690 pandecode_format_component_count(enum mali_format fmt
)
692 /* Mask out the format class */
693 unsigned top
= fmt
& 0b11100000;
696 case MALI_FORMAT_SNORM
:
697 case MALI_FORMAT_UINT
:
698 case MALI_FORMAT_UNORM
:
699 case MALI_FORMAT_SINT
:
700 return ((fmt
>> 3) & 3) + 1;
707 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
710 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
713 assert(MALI_CHANNEL_R
== 0);
715 for (unsigned c
= 0; c
< 4; ++c
) {
716 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
718 if (chan
<= MALI_CHANNEL_A
)
725 /* Validates that a (format, swizzle) pair is valid, in the sense that the
726 * swizzle doesn't access any components that are undefined in the format.
727 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
731 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
733 unsigned nr_comp
= pandecode_format_component_count(fmt
);
734 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
735 unsigned valid_mask
= (1 << nr_comp
) - 1;
736 unsigned invalid_mask
= ~valid_mask
;
738 if (access_mask
& invalid_mask
) {
739 pandecode_msg("XXX: invalid components accessed\n");
743 /* Check for the default non-swizzling swizzle so we can suppress
744 * useless printing for the defaults */
746 unsigned default_swizzles
[4] = {
747 MALI_CHANNEL_R
| (MALI_CHANNEL_0
<< 3) | (MALI_CHANNEL_0
<< 6) | (MALI_CHANNEL_1
<< 9),
748 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_0
<< 6) | (MALI_CHANNEL_1
<< 9),
749 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_B
<< 6) | (MALI_CHANNEL_1
<< 9),
750 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_B
<< 6) | (MALI_CHANNEL_A
<< 9)
753 return (swizzle
== default_swizzles
[nr_comp
- 1]);
757 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
759 /* First, do some validation */
760 bool trivial_swizzle
= pandecode_validate_format_swizzle(
766 /* Next, print the swizzle */
767 pandecode_log_cont(".");
769 static const char components
[] = "rgba01";
771 for (unsigned c
= 0; c
< 4; ++c
) {
772 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
774 if (chan
> MALI_CHANNEL_1
) {
775 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
778 pandecode_log_cont("%c", components
[chan
]);
783 pandecode_rt_format(struct mali_rt_format format
)
785 pandecode_log(".format = {\n");
788 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
789 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
790 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
791 pandecode_prop("unk4 = 0x%" PRIx32
, format
.unk4
);
793 pandecode_prop("block = %s", mali_block_format_as_str(format
.block
));
795 /* TODO: Map formats so we can check swizzles and print nicely */
796 pandecode_log("swizzle");
797 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
798 pandecode_log_cont(",\n");
800 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
801 (format
.nr_channels
+ 1));
803 pandecode_log(".flags = ");
804 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
805 pandecode_log_cont(",\n");
807 pandecode_prop("msaa = %s", mali_msaa_as_str(format
.msaa
));
809 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
810 * which is a faster hardware-based alternative to the wallpaper method
811 * to preserve framebuffer contents across frames. In practice, MFBD
812 * preload is buggy on Midgard, and so this is a chicken bit. If this
813 * bit isn't set, most likely something broke unrelated to preload */
815 if (!format
.no_preload
) {
816 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
817 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
821 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
824 pandecode_log("},\n");
828 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct mali_framebuffer
*fb
)
830 pandecode_log("struct mali_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
833 for (int i
= 0; i
< (fb
->rt_count_1
+ 1); i
++) {
834 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct mali_render_target
);
835 struct pandecode_mapped_memory
*mem
=
836 pandecode_find_mapped_gpu_mem_containing(rt_va
);
837 const struct mali_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
839 pandecode_log("{\n");
842 pandecode_rt_format(rt
->format
);
844 if (rt
->format
.block
== MALI_BLOCK_FORMAT_AFBC
) {
845 pandecode_log(".afbc = {\n");
848 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
849 pandecode_prop("metadata = %s", a
);
852 pandecode_prop("stride = %d", rt
->afbc
.stride
);
854 pandecode_log(".flags = ");
855 pandecode_log_decoded_flags(afbc_fmt_flag_info
, rt
->afbc
.flags
);
856 pandecode_log_cont(",\n");
859 pandecode_log("},\n");
860 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.flags
) {
861 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
867 MEMORY_PROP(rt
, framebuffer
);
868 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
870 if (rt
->layer_stride
)
871 pandecode_prop("layer_stride = %d", rt
->layer_stride
);
873 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
874 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
875 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
876 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
877 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
880 if (rt
->zero1
|| rt
->zero2
) {
881 pandecode_msg("XXX: render target zeros tripped\n");
882 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
883 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
887 pandecode_log("},\n");
891 pandecode_log("};\n");
894 static struct pandecode_fbd
895 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
, bool is_compute
, bool is_bifrost
)
897 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
898 const struct mali_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
900 struct pandecode_fbd info
;
902 if (is_bifrost
&& fb
->msaa
.sample_locations
) {
903 /* The blob stores all possible sample locations in a single buffer
904 * allocated on startup, and just switches the pointer when switching
905 * MSAA state. For now, we just put the data into the cmdstream, but we
906 * should do something like what the blob does with a real driver.
908 * There seem to be 32 slots for sample locations, followed by another
909 * 16. The second 16 is just the center location followed by 15 zeros
910 * in all the cases I've identified (maybe shader vs. depth/color
914 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->msaa
.sample_locations
);
916 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->msaa
.sample_locations
);
918 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
921 for (int i
= 0; i
< 32 + 16; i
++) {
922 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
926 pandecode_log("};\n");
929 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
933 pandecode_log(".msaa = {\n");
936 if (fb
->msaa
.sample_locations
)
937 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
939 pandecode_msg("XXX: sample_locations missing\n");
941 if (fb
->msaa
.zero1
|| fb
->msaa
.zero2
|| fb
->msaa
.zero4
) {
942 pandecode_msg("XXX: multisampling zero tripped\n");
943 pandecode_prop("zero1 = %" PRIx64
, fb
->msaa
.zero1
);
944 pandecode_prop("zero2 = %" PRIx64
, fb
->msaa
.zero2
);
945 pandecode_prop("zero4 = %" PRIx64
, fb
->msaa
.zero4
);
949 pandecode_log("},\n");
951 pandecode_log(".shared_memory = {\n");
953 pandecode_shared_memory(&fb
->shared_memory
, is_compute
);
955 pandecode_log("},\n");
958 info
.width
= fb
->width1
+ 1;
959 info
.height
= fb
->height1
+ 1;
960 info
.rt_count
= fb
->rt_count_1
+ 1;
962 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
963 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
964 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
965 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
967 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
968 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
969 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
970 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
972 pandecode_log(".mfbd_flags = ");
973 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
974 pandecode_log_cont(",\n");
976 if (fb
->clear_stencil
)
977 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
980 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
984 pandecode_bifrost_tiler_descriptor(fb
);
986 const struct midgard_tiler_descriptor t
= fb
->tiler
;
987 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
, true);
990 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
992 if (fb
->zero3
|| fb
->zero4
) {
993 pandecode_msg("XXX: framebuffer zeros tripped\n");
994 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
995 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
999 pandecode_log("};\n");
1001 gpu_va
+= sizeof(struct mali_framebuffer
);
1003 info
.has_extra
= (fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
;
1005 if (info
.has_extra
) {
1006 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1007 const struct mali_framebuffer_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
1009 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1012 MEMORY_PROP(fbx
, checksum
);
1014 if (fbx
->checksum_stride
)
1015 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
1017 pandecode_log(".flags_hi = ");
1018 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info
, fbx
->flags_hi
);
1019 pandecode_log_cont(",\n");
1021 pandecode_log(".flags_lo = ");
1022 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info
, fbx
->flags_lo
);
1023 pandecode_log_cont(",\n");
1025 pandecode_prop("zs_block = %s", mali_block_format_as_str(fbx
->zs_block
));
1026 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx
->zs_samples
+ 1);
1028 if (fbx
->zs_block
== MALI_BLOCK_FORMAT_AFBC
) {
1029 pandecode_log(".ds_afbc = {\n");
1032 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
1033 pandecode_prop("depth_stencil_afbc_stride = %d",
1034 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
1035 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
1037 pandecode_log(".flags = ");
1038 pandecode_log_decoded_flags(afbc_fmt_flag_info
, fbx
->ds_afbc
.flags
);
1039 pandecode_log_cont(",\n");
1041 if (fbx
->ds_afbc
.padding
) {
1042 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1043 pandecode_prop("padding = 0x%" PRIx64
, fbx
->ds_afbc
.padding
);
1047 pandecode_log("},\n");
1049 pandecode_log(".ds_linear = {\n");
1052 if (fbx
->ds_linear
.depth
) {
1053 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
1054 pandecode_prop("depth_stride = %d",
1055 fbx
->ds_linear
.depth_stride
);
1056 pandecode_prop("depth_layer_stride = %d",
1057 fbx
->ds_linear
.depth_layer_stride
);
1058 } else if (fbx
->ds_linear
.depth_stride
|| fbx
->ds_linear
.depth_layer_stride
) {
1059 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx
->ds_linear
.depth_stride
, fbx
->ds_linear
.depth_layer_stride
);
1062 if (fbx
->ds_linear
.stencil
) {
1063 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1064 pandecode_prop("stencil_stride = %d",
1065 fbx
->ds_linear
.stencil_stride
);
1066 pandecode_prop("stencil_layer_stride = %d",
1067 fbx
->ds_linear
.stencil_layer_stride
);
1068 } else if (fbx
->ds_linear
.stencil_stride
|| fbx
->ds_linear
.stencil_layer_stride
) {
1069 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx
->ds_linear
.stencil_stride
, fbx
->ds_linear
.stencil_layer_stride
);
1072 if (fbx
->ds_linear
.depth_stride_zero
||
1073 fbx
->ds_linear
.stencil_stride_zero
) {
1074 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1075 pandecode_prop("depth_stride_zero = 0x%x",
1076 fbx
->ds_linear
.depth_stride_zero
);
1077 pandecode_prop("stencil_stride_zero = 0x%x",
1078 fbx
->ds_linear
.stencil_stride_zero
);
1082 pandecode_log("},\n");
1085 if (fbx
->clear_color_1
| fbx
->clear_color_2
) {
1086 pandecode_prop("clear_color_1 = 0x%" PRIx32
, fbx
->clear_color_1
);
1087 pandecode_prop("clear_color_2 = 0x%" PRIx32
, fbx
->clear_color_2
);
1091 pandecode_msg("XXX: fb_extra zeros tripped\n");
1092 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1096 pandecode_log("};\n");
1098 gpu_va
+= sizeof(struct mali_framebuffer_extra
);
1102 pandecode_render_target(gpu_va
, job_no
, fb
);
1108 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1109 mali_ptr addr
, int job_no
, char *suffix
,
1110 int count
, bool varying
, enum mali_job_type job_type
)
1112 char *prefix
= varying
? "Varying" : "Attribute";
1116 pandecode_msg("warn: No %s records\n", prefix
);
1120 MAP_ADDR(ATTRIBUTE_BUFFER
, addr
, cl
);
1122 for (int i
= 0; i
< count
; ++i
) {
1123 fprintf(pandecode_dump_stream
, "%s\n", prefix
);
1125 struct MALI_ATTRIBUTE_BUFFER temp
;
1126 MALI_ATTRIBUTE_BUFFER_unpack(cl
+ i
* MALI_ATTRIBUTE_BUFFER_LENGTH
, &temp
);
1127 MALI_ATTRIBUTE_BUFFER_print(pandecode_dump_stream
, &temp
, 2);
1129 if (temp
.type
== MALI_ATTRIBUTE_TYPE_1D_NPOT_DIVISOR
) {
1130 struct MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT temp2
;
1131 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_unpack(cl
+ (i
+ 1) * MALI_ATTRIBUTE_BUFFER_LENGTH
, &temp2
);
1132 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_print(pandecode_dump_stream
, &temp2
, 2);
1138 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1140 /* TODO: Decode flags */
1141 mali_ptr shader_ptr
= ptr
& ~15;
1143 char *a
= pointer_as_memory_reference(shader_ptr
);
1144 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1151 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1154 pandecode_msg("XXX: blend zero tripped: %X\n", blend
->zero1
);
1156 pandecode_log(".equation = {\n");
1159 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1160 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1162 pandecode_log(".color_mask = ");
1163 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1164 pandecode_log_cont(",\n");
1167 pandecode_log("},\n");
1170 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1173 decode_bifrost_constant(u16 constant
)
1175 float lo
= (float) (constant
& 0xFF);
1176 float hi
= (float) (constant
>> 8);
1178 return (hi
/ 255.0) + (lo
/ 65535.0);
1182 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1184 struct bifrost_blend_rt
*b
=
1185 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1187 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1190 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1191 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1192 b
->constant
, decode_bifrost_constant(b
->constant
));
1194 /* TODO figure out blend shader enable bit */
1195 pandecode_blend_equation(&b
->equation
);
1197 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1198 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1200 pandecode_log(".format = %s", mali_format_as_str(b
->format
));
1201 pandecode_swizzle(b
->swizzle
, b
->format
);
1202 pandecode_log_cont(",\n");
1204 pandecode_prop("swizzle = 0x%" PRIx32
, b
->swizzle
);
1205 pandecode_prop("format = 0x%" PRIx32
, b
->format
);
1208 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1209 pandecode_prop("zero1 = 0x%" PRIx32
, b
->zero1
);
1212 pandecode_log(".shader_type = ");
1213 switch(b
->shader_type
) {
1214 case BIFROST_BLEND_F16
:
1215 pandecode_log_cont("BIFROST_BLEND_F16");
1217 case BIFROST_BLEND_F32
:
1218 pandecode_log_cont("BIFROST_BLEND_F32");
1220 case BIFROST_BLEND_I32
:
1221 pandecode_log_cont("BIFROST_BLEND_I32");
1223 case BIFROST_BLEND_U32
:
1224 pandecode_log_cont("BIFROST_BLEND_U32");
1226 case BIFROST_BLEND_I16
:
1227 pandecode_log_cont("BIFROST_BLEND_I16");
1229 case BIFROST_BLEND_U16
:
1230 pandecode_log_cont("BIFROST_BLEND_U16");
1233 pandecode_log_cont(",\n");
1236 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1237 pandecode_prop("zero2 = 0x%" PRIx32
, b
->zero2
);
1240 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1243 pandecode_log("},\n");
1249 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1251 /* constant/equation is in a union */
1255 pandecode_log(".blend = {\n");
1259 pandecode_shader_address("shader", blend
->shader
);
1261 pandecode_blend_equation(&blend
->equation
);
1262 pandecode_prop("constant = %f", blend
->constant
);
1266 pandecode_log("},\n");
1268 /* Return blend shader to disassemble if present */
1269 return is_shader
? (blend
->shader
& ~0xF) : 0;
1273 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1275 struct midgard_blend_rt
*b
=
1276 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1278 /* Flags determine presence of blend shader */
1279 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1281 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1284 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1286 union midgard_blend blend
= b
->blend
;
1287 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1290 pandecode_log("};\n");
1295 /* Attributes and varyings have descriptor records, which contain information
1296 * about their format and ordering with the attribute/varying buffers. We'll
1297 * want to validate that the combinations specified are self-consistent.
1301 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1303 const char *prefix
= varying
? "Varying" : "Attribute";
1304 mali_ptr p
= varying
? v
->varying_meta
: v
->attribute_meta
;
1306 for (int i
= 0; i
< count
; ++i
, p
+= MALI_ATTRIBUTE_LENGTH
)
1307 DUMP_ADDR(prefix
, ATTRIBUTE
, p
, 1);
1312 /* return bits [lo, hi) of word */
1314 bits(u32 word
, u32 lo
, u32 hi
)
1317 return word
; // avoid undefined behavior with the shift
1319 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1323 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool graphics
)
1325 pandecode_log(".prefix = {\n");
1328 /* Decode invocation_count. See the comment before the definition of
1329 * invocation_count for an explanation.
1332 unsigned size_y_shift
= bits(p
->invocation_shifts
, 0, 5);
1333 unsigned size_z_shift
= bits(p
->invocation_shifts
, 5, 10);
1334 unsigned workgroups_x_shift
= bits(p
->invocation_shifts
, 10, 16);
1335 unsigned workgroups_y_shift
= bits(p
->invocation_shifts
, 16, 22);
1336 unsigned workgroups_z_shift
= bits(p
->invocation_shifts
, 22, 28);
1337 unsigned workgroups_x_shift_2
= bits(p
->invocation_shifts
, 28, 32);
1339 unsigned size_x
= bits(p
->invocation_count
, 0, size_y_shift
) + 1;
1340 unsigned size_y
= bits(p
->invocation_count
, size_y_shift
, size_z_shift
) + 1;
1341 unsigned size_z
= bits(p
->invocation_count
, size_z_shift
, workgroups_x_shift
) + 1;
1343 unsigned groups_x
= bits(p
->invocation_count
, workgroups_x_shift
, workgroups_y_shift
) + 1;
1344 unsigned groups_y
= bits(p
->invocation_count
, workgroups_y_shift
, workgroups_z_shift
) + 1;
1345 unsigned groups_z
= bits(p
->invocation_count
, workgroups_z_shift
, 32) + 1;
1347 /* Even though we have this decoded, we want to ensure that the
1348 * representation is "unique" so we don't lose anything by printing only
1349 * the final result. More specifically, we need to check that we were
1350 * passed something in canonical form, since the definition per the
1351 * hardware is inherently not unique. How? Well, take the resulting
1352 * decode and pack it ourselves! If it is bit exact with what we
1353 * decoded, we're good to go. */
1355 struct mali_vertex_tiler_prefix ref
;
1356 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, graphics
);
1359 (p
->invocation_count
== ref
.invocation_count
) &&
1360 (p
->invocation_shifts
== ref
.invocation_shifts
);
1363 pandecode_msg("XXX: non-canonical workgroups packing\n");
1364 pandecode_msg("expected: %X, %X",
1365 ref
.invocation_count
,
1366 ref
.invocation_shifts
);
1368 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1369 pandecode_prop("size_y_shift = %d", size_y_shift
);
1370 pandecode_prop("size_z_shift = %d", size_z_shift
);
1371 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift
);
1372 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift
);
1373 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift
);
1374 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2
);
1377 /* Regardless, print the decode */
1378 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1379 size_x
, size_y
, size_z
,
1380 groups_x
, groups_y
, groups_z
);
1383 if (p
->unknown_draw
)
1384 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1386 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1388 if (p
->draw_mode
!= MALI_DRAW_MODE_NONE
)
1389 pandecode_prop("draw_mode = %s", mali_draw_mode_as_str(p
->draw_mode
));
1391 /* Index count only exists for tiler jobs anyway */
1394 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1397 unsigned index_raw_size
= (p
->unknown_draw
& MALI_DRAW_INDEXED_SIZE
);
1398 index_raw_size
>>= MALI_DRAW_INDEXED_SHIFT
;
1400 /* Validate an index buffer is present if we need one. TODO: verify
1401 * relationship between invocation_count and index_count */
1404 unsigned count
= p
->index_count
;
1407 unsigned size
= (index_raw_size
== 0x3) ? 4 : index_raw_size
;
1409 /* Ensure we got a size, and if so, validate the index buffer
1410 * is large enough to hold a full set of indices of the given
1413 if (!index_raw_size
)
1414 pandecode_msg("XXX: index size missing\n");
1416 pandecode_validate_buffer(p
->indices
, count
* size
);
1417 } else if (index_raw_size
)
1418 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size
);
1420 if (p
->offset_bias_correction
)
1421 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1423 /* TODO: Figure out what this is. It's not zero */
1424 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1427 pandecode_log("},\n");
1431 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1433 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1434 uint64_t *PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1436 for (int i
= 0; i
< ubufs_count
; i
++) {
1437 unsigned size
= (ubufs
[i
] & ((1 << 10) - 1)) * 16;
1438 mali_ptr addr
= (ubufs
[i
] >> 10) << 2;
1440 pandecode_validate_buffer(addr
, size
);
1442 char *ptr
= pointer_as_memory_reference(addr
);
1443 pandecode_log("ubuf_%d[%u] = %s;\n", i
, size
, ptr
);
1447 pandecode_log("\n");
1451 pandecode_uniforms(mali_ptr uniforms
, unsigned uniform_count
)
1453 pandecode_validate_buffer(uniforms
, uniform_count
* 16);
1455 char *ptr
= pointer_as_memory_reference(uniforms
);
1456 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count
, ptr
);
1461 shader_type_for_job(unsigned type
)
1464 case MALI_JOB_TYPE_VERTEX
: return "VERTEX";
1465 case MALI_JOB_TYPE_TILER
: return "FRAGMENT";
1466 case MALI_JOB_TYPE_COMPUTE
: return "COMPUTE";
1472 static unsigned shader_id
= 0;
1474 static struct midgard_disasm_stats
1475 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1476 bool is_bifrost
, unsigned gpu_id
)
1478 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1479 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1481 /* Compute maximum possible size */
1482 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1484 /* Print some boilerplate to clearly denote the assembly (which doesn't
1485 * obey indentation rules), and actually do the disassembly! */
1487 pandecode_log_cont("\n\n");
1489 struct midgard_disasm_stats stats
;
1492 disassemble_bifrost(pandecode_dump_stream
, code
, sz
, true);
1494 /* TODO: Extend stats to Bifrost */
1495 stats
.texture_count
= -128;
1496 stats
.sampler_count
= -128;
1497 stats
.attribute_count
= -128;
1498 stats
.varying_count
= -128;
1499 stats
.uniform_count
= -128;
1500 stats
.uniform_buffer_count
= -128;
1501 stats
.work_count
= -128;
1503 stats
.instruction_count
= 0;
1504 stats
.bundle_count
= 0;
1505 stats
.quadword_count
= 0;
1506 stats
.helper_invocations
= false;
1508 stats
= disassemble_midgard(pandecode_dump_stream
,
1510 type
== MALI_JOB_TYPE_TILER
?
1511 MESA_SHADER_FRAGMENT
: MESA_SHADER_VERTEX
);
1514 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1515 * driver-internal purposes with the blob and interfere */
1517 bool should_shaderdb
= type
!= MALI_JOB_TYPE_COMPUTE
;
1519 if (should_shaderdb
) {
1520 unsigned nr_threads
=
1521 (stats
.work_count
<= 4) ? 4 :
1522 (stats
.work_count
<= 8) ? 2 :
1525 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
1526 "%u inst, %u bundles, %u quadwords, "
1527 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1529 shader_type_for_job(type
),
1530 stats
.instruction_count
, stats
.bundle_count
, stats
.quadword_count
,
1531 stats
.work_count
, nr_threads
);
1539 pandecode_texture_payload(mali_ptr payload
,
1540 enum mali_texture_dimension dim
,
1541 enum mali_texture_layout layout
,
1545 uint16_t array_size
,
1546 struct pandecode_mapped_memory
*tmem
)
1548 pandecode_log(".payload = {\n");
1551 /* A bunch of bitmap pointers follow.
1552 * We work out the correct number,
1553 * based on the mipmap/cubemap
1554 * properties, but dump extra
1555 * possibilities to futureproof */
1557 int bitmap_count
= levels
+ 1;
1559 /* Miptree for each face */
1560 if (dim
== MALI_TEXTURE_DIMENSION_CUBE
)
1563 /* Array of layers */
1564 bitmap_count
*= depth
;
1566 /* Array of textures */
1567 bitmap_count
*= array_size
;
1569 /* Stride for each element */
1573 mali_ptr
*pointers_and_strides
= pandecode_fetch_gpu_mem(tmem
,
1574 payload
, sizeof(mali_ptr
) * bitmap_count
);
1575 for (int i
= 0; i
< bitmap_count
; ++i
) {
1576 /* How we dump depends if this is a stride or a pointer */
1578 if (manual_stride
&& (i
& 1)) {
1579 /* signed 32-bit snuck in as a 64-bit pointer */
1580 uint64_t stride_set
= pointers_and_strides
[i
];
1581 uint32_t clamped_stride
= stride_set
;
1582 int32_t stride
= clamped_stride
;
1583 assert(stride_set
== clamped_stride
);
1584 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
1586 char *a
= pointer_as_memory_reference(pointers_and_strides
[i
]);
1587 pandecode_log("%s, \n", a
);
1593 pandecode_log("},\n");
1597 pandecode_texture(mali_ptr u
,
1598 struct pandecode_mapped_memory
*tmem
,
1599 unsigned job_no
, unsigned tex
)
1601 struct pandecode_mapped_memory
*mapped_mem
= pandecode_find_mapped_gpu_mem_containing(u
);
1602 const uint8_t *cl
= pandecode_fetch_gpu_mem(mapped_mem
, u
, MALI_MIDGARD_TEXTURE_LENGTH
);
1604 struct MALI_MIDGARD_TEXTURE temp
;
1605 MALI_MIDGARD_TEXTURE_unpack(cl
, &temp
);
1606 MALI_MIDGARD_TEXTURE_print(pandecode_dump_stream
, &temp
, 2);
1608 pandecode_texture_payload(u
+ MALI_MIDGARD_TEXTURE_LENGTH
,
1609 temp
.dimension
, temp
.texel_ordering
, temp
.manual_stride
,
1610 temp
.levels
, temp
.depth
, temp
.array_size
, mapped_mem
);
1614 pandecode_bifrost_texture(
1619 struct MALI_BIFROST_TEXTURE temp
;
1620 MALI_BIFROST_TEXTURE_unpack(cl
, &temp
);
1621 MALI_BIFROST_TEXTURE_print(pandecode_dump_stream
, &temp
, 2);
1623 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(temp
.surfaces
);
1624 pandecode_texture_payload(temp
.surfaces
, temp
.dimension
, temp
.texel_ordering
,
1625 true, temp
.levels
, 1, 1, tmem
);
1628 /* 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 */
1631 pandecode_shader_prop(const char *name
, unsigned claim
, signed truth
, bool fuzzy
)
1637 if (fuzzy
&& (truth
< 0))
1638 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name
, claim
, truth
);
1640 if ((truth
>= 0) && !fuzzy
) {
1641 pandecode_msg("%s: expected %s = %d, claimed %u\n",
1642 (truth
< claim
) ? "warn" : "XXX",
1643 name
, truth
, claim
);
1644 } else if ((claim
> -truth
) && !fuzzy
) {
1645 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
1646 name
, -truth
, claim
);
1647 } else if (fuzzy
&& (claim
< truth
))
1648 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
1649 name
, truth
, claim
);
1651 pandecode_log(".%s = %" PRId16
, name
, claim
);
1654 pandecode_log_cont(" /* %u used */", truth
);
1656 pandecode_log_cont(",\n");
1660 pandecode_blend_shader_disassemble(mali_ptr shader
, int job_no
, int job_type
,
1661 bool is_bifrost
, unsigned gpu_id
)
1663 struct midgard_disasm_stats stats
=
1664 pandecode_shader_disassemble(shader
, job_no
, job_type
, is_bifrost
, gpu_id
);
1666 bool has_texture
= (stats
.texture_count
> 0);
1667 bool has_sampler
= (stats
.sampler_count
> 0);
1668 bool has_attribute
= (stats
.attribute_count
> 0);
1669 bool has_varying
= (stats
.varying_count
> 0);
1670 bool has_uniform
= (stats
.uniform_count
> 0);
1671 bool has_ubo
= (stats
.uniform_buffer_count
> 0);
1673 if (has_texture
|| has_sampler
)
1674 pandecode_msg("XXX: blend shader accessing textures\n");
1676 if (has_attribute
|| has_varying
)
1677 pandecode_msg("XXX: blend shader accessing interstage\n");
1679 if (has_uniform
|| has_ubo
)
1680 pandecode_msg("XXX: blend shader accessing uniforms\n");
1684 pandecode_textures(mali_ptr textures
, unsigned texture_count
, int job_no
, bool is_bifrost
)
1686 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(textures
);
1691 pandecode_log("Textures (%"PRIx64
"):\n", textures
);
1694 const void *cl
= pandecode_fetch_gpu_mem(mmem
,
1695 textures
, MALI_BIFROST_TEXTURE_LENGTH
*
1698 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
1699 pandecode_bifrost_texture(cl
+
1700 MALI_BIFROST_TEXTURE_LENGTH
* tex
,
1704 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
);
1706 for (int tex
= 0; tex
< texture_count
; ++tex
) {
1707 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
1708 char *a
= pointer_as_memory_reference(*u
);
1709 pandecode_log("%s,\n", a
);
1713 /* Now, finally, descend down into the texture descriptor */
1714 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
1715 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
1716 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
1718 pandecode_texture(*u
, tmem
, job_no
, tex
);
1724 pandecode_samplers(mali_ptr samplers
, unsigned sampler_count
, int job_no
, bool is_bifrost
)
1726 for (int i
= 0; i
< sampler_count
; ++i
) {
1728 DUMP_ADDR("Sampler", BIFROST_SAMPLER
, samplers
+ (MALI_BIFROST_SAMPLER_LENGTH
* i
), 1);
1730 DUMP_ADDR("Sampler", MIDGARD_SAMPLER
, samplers
+ (MALI_MIDGARD_SAMPLER_LENGTH
* i
), 1);
1736 pandecode_vertex_tiler_postfix_pre(
1737 const struct mali_vertex_tiler_postfix
*p
,
1738 int job_no
, enum mali_job_type job_type
,
1739 char *suffix
, bool is_bifrost
, unsigned gpu_id
)
1741 struct pandecode_mapped_memory
*attr_mem
;
1743 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1744 * are the only things actually needed from the FBD, vertex/tiler jobs
1745 * no longer reference the FBD -- instead, this field points to some
1746 * info about the scratchpad.
1749 struct pandecode_fbd fbd_info
= {
1750 /* Default for Bifrost */
1755 pandecode_log_cont("\t/* %X %/\n", p
->shared_memory
& 1);
1756 pandecode_compute_fbd(p
->shared_memory
& ~1, job_no
);
1757 } else if (p
->shared_memory
& MALI_MFBD
)
1758 fbd_info
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->shared_memory
) & FBD_MASK
, job_no
, false, job_type
== MALI_JOB_TYPE_COMPUTE
, false);
1759 else if (job_type
== MALI_JOB_TYPE_COMPUTE
)
1760 pandecode_compute_fbd((u64
) (uintptr_t) p
->shared_memory
, job_no
);
1762 fbd_info
= pandecode_sfbd((u64
) (uintptr_t) p
->shared_memory
, job_no
, false, gpu_id
);
1764 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
1765 int texture_count
= 0, sampler_count
= 0;
1768 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->shader
);
1769 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, p
->shader
);
1771 /* Disassemble ahead-of-time to get stats. Initialize with
1772 * stats for the missing-shader case so we get validation
1775 struct midgard_disasm_stats info
= {
1778 .attribute_count
= 0,
1782 .uniform_count
= -128,
1783 .uniform_buffer_count
= 0
1786 if (s
->shader
& ~0xF)
1787 info
= pandecode_shader_disassemble(s
->shader
& ~0xF, job_no
, job_type
, is_bifrost
, gpu_id
);
1789 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", p
->shader
, job_no
, suffix
);
1792 /* Save for dumps */
1793 attribute_count
= s
->attribute_count
;
1794 varying_count
= s
->varying_count
;
1795 texture_count
= s
->texture_count
;
1796 sampler_count
= s
->sampler_count
;
1799 uniform_count
= s
->bifrost2
.uniform_count
;
1800 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
1802 uniform_count
= s
->midgard1
.uniform_count
;
1803 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
1806 pandecode_shader_address("shader", s
->shader
);
1808 pandecode_shader_prop("texture_count", s
->texture_count
, info
.texture_count
, false);
1809 pandecode_shader_prop("sampler_count", s
->sampler_count
, info
.sampler_count
, false);
1810 pandecode_shader_prop("attribute_count", s
->attribute_count
, info
.attribute_count
, false);
1811 pandecode_shader_prop("varying_count", s
->varying_count
, info
.varying_count
, false);
1812 pandecode_shader_prop("uniform_buffer_count",
1813 uniform_buffer_count
,
1814 info
.uniform_buffer_count
, true);
1817 pandecode_shader_prop("uniform_count",
1819 info
.uniform_count
, false);
1821 pandecode_shader_prop("work_count",
1822 s
->midgard1
.work_count
, info
.work_count
, false);
1826 pandecode_log("bifrost1.unk1 = ");
1827 pandecode_log_decoded_flags(shader_bifrost_info
, s
->bifrost1
.unk1
);
1828 pandecode_log_cont(",\n");
1830 bool helpers
= s
->midgard1
.flags_lo
& MALI_HELPER_INVOCATIONS
;
1832 if (helpers
!= info
.helper_invocations
) {
1833 pandecode_msg("XXX: expected helpers %u but got %u\n",
1834 info
.helper_invocations
, helpers
);
1837 pandecode_log(".midgard1.flags_lo = ");
1838 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info
,
1839 s
->midgard1
.flags_lo
& ~MALI_HELPER_INVOCATIONS
);
1840 pandecode_log_cont(",\n");
1842 pandecode_log(".midgard1.flags_hi = ");
1843 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info
, s
->midgard1
.flags_hi
);
1844 pandecode_log_cont(",\n");
1847 if (s
->depth_units
|| s
->depth_factor
) {
1848 pandecode_prop("depth_factor = %f", s
->depth_factor
);
1849 pandecode_prop("depth_units = %f", s
->depth_units
);
1852 if (s
->coverage_mask
)
1853 pandecode_prop("coverage_mask = 0x%X", s
->coverage_mask
);
1856 pandecode_prop(".unknown2_2 = %X", s
->unknown2_2
);
1858 if (s
->unknown2_3
|| s
->unknown2_4
) {
1859 pandecode_log(".unknown2_3 = ");
1861 int unknown2_3
= s
->unknown2_3
;
1862 int unknown2_4
= s
->unknown2_4
;
1864 /* We're not quite sure what these flags mean without the depth test, if anything */
1866 if (unknown2_3
& (MALI_DEPTH_WRITEMASK
| MALI_DEPTH_FUNC_MASK
)) {
1867 const char *func
= mali_func_as_str(MALI_GET_DEPTH_FUNC(unknown2_3
));
1868 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
1870 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
1873 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
1874 pandecode_log_cont(",\n");
1876 pandecode_log(".unknown2_4 = ");
1877 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
1878 pandecode_log_cont(",\n");
1881 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
1882 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
1883 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
1886 DUMP_CL("Stencil front", STENCIL
, &s
->stencil_front
, 1);
1887 DUMP_CL("Stencil back", STENCIL
, &s
->stencil_back
, 1);
1890 pandecode_log(".bifrost2 = {\n");
1893 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
1894 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
1895 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
1896 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
1899 pandecode_log("},\n");
1900 } else if (s
->midgard2
.unknown2_7
) {
1901 pandecode_log(".midgard2 = {\n");
1904 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
1906 pandecode_log("},\n");
1910 pandecode_msg("XXX: shader padding tripped\n");
1911 pandecode_prop("padding = 0x%" PRIx32
, s
->padding
);
1915 /* TODO: Blend shaders routing/disasm */
1916 union midgard_blend blend
= s
->blend
;
1917 mali_ptr shader
= pandecode_midgard_blend(&blend
, s
->unknown2_3
& MALI_HAS_BLEND_SHADER
);
1919 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
1921 pandecode_msg("mdg_blend = %" PRIx64
"\n", s
->blend
.shader
);
1925 pandecode_log("};\n");
1927 /* MRT blend fields are used whenever MFBD is used, with
1928 * per-RT descriptors */
1930 if (job_type
== MALI_JOB_TYPE_TILER
&& (is_bifrost
|| p
->shared_memory
& MALI_MFBD
)) {
1931 void* blend_base
= (void *) (s
+ 1);
1933 for (unsigned i
= 0; i
< fbd_info
.rt_count
; i
++) {
1934 mali_ptr shader
= 0;
1937 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
1939 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
1942 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
1947 pandecode_msg("XXX: missing shader descriptor\n");
1950 DUMP_ADDR("Viewport", VIEWPORT
, p
->viewport
, 1);
1952 unsigned max_attr_index
= 0;
1954 if (p
->attribute_meta
)
1955 max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
1957 if (p
->attributes
) {
1958 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
1959 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false, job_type
);
1962 /* Varyings are encoded like attributes but not actually sent; we just
1963 * pass a zero buffer with the right stride/size set, (or whatever)
1964 * since the GPU will write to it itself */
1966 if (p
->varying_meta
) {
1967 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
1971 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
1973 /* Number of descriptors depends on whether there are
1974 * non-internal varyings */
1976 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true, job_type
);
1979 if (p
->uniform_buffers
) {
1980 if (uniform_buffer_count
)
1981 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
1983 pandecode_msg("warn: UBOs specified but not referenced\n");
1984 } else if (uniform_buffer_count
)
1985 pandecode_msg("XXX: UBOs referenced but not specified\n");
1987 /* We don't want to actually dump uniforms, but we do need to validate
1988 * that the counts we were given are sane */
1992 pandecode_uniforms(p
->uniforms
, uniform_count
);
1994 pandecode_msg("warn: Uniforms specified but not referenced\n");
1995 } else if (uniform_count
)
1996 pandecode_msg("XXX: Uniforms referenced but not specified\n");
1999 pandecode_textures(p
->textures
, texture_count
, job_no
, is_bifrost
);
2001 if (p
->sampler_descriptor
)
2002 pandecode_samplers(p
->sampler_descriptor
, sampler_count
, job_no
, is_bifrost
);
2006 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2008 pandecode_log(".gl_enables = ");
2010 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2012 pandecode_log_cont(",\n");
2016 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2018 if (p
->shader
& 0xF)
2019 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p
->shader
& 0xF));
2021 pandecode_log(".postfix = {\n");
2024 pandecode_gl_enables(p
->gl_enables
, MALI_JOB_TYPE_TILER
);
2025 pandecode_prop("instance_shift = 0x%x", p
->instance_shift
);
2026 pandecode_prop("instance_odd = 0x%x", p
->instance_odd
);
2029 pandecode_msg("XXX: vertex only zero tripped");
2030 pandecode_prop("zero4 = 0x%" PRIx32
, p
->zero4
);
2033 pandecode_prop("offset_start = 0x%x", p
->offset_start
);
2036 pandecode_msg("XXX: vertex only zero tripped");
2037 pandecode_prop("zero5 = 0x%" PRIx64
, p
->zero5
);
2040 MEMORY_PROP(p
, position_varying
);
2041 MEMORY_PROP(p
, occlusion_counter
);
2044 pandecode_log("},\n");
2048 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2050 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2051 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2053 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2057 pandecode_msg("XXX: tiler heap zero tripped\n");
2058 pandecode_prop("zero = 0x%x", h
->zero
);
2061 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2062 MEMORY_PROP(h
, tiler_heap_start
);
2063 MEMORY_PROP(h
, tiler_heap_free
);
2065 /* this might point to the beginning of another buffer, when it's
2066 * really the end of the tiler heap buffer, so we have to be careful
2067 * here. but for zero length, we need the same pointer.
2070 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2071 MEMORY_PROP(h
, tiler_heap_start
);
2073 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2074 pandecode_prop("tiler_heap_end = %s + 1", a
);
2078 for (int i
= 0; i
< 10; i
++) {
2079 if (h
->zeros
[i
] != 0) {
2080 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2085 if (h
->unk1
!= 0x1) {
2086 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2087 pandecode_prop("unk1 = 0x%x", h
->unk1
);
2090 if (h
->unk7e007e
!= 0x7e007e) {
2091 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
2092 pandecode_prop("unk7e007e = 0x%x", h
->unk7e007e
);
2096 pandecode_log("};\n");
2100 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2102 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2103 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2105 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2107 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2110 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32
, t
->tiler_heap_next_start
);
2111 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32
, t
->used_hierarchy_mask
);
2113 if (t
->hierarchy_mask
!= 0xa &&
2114 t
->hierarchy_mask
!= 0x14 &&
2115 t
->hierarchy_mask
!= 0x28 &&
2116 t
->hierarchy_mask
!= 0x50 &&
2117 t
->hierarchy_mask
!= 0xa0)
2118 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
2120 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2122 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2124 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2125 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2128 pandecode_msg("XXX: tiler meta zero tripped\n");
2129 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2132 for (int i
= 0; i
< 12; i
++) {
2133 if (t
->zeros
[i
] != 0) {
2134 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2140 pandecode_log("};\n");
2144 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2146 if (u
.pointer
== 0x0)
2149 pandecode_log(".primitive_size = {\n");
2153 pandecode_prop("constant = %f", u
.constant
);
2155 MEMORY_PROP((&u
), pointer
);
2159 pandecode_log("},\n");
2163 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2165 pandecode_log_cont("{\n");
2168 /* TODO: gl_PointSize on Bifrost */
2169 pandecode_primitive_size(t
->primitive_size
, true);
2171 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2173 pandecode_msg("XXX: tiler only zero tripped\n");
2174 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2175 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2176 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2177 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2178 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2179 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2183 pandecode_log("},\n");
2187 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2188 const struct pandecode_mapped_memory
*mem
,
2189 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2191 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2193 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2195 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2198 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2199 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2202 pandecode_log("};\n");
2208 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2209 const struct pandecode_mapped_memory
*mem
,
2210 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2212 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2214 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2215 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2217 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2220 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2222 pandecode_log(".tiler = ");
2223 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2225 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2228 pandecode_log("};\n");
2234 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2235 const struct pandecode_mapped_memory
*mem
,
2236 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2238 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2239 bool is_graphics
= (h
->job_type
== MALI_JOB_TYPE_VERTEX
) || (h
->job_type
== MALI_JOB_TYPE_TILER
);
2241 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false, gpu_id
);
2243 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2246 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, is_graphics
);
2247 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2249 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2250 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2253 pandecode_log("};\n");
2259 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2260 mali_ptr payload
, int job_no
,
2261 bool is_bifrost
, unsigned gpu_id
)
2263 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2265 bool is_mfbd
= s
->framebuffer
& MALI_MFBD
;
2267 if (!is_mfbd
&& is_bifrost
)
2268 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2270 struct pandecode_fbd info
;
2273 info
= pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true, false, is_bifrost
);
2275 info
= pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true, gpu_id
);
2277 /* Compute the tag for the tagged pointer. This contains the type of
2278 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2279 * additional structures follow the MFBD header (an extra payload or
2280 * not, as well as a count of render targets) */
2282 unsigned expected_tag
= is_mfbd
? MALI_MFBD
: 0;
2286 expected_tag
|= MALI_MFBD_TAG_EXTRA
;
2288 expected_tag
|= (MALI_POSITIVE(info
.rt_count
) << 2);
2291 if ((s
->min_tile_coord
| s
->max_tile_coord
) & ~(MALI_X_COORD_MASK
| MALI_Y_COORD_MASK
)) {
2292 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2293 pandecode_prop("min_tile_coord = 0x%X\n", s
->min_tile_coord
);
2294 pandecode_prop("max_tile_coord = 0x%X\n", s
->max_tile_coord
);
2297 /* Extract tile coordinates */
2299 unsigned min_x
= MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2300 unsigned min_y
= MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2302 unsigned max_x
= (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2303 unsigned max_y
= (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2305 /* For the max, we also want the floored (rather than ceiled) version for checking */
2307 unsigned max_x_f
= (MALI_TILE_COORD_X(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2308 unsigned max_y_f
= (MALI_TILE_COORD_Y(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2310 /* Validate the coordinates are well-ordered */
2313 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2314 else if (min_x
> max_x
)
2315 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2318 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2319 else if (min_y
> max_y
)
2320 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2322 /* Validate the coordinates fit inside the framebuffer. We use floor,
2323 * rather than ceil, for the max coordinates, since the tile
2324 * coordinates for something like an 800x600 framebuffer will actually
2325 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2327 if ((min_x
> info
.width
) || (max_x_f
> info
.width
))
2328 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2330 if ((min_y
> info
.height
) || (max_y_f
> info
.height
))
2331 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2333 /* After validation, we print */
2335 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x
, min_y
, max_x
, max_y
);
2337 /* The FBD is a tagged pointer */
2339 unsigned tag
= (s
->framebuffer
& ~FBD_MASK
);
2341 if (tag
!= expected_tag
)
2342 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag
, tag
);
2347 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
2348 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
2349 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
2350 * might wish to combine this with the bifrost parameter) because some details
2351 * are model-specific even within a particular architecture. Minimal traces
2352 * *only* examine the job descriptors, skipping printing entirely if there is
2353 * no faults, and only descends into the payload if there are faults. This is
2354 * useful for looking for faults without the overhead of invasive traces. */
2357 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
, unsigned gpu_id
, bool minimal
)
2359 pandecode_dump_file_open();
2361 struct mali_job_descriptor_header
*h
;
2362 unsigned job_descriptor_number
= 0;
2365 struct pandecode_mapped_memory
*mem
=
2366 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2370 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2372 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2373 * high 32-bits of the 64-bit pointer are reused to store
2376 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
2377 h
->job_type
!= MALI_JOB_TYPE_FRAGMENT
? 4 : 0;
2378 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
2380 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
, 256);
2382 int job_no
= job_descriptor_number
++;
2384 /* If the job is good to go, skip it in minimal mode */
2385 if (minimal
&& (h
->exception_status
== 0x0 || h
->exception_status
== 0x1))
2388 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2391 pandecode_prop("job_type = %s", mali_job_type_as_str(h
->job_type
));
2393 if (h
->job_descriptor_size
)
2394 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2396 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2397 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2398 h
->exception_status
,
2399 (h
->exception_status
>> 16) & 0xFFFF,
2400 mali_exception_access_as_str((h
->exception_status
>> 8) & 0x3),
2401 h
->exception_status
& 0xFF);
2403 if (h
->first_incomplete_task
)
2404 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2406 if (h
->fault_pointer
)
2407 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2410 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2412 pandecode_prop("job_index = %d", h
->job_index
);
2414 if (h
->unknown_flags
)
2415 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2417 if (h
->job_dependency_index_1
)
2418 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2420 if (h
->job_dependency_index_2
)
2421 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2424 pandecode_log("};\n");
2426 switch (h
->job_type
) {
2427 case MALI_JOB_TYPE_WRITE_VALUE
: {
2428 struct mali_payload_write_value
*s
= payload
;
2429 pandecode_log("struct mali_payload_write_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2431 MEMORY_PROP(s
, address
);
2433 if (s
->value_descriptor
!= MALI_WRITE_VALUE_ZERO
) {
2434 pandecode_msg("XXX: unknown value descriptor\n");
2435 pandecode_prop("value_descriptor = 0x%" PRIX32
, s
->value_descriptor
);
2439 pandecode_msg("XXX: set value tripped\n");
2440 pandecode_prop("reserved = 0x%" PRIX32
, s
->reserved
);
2443 pandecode_prop("immediate = 0x%" PRIX64
, s
->immediate
);
2445 pandecode_log("};\n");
2450 case MALI_JOB_TYPE_TILER
:
2451 case MALI_JOB_TYPE_VERTEX
:
2452 case MALI_JOB_TYPE_COMPUTE
:
2454 if (h
->job_type
== MALI_JOB_TYPE_TILER
)
2455 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2457 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2459 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2463 case MALI_JOB_TYPE_FRAGMENT
:
2464 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
, gpu_id
);
2470 } while ((jc_gpu_va
= h
->next_job
));
2472 pandecode_map_read_write();