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_##flag, "MALI_" #flag }
232 static const struct pandecode_flag_info u3_flag_info
[] = {
234 FLAG_INFO(PER_SAMPLE
),
235 FLAG_INFO(CAN_DISCARD
),
236 FLAG_INFO(HAS_BLEND_SHADER
),
237 FLAG_INFO(DEPTH_WRITEMASK
),
238 FLAG_INFO(DEPTH_CLIP_NEAR
),
239 FLAG_INFO(DEPTH_CLIP_FAR
),
243 static const struct pandecode_flag_info u4_flag_info
[] = {
245 FLAG_INFO(NO_DITHER
),
246 FLAG_INFO(DEPTH_RANGE_A
),
247 FLAG_INFO(DEPTH_RANGE_B
),
248 FLAG_INFO(STENCIL_TEST
),
249 FLAG_INFO(ALPHA_TO_COVERAGE
),
254 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
255 static const struct pandecode_flag_info mfbd_fmt_flag_info
[] = {
261 #define FLAG_INFO(flag) { MALI_AFBC_##flag, "MALI_AFBC_" #flag }
262 static const struct pandecode_flag_info afbc_fmt_flag_info
[] = {
268 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
269 static const struct pandecode_flag_info mfbd_extra_flag_hi_info
[] = {
275 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
276 static const struct pandecode_flag_info mfbd_extra_flag_lo_info
[] = {
282 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
283 static const struct pandecode_flag_info shader_midgard1_flag_lo_info
[] = {
286 FLAG_INFO(READS_TILEBUFFER
),
287 FLAG_INFO(WRITES_GLOBAL
),
292 static const struct pandecode_flag_info shader_midgard1_flag_hi_info
[] = {
294 FLAG_INFO(SUPPRESS_INF_NAN
),
299 #define FLAG_INFO(flag) { MALI_BIFROST_##flag, "MALI_BIFROST_" #flag }
300 static const struct pandecode_flag_info shader_bifrost_info
[] = {
301 FLAG_INFO(FULL_THREAD
),
303 FLAG_INFO(FIRST_ATEST
),
309 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
310 static const struct pandecode_flag_info mfbd_flag_info
[] = {
311 FLAG_INFO(DEPTH_WRITE
),
317 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
318 static const struct pandecode_flag_info sfbd_unk1_info
[] = {
325 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
326 static const struct pandecode_flag_info sfbd_unk2_info
[] = {
333 /* Midgard's tiler descriptor is embedded within the
337 pandecode_midgard_tiler_descriptor(
338 const struct midgard_tiler_descriptor
*t
,
344 pandecode_log(".tiler = {\n");
347 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
348 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
350 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
352 /* We know this name from the kernel, but we never see it nonzero */
355 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
357 MEMORY_PROP(t
, polygon_list
);
359 /* The body is offset from the base of the polygon list */
360 //assert(t->polygon_list_body > t->polygon_list);
361 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
363 /* It needs to fit inside the reported size */
364 //assert(t->polygon_list_size >= body_offset);
366 /* Now that we've sanity checked, we'll try to calculate the sizes
367 * ourselves for comparison */
369 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
370 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
372 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
373 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
374 ref_header
, ref_size
);
375 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
376 pandecode_msg("body offset %d\n", body_offset
);
379 /* The tiler heap has a start and end specified -- it should be
380 * identical to what we have in the BO. The exception is if tiling is
383 MEMORY_PROP(t
, heap_start
);
384 assert(t
->heap_end
>= t
->heap_start
);
386 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
388 /* Tiling is enabled with a special flag */
389 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
390 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
392 bool tiling_enabled
= hierarchy_mask
;
394 if (tiling_enabled
) {
395 /* We should also have no other flags */
397 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
399 /* When tiling is disabled, we should have that flag and no others */
401 if (tiler_flags
!= MALI_TILER_DISABLED
) {
402 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
406 /* We should also have an empty heap */
408 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
412 /* Disabled tiling is used only for clear-only jobs, which are
413 * purely FRAGMENT, so we should never see this for
414 * non-FRAGMENT descriptors. */
417 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
420 /* We've never seen weights used in practice, but we know from the
421 * kernel these fields is there */
423 bool nonzero_weights
= false;
425 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
426 nonzero_weights
|= t
->weights
[w
] != 0x0;
429 if (nonzero_weights
) {
430 pandecode_log(".weights = { ");
432 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
433 pandecode_log_cont("%d, ", t
->weights
[w
]);
440 pandecode_log("}\n");
443 /* TODO: The Bifrost tiler is not understood at all yet */
446 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer
*fb
)
448 pandecode_log(".tiler = {\n");
451 MEMORY_PROP(fb
, tiler_meta
);
453 for (int i
= 0; i
< 16; i
++) {
454 if (fb
->zeros
[i
] != 0) {
455 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
460 pandecode_log("},\n");
463 pandecode_log("}\n");
467 /* Information about the framebuffer passed back for
468 * additional analysis */
470 struct pandecode_fbd
{
478 pandecode_sfbd_format(struct mali_sfbd_format format
)
480 pandecode_log(".format = {\n");
483 pandecode_log(".unk1 = ");
484 pandecode_log_decoded_flags(sfbd_unk1_info
, format
.unk1
);
485 pandecode_log_cont(",\n");
487 /* TODO: Map formats so we can check swizzles and print nicely */
488 pandecode_log("swizzle");
489 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
490 pandecode_log_cont(",\n");
492 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
493 (format
.nr_channels
+ 1));
495 pandecode_log(".unk2 = ");
496 pandecode_log_decoded_flags(sfbd_unk2_info
, format
.unk2
);
497 pandecode_log_cont(",\n");
499 pandecode_prop("block = %s", mali_block_format_as_str(format
.block
));
501 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
504 pandecode_log("},\n");
508 pandecode_shared_memory(const struct mali_shared_memory
*desc
, bool is_compute
)
510 pandecode_prop("stack_shift = 0x%x", desc
->stack_shift
);
513 pandecode_prop("unk0 = 0x%x", desc
->unk0
);
515 if (desc
->shared_workgroup_count
!= 0x1F) {
516 pandecode_prop("shared_workgroup_count = %d", desc
->shared_workgroup_count
);
518 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
521 if (desc
->shared_unk1
|| desc
->shared_shift
) {
522 pandecode_prop("shared_unk1 = %X", desc
->shared_unk1
);
523 pandecode_prop("shared_shift = %X", desc
->shared_shift
);
526 pandecode_msg("XXX: shared memory configured in noncompute shader");
529 if (desc
->shared_zero
) {
530 pandecode_msg("XXX: shared memory zero tripped\n");
531 pandecode_prop("shared_zero = 0x%" PRIx32
, desc
->shared_zero
);
534 if (desc
->shared_memory
&& !is_compute
)
535 pandecode_msg("XXX: shared memory used in noncompute shader\n");
537 MEMORY_PROP(desc
, scratchpad
);
538 MEMORY_PROP(desc
, shared_memory
);
539 MEMORY_PROP(desc
, unknown1
);
541 if (desc
->scratchpad
) {
542 struct pandecode_mapped_memory
*smem
=
543 pandecode_find_mapped_gpu_mem_containing(desc
->scratchpad
);
545 pandecode_msg("scratchpad size %u\n", smem
->length
);
550 static struct pandecode_fbd
551 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
, unsigned gpu_id
)
553 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
554 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
556 struct pandecode_fbd info
= {
561 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
564 pandecode_log(".shared_memory = {\n");
566 pandecode_shared_memory(&s
->shared_memory
, false);
568 pandecode_log("},\n");
570 pandecode_sfbd_format(s
->format
);
572 info
.width
= s
->width
+ 1;
573 info
.height
= s
->height
+ 1;
575 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", info
.width
);
576 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", info
.height
);
578 MEMORY_PROP(s
, checksum
);
580 if (s
->checksum_stride
)
581 pandecode_prop("checksum_stride = %d", s
->checksum_stride
);
583 MEMORY_PROP(s
, framebuffer
);
584 pandecode_prop("stride = %d", s
->stride
);
586 /* Earlier in the actual commandstream -- right before width -- but we
587 * delay to flow nicer */
589 pandecode_log(".clear_flags = ");
590 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
591 pandecode_log_cont(",\n");
593 if (s
->depth_buffer
) {
594 MEMORY_PROP(s
, depth_buffer
);
595 pandecode_prop("depth_stride = %d", s
->depth_stride
);
598 if (s
->stencil_buffer
) {
599 MEMORY_PROP(s
, stencil_buffer
);
600 pandecode_prop("stencil_stride = %d", s
->stencil_stride
);
603 if (s
->depth_stride_zero
||
604 s
->stencil_stride_zero
||
605 s
->zero7
|| s
->zero8
) {
606 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
607 pandecode_prop("depth_stride_zero = 0x%x",
608 s
->depth_stride_zero
);
609 pandecode_prop("stencil_stride_zero = 0x%x",
610 s
->stencil_stride_zero
);
611 pandecode_prop("zero7 = 0x%" PRIx32
,
613 pandecode_prop("zero8 = 0x%" PRIx32
,
617 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
618 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
619 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
620 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
621 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
624 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
625 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
626 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
627 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
628 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
631 if (s
->clear_stencil
) {
632 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
635 const struct midgard_tiler_descriptor t
= s
->tiler
;
637 bool has_hierarchy
= !(gpu_id
== 0x0720 || gpu_id
== 0x0820 || gpu_id
== 0x0830);
638 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
, has_hierarchy
);
641 pandecode_log("};\n");
643 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
644 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
645 pandecode_prop("zero5 = 0x%" PRIx32
, s
->zero5
);
647 pandecode_log_cont(".zero3 = {");
649 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
650 pandecode_log_cont("%X, ", s
->zero3
[i
]);
652 pandecode_log_cont("},\n");
654 pandecode_log_cont(".zero6 = {");
656 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
657 pandecode_log_cont("%X, ", s
->zero6
[i
]);
659 pandecode_log_cont("},\n");
665 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
667 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
668 const struct mali_shared_memory
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
670 pandecode_log("struct mali_shared_memory shared_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
672 pandecode_shared_memory(s
, true);
674 pandecode_log("},\n");
677 /* Extracts the number of components associated with a Mali format */
680 pandecode_format_component_count(enum mali_format fmt
)
682 /* Mask out the format class */
683 unsigned top
= fmt
& 0b11100000;
686 case MALI_FORMAT_SNORM
:
687 case MALI_FORMAT_UINT
:
688 case MALI_FORMAT_UNORM
:
689 case MALI_FORMAT_SINT
:
690 return ((fmt
>> 3) & 3) + 1;
697 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
700 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
703 assert(MALI_CHANNEL_R
== 0);
705 for (unsigned c
= 0; c
< 4; ++c
) {
706 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
708 if (chan
<= MALI_CHANNEL_A
)
715 /* Validates that a (format, swizzle) pair is valid, in the sense that the
716 * swizzle doesn't access any components that are undefined in the format.
717 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
721 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
723 unsigned nr_comp
= pandecode_format_component_count(fmt
);
724 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
725 unsigned valid_mask
= (1 << nr_comp
) - 1;
726 unsigned invalid_mask
= ~valid_mask
;
728 if (access_mask
& invalid_mask
) {
729 pandecode_msg("XXX: invalid components accessed\n");
733 /* Check for the default non-swizzling swizzle so we can suppress
734 * useless printing for the defaults */
736 unsigned default_swizzles
[4] = {
737 MALI_CHANNEL_R
| (MALI_CHANNEL_0
<< 3) | (MALI_CHANNEL_0
<< 6) | (MALI_CHANNEL_1
<< 9),
738 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_0
<< 6) | (MALI_CHANNEL_1
<< 9),
739 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_B
<< 6) | (MALI_CHANNEL_1
<< 9),
740 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_B
<< 6) | (MALI_CHANNEL_A
<< 9)
743 return (swizzle
== default_swizzles
[nr_comp
- 1]);
747 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
749 /* First, do some validation */
750 bool trivial_swizzle
= pandecode_validate_format_swizzle(
756 /* Next, print the swizzle */
757 pandecode_log_cont(".");
759 static const char components
[] = "rgba01";
761 for (unsigned c
= 0; c
< 4; ++c
) {
762 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
764 if (chan
> MALI_CHANNEL_1
) {
765 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
768 pandecode_log_cont("%c", components
[chan
]);
773 pandecode_rt_format(struct mali_rt_format format
)
775 pandecode_log(".format = {\n");
778 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
779 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
780 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
781 pandecode_prop("unk4 = 0x%" PRIx32
, format
.unk4
);
783 pandecode_prop("block = %s", mali_block_format_as_str(format
.block
));
785 /* TODO: Map formats so we can check swizzles and print nicely */
786 pandecode_log("swizzle");
787 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
788 pandecode_log_cont(",\n");
790 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
791 (format
.nr_channels
+ 1));
793 pandecode_log(".flags = ");
794 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
795 pandecode_log_cont(",\n");
797 pandecode_prop("msaa = %s", mali_msaa_as_str(format
.msaa
));
799 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
800 * which is a faster hardware-based alternative to the wallpaper method
801 * to preserve framebuffer contents across frames. In practice, MFBD
802 * preload is buggy on Midgard, and so this is a chicken bit. If this
803 * bit isn't set, most likely something broke unrelated to preload */
805 if (!format
.no_preload
) {
806 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
807 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
811 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
814 pandecode_log("},\n");
818 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct mali_framebuffer
*fb
)
820 pandecode_log("struct mali_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
823 for (int i
= 0; i
< (fb
->rt_count_1
+ 1); i
++) {
824 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct mali_render_target
);
825 struct pandecode_mapped_memory
*mem
=
826 pandecode_find_mapped_gpu_mem_containing(rt_va
);
827 const struct mali_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
829 pandecode_log("{\n");
832 pandecode_rt_format(rt
->format
);
834 if (rt
->format
.block
== MALI_BLOCK_FORMAT_AFBC
) {
835 pandecode_log(".afbc = {\n");
838 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
839 pandecode_prop("metadata = %s", a
);
842 pandecode_prop("stride = %d", rt
->afbc
.stride
);
844 pandecode_log(".flags = ");
845 pandecode_log_decoded_flags(afbc_fmt_flag_info
, rt
->afbc
.flags
);
846 pandecode_log_cont(",\n");
849 pandecode_log("},\n");
850 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.flags
) {
851 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
857 MEMORY_PROP(rt
, framebuffer
);
858 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
860 if (rt
->layer_stride
)
861 pandecode_prop("layer_stride = %d", rt
->layer_stride
);
863 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
864 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
865 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
866 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
867 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
870 if (rt
->zero1
|| rt
->zero2
) {
871 pandecode_msg("XXX: render target zeros tripped\n");
872 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
873 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
877 pandecode_log("},\n");
881 pandecode_log("};\n");
884 static struct pandecode_fbd
885 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
, bool is_compute
, bool is_bifrost
)
887 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
888 const struct mali_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
890 struct pandecode_fbd info
;
892 if (is_bifrost
&& fb
->msaa
.sample_locations
) {
893 /* The blob stores all possible sample locations in a single buffer
894 * allocated on startup, and just switches the pointer when switching
895 * MSAA state. For now, we just put the data into the cmdstream, but we
896 * should do something like what the blob does with a real driver.
898 * There seem to be 32 slots for sample locations, followed by another
899 * 16. The second 16 is just the center location followed by 15 zeros
900 * in all the cases I've identified (maybe shader vs. depth/color
904 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->msaa
.sample_locations
);
906 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->msaa
.sample_locations
);
908 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
911 for (int i
= 0; i
< 32 + 16; i
++) {
912 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
916 pandecode_log("};\n");
919 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
923 pandecode_log(".msaa = {\n");
926 if (fb
->msaa
.sample_locations
)
927 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
929 pandecode_msg("XXX: sample_locations missing\n");
931 if (fb
->msaa
.zero1
|| fb
->msaa
.zero2
|| fb
->msaa
.zero4
) {
932 pandecode_msg("XXX: multisampling zero tripped\n");
933 pandecode_prop("zero1 = %" PRIx64
, fb
->msaa
.zero1
);
934 pandecode_prop("zero2 = %" PRIx64
, fb
->msaa
.zero2
);
935 pandecode_prop("zero4 = %" PRIx64
, fb
->msaa
.zero4
);
939 pandecode_log("},\n");
941 pandecode_log(".shared_memory = {\n");
943 pandecode_shared_memory(&fb
->shared_memory
, is_compute
);
945 pandecode_log("},\n");
948 info
.width
= fb
->width1
+ 1;
949 info
.height
= fb
->height1
+ 1;
950 info
.rt_count
= fb
->rt_count_1
+ 1;
952 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
953 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
954 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
955 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
957 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
958 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
959 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
960 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
962 pandecode_log(".mfbd_flags = ");
963 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
964 pandecode_log_cont(",\n");
966 if (fb
->clear_stencil
)
967 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
970 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
974 pandecode_bifrost_tiler_descriptor(fb
);
976 const struct midgard_tiler_descriptor t
= fb
->tiler
;
977 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
, true);
980 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
982 if (fb
->zero3
|| fb
->zero4
) {
983 pandecode_msg("XXX: framebuffer zeros tripped\n");
984 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
985 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
989 pandecode_log("};\n");
991 gpu_va
+= sizeof(struct mali_framebuffer
);
993 info
.has_extra
= (fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
;
995 if (info
.has_extra
) {
996 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
997 const struct mali_framebuffer_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
999 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1002 MEMORY_PROP(fbx
, checksum
);
1004 if (fbx
->checksum_stride
)
1005 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
1007 pandecode_log(".flags_hi = ");
1008 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info
, fbx
->flags_hi
);
1009 pandecode_log_cont(",\n");
1011 pandecode_log(".flags_lo = ");
1012 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info
, fbx
->flags_lo
);
1013 pandecode_log_cont(",\n");
1015 pandecode_prop("zs_block = %s", mali_block_format_as_str(fbx
->zs_block
));
1016 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx
->zs_samples
+ 1);
1018 if (fbx
->zs_block
== MALI_BLOCK_FORMAT_AFBC
) {
1019 pandecode_log(".ds_afbc = {\n");
1022 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
1023 pandecode_prop("depth_stencil_afbc_stride = %d",
1024 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
1025 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
1027 pandecode_log(".flags = ");
1028 pandecode_log_decoded_flags(afbc_fmt_flag_info
, fbx
->ds_afbc
.flags
);
1029 pandecode_log_cont(",\n");
1031 if (fbx
->ds_afbc
.padding
) {
1032 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1033 pandecode_prop("padding = 0x%" PRIx64
, fbx
->ds_afbc
.padding
);
1037 pandecode_log("},\n");
1039 pandecode_log(".ds_linear = {\n");
1042 if (fbx
->ds_linear
.depth
) {
1043 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
1044 pandecode_prop("depth_stride = %d",
1045 fbx
->ds_linear
.depth_stride
);
1046 pandecode_prop("depth_layer_stride = %d",
1047 fbx
->ds_linear
.depth_layer_stride
);
1048 } else if (fbx
->ds_linear
.depth_stride
|| fbx
->ds_linear
.depth_layer_stride
) {
1049 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx
->ds_linear
.depth_stride
, fbx
->ds_linear
.depth_layer_stride
);
1052 if (fbx
->ds_linear
.stencil
) {
1053 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1054 pandecode_prop("stencil_stride = %d",
1055 fbx
->ds_linear
.stencil_stride
);
1056 pandecode_prop("stencil_layer_stride = %d",
1057 fbx
->ds_linear
.stencil_layer_stride
);
1058 } else if (fbx
->ds_linear
.stencil_stride
|| fbx
->ds_linear
.stencil_layer_stride
) {
1059 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx
->ds_linear
.stencil_stride
, fbx
->ds_linear
.stencil_layer_stride
);
1062 if (fbx
->ds_linear
.depth_stride_zero
||
1063 fbx
->ds_linear
.stencil_stride_zero
) {
1064 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1065 pandecode_prop("depth_stride_zero = 0x%x",
1066 fbx
->ds_linear
.depth_stride_zero
);
1067 pandecode_prop("stencil_stride_zero = 0x%x",
1068 fbx
->ds_linear
.stencil_stride_zero
);
1072 pandecode_log("},\n");
1075 if (fbx
->clear_color_1
| fbx
->clear_color_2
) {
1076 pandecode_prop("clear_color_1 = 0x%" PRIx32
, fbx
->clear_color_1
);
1077 pandecode_prop("clear_color_2 = 0x%" PRIx32
, fbx
->clear_color_2
);
1081 pandecode_msg("XXX: fb_extra zeros tripped\n");
1082 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1086 pandecode_log("};\n");
1088 gpu_va
+= sizeof(struct mali_framebuffer_extra
);
1092 pandecode_render_target(gpu_va
, job_no
, fb
);
1098 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1099 mali_ptr addr
, int job_no
, char *suffix
,
1100 int count
, bool varying
, enum mali_job_type job_type
)
1102 char *prefix
= varying
? "Varying" : "Attribute";
1106 pandecode_msg("warn: No %s records\n", prefix
);
1110 MAP_ADDR(ATTRIBUTE_BUFFER
, addr
, cl
);
1112 for (int i
= 0; i
< count
; ++i
) {
1113 fprintf(pandecode_dump_stream
, "%s\n", prefix
);
1115 struct MALI_ATTRIBUTE_BUFFER temp
;
1116 MALI_ATTRIBUTE_BUFFER_unpack(cl
+ i
* MALI_ATTRIBUTE_BUFFER_LENGTH
, &temp
);
1117 MALI_ATTRIBUTE_BUFFER_print(pandecode_dump_stream
, &temp
, 2);
1119 if (temp
.type
== MALI_ATTRIBUTE_TYPE_1D_NPOT_DIVISOR
) {
1120 struct MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT temp2
;
1121 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_unpack(cl
+ (i
+ 1) * MALI_ATTRIBUTE_BUFFER_LENGTH
, &temp2
);
1122 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_print(pandecode_dump_stream
, &temp2
, 2);
1128 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1130 /* TODO: Decode flags */
1131 mali_ptr shader_ptr
= ptr
& ~15;
1133 char *a
= pointer_as_memory_reference(shader_ptr
);
1134 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1140 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1143 decode_bifrost_constant(u16 constant
)
1145 float lo
= (float) (constant
& 0xFF);
1146 float hi
= (float) (constant
>> 8);
1148 return (hi
/ 255.0) + (lo
/ 65535.0);
1152 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1154 struct bifrost_blend_rt
*b
=
1155 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1157 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1160 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1161 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1162 b
->constant
, decode_bifrost_constant(b
->constant
));
1164 /* TODO figure out blend shader enable bit */
1165 DUMP_CL("Equation", BLEND_EQUATION
, &b
->equation
, 2);
1167 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1168 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1170 pandecode_log(".format = %s", mali_format_as_str(b
->format
));
1171 pandecode_swizzle(b
->swizzle
, b
->format
);
1172 pandecode_log_cont(",\n");
1174 pandecode_prop("swizzle = 0x%" PRIx32
, b
->swizzle
);
1175 pandecode_prop("format = 0x%" PRIx32
, b
->format
);
1178 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1179 pandecode_prop("zero1 = 0x%" PRIx32
, b
->zero1
);
1182 pandecode_log(".shader_type = ");
1183 switch(b
->shader_type
) {
1184 case BIFROST_BLEND_F16
:
1185 pandecode_log_cont("BIFROST_BLEND_F16");
1187 case BIFROST_BLEND_F32
:
1188 pandecode_log_cont("BIFROST_BLEND_F32");
1190 case BIFROST_BLEND_I32
:
1191 pandecode_log_cont("BIFROST_BLEND_I32");
1193 case BIFROST_BLEND_U32
:
1194 pandecode_log_cont("BIFROST_BLEND_U32");
1196 case BIFROST_BLEND_I16
:
1197 pandecode_log_cont("BIFROST_BLEND_I16");
1199 case BIFROST_BLEND_U16
:
1200 pandecode_log_cont("BIFROST_BLEND_U16");
1203 pandecode_log_cont(",\n");
1206 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1207 pandecode_prop("zero2 = 0x%" PRIx32
, b
->zero2
);
1210 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1213 pandecode_log("},\n");
1219 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1221 /* constant/equation is in a union */
1225 pandecode_log(".blend = {\n");
1229 pandecode_shader_address("shader", blend
->shader
);
1231 DUMP_CL("Equation", BLEND_EQUATION
, &blend
->equation
, 2);
1232 pandecode_prop("constant = %f", blend
->constant
);
1236 pandecode_log("},\n");
1238 /* Return blend shader to disassemble if present */
1239 return is_shader
? (blend
->shader
& ~0xF) : 0;
1243 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1245 struct midgard_blend_rt
*b
=
1246 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1248 /* Flags determine presence of blend shader */
1249 bool is_shader
= b
->flags
.opaque
[0] & 0x2;
1251 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1254 DUMP_CL("Flags", BLEND_FLAGS
, &b
->flags
, 2);
1256 union midgard_blend blend
= b
->blend
;
1257 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1260 pandecode_log("};\n");
1265 /* Attributes and varyings have descriptor records, which contain information
1266 * about their format and ordering with the attribute/varying buffers. We'll
1267 * want to validate that the combinations specified are self-consistent.
1271 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1273 const char *prefix
= varying
? "Varying" : "Attribute";
1274 mali_ptr p
= varying
? v
->varying_meta
: v
->attribute_meta
;
1276 for (int i
= 0; i
< count
; ++i
, p
+= MALI_ATTRIBUTE_LENGTH
)
1277 DUMP_ADDR(prefix
, ATTRIBUTE
, p
, 1);
1282 /* return bits [lo, hi) of word */
1284 bits(u32 word
, u32 lo
, u32 hi
)
1287 return word
; // avoid undefined behavior with the shift
1289 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1293 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool graphics
)
1295 pandecode_log(".prefix = {\n");
1298 /* Decode invocation_count. See the comment before the definition of
1299 * invocation_count for an explanation.
1302 unsigned size_y_shift
= bits(p
->invocation_shifts
, 0, 5);
1303 unsigned size_z_shift
= bits(p
->invocation_shifts
, 5, 10);
1304 unsigned workgroups_x_shift
= bits(p
->invocation_shifts
, 10, 16);
1305 unsigned workgroups_y_shift
= bits(p
->invocation_shifts
, 16, 22);
1306 unsigned workgroups_z_shift
= bits(p
->invocation_shifts
, 22, 28);
1307 unsigned workgroups_x_shift_2
= bits(p
->invocation_shifts
, 28, 32);
1309 unsigned size_x
= bits(p
->invocation_count
, 0, size_y_shift
) + 1;
1310 unsigned size_y
= bits(p
->invocation_count
, size_y_shift
, size_z_shift
) + 1;
1311 unsigned size_z
= bits(p
->invocation_count
, size_z_shift
, workgroups_x_shift
) + 1;
1313 unsigned groups_x
= bits(p
->invocation_count
, workgroups_x_shift
, workgroups_y_shift
) + 1;
1314 unsigned groups_y
= bits(p
->invocation_count
, workgroups_y_shift
, workgroups_z_shift
) + 1;
1315 unsigned groups_z
= bits(p
->invocation_count
, workgroups_z_shift
, 32) + 1;
1317 /* Even though we have this decoded, we want to ensure that the
1318 * representation is "unique" so we don't lose anything by printing only
1319 * the final result. More specifically, we need to check that we were
1320 * passed something in canonical form, since the definition per the
1321 * hardware is inherently not unique. How? Well, take the resulting
1322 * decode and pack it ourselves! If it is bit exact with what we
1323 * decoded, we're good to go. */
1325 struct mali_vertex_tiler_prefix ref
;
1326 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, graphics
);
1329 (p
->invocation_count
== ref
.invocation_count
) &&
1330 (p
->invocation_shifts
== ref
.invocation_shifts
);
1333 pandecode_msg("XXX: non-canonical workgroups packing\n");
1334 pandecode_msg("expected: %X, %X",
1335 ref
.invocation_count
,
1336 ref
.invocation_shifts
);
1338 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1339 pandecode_prop("size_y_shift = %d", size_y_shift
);
1340 pandecode_prop("size_z_shift = %d", size_z_shift
);
1341 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift
);
1342 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift
);
1343 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift
);
1344 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2
);
1347 /* Regardless, print the decode */
1348 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1349 size_x
, size_y
, size_z
,
1350 groups_x
, groups_y
, groups_z
);
1353 if (p
->unknown_draw
)
1354 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1356 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1358 if (p
->draw_mode
!= MALI_DRAW_MODE_NONE
)
1359 pandecode_prop("draw_mode = %s", mali_draw_mode_as_str(p
->draw_mode
));
1361 /* Index count only exists for tiler jobs anyway */
1364 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1367 unsigned index_raw_size
= (p
->unknown_draw
& MALI_DRAW_INDEXED_SIZE
);
1368 index_raw_size
>>= MALI_DRAW_INDEXED_SHIFT
;
1370 /* Validate an index buffer is present if we need one. TODO: verify
1371 * relationship between invocation_count and index_count */
1374 unsigned count
= p
->index_count
;
1377 unsigned size
= (index_raw_size
== 0x3) ? 4 : index_raw_size
;
1379 /* Ensure we got a size, and if so, validate the index buffer
1380 * is large enough to hold a full set of indices of the given
1383 if (!index_raw_size
)
1384 pandecode_msg("XXX: index size missing\n");
1386 pandecode_validate_buffer(p
->indices
, count
* size
);
1387 } else if (index_raw_size
)
1388 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size
);
1390 if (p
->offset_bias_correction
)
1391 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1393 /* TODO: Figure out what this is. It's not zero */
1394 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1397 pandecode_log("},\n");
1401 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1403 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1404 uint64_t *PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1406 for (int i
= 0; i
< ubufs_count
; i
++) {
1407 unsigned size
= (ubufs
[i
] & ((1 << 10) - 1)) * 16;
1408 mali_ptr addr
= (ubufs
[i
] >> 10) << 2;
1410 pandecode_validate_buffer(addr
, size
);
1412 char *ptr
= pointer_as_memory_reference(addr
);
1413 pandecode_log("ubuf_%d[%u] = %s;\n", i
, size
, ptr
);
1417 pandecode_log("\n");
1421 pandecode_uniforms(mali_ptr uniforms
, unsigned uniform_count
)
1423 pandecode_validate_buffer(uniforms
, uniform_count
* 16);
1425 char *ptr
= pointer_as_memory_reference(uniforms
);
1426 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count
, ptr
);
1431 shader_type_for_job(unsigned type
)
1434 case MALI_JOB_TYPE_VERTEX
: return "VERTEX";
1435 case MALI_JOB_TYPE_TILER
: return "FRAGMENT";
1436 case MALI_JOB_TYPE_COMPUTE
: return "COMPUTE";
1442 static unsigned shader_id
= 0;
1444 static struct midgard_disasm_stats
1445 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1446 bool is_bifrost
, unsigned gpu_id
)
1448 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1449 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1451 /* Compute maximum possible size */
1452 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1454 /* Print some boilerplate to clearly denote the assembly (which doesn't
1455 * obey indentation rules), and actually do the disassembly! */
1457 pandecode_log_cont("\n\n");
1459 struct midgard_disasm_stats stats
;
1462 disassemble_bifrost(pandecode_dump_stream
, code
, sz
, true);
1464 /* TODO: Extend stats to Bifrost */
1465 stats
.texture_count
= -128;
1466 stats
.sampler_count
= -128;
1467 stats
.attribute_count
= -128;
1468 stats
.varying_count
= -128;
1469 stats
.uniform_count
= -128;
1470 stats
.uniform_buffer_count
= -128;
1471 stats
.work_count
= -128;
1473 stats
.instruction_count
= 0;
1474 stats
.bundle_count
= 0;
1475 stats
.quadword_count
= 0;
1476 stats
.helper_invocations
= false;
1478 stats
= disassemble_midgard(pandecode_dump_stream
,
1480 type
== MALI_JOB_TYPE_TILER
?
1481 MESA_SHADER_FRAGMENT
: MESA_SHADER_VERTEX
);
1484 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1485 * driver-internal purposes with the blob and interfere */
1487 bool should_shaderdb
= type
!= MALI_JOB_TYPE_COMPUTE
;
1489 if (should_shaderdb
) {
1490 unsigned nr_threads
=
1491 (stats
.work_count
<= 4) ? 4 :
1492 (stats
.work_count
<= 8) ? 2 :
1495 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
1496 "%u inst, %u bundles, %u quadwords, "
1497 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1499 shader_type_for_job(type
),
1500 stats
.instruction_count
, stats
.bundle_count
, stats
.quadword_count
,
1501 stats
.work_count
, nr_threads
);
1509 pandecode_texture_payload(mali_ptr payload
,
1510 enum mali_texture_dimension dim
,
1511 enum mali_texture_layout layout
,
1515 uint16_t array_size
,
1516 struct pandecode_mapped_memory
*tmem
)
1518 pandecode_log(".payload = {\n");
1521 /* A bunch of bitmap pointers follow.
1522 * We work out the correct number,
1523 * based on the mipmap/cubemap
1524 * properties, but dump extra
1525 * possibilities to futureproof */
1527 int bitmap_count
= levels
+ 1;
1529 /* Miptree for each face */
1530 if (dim
== MALI_TEXTURE_DIMENSION_CUBE
)
1533 /* Array of layers */
1534 bitmap_count
*= depth
;
1536 /* Array of textures */
1537 bitmap_count
*= array_size
;
1539 /* Stride for each element */
1543 mali_ptr
*pointers_and_strides
= pandecode_fetch_gpu_mem(tmem
,
1544 payload
, sizeof(mali_ptr
) * bitmap_count
);
1545 for (int i
= 0; i
< bitmap_count
; ++i
) {
1546 /* How we dump depends if this is a stride or a pointer */
1548 if (manual_stride
&& (i
& 1)) {
1549 /* signed 32-bit snuck in as a 64-bit pointer */
1550 uint64_t stride_set
= pointers_and_strides
[i
];
1551 uint32_t clamped_stride
= stride_set
;
1552 int32_t stride
= clamped_stride
;
1553 assert(stride_set
== clamped_stride
);
1554 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
1556 char *a
= pointer_as_memory_reference(pointers_and_strides
[i
]);
1557 pandecode_log("%s, \n", a
);
1563 pandecode_log("},\n");
1567 pandecode_texture(mali_ptr u
,
1568 struct pandecode_mapped_memory
*tmem
,
1569 unsigned job_no
, unsigned tex
)
1571 struct pandecode_mapped_memory
*mapped_mem
= pandecode_find_mapped_gpu_mem_containing(u
);
1572 const uint8_t *cl
= pandecode_fetch_gpu_mem(mapped_mem
, u
, MALI_MIDGARD_TEXTURE_LENGTH
);
1574 struct MALI_MIDGARD_TEXTURE temp
;
1575 MALI_MIDGARD_TEXTURE_unpack(cl
, &temp
);
1576 MALI_MIDGARD_TEXTURE_print(pandecode_dump_stream
, &temp
, 2);
1578 pandecode_texture_payload(u
+ MALI_MIDGARD_TEXTURE_LENGTH
,
1579 temp
.dimension
, temp
.texel_ordering
, temp
.manual_stride
,
1580 temp
.levels
, temp
.depth
, temp
.array_size
, mapped_mem
);
1584 pandecode_bifrost_texture(
1589 struct MALI_BIFROST_TEXTURE temp
;
1590 MALI_BIFROST_TEXTURE_unpack(cl
, &temp
);
1591 MALI_BIFROST_TEXTURE_print(pandecode_dump_stream
, &temp
, 2);
1593 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(temp
.surfaces
);
1594 pandecode_texture_payload(temp
.surfaces
, temp
.dimension
, temp
.texel_ordering
,
1595 true, temp
.levels
, 1, 1, tmem
);
1598 /* 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 */
1601 pandecode_shader_prop(const char *name
, unsigned claim
, signed truth
, bool fuzzy
)
1607 if (fuzzy
&& (truth
< 0))
1608 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name
, claim
, truth
);
1610 if ((truth
>= 0) && !fuzzy
) {
1611 pandecode_msg("%s: expected %s = %d, claimed %u\n",
1612 (truth
< claim
) ? "warn" : "XXX",
1613 name
, truth
, claim
);
1614 } else if ((claim
> -truth
) && !fuzzy
) {
1615 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
1616 name
, -truth
, claim
);
1617 } else if (fuzzy
&& (claim
< truth
))
1618 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
1619 name
, truth
, claim
);
1621 pandecode_log(".%s = %" PRId16
, name
, claim
);
1624 pandecode_log_cont(" /* %u used */", truth
);
1626 pandecode_log_cont(",\n");
1630 pandecode_blend_shader_disassemble(mali_ptr shader
, int job_no
, int job_type
,
1631 bool is_bifrost
, unsigned gpu_id
)
1633 struct midgard_disasm_stats stats
=
1634 pandecode_shader_disassemble(shader
, job_no
, job_type
, is_bifrost
, gpu_id
);
1636 bool has_texture
= (stats
.texture_count
> 0);
1637 bool has_sampler
= (stats
.sampler_count
> 0);
1638 bool has_attribute
= (stats
.attribute_count
> 0);
1639 bool has_varying
= (stats
.varying_count
> 0);
1640 bool has_uniform
= (stats
.uniform_count
> 0);
1641 bool has_ubo
= (stats
.uniform_buffer_count
> 0);
1643 if (has_texture
|| has_sampler
)
1644 pandecode_msg("XXX: blend shader accessing textures\n");
1646 if (has_attribute
|| has_varying
)
1647 pandecode_msg("XXX: blend shader accessing interstage\n");
1649 if (has_uniform
|| has_ubo
)
1650 pandecode_msg("XXX: blend shader accessing uniforms\n");
1654 pandecode_textures(mali_ptr textures
, unsigned texture_count
, int job_no
, bool is_bifrost
)
1656 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(textures
);
1661 pandecode_log("Textures (%"PRIx64
"):\n", textures
);
1664 const void *cl
= pandecode_fetch_gpu_mem(mmem
,
1665 textures
, MALI_BIFROST_TEXTURE_LENGTH
*
1668 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
1669 pandecode_bifrost_texture(cl
+
1670 MALI_BIFROST_TEXTURE_LENGTH
* tex
,
1674 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
);
1676 for (int tex
= 0; tex
< texture_count
; ++tex
) {
1677 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
1678 char *a
= pointer_as_memory_reference(*u
);
1679 pandecode_log("%s,\n", a
);
1683 /* Now, finally, descend down into the texture descriptor */
1684 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
1685 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
1686 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
1688 pandecode_texture(*u
, tmem
, job_no
, tex
);
1694 pandecode_samplers(mali_ptr samplers
, unsigned sampler_count
, int job_no
, bool is_bifrost
)
1696 for (int i
= 0; i
< sampler_count
; ++i
) {
1698 DUMP_ADDR("Sampler", BIFROST_SAMPLER
, samplers
+ (MALI_BIFROST_SAMPLER_LENGTH
* i
), 1);
1700 DUMP_ADDR("Sampler", MIDGARD_SAMPLER
, samplers
+ (MALI_MIDGARD_SAMPLER_LENGTH
* i
), 1);
1706 pandecode_vertex_tiler_postfix_pre(
1707 const struct mali_vertex_tiler_postfix
*p
,
1708 int job_no
, enum mali_job_type job_type
,
1709 char *suffix
, bool is_bifrost
, unsigned gpu_id
)
1711 struct pandecode_mapped_memory
*attr_mem
;
1713 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1714 * are the only things actually needed from the FBD, vertex/tiler jobs
1715 * no longer reference the FBD -- instead, this field points to some
1716 * info about the scratchpad.
1719 struct pandecode_fbd fbd_info
= {
1720 /* Default for Bifrost */
1725 pandecode_log_cont("\t/* %X %/\n", p
->shared_memory
& 1);
1726 pandecode_compute_fbd(p
->shared_memory
& ~1, job_no
);
1727 } else if (p
->shared_memory
& MALI_MFBD
)
1728 fbd_info
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->shared_memory
) & FBD_MASK
, job_no
, false, job_type
== MALI_JOB_TYPE_COMPUTE
, false);
1729 else if (job_type
== MALI_JOB_TYPE_COMPUTE
)
1730 pandecode_compute_fbd((u64
) (uintptr_t) p
->shared_memory
, job_no
);
1732 fbd_info
= pandecode_sfbd((u64
) (uintptr_t) p
->shared_memory
, job_no
, false, gpu_id
);
1734 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
1735 int texture_count
= 0, sampler_count
= 0;
1738 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->shader
);
1739 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, p
->shader
);
1741 /* Disassemble ahead-of-time to get stats. Initialize with
1742 * stats for the missing-shader case so we get validation
1745 struct midgard_disasm_stats info
= {
1748 .attribute_count
= 0,
1752 .uniform_count
= -128,
1753 .uniform_buffer_count
= 0
1756 if (s
->shader
& ~0xF)
1757 info
= pandecode_shader_disassemble(s
->shader
& ~0xF, job_no
, job_type
, is_bifrost
, gpu_id
);
1759 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", p
->shader
, job_no
, suffix
);
1762 /* Save for dumps */
1763 attribute_count
= s
->attribute_count
;
1764 varying_count
= s
->varying_count
;
1765 texture_count
= s
->texture_count
;
1766 sampler_count
= s
->sampler_count
;
1769 uniform_count
= s
->bifrost2
.uniform_count
;
1770 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
1772 uniform_count
= s
->midgard1
.uniform_count
;
1773 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
1776 pandecode_shader_address("shader", s
->shader
);
1778 pandecode_shader_prop("texture_count", s
->texture_count
, info
.texture_count
, false);
1779 pandecode_shader_prop("sampler_count", s
->sampler_count
, info
.sampler_count
, false);
1780 pandecode_shader_prop("attribute_count", s
->attribute_count
, info
.attribute_count
, false);
1781 pandecode_shader_prop("varying_count", s
->varying_count
, info
.varying_count
, false);
1782 pandecode_shader_prop("uniform_buffer_count",
1783 uniform_buffer_count
,
1784 info
.uniform_buffer_count
, true);
1787 pandecode_shader_prop("uniform_count",
1789 info
.uniform_count
, false);
1791 pandecode_shader_prop("work_count",
1792 s
->midgard1
.work_count
, info
.work_count
, false);
1796 pandecode_log("bifrost1.unk1 = ");
1797 pandecode_log_decoded_flags(shader_bifrost_info
, s
->bifrost1
.unk1
);
1798 pandecode_log_cont(",\n");
1800 bool helpers
= s
->midgard1
.flags_lo
& MALI_HELPER_INVOCATIONS
;
1802 if (helpers
!= info
.helper_invocations
) {
1803 pandecode_msg("XXX: expected helpers %u but got %u\n",
1804 info
.helper_invocations
, helpers
);
1807 pandecode_log(".midgard1.flags_lo = ");
1808 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info
,
1809 s
->midgard1
.flags_lo
& ~MALI_HELPER_INVOCATIONS
);
1810 pandecode_log_cont(",\n");
1812 pandecode_log(".midgard1.flags_hi = ");
1813 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info
, s
->midgard1
.flags_hi
);
1814 pandecode_log_cont(",\n");
1817 if (s
->depth_units
|| s
->depth_factor
) {
1818 pandecode_prop("depth_factor = %f", s
->depth_factor
);
1819 pandecode_prop("depth_units = %f", s
->depth_units
);
1822 if (s
->coverage_mask
)
1823 pandecode_prop("coverage_mask = 0x%X", s
->coverage_mask
);
1826 pandecode_prop(".unknown2_2 = %X", s
->unknown2_2
);
1828 if (s
->unknown2_3
|| s
->unknown2_4
) {
1829 pandecode_log(".unknown2_3 = ");
1831 int unknown2_3
= s
->unknown2_3
;
1832 int unknown2_4
= s
->unknown2_4
;
1834 /* We're not quite sure what these flags mean without the depth test, if anything */
1836 if (unknown2_3
& (MALI_DEPTH_WRITEMASK
| MALI_DEPTH_FUNC_MASK
)) {
1837 const char *func
= mali_func_as_str(MALI_GET_DEPTH_FUNC(unknown2_3
));
1838 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
1840 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
1843 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
1844 pandecode_log_cont(",\n");
1846 pandecode_log(".unknown2_4 = ");
1847 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
1848 pandecode_log_cont(",\n");
1851 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
1852 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
1853 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
1856 DUMP_CL("Stencil front", STENCIL
, &s
->stencil_front
, 1);
1857 DUMP_CL("Stencil back", STENCIL
, &s
->stencil_back
, 1);
1860 pandecode_log(".bifrost2 = {\n");
1863 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
1864 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
1865 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
1866 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
1869 pandecode_log("},\n");
1870 } else if (s
->midgard2
.unknown2_7
) {
1871 pandecode_log(".midgard2 = {\n");
1874 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
1876 pandecode_log("},\n");
1880 pandecode_msg("XXX: shader padding tripped\n");
1881 pandecode_prop("padding = 0x%" PRIx32
, s
->padding
);
1885 /* TODO: Blend shaders routing/disasm */
1886 union midgard_blend blend
= s
->blend
;
1887 mali_ptr shader
= pandecode_midgard_blend(&blend
, s
->unknown2_3
& MALI_HAS_BLEND_SHADER
);
1889 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
1891 pandecode_msg("mdg_blend = %" PRIx64
"\n", s
->blend
.shader
);
1895 pandecode_log("};\n");
1897 /* MRT blend fields are used whenever MFBD is used, with
1898 * per-RT descriptors */
1900 if (job_type
== MALI_JOB_TYPE_TILER
&& (is_bifrost
|| p
->shared_memory
& MALI_MFBD
)) {
1901 void* blend_base
= (void *) (s
+ 1);
1903 for (unsigned i
= 0; i
< fbd_info
.rt_count
; i
++) {
1904 mali_ptr shader
= 0;
1907 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
1909 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
1912 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
1917 pandecode_msg("XXX: missing shader descriptor\n");
1920 DUMP_ADDR("Viewport", VIEWPORT
, p
->viewport
, 1);
1922 unsigned max_attr_index
= 0;
1924 if (p
->attribute_meta
)
1925 max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
1927 if (p
->attributes
) {
1928 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
1929 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false, job_type
);
1932 /* Varyings are encoded like attributes but not actually sent; we just
1933 * pass a zero buffer with the right stride/size set, (or whatever)
1934 * since the GPU will write to it itself */
1936 if (p
->varying_meta
) {
1937 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
1941 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
1943 /* Number of descriptors depends on whether there are
1944 * non-internal varyings */
1946 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true, job_type
);
1949 if (p
->uniform_buffers
) {
1950 if (uniform_buffer_count
)
1951 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
1953 pandecode_msg("warn: UBOs specified but not referenced\n");
1954 } else if (uniform_buffer_count
)
1955 pandecode_msg("XXX: UBOs referenced but not specified\n");
1957 /* We don't want to actually dump uniforms, but we do need to validate
1958 * that the counts we were given are sane */
1962 pandecode_uniforms(p
->uniforms
, uniform_count
);
1964 pandecode_msg("warn: Uniforms specified but not referenced\n");
1965 } else if (uniform_count
)
1966 pandecode_msg("XXX: Uniforms referenced but not specified\n");
1969 pandecode_textures(p
->textures
, texture_count
, job_no
, is_bifrost
);
1971 if (p
->sampler_descriptor
)
1972 pandecode_samplers(p
->sampler_descriptor
, sampler_count
, job_no
, is_bifrost
);
1976 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
1978 pandecode_log(".gl_enables = ");
1980 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
1982 pandecode_log_cont(",\n");
1986 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
1988 if (p
->shader
& 0xF)
1989 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p
->shader
& 0xF));
1991 pandecode_log(".postfix = {\n");
1994 pandecode_gl_enables(p
->gl_enables
, MALI_JOB_TYPE_TILER
);
1995 pandecode_prop("instance_shift = 0x%x", p
->instance_shift
);
1996 pandecode_prop("instance_odd = 0x%x", p
->instance_odd
);
1999 pandecode_msg("XXX: vertex only zero tripped");
2000 pandecode_prop("zero4 = 0x%" PRIx32
, p
->zero4
);
2003 pandecode_prop("offset_start = 0x%x", p
->offset_start
);
2006 pandecode_msg("XXX: vertex only zero tripped");
2007 pandecode_prop("zero5 = 0x%" PRIx64
, p
->zero5
);
2010 MEMORY_PROP(p
, position_varying
);
2011 MEMORY_PROP(p
, occlusion_counter
);
2014 pandecode_log("},\n");
2018 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2020 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2021 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2023 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2027 pandecode_msg("XXX: tiler heap zero tripped\n");
2028 pandecode_prop("zero = 0x%x", h
->zero
);
2031 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2032 MEMORY_PROP(h
, tiler_heap_start
);
2033 MEMORY_PROP(h
, tiler_heap_free
);
2035 /* this might point to the beginning of another buffer, when it's
2036 * really the end of the tiler heap buffer, so we have to be careful
2037 * here. but for zero length, we need the same pointer.
2040 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2041 MEMORY_PROP(h
, tiler_heap_start
);
2043 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2044 pandecode_prop("tiler_heap_end = %s + 1", a
);
2048 for (int i
= 0; i
< 10; i
++) {
2049 if (h
->zeros
[i
] != 0) {
2050 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2055 if (h
->unk1
!= 0x1) {
2056 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2057 pandecode_prop("unk1 = 0x%x", h
->unk1
);
2060 if (h
->unk7e007e
!= 0x7e007e) {
2061 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
2062 pandecode_prop("unk7e007e = 0x%x", h
->unk7e007e
);
2066 pandecode_log("};\n");
2070 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2072 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2073 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2075 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2077 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2080 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32
, t
->tiler_heap_next_start
);
2081 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32
, t
->used_hierarchy_mask
);
2083 if (t
->hierarchy_mask
!= 0xa &&
2084 t
->hierarchy_mask
!= 0x14 &&
2085 t
->hierarchy_mask
!= 0x28 &&
2086 t
->hierarchy_mask
!= 0x50 &&
2087 t
->hierarchy_mask
!= 0xa0)
2088 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
2090 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2092 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2094 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2095 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2098 pandecode_msg("XXX: tiler meta zero tripped\n");
2099 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2102 for (int i
= 0; i
< 12; i
++) {
2103 if (t
->zeros
[i
] != 0) {
2104 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2110 pandecode_log("};\n");
2114 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2116 if (u
.pointer
== 0x0)
2119 pandecode_log(".primitive_size = {\n");
2123 pandecode_prop("constant = %f", u
.constant
);
2125 MEMORY_PROP((&u
), pointer
);
2129 pandecode_log("},\n");
2133 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2135 pandecode_log_cont("{\n");
2138 /* TODO: gl_PointSize on Bifrost */
2139 pandecode_primitive_size(t
->primitive_size
, true);
2141 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2143 pandecode_msg("XXX: tiler only zero tripped\n");
2144 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2145 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2146 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2147 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2148 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2149 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2153 pandecode_log("},\n");
2157 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2158 const struct pandecode_mapped_memory
*mem
,
2159 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2161 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2163 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2165 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2168 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2169 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2172 pandecode_log("};\n");
2178 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2179 const struct pandecode_mapped_memory
*mem
,
2180 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2182 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2184 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2185 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2187 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2190 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2192 pandecode_log(".tiler = ");
2193 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2195 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2198 pandecode_log("};\n");
2204 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2205 const struct pandecode_mapped_memory
*mem
,
2206 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2208 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2209 bool is_graphics
= (h
->job_type
== MALI_JOB_TYPE_VERTEX
) || (h
->job_type
== MALI_JOB_TYPE_TILER
);
2211 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false, gpu_id
);
2213 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2216 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, is_graphics
);
2217 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2219 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2220 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2223 pandecode_log("};\n");
2229 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2230 mali_ptr payload
, int job_no
,
2231 bool is_bifrost
, unsigned gpu_id
)
2233 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2235 bool is_mfbd
= s
->framebuffer
& MALI_MFBD
;
2237 if (!is_mfbd
&& is_bifrost
)
2238 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2240 struct pandecode_fbd info
;
2243 info
= pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true, false, is_bifrost
);
2245 info
= pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true, gpu_id
);
2247 /* Compute the tag for the tagged pointer. This contains the type of
2248 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2249 * additional structures follow the MFBD header (an extra payload or
2250 * not, as well as a count of render targets) */
2252 unsigned expected_tag
= is_mfbd
? MALI_MFBD
: 0;
2256 expected_tag
|= MALI_MFBD_TAG_EXTRA
;
2258 expected_tag
|= (MALI_POSITIVE(info
.rt_count
) << 2);
2261 if ((s
->min_tile_coord
| s
->max_tile_coord
) & ~(MALI_X_COORD_MASK
| MALI_Y_COORD_MASK
)) {
2262 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2263 pandecode_prop("min_tile_coord = 0x%X\n", s
->min_tile_coord
);
2264 pandecode_prop("max_tile_coord = 0x%X\n", s
->max_tile_coord
);
2267 /* Extract tile coordinates */
2269 unsigned min_x
= MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2270 unsigned min_y
= MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2272 unsigned max_x
= (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2273 unsigned max_y
= (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2275 /* For the max, we also want the floored (rather than ceiled) version for checking */
2277 unsigned max_x_f
= (MALI_TILE_COORD_X(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2278 unsigned max_y_f
= (MALI_TILE_COORD_Y(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2280 /* Validate the coordinates are well-ordered */
2283 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2284 else if (min_x
> max_x
)
2285 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2288 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2289 else if (min_y
> max_y
)
2290 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2292 /* Validate the coordinates fit inside the framebuffer. We use floor,
2293 * rather than ceil, for the max coordinates, since the tile
2294 * coordinates for something like an 800x600 framebuffer will actually
2295 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2297 if ((min_x
> info
.width
) || (max_x_f
> info
.width
))
2298 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2300 if ((min_y
> info
.height
) || (max_y_f
> info
.height
))
2301 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2303 /* After validation, we print */
2305 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x
, min_y
, max_x
, max_y
);
2307 /* The FBD is a tagged pointer */
2309 unsigned tag
= (s
->framebuffer
& ~FBD_MASK
);
2311 if (tag
!= expected_tag
)
2312 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag
, tag
);
2317 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
2318 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
2319 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
2320 * might wish to combine this with the bifrost parameter) because some details
2321 * are model-specific even within a particular architecture. Minimal traces
2322 * *only* examine the job descriptors, skipping printing entirely if there is
2323 * no faults, and only descends into the payload if there are faults. This is
2324 * useful for looking for faults without the overhead of invasive traces. */
2327 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
, unsigned gpu_id
, bool minimal
)
2329 pandecode_dump_file_open();
2331 struct mali_job_descriptor_header
*h
;
2332 unsigned job_descriptor_number
= 0;
2335 struct pandecode_mapped_memory
*mem
=
2336 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2340 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2342 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
);
2343 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
, 256);
2345 int job_no
= job_descriptor_number
++;
2347 /* If the job is good to go, skip it in minimal mode */
2348 if (minimal
&& (h
->exception_status
== 0x0 || h
->exception_status
== 0x1))
2351 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2354 pandecode_prop("job_type = %s", mali_job_type_as_str(h
->job_type
));
2356 if (h
->job_descriptor_size
)
2357 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2359 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2360 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2361 h
->exception_status
,
2362 (h
->exception_status
>> 16) & 0xFFFF,
2363 mali_exception_access_as_str((h
->exception_status
>> 8) & 0x3),
2364 h
->exception_status
& 0xFF);
2366 if (h
->first_incomplete_task
)
2367 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2369 if (h
->fault_pointer
)
2370 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2373 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2375 pandecode_prop("job_index = %d", h
->job_index
);
2377 if (h
->unknown_flags
)
2378 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2380 if (h
->job_dependency_index_1
)
2381 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2383 if (h
->job_dependency_index_2
)
2384 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2387 pandecode_log("};\n");
2389 switch (h
->job_type
) {
2390 case MALI_JOB_TYPE_WRITE_VALUE
: {
2391 struct mali_payload_write_value
*s
= payload
;
2392 pandecode_log("struct mali_payload_write_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2394 MEMORY_PROP(s
, address
);
2396 if (s
->value_descriptor
!= MALI_WRITE_VALUE_ZERO
) {
2397 pandecode_msg("XXX: unknown value descriptor\n");
2398 pandecode_prop("value_descriptor = 0x%" PRIX32
, s
->value_descriptor
);
2402 pandecode_msg("XXX: set value tripped\n");
2403 pandecode_prop("reserved = 0x%" PRIX32
, s
->reserved
);
2406 pandecode_prop("immediate = 0x%" PRIX64
, s
->immediate
);
2408 pandecode_log("};\n");
2413 case MALI_JOB_TYPE_TILER
:
2414 case MALI_JOB_TYPE_VERTEX
:
2415 case MALI_JOB_TYPE_COMPUTE
:
2417 if (h
->job_type
== MALI_JOB_TYPE_TILER
)
2418 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2420 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2422 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2426 case MALI_JOB_TYPE_FRAGMENT
:
2427 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
, gpu_id
);
2433 } while ((jc_gpu_va
= h
->next_job
));
2435 pandecode_map_read_write();