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_MFBD_##flag, "MALI_MFBD_" #flag }
283 static const struct pandecode_flag_info mfbd_flag_info
[] = {
284 FLAG_INFO(DEPTH_WRITE
),
290 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
291 static const struct pandecode_flag_info sfbd_unk1_info
[] = {
298 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
299 static const struct pandecode_flag_info sfbd_unk2_info
[] = {
306 /* Midgard's tiler descriptor is embedded within the
310 pandecode_midgard_tiler_descriptor(
311 const struct midgard_tiler_descriptor
*t
,
317 pandecode_log(".tiler = {\n");
320 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
321 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
323 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
325 /* We know this name from the kernel, but we never see it nonzero */
328 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
330 MEMORY_PROP(t
, polygon_list
);
332 /* The body is offset from the base of the polygon list */
333 //assert(t->polygon_list_body > t->polygon_list);
334 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
336 /* It needs to fit inside the reported size */
337 //assert(t->polygon_list_size >= body_offset);
339 /* Now that we've sanity checked, we'll try to calculate the sizes
340 * ourselves for comparison */
342 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
343 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
345 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
346 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
347 ref_header
, ref_size
);
348 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
349 pandecode_msg("body offset %d\n", body_offset
);
352 /* The tiler heap has a start and end specified -- it should be
353 * identical to what we have in the BO. The exception is if tiling is
356 MEMORY_PROP(t
, heap_start
);
357 assert(t
->heap_end
>= t
->heap_start
);
359 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
361 /* Tiling is enabled with a special flag */
362 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
363 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
365 bool tiling_enabled
= hierarchy_mask
;
367 if (tiling_enabled
) {
368 /* We should also have no other flags */
370 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
372 /* When tiling is disabled, we should have that flag and no others */
374 if (tiler_flags
!= MALI_TILER_DISABLED
) {
375 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
379 /* We should also have an empty heap */
381 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
385 /* Disabled tiling is used only for clear-only jobs, which are
386 * purely FRAGMENT, so we should never see this for
387 * non-FRAGMENT descriptors. */
390 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
393 /* We've never seen weights used in practice, but we know from the
394 * kernel these fields is there */
396 bool nonzero_weights
= false;
398 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
399 nonzero_weights
|= t
->weights
[w
] != 0x0;
402 if (nonzero_weights
) {
403 pandecode_log(".weights = { ");
405 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
406 pandecode_log_cont("%d, ", t
->weights
[w
]);
413 pandecode_log("}\n");
416 /* TODO: The Bifrost tiler is not understood at all yet */
419 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer
*fb
)
421 pandecode_log(".tiler = {\n");
424 MEMORY_PROP(fb
, tiler_meta
);
426 for (int i
= 0; i
< 16; i
++) {
427 if (fb
->zeros
[i
] != 0) {
428 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
433 pandecode_log("},\n");
436 pandecode_log("}\n");
440 /* Information about the framebuffer passed back for
441 * additional analysis */
443 struct pandecode_fbd
{
451 pandecode_sfbd_format(struct mali_sfbd_format format
)
453 pandecode_log(".format = {\n");
456 pandecode_log(".unk1 = ");
457 pandecode_log_decoded_flags(sfbd_unk1_info
, format
.unk1
);
458 pandecode_log_cont(",\n");
460 /* TODO: Map formats so we can check swizzles and print nicely */
461 pandecode_log("swizzle");
462 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
463 pandecode_log_cont(",\n");
465 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
466 (format
.nr_channels
+ 1));
468 pandecode_log(".unk2 = ");
469 pandecode_log_decoded_flags(sfbd_unk2_info
, format
.unk2
);
470 pandecode_log_cont(",\n");
472 pandecode_prop("block = %s", mali_block_format_as_str(format
.block
));
474 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
477 pandecode_log("},\n");
481 pandecode_shared_memory(const struct mali_shared_memory
*desc
, bool is_compute
)
483 pandecode_prop("stack_shift = 0x%x", desc
->stack_shift
);
486 pandecode_prop("unk0 = 0x%x", desc
->unk0
);
488 if (desc
->shared_workgroup_count
!= 0x1F) {
489 pandecode_prop("shared_workgroup_count = %d", desc
->shared_workgroup_count
);
491 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
494 if (desc
->shared_unk1
|| desc
->shared_shift
) {
495 pandecode_prop("shared_unk1 = %X", desc
->shared_unk1
);
496 pandecode_prop("shared_shift = %X", desc
->shared_shift
);
499 pandecode_msg("XXX: shared memory configured in noncompute shader");
502 if (desc
->shared_zero
) {
503 pandecode_msg("XXX: shared memory zero tripped\n");
504 pandecode_prop("shared_zero = 0x%" PRIx32
, desc
->shared_zero
);
507 if (desc
->shared_memory
&& !is_compute
)
508 pandecode_msg("XXX: shared memory used in noncompute shader\n");
510 MEMORY_PROP(desc
, scratchpad
);
511 MEMORY_PROP(desc
, shared_memory
);
512 MEMORY_PROP(desc
, unknown1
);
514 if (desc
->scratchpad
) {
515 struct pandecode_mapped_memory
*smem
=
516 pandecode_find_mapped_gpu_mem_containing(desc
->scratchpad
);
518 pandecode_msg("scratchpad size %u\n", smem
->length
);
523 static struct pandecode_fbd
524 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
, unsigned gpu_id
)
526 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
527 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
529 struct pandecode_fbd info
= {
534 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
537 pandecode_log(".shared_memory = {\n");
539 pandecode_shared_memory(&s
->shared_memory
, false);
541 pandecode_log("},\n");
543 pandecode_sfbd_format(s
->format
);
545 info
.width
= s
->width
+ 1;
546 info
.height
= s
->height
+ 1;
548 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", info
.width
);
549 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", info
.height
);
551 MEMORY_PROP(s
, checksum
);
553 if (s
->checksum_stride
)
554 pandecode_prop("checksum_stride = %d", s
->checksum_stride
);
556 MEMORY_PROP(s
, framebuffer
);
557 pandecode_prop("stride = %d", s
->stride
);
559 /* Earlier in the actual commandstream -- right before width -- but we
560 * delay to flow nicer */
562 pandecode_log(".clear_flags = ");
563 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
564 pandecode_log_cont(",\n");
566 if (s
->depth_buffer
) {
567 MEMORY_PROP(s
, depth_buffer
);
568 pandecode_prop("depth_stride = %d", s
->depth_stride
);
571 if (s
->stencil_buffer
) {
572 MEMORY_PROP(s
, stencil_buffer
);
573 pandecode_prop("stencil_stride = %d", s
->stencil_stride
);
576 if (s
->depth_stride_zero
||
577 s
->stencil_stride_zero
||
578 s
->zero7
|| s
->zero8
) {
579 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
580 pandecode_prop("depth_stride_zero = 0x%x",
581 s
->depth_stride_zero
);
582 pandecode_prop("stencil_stride_zero = 0x%x",
583 s
->stencil_stride_zero
);
584 pandecode_prop("zero7 = 0x%" PRIx32
,
586 pandecode_prop("zero8 = 0x%" PRIx32
,
590 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
591 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
592 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
593 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
594 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
597 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
598 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
599 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
600 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
601 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
604 if (s
->clear_stencil
) {
605 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
608 const struct midgard_tiler_descriptor t
= s
->tiler
;
610 bool has_hierarchy
= !(gpu_id
== 0x0720 || gpu_id
== 0x0820 || gpu_id
== 0x0830);
611 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
, has_hierarchy
);
614 pandecode_log("};\n");
616 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
617 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
618 pandecode_prop("zero5 = 0x%" PRIx32
, s
->zero5
);
620 pandecode_log_cont(".zero3 = {");
622 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
623 pandecode_log_cont("%X, ", s
->zero3
[i
]);
625 pandecode_log_cont("},\n");
627 pandecode_log_cont(".zero6 = {");
629 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
630 pandecode_log_cont("%X, ", s
->zero6
[i
]);
632 pandecode_log_cont("},\n");
638 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
640 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
641 const struct mali_shared_memory
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
643 pandecode_log("struct mali_shared_memory shared_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
645 pandecode_shared_memory(s
, true);
647 pandecode_log("},\n");
650 /* Extracts the number of components associated with a Mali format */
653 pandecode_format_component_count(enum mali_format fmt
)
655 /* Mask out the format class */
656 unsigned top
= fmt
& 0b11100000;
659 case MALI_FORMAT_SNORM
:
660 case MALI_FORMAT_UINT
:
661 case MALI_FORMAT_UNORM
:
662 case MALI_FORMAT_SINT
:
663 return ((fmt
>> 3) & 3) + 1;
670 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
673 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
676 assert(MALI_CHANNEL_R
== 0);
678 for (unsigned c
= 0; c
< 4; ++c
) {
679 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
681 if (chan
<= MALI_CHANNEL_A
)
688 /* Validates that a (format, swizzle) pair is valid, in the sense that the
689 * swizzle doesn't access any components that are undefined in the format.
690 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
694 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
696 unsigned nr_comp
= pandecode_format_component_count(fmt
);
697 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
698 unsigned valid_mask
= (1 << nr_comp
) - 1;
699 unsigned invalid_mask
= ~valid_mask
;
701 if (access_mask
& invalid_mask
) {
702 pandecode_msg("XXX: invalid components accessed\n");
706 /* Check for the default non-swizzling swizzle so we can suppress
707 * useless printing for the defaults */
709 unsigned default_swizzles
[4] = {
710 MALI_CHANNEL_R
| (MALI_CHANNEL_0
<< 3) | (MALI_CHANNEL_0
<< 6) | (MALI_CHANNEL_1
<< 9),
711 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_0
<< 6) | (MALI_CHANNEL_1
<< 9),
712 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_B
<< 6) | (MALI_CHANNEL_1
<< 9),
713 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_B
<< 6) | (MALI_CHANNEL_A
<< 9)
716 return (swizzle
== default_swizzles
[nr_comp
- 1]);
720 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
722 /* First, do some validation */
723 bool trivial_swizzle
= pandecode_validate_format_swizzle(
729 /* Next, print the swizzle */
730 pandecode_log_cont(".");
732 static const char components
[] = "rgba01";
734 for (unsigned c
= 0; c
< 4; ++c
) {
735 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
737 if (chan
> MALI_CHANNEL_1
) {
738 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
741 pandecode_log_cont("%c", components
[chan
]);
746 pandecode_rt_format(struct mali_rt_format format
)
748 pandecode_log(".format = {\n");
751 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
752 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
753 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
754 pandecode_prop("unk4 = 0x%" PRIx32
, format
.unk4
);
756 pandecode_prop("block = %s", mali_block_format_as_str(format
.block
));
758 /* TODO: Map formats so we can check swizzles and print nicely */
759 pandecode_log("swizzle");
760 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
761 pandecode_log_cont(",\n");
763 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
764 (format
.nr_channels
+ 1));
766 pandecode_log(".flags = ");
767 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
768 pandecode_log_cont(",\n");
770 pandecode_prop("msaa = %s", mali_msaa_as_str(format
.msaa
));
772 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
773 * which is a faster hardware-based alternative to the wallpaper method
774 * to preserve framebuffer contents across frames. In practice, MFBD
775 * preload is buggy on Midgard, and so this is a chicken bit. If this
776 * bit isn't set, most likely something broke unrelated to preload */
778 if (!format
.no_preload
) {
779 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
780 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
784 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
787 pandecode_log("},\n");
791 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct mali_framebuffer
*fb
)
793 pandecode_log("struct mali_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
796 for (int i
= 0; i
< (fb
->rt_count_1
+ 1); i
++) {
797 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct mali_render_target
);
798 struct pandecode_mapped_memory
*mem
=
799 pandecode_find_mapped_gpu_mem_containing(rt_va
);
800 const struct mali_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
802 pandecode_log("{\n");
805 pandecode_rt_format(rt
->format
);
807 if (rt
->format
.block
== MALI_BLOCK_FORMAT_AFBC
) {
808 pandecode_log(".afbc = {\n");
811 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
812 pandecode_prop("metadata = %s", a
);
815 pandecode_prop("stride = %d", rt
->afbc
.stride
);
817 pandecode_log(".flags = ");
818 pandecode_log_decoded_flags(afbc_fmt_flag_info
, rt
->afbc
.flags
);
819 pandecode_log_cont(",\n");
822 pandecode_log("},\n");
823 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.flags
) {
824 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
830 MEMORY_PROP(rt
, framebuffer
);
831 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
833 if (rt
->layer_stride
)
834 pandecode_prop("layer_stride = %d", rt
->layer_stride
);
836 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
837 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
838 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
839 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
840 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
843 if (rt
->zero1
|| rt
->zero2
) {
844 pandecode_msg("XXX: render target zeros tripped\n");
845 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
846 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
850 pandecode_log("},\n");
854 pandecode_log("};\n");
857 static struct pandecode_fbd
858 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
, bool is_compute
, bool is_bifrost
)
860 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
861 const struct mali_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
863 struct pandecode_fbd info
;
865 if (is_bifrost
&& fb
->msaa
.sample_locations
) {
866 /* The blob stores all possible sample locations in a single buffer
867 * allocated on startup, and just switches the pointer when switching
868 * MSAA state. For now, we just put the data into the cmdstream, but we
869 * should do something like what the blob does with a real driver.
871 * There seem to be 32 slots for sample locations, followed by another
872 * 16. The second 16 is just the center location followed by 15 zeros
873 * in all the cases I've identified (maybe shader vs. depth/color
877 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->msaa
.sample_locations
);
879 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->msaa
.sample_locations
);
881 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
884 for (int i
= 0; i
< 32 + 16; i
++) {
885 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
889 pandecode_log("};\n");
892 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
896 pandecode_log(".msaa = {\n");
899 if (fb
->msaa
.sample_locations
)
900 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
902 pandecode_msg("XXX: sample_locations missing\n");
904 if (fb
->msaa
.zero1
|| fb
->msaa
.zero2
|| fb
->msaa
.zero4
) {
905 pandecode_msg("XXX: multisampling zero tripped\n");
906 pandecode_prop("zero1 = %" PRIx64
, fb
->msaa
.zero1
);
907 pandecode_prop("zero2 = %" PRIx64
, fb
->msaa
.zero2
);
908 pandecode_prop("zero4 = %" PRIx64
, fb
->msaa
.zero4
);
912 pandecode_log("},\n");
914 pandecode_log(".shared_memory = {\n");
916 pandecode_shared_memory(&fb
->shared_memory
, is_compute
);
918 pandecode_log("},\n");
921 info
.width
= fb
->width1
+ 1;
922 info
.height
= fb
->height1
+ 1;
923 info
.rt_count
= fb
->rt_count_1
+ 1;
925 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
926 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
927 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
928 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
930 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
931 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
932 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
933 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
935 pandecode_log(".mfbd_flags = ");
936 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
937 pandecode_log_cont(",\n");
939 if (fb
->clear_stencil
)
940 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
943 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
947 pandecode_bifrost_tiler_descriptor(fb
);
949 const struct midgard_tiler_descriptor t
= fb
->tiler
;
950 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
, true);
953 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
955 if (fb
->zero3
|| fb
->zero4
) {
956 pandecode_msg("XXX: framebuffer zeros tripped\n");
957 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
958 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
962 pandecode_log("};\n");
964 gpu_va
+= sizeof(struct mali_framebuffer
);
966 info
.has_extra
= (fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
;
968 if (info
.has_extra
) {
969 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
970 const struct mali_framebuffer_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
972 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
975 MEMORY_PROP(fbx
, checksum
);
977 if (fbx
->checksum_stride
)
978 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
980 pandecode_log(".flags_hi = ");
981 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info
, fbx
->flags_hi
);
982 pandecode_log_cont(",\n");
984 pandecode_log(".flags_lo = ");
985 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info
, fbx
->flags_lo
);
986 pandecode_log_cont(",\n");
988 pandecode_prop("zs_block = %s", mali_block_format_as_str(fbx
->zs_block
));
989 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx
->zs_samples
+ 1);
991 if (fbx
->zs_block
== MALI_BLOCK_FORMAT_AFBC
) {
992 pandecode_log(".ds_afbc = {\n");
995 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
996 pandecode_prop("depth_stencil_afbc_stride = %d",
997 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
998 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
1000 pandecode_log(".flags = ");
1001 pandecode_log_decoded_flags(afbc_fmt_flag_info
, fbx
->ds_afbc
.flags
);
1002 pandecode_log_cont(",\n");
1004 if (fbx
->ds_afbc
.padding
) {
1005 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1006 pandecode_prop("padding = 0x%" PRIx64
, fbx
->ds_afbc
.padding
);
1010 pandecode_log("},\n");
1012 pandecode_log(".ds_linear = {\n");
1015 if (fbx
->ds_linear
.depth
) {
1016 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
1017 pandecode_prop("depth_stride = %d",
1018 fbx
->ds_linear
.depth_stride
);
1019 pandecode_prop("depth_layer_stride = %d",
1020 fbx
->ds_linear
.depth_layer_stride
);
1021 } else if (fbx
->ds_linear
.depth_stride
|| fbx
->ds_linear
.depth_layer_stride
) {
1022 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx
->ds_linear
.depth_stride
, fbx
->ds_linear
.depth_layer_stride
);
1025 if (fbx
->ds_linear
.stencil
) {
1026 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1027 pandecode_prop("stencil_stride = %d",
1028 fbx
->ds_linear
.stencil_stride
);
1029 pandecode_prop("stencil_layer_stride = %d",
1030 fbx
->ds_linear
.stencil_layer_stride
);
1031 } else if (fbx
->ds_linear
.stencil_stride
|| fbx
->ds_linear
.stencil_layer_stride
) {
1032 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx
->ds_linear
.stencil_stride
, fbx
->ds_linear
.stencil_layer_stride
);
1035 if (fbx
->ds_linear
.depth_stride_zero
||
1036 fbx
->ds_linear
.stencil_stride_zero
) {
1037 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1038 pandecode_prop("depth_stride_zero = 0x%x",
1039 fbx
->ds_linear
.depth_stride_zero
);
1040 pandecode_prop("stencil_stride_zero = 0x%x",
1041 fbx
->ds_linear
.stencil_stride_zero
);
1045 pandecode_log("},\n");
1048 if (fbx
->clear_color_1
| fbx
->clear_color_2
) {
1049 pandecode_prop("clear_color_1 = 0x%" PRIx32
, fbx
->clear_color_1
);
1050 pandecode_prop("clear_color_2 = 0x%" PRIx32
, fbx
->clear_color_2
);
1054 pandecode_msg("XXX: fb_extra zeros tripped\n");
1055 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1059 pandecode_log("};\n");
1061 gpu_va
+= sizeof(struct mali_framebuffer_extra
);
1065 pandecode_render_target(gpu_va
, job_no
, fb
);
1071 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1072 mali_ptr addr
, int job_no
, char *suffix
,
1073 int count
, bool varying
, enum mali_job_type job_type
)
1075 char *prefix
= varying
? "Varying" : "Attribute";
1079 pandecode_msg("warn: No %s records\n", prefix
);
1083 MAP_ADDR(ATTRIBUTE_BUFFER
, addr
, cl
);
1085 for (int i
= 0; i
< count
; ++i
) {
1086 fprintf(pandecode_dump_stream
, "%s\n", prefix
);
1088 struct MALI_ATTRIBUTE_BUFFER temp
;
1089 MALI_ATTRIBUTE_BUFFER_unpack(cl
+ i
* MALI_ATTRIBUTE_BUFFER_LENGTH
, &temp
);
1090 MALI_ATTRIBUTE_BUFFER_print(pandecode_dump_stream
, &temp
, 2);
1092 if (temp
.type
== MALI_ATTRIBUTE_TYPE_1D_NPOT_DIVISOR
) {
1093 struct MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT temp2
;
1094 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_unpack(cl
+ (i
+ 1) * MALI_ATTRIBUTE_BUFFER_LENGTH
, &temp2
);
1095 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_print(pandecode_dump_stream
, &temp2
, 2);
1101 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1103 /* TODO: Decode flags */
1104 mali_ptr shader_ptr
= ptr
& ~15;
1106 char *a
= pointer_as_memory_reference(shader_ptr
);
1107 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1113 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1116 decode_bifrost_constant(u16 constant
)
1118 float lo
= (float) (constant
& 0xFF);
1119 float hi
= (float) (constant
>> 8);
1121 return (hi
/ 255.0) + (lo
/ 65535.0);
1125 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1127 struct bifrost_blend_rt
*b
=
1128 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1130 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1133 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1134 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1135 b
->constant
, decode_bifrost_constant(b
->constant
));
1137 /* TODO figure out blend shader enable bit */
1138 DUMP_CL("Equation", BLEND_EQUATION
, &b
->equation
, 2);
1140 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1141 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1143 pandecode_log(".format = %s", mali_format_as_str(b
->format
));
1144 pandecode_swizzle(b
->swizzle
, b
->format
);
1145 pandecode_log_cont(",\n");
1147 pandecode_prop("swizzle = 0x%" PRIx32
, b
->swizzle
);
1148 pandecode_prop("format = 0x%" PRIx32
, b
->format
);
1151 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1152 pandecode_prop("zero1 = 0x%" PRIx32
, b
->zero1
);
1155 pandecode_log(".shader_type = ");
1156 switch(b
->shader_type
) {
1157 case BIFROST_BLEND_F16
:
1158 pandecode_log_cont("BIFROST_BLEND_F16");
1160 case BIFROST_BLEND_F32
:
1161 pandecode_log_cont("BIFROST_BLEND_F32");
1163 case BIFROST_BLEND_I32
:
1164 pandecode_log_cont("BIFROST_BLEND_I32");
1166 case BIFROST_BLEND_U32
:
1167 pandecode_log_cont("BIFROST_BLEND_U32");
1169 case BIFROST_BLEND_I16
:
1170 pandecode_log_cont("BIFROST_BLEND_I16");
1172 case BIFROST_BLEND_U16
:
1173 pandecode_log_cont("BIFROST_BLEND_U16");
1176 pandecode_log_cont(",\n");
1179 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1180 pandecode_prop("zero2 = 0x%" PRIx32
, b
->zero2
);
1183 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1186 pandecode_log("},\n");
1192 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1194 /* constant/equation is in a union */
1198 pandecode_log(".blend = {\n");
1202 pandecode_shader_address("shader", blend
->shader
);
1204 DUMP_CL("Equation", BLEND_EQUATION
, &blend
->equation
, 2);
1205 pandecode_prop("constant = %f", blend
->constant
);
1209 pandecode_log("},\n");
1211 /* Return blend shader to disassemble if present */
1212 return is_shader
? (blend
->shader
& ~0xF) : 0;
1216 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1218 struct midgard_blend_rt
*b
=
1219 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1221 /* Flags determine presence of blend shader */
1222 bool is_shader
= b
->flags
.opaque
[0] & 0x2;
1224 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1227 DUMP_CL("Flags", BLEND_FLAGS
, &b
->flags
, 2);
1229 union midgard_blend blend
= b
->blend
;
1230 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1233 pandecode_log("};\n");
1238 /* Attributes and varyings have descriptor records, which contain information
1239 * about their format and ordering with the attribute/varying buffers. We'll
1240 * want to validate that the combinations specified are self-consistent.
1244 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1246 const char *prefix
= varying
? "Varying" : "Attribute";
1247 mali_ptr p
= varying
? v
->varying_meta
: v
->attribute_meta
;
1249 for (int i
= 0; i
< count
; ++i
, p
+= MALI_ATTRIBUTE_LENGTH
)
1250 DUMP_ADDR(prefix
, ATTRIBUTE
, p
, 1);
1255 /* return bits [lo, hi) of word */
1257 bits(u32 word
, u32 lo
, u32 hi
)
1260 return word
; // avoid undefined behavior with the shift
1262 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1266 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool graphics
)
1268 pandecode_log(".prefix = {\n");
1271 /* Decode invocation_count. See the comment before the definition of
1272 * invocation_count for an explanation.
1275 unsigned size_y_shift
= bits(p
->invocation_shifts
, 0, 5);
1276 unsigned size_z_shift
= bits(p
->invocation_shifts
, 5, 10);
1277 unsigned workgroups_x_shift
= bits(p
->invocation_shifts
, 10, 16);
1278 unsigned workgroups_y_shift
= bits(p
->invocation_shifts
, 16, 22);
1279 unsigned workgroups_z_shift
= bits(p
->invocation_shifts
, 22, 28);
1280 unsigned workgroups_x_shift_2
= bits(p
->invocation_shifts
, 28, 32);
1282 unsigned size_x
= bits(p
->invocation_count
, 0, size_y_shift
) + 1;
1283 unsigned size_y
= bits(p
->invocation_count
, size_y_shift
, size_z_shift
) + 1;
1284 unsigned size_z
= bits(p
->invocation_count
, size_z_shift
, workgroups_x_shift
) + 1;
1286 unsigned groups_x
= bits(p
->invocation_count
, workgroups_x_shift
, workgroups_y_shift
) + 1;
1287 unsigned groups_y
= bits(p
->invocation_count
, workgroups_y_shift
, workgroups_z_shift
) + 1;
1288 unsigned groups_z
= bits(p
->invocation_count
, workgroups_z_shift
, 32) + 1;
1290 /* Even though we have this decoded, we want to ensure that the
1291 * representation is "unique" so we don't lose anything by printing only
1292 * the final result. More specifically, we need to check that we were
1293 * passed something in canonical form, since the definition per the
1294 * hardware is inherently not unique. How? Well, take the resulting
1295 * decode and pack it ourselves! If it is bit exact with what we
1296 * decoded, we're good to go. */
1298 struct mali_vertex_tiler_prefix ref
;
1299 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, graphics
);
1302 (p
->invocation_count
== ref
.invocation_count
) &&
1303 (p
->invocation_shifts
== ref
.invocation_shifts
);
1306 pandecode_msg("XXX: non-canonical workgroups packing\n");
1307 pandecode_msg("expected: %X, %X",
1308 ref
.invocation_count
,
1309 ref
.invocation_shifts
);
1311 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1312 pandecode_prop("size_y_shift = %d", size_y_shift
);
1313 pandecode_prop("size_z_shift = %d", size_z_shift
);
1314 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift
);
1315 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift
);
1316 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift
);
1317 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2
);
1320 /* Regardless, print the decode */
1321 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1322 size_x
, size_y
, size_z
,
1323 groups_x
, groups_y
, groups_z
);
1326 if (p
->unknown_draw
)
1327 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1329 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1331 if (p
->draw_mode
!= MALI_DRAW_MODE_NONE
)
1332 pandecode_prop("draw_mode = %s", mali_draw_mode_as_str(p
->draw_mode
));
1334 /* Index count only exists for tiler jobs anyway */
1337 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1340 unsigned index_raw_size
= (p
->unknown_draw
& MALI_DRAW_INDEXED_SIZE
);
1341 index_raw_size
>>= MALI_DRAW_INDEXED_SHIFT
;
1343 /* Validate an index buffer is present if we need one. TODO: verify
1344 * relationship between invocation_count and index_count */
1347 unsigned count
= p
->index_count
;
1350 unsigned size
= (index_raw_size
== 0x3) ? 4 : index_raw_size
;
1352 /* Ensure we got a size, and if so, validate the index buffer
1353 * is large enough to hold a full set of indices of the given
1356 if (!index_raw_size
)
1357 pandecode_msg("XXX: index size missing\n");
1359 pandecode_validate_buffer(p
->indices
, count
* size
);
1360 } else if (index_raw_size
)
1361 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size
);
1363 if (p
->offset_bias_correction
)
1364 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1366 /* TODO: Figure out what this is. It's not zero */
1367 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1370 pandecode_log("},\n");
1374 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1376 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1377 uint64_t *PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1379 for (int i
= 0; i
< ubufs_count
; i
++) {
1380 unsigned size
= (ubufs
[i
] & ((1 << 10) - 1)) * 16;
1381 mali_ptr addr
= (ubufs
[i
] >> 10) << 2;
1383 pandecode_validate_buffer(addr
, size
);
1385 char *ptr
= pointer_as_memory_reference(addr
);
1386 pandecode_log("ubuf_%d[%u] = %s;\n", i
, size
, ptr
);
1390 pandecode_log("\n");
1394 pandecode_uniforms(mali_ptr uniforms
, unsigned uniform_count
)
1396 pandecode_validate_buffer(uniforms
, uniform_count
* 16);
1398 char *ptr
= pointer_as_memory_reference(uniforms
);
1399 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count
, ptr
);
1404 shader_type_for_job(unsigned type
)
1407 case MALI_JOB_TYPE_VERTEX
: return "VERTEX";
1408 case MALI_JOB_TYPE_TILER
: return "FRAGMENT";
1409 case MALI_JOB_TYPE_COMPUTE
: return "COMPUTE";
1415 static unsigned shader_id
= 0;
1417 static struct midgard_disasm_stats
1418 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1419 bool is_bifrost
, unsigned gpu_id
)
1421 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1422 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1424 /* Compute maximum possible size */
1425 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1427 /* Print some boilerplate to clearly denote the assembly (which doesn't
1428 * obey indentation rules), and actually do the disassembly! */
1430 pandecode_log_cont("\n\n");
1432 struct midgard_disasm_stats stats
;
1435 disassemble_bifrost(pandecode_dump_stream
, code
, sz
, true);
1437 /* TODO: Extend stats to Bifrost */
1438 stats
.texture_count
= -128;
1439 stats
.sampler_count
= -128;
1440 stats
.attribute_count
= -128;
1441 stats
.varying_count
= -128;
1442 stats
.uniform_count
= -128;
1443 stats
.uniform_buffer_count
= -128;
1444 stats
.work_count
= -128;
1446 stats
.instruction_count
= 0;
1447 stats
.bundle_count
= 0;
1448 stats
.quadword_count
= 0;
1449 stats
.helper_invocations
= false;
1451 stats
= disassemble_midgard(pandecode_dump_stream
,
1453 type
== MALI_JOB_TYPE_TILER
?
1454 MESA_SHADER_FRAGMENT
: MESA_SHADER_VERTEX
);
1457 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1458 * driver-internal purposes with the blob and interfere */
1460 bool should_shaderdb
= type
!= MALI_JOB_TYPE_COMPUTE
;
1462 if (should_shaderdb
) {
1463 unsigned nr_threads
=
1464 (stats
.work_count
<= 4) ? 4 :
1465 (stats
.work_count
<= 8) ? 2 :
1468 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
1469 "%u inst, %u bundles, %u quadwords, "
1470 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1472 shader_type_for_job(type
),
1473 stats
.instruction_count
, stats
.bundle_count
, stats
.quadword_count
,
1474 stats
.work_count
, nr_threads
);
1482 pandecode_texture_payload(mali_ptr payload
,
1483 enum mali_texture_dimension dim
,
1484 enum mali_texture_layout layout
,
1488 uint16_t array_size
,
1489 struct pandecode_mapped_memory
*tmem
)
1491 pandecode_log(".payload = {\n");
1494 /* A bunch of bitmap pointers follow.
1495 * We work out the correct number,
1496 * based on the mipmap/cubemap
1497 * properties, but dump extra
1498 * possibilities to futureproof */
1500 int bitmap_count
= levels
+ 1;
1502 /* Miptree for each face */
1503 if (dim
== MALI_TEXTURE_DIMENSION_CUBE
)
1506 /* Array of layers */
1507 bitmap_count
*= depth
;
1509 /* Array of textures */
1510 bitmap_count
*= array_size
;
1512 /* Stride for each element */
1516 mali_ptr
*pointers_and_strides
= pandecode_fetch_gpu_mem(tmem
,
1517 payload
, sizeof(mali_ptr
) * bitmap_count
);
1518 for (int i
= 0; i
< bitmap_count
; ++i
) {
1519 /* How we dump depends if this is a stride or a pointer */
1521 if (manual_stride
&& (i
& 1)) {
1522 /* signed 32-bit snuck in as a 64-bit pointer */
1523 uint64_t stride_set
= pointers_and_strides
[i
];
1524 uint32_t clamped_stride
= stride_set
;
1525 int32_t stride
= clamped_stride
;
1526 assert(stride_set
== clamped_stride
);
1527 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
1529 char *a
= pointer_as_memory_reference(pointers_and_strides
[i
]);
1530 pandecode_log("%s, \n", a
);
1536 pandecode_log("},\n");
1540 pandecode_texture(mali_ptr u
,
1541 struct pandecode_mapped_memory
*tmem
,
1542 unsigned job_no
, unsigned tex
)
1544 struct pandecode_mapped_memory
*mapped_mem
= pandecode_find_mapped_gpu_mem_containing(u
);
1545 const uint8_t *cl
= pandecode_fetch_gpu_mem(mapped_mem
, u
, MALI_MIDGARD_TEXTURE_LENGTH
);
1547 struct MALI_MIDGARD_TEXTURE temp
;
1548 MALI_MIDGARD_TEXTURE_unpack(cl
, &temp
);
1549 MALI_MIDGARD_TEXTURE_print(pandecode_dump_stream
, &temp
, 2);
1551 pandecode_texture_payload(u
+ MALI_MIDGARD_TEXTURE_LENGTH
,
1552 temp
.dimension
, temp
.texel_ordering
, temp
.manual_stride
,
1553 temp
.levels
, temp
.depth
, temp
.array_size
, mapped_mem
);
1557 pandecode_bifrost_texture(
1562 struct MALI_BIFROST_TEXTURE temp
;
1563 MALI_BIFROST_TEXTURE_unpack(cl
, &temp
);
1564 MALI_BIFROST_TEXTURE_print(pandecode_dump_stream
, &temp
, 2);
1566 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(temp
.surfaces
);
1567 pandecode_texture_payload(temp
.surfaces
, temp
.dimension
, temp
.texel_ordering
,
1568 true, temp
.levels
, 1, 1, tmem
);
1571 /* 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 */
1574 pandecode_shader_prop(const char *name
, unsigned claim
, signed truth
, bool fuzzy
)
1580 if (fuzzy
&& (truth
< 0))
1581 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name
, claim
, truth
);
1583 if ((truth
>= 0) && !fuzzy
) {
1584 pandecode_msg("%s: expected %s = %d, claimed %u\n",
1585 (truth
< claim
) ? "warn" : "XXX",
1586 name
, truth
, claim
);
1587 } else if ((claim
> -truth
) && !fuzzy
) {
1588 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
1589 name
, -truth
, claim
);
1590 } else if (fuzzy
&& (claim
< truth
))
1591 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
1592 name
, truth
, claim
);
1594 pandecode_log(".%s = %" PRId16
, name
, claim
);
1597 pandecode_log_cont(" /* %u used */", truth
);
1599 pandecode_log_cont(",\n");
1603 pandecode_blend_shader_disassemble(mali_ptr shader
, int job_no
, int job_type
,
1604 bool is_bifrost
, unsigned gpu_id
)
1606 struct midgard_disasm_stats stats
=
1607 pandecode_shader_disassemble(shader
, job_no
, job_type
, is_bifrost
, gpu_id
);
1609 bool has_texture
= (stats
.texture_count
> 0);
1610 bool has_sampler
= (stats
.sampler_count
> 0);
1611 bool has_attribute
= (stats
.attribute_count
> 0);
1612 bool has_varying
= (stats
.varying_count
> 0);
1613 bool has_uniform
= (stats
.uniform_count
> 0);
1614 bool has_ubo
= (stats
.uniform_buffer_count
> 0);
1616 if (has_texture
|| has_sampler
)
1617 pandecode_msg("XXX: blend shader accessing textures\n");
1619 if (has_attribute
|| has_varying
)
1620 pandecode_msg("XXX: blend shader accessing interstage\n");
1622 if (has_uniform
|| has_ubo
)
1623 pandecode_msg("XXX: blend shader accessing uniforms\n");
1627 pandecode_textures(mali_ptr textures
, unsigned texture_count
, int job_no
, bool is_bifrost
)
1629 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(textures
);
1634 pandecode_log("Textures (%"PRIx64
"):\n", textures
);
1637 const void *cl
= pandecode_fetch_gpu_mem(mmem
,
1638 textures
, MALI_BIFROST_TEXTURE_LENGTH
*
1641 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
1642 pandecode_bifrost_texture(cl
+
1643 MALI_BIFROST_TEXTURE_LENGTH
* tex
,
1647 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
);
1649 for (int tex
= 0; tex
< texture_count
; ++tex
) {
1650 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
1651 char *a
= pointer_as_memory_reference(*u
);
1652 pandecode_log("%s,\n", a
);
1656 /* Now, finally, descend down into the texture descriptor */
1657 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
1658 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
1659 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
1661 pandecode_texture(*u
, tmem
, job_no
, tex
);
1667 pandecode_samplers(mali_ptr samplers
, unsigned sampler_count
, int job_no
, bool is_bifrost
)
1669 for (int i
= 0; i
< sampler_count
; ++i
) {
1671 DUMP_ADDR("Sampler", BIFROST_SAMPLER
, samplers
+ (MALI_BIFROST_SAMPLER_LENGTH
* i
), 1);
1673 DUMP_ADDR("Sampler", MIDGARD_SAMPLER
, samplers
+ (MALI_MIDGARD_SAMPLER_LENGTH
* i
), 1);
1679 pandecode_vertex_tiler_postfix_pre(
1680 const struct mali_vertex_tiler_postfix
*p
,
1681 int job_no
, enum mali_job_type job_type
,
1682 char *suffix
, bool is_bifrost
, unsigned gpu_id
)
1684 struct pandecode_mapped_memory
*attr_mem
;
1686 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1687 * are the only things actually needed from the FBD, vertex/tiler jobs
1688 * no longer reference the FBD -- instead, this field points to some
1689 * info about the scratchpad.
1692 struct pandecode_fbd fbd_info
= {
1693 /* Default for Bifrost */
1698 pandecode_log_cont("\t/* %X %/\n", p
->shared_memory
& 1);
1699 pandecode_compute_fbd(p
->shared_memory
& ~1, job_no
);
1700 } else if (p
->shared_memory
& MALI_MFBD
)
1701 fbd_info
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->shared_memory
) & FBD_MASK
, job_no
, false, job_type
== MALI_JOB_TYPE_COMPUTE
, false);
1702 else if (job_type
== MALI_JOB_TYPE_COMPUTE
)
1703 pandecode_compute_fbd((u64
) (uintptr_t) p
->shared_memory
, job_no
);
1705 fbd_info
= pandecode_sfbd((u64
) (uintptr_t) p
->shared_memory
, job_no
, false, gpu_id
);
1707 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
1708 int texture_count
= 0, sampler_count
= 0;
1711 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->shader
);
1712 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, p
->shader
);
1714 /* Disassemble ahead-of-time to get stats. Initialize with
1715 * stats for the missing-shader case so we get validation
1718 struct midgard_disasm_stats info
= {
1721 .attribute_count
= 0,
1725 .uniform_count
= -128,
1726 .uniform_buffer_count
= 0
1729 struct MALI_SHADER shader
;
1730 struct MALI_MIDGARD_PROPERTIES midg_props
;
1731 struct MALI_BIFROST_PROPERTIES bi_props
;
1732 struct MALI_PRELOAD bi_preload
;
1734 struct mali_shader_packed shader_packed
= s
->shader
;
1735 MALI_SHADER_unpack((const uint8_t *) &shader_packed
, &shader
);
1737 if (shader
.shader
& ~0xF)
1738 info
= pandecode_shader_disassemble(shader
.shader
& ~0xF, job_no
, job_type
, is_bifrost
, gpu_id
);
1741 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", p
->shader
, job_no
, suffix
);
1744 /* Save for dumps */
1745 attribute_count
= shader
.attribute_count
;
1746 varying_count
= shader
.varying_count
;
1747 texture_count
= shader
.texture_count
;
1748 sampler_count
= shader
.sampler_count
;
1751 uint32_t opaque
= s
->bifrost_props
.opaque
[0];
1752 MALI_BIFROST_PROPERTIES_unpack((const uint8_t *) &opaque
, &bi_props
);
1754 opaque
= s
->bifrost_preload
.opaque
[0];
1755 MALI_PRELOAD_unpack((const uint8_t *) &opaque
, &bi_preload
);
1757 uniform_count
= bi_preload
.uniform_count
;
1758 uniform_buffer_count
= bi_props
.uniform_buffer_count
;
1760 uint32_t opaque
= s
->midgard_props
.opaque
[0];
1761 MALI_MIDGARD_PROPERTIES_unpack((const uint8_t *) &opaque
, &midg_props
);
1763 uniform_count
= midg_props
.uniform_count
;
1764 uniform_buffer_count
= midg_props
.uniform_buffer_count
;
1767 pandecode_shader_address("shader", shader
.shader
);
1769 pandecode_shader_prop("texture_count", texture_count
, info
.texture_count
, false);
1770 pandecode_shader_prop("sampler_count", sampler_count
, info
.sampler_count
, false);
1771 pandecode_shader_prop("attribute_count", attribute_count
, info
.attribute_count
, false);
1772 pandecode_shader_prop("varying_count", varying_count
, info
.varying_count
, false);
1775 MALI_BIFROST_PROPERTIES_print(pandecode_dump_stream
, &bi_props
, 2);
1777 MALI_MIDGARD_PROPERTIES_print(pandecode_dump_stream
, &midg_props
, 2);
1780 uint32_t opaque
= s
->bifrost_preload
.opaque
[0];
1782 case MALI_JOB_TYPE_VERTEX
:
1783 DUMP_CL("Preload", PRELOAD_VERTEX
, &opaque
, 2);
1785 case MALI_JOB_TYPE_TILER
:
1786 DUMP_CL("Preload", PRELOAD_FRAGMENT
, &opaque
, 2);
1788 case MALI_JOB_TYPE_COMPUTE
:
1789 DUMP_CL("Preload", PRELOAD_COMPUTE
, &opaque
, 2);
1792 DUMP_CL("Preload", PRELOAD
, &opaque
, 2);
1797 if (s
->depth_units
|| s
->depth_factor
) {
1798 pandecode_prop("depth_factor = %f", s
->depth_factor
);
1799 pandecode_prop("depth_units = %f", s
->depth_units
);
1802 if (s
->coverage_mask
)
1803 pandecode_prop("coverage_mask = 0x%X", s
->coverage_mask
);
1806 pandecode_prop(".unknown2_2 = %X", s
->unknown2_2
);
1808 if (s
->unknown2_3
|| s
->unknown2_4
) {
1809 pandecode_log(".unknown2_3 = ");
1811 int unknown2_3
= s
->unknown2_3
;
1812 int unknown2_4
= s
->unknown2_4
;
1814 /* We're not quite sure what these flags mean without the depth test, if anything */
1816 if (unknown2_3
& (MALI_DEPTH_WRITEMASK
| MALI_DEPTH_FUNC_MASK
)) {
1817 const char *func
= mali_func_as_str(MALI_GET_DEPTH_FUNC(unknown2_3
));
1818 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
1820 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
1823 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
1824 pandecode_log_cont(",\n");
1826 pandecode_log(".unknown2_4 = ");
1827 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
1828 pandecode_log_cont(",\n");
1831 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
1832 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
1833 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
1836 DUMP_CL("Stencil front", STENCIL
, &s
->stencil_front
, 1);
1837 DUMP_CL("Stencil back", STENCIL
, &s
->stencil_back
, 1);
1839 if (!is_bifrost
&& s
->midgard2
.unknown2_7
) {
1840 pandecode_log(".midgard2 = {\n");
1843 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
1845 pandecode_log("},\n");
1849 pandecode_msg("XXX: shader padding tripped\n");
1850 pandecode_prop("padding = 0x%" PRIx32
, s
->padding
);
1854 /* TODO: Blend shaders routing/disasm */
1855 union midgard_blend blend
= s
->blend
;
1856 mali_ptr shader
= pandecode_midgard_blend(&blend
, s
->unknown2_3
& MALI_HAS_BLEND_SHADER
);
1858 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
1860 pandecode_msg("mdg_blend = %" PRIx64
"\n", s
->blend
.shader
);
1864 pandecode_log("};\n");
1866 /* MRT blend fields are used whenever MFBD is used, with
1867 * per-RT descriptors */
1869 if (job_type
== MALI_JOB_TYPE_TILER
&& (is_bifrost
|| p
->shared_memory
& MALI_MFBD
)) {
1870 void* blend_base
= (void *) (s
+ 1);
1872 for (unsigned i
= 0; i
< fbd_info
.rt_count
; i
++) {
1873 mali_ptr shader
= 0;
1876 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
1878 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
1881 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
1886 pandecode_msg("XXX: missing shader descriptor\n");
1889 DUMP_ADDR("Viewport", VIEWPORT
, p
->viewport
, 1);
1891 unsigned max_attr_index
= 0;
1893 if (p
->attribute_meta
)
1894 max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
1896 if (p
->attributes
) {
1897 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
1898 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false, job_type
);
1901 /* Varyings are encoded like attributes but not actually sent; we just
1902 * pass a zero buffer with the right stride/size set, (or whatever)
1903 * since the GPU will write to it itself */
1905 if (p
->varying_meta
) {
1906 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
1910 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
1912 /* Number of descriptors depends on whether there are
1913 * non-internal varyings */
1915 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true, job_type
);
1918 if (p
->uniform_buffers
) {
1919 if (uniform_buffer_count
)
1920 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
1922 pandecode_msg("warn: UBOs specified but not referenced\n");
1923 } else if (uniform_buffer_count
)
1924 pandecode_msg("XXX: UBOs referenced but not specified\n");
1926 /* We don't want to actually dump uniforms, but we do need to validate
1927 * that the counts we were given are sane */
1931 pandecode_uniforms(p
->uniforms
, uniform_count
);
1933 pandecode_msg("warn: Uniforms specified but not referenced\n");
1934 } else if (uniform_count
)
1935 pandecode_msg("XXX: Uniforms referenced but not specified\n");
1938 pandecode_textures(p
->textures
, texture_count
, job_no
, is_bifrost
);
1940 if (p
->sampler_descriptor
)
1941 pandecode_samplers(p
->sampler_descriptor
, sampler_count
, job_no
, is_bifrost
);
1945 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
1947 pandecode_log(".gl_enables = ");
1949 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
1951 pandecode_log_cont(",\n");
1955 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
1957 if (p
->shader
& 0xF)
1958 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p
->shader
& 0xF));
1960 pandecode_log(".postfix = {\n");
1963 pandecode_gl_enables(p
->gl_enables
, MALI_JOB_TYPE_TILER
);
1964 pandecode_prop("instance_shift = 0x%x", p
->instance_shift
);
1965 pandecode_prop("instance_odd = 0x%x", p
->instance_odd
);
1968 pandecode_msg("XXX: vertex only zero tripped");
1969 pandecode_prop("zero4 = 0x%" PRIx32
, p
->zero4
);
1972 pandecode_prop("offset_start = 0x%x", p
->offset_start
);
1975 pandecode_msg("XXX: vertex only zero tripped");
1976 pandecode_prop("zero5 = 0x%" PRIx64
, p
->zero5
);
1979 MEMORY_PROP(p
, position_varying
);
1980 MEMORY_PROP(p
, occlusion_counter
);
1983 pandecode_log("},\n");
1987 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
1989 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1990 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
1992 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1996 pandecode_msg("XXX: tiler heap zero tripped\n");
1997 pandecode_prop("zero = 0x%x", h
->zero
);
2000 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2001 MEMORY_PROP(h
, tiler_heap_start
);
2002 MEMORY_PROP(h
, tiler_heap_free
);
2004 /* this might point to the beginning of another buffer, when it's
2005 * really the end of the tiler heap buffer, so we have to be careful
2006 * here. but for zero length, we need the same pointer.
2009 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2010 MEMORY_PROP(h
, tiler_heap_start
);
2012 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2013 pandecode_prop("tiler_heap_end = %s + 1", a
);
2017 for (int i
= 0; i
< 10; i
++) {
2018 if (h
->zeros
[i
] != 0) {
2019 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2024 if (h
->unk1
!= 0x1) {
2025 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2026 pandecode_prop("unk1 = 0x%x", h
->unk1
);
2029 if (h
->unk7e007e
!= 0x7e007e) {
2030 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
2031 pandecode_prop("unk7e007e = 0x%x", h
->unk7e007e
);
2035 pandecode_log("};\n");
2039 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2041 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2042 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2044 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2046 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2049 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32
, t
->tiler_heap_next_start
);
2050 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32
, t
->used_hierarchy_mask
);
2052 if (t
->hierarchy_mask
!= 0xa &&
2053 t
->hierarchy_mask
!= 0x14 &&
2054 t
->hierarchy_mask
!= 0x28 &&
2055 t
->hierarchy_mask
!= 0x50 &&
2056 t
->hierarchy_mask
!= 0xa0)
2057 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
2059 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2061 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2063 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2064 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2067 pandecode_msg("XXX: tiler meta zero tripped\n");
2068 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2071 for (int i
= 0; i
< 12; i
++) {
2072 if (t
->zeros
[i
] != 0) {
2073 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2079 pandecode_log("};\n");
2083 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2085 if (u
.pointer
== 0x0)
2088 pandecode_log(".primitive_size = {\n");
2092 pandecode_prop("constant = %f", u
.constant
);
2094 MEMORY_PROP((&u
), pointer
);
2098 pandecode_log("},\n");
2102 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2104 pandecode_log_cont("{\n");
2107 /* TODO: gl_PointSize on Bifrost */
2108 pandecode_primitive_size(t
->primitive_size
, true);
2110 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2112 pandecode_msg("XXX: tiler only zero tripped\n");
2113 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2114 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2115 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2116 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2117 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2118 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2122 pandecode_log("},\n");
2126 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2127 const struct pandecode_mapped_memory
*mem
,
2128 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2130 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2132 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2134 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2137 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2138 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2141 pandecode_log("};\n");
2147 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2148 const struct pandecode_mapped_memory
*mem
,
2149 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2151 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2153 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2154 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2156 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2159 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2161 pandecode_log(".tiler = ");
2162 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2164 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2167 pandecode_log("};\n");
2173 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2174 const struct pandecode_mapped_memory
*mem
,
2175 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2177 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2178 bool is_graphics
= (h
->job_type
== MALI_JOB_TYPE_VERTEX
) || (h
->job_type
== MALI_JOB_TYPE_TILER
);
2180 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false, gpu_id
);
2182 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2185 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, is_graphics
);
2186 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2188 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2189 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2192 pandecode_log("};\n");
2198 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2199 mali_ptr payload
, int job_no
,
2200 bool is_bifrost
, unsigned gpu_id
)
2202 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2204 bool is_mfbd
= s
->framebuffer
& MALI_MFBD
;
2206 if (!is_mfbd
&& is_bifrost
)
2207 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2209 struct pandecode_fbd info
;
2212 info
= pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true, false, is_bifrost
);
2214 info
= pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true, gpu_id
);
2216 /* Compute the tag for the tagged pointer. This contains the type of
2217 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2218 * additional structures follow the MFBD header (an extra payload or
2219 * not, as well as a count of render targets) */
2221 unsigned expected_tag
= is_mfbd
? MALI_MFBD
: 0;
2225 expected_tag
|= MALI_MFBD_TAG_EXTRA
;
2227 expected_tag
|= (MALI_POSITIVE(info
.rt_count
) << 2);
2230 if ((s
->min_tile_coord
| s
->max_tile_coord
) & ~(MALI_X_COORD_MASK
| MALI_Y_COORD_MASK
)) {
2231 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2232 pandecode_prop("min_tile_coord = 0x%X\n", s
->min_tile_coord
);
2233 pandecode_prop("max_tile_coord = 0x%X\n", s
->max_tile_coord
);
2236 /* Extract tile coordinates */
2238 unsigned min_x
= MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2239 unsigned min_y
= MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2241 unsigned max_x
= (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2242 unsigned max_y
= (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2244 /* For the max, we also want the floored (rather than ceiled) version for checking */
2246 unsigned max_x_f
= (MALI_TILE_COORD_X(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2247 unsigned max_y_f
= (MALI_TILE_COORD_Y(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2249 /* Validate the coordinates are well-ordered */
2252 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2253 else if (min_x
> max_x
)
2254 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2257 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2258 else if (min_y
> max_y
)
2259 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2261 /* Validate the coordinates fit inside the framebuffer. We use floor,
2262 * rather than ceil, for the max coordinates, since the tile
2263 * coordinates for something like an 800x600 framebuffer will actually
2264 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2266 if ((min_x
> info
.width
) || (max_x_f
> info
.width
))
2267 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2269 if ((min_y
> info
.height
) || (max_y_f
> info
.height
))
2270 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2272 /* After validation, we print */
2274 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x
, min_y
, max_x
, max_y
);
2276 /* The FBD is a tagged pointer */
2278 unsigned tag
= (s
->framebuffer
& ~FBD_MASK
);
2280 if (tag
!= expected_tag
)
2281 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag
, tag
);
2286 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
2287 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
2288 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
2289 * might wish to combine this with the bifrost parameter) because some details
2290 * are model-specific even within a particular architecture. Minimal traces
2291 * *only* examine the job descriptors, skipping printing entirely if there is
2292 * no faults, and only descends into the payload if there are faults. This is
2293 * useful for looking for faults without the overhead of invasive traces. */
2296 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
, unsigned gpu_id
, bool minimal
)
2298 pandecode_dump_file_open();
2300 struct mali_job_descriptor_header
*h
;
2301 unsigned job_descriptor_number
= 0;
2304 struct pandecode_mapped_memory
*mem
=
2305 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2309 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2311 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
);
2312 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
, 64);
2314 int job_no
= job_descriptor_number
++;
2316 /* If the job is good to go, skip it in minimal mode */
2317 if (minimal
&& (h
->exception_status
== 0x0 || h
->exception_status
== 0x1))
2320 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2323 pandecode_prop("job_type = %s", mali_job_type_as_str(h
->job_type
));
2325 if (h
->job_descriptor_size
)
2326 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2328 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2329 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2330 h
->exception_status
,
2331 (h
->exception_status
>> 16) & 0xFFFF,
2332 mali_exception_access_as_str((h
->exception_status
>> 8) & 0x3),
2333 h
->exception_status
& 0xFF);
2335 if (h
->first_incomplete_task
)
2336 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2338 if (h
->fault_pointer
)
2339 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2342 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2344 pandecode_prop("job_index = %d", h
->job_index
);
2346 if (h
->unknown_flags
)
2347 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2349 if (h
->job_dependency_index_1
)
2350 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2352 if (h
->job_dependency_index_2
)
2353 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2356 pandecode_log("};\n");
2358 switch (h
->job_type
) {
2359 case MALI_JOB_TYPE_WRITE_VALUE
: {
2360 struct mali_payload_write_value
*s
= payload
;
2361 pandecode_log("struct mali_payload_write_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2363 MEMORY_PROP(s
, address
);
2365 if (s
->value_descriptor
!= MALI_WRITE_VALUE_ZERO
) {
2366 pandecode_msg("XXX: unknown value descriptor\n");
2367 pandecode_prop("value_descriptor = 0x%" PRIX32
, s
->value_descriptor
);
2371 pandecode_msg("XXX: set value tripped\n");
2372 pandecode_prop("reserved = 0x%" PRIX32
, s
->reserved
);
2375 pandecode_prop("immediate = 0x%" PRIX64
, s
->immediate
);
2377 pandecode_log("};\n");
2382 case MALI_JOB_TYPE_TILER
:
2383 case MALI_JOB_TYPE_VERTEX
:
2384 case MALI_JOB_TYPE_COMPUTE
:
2386 if (h
->job_type
== MALI_JOB_TYPE_TILER
)
2387 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2389 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2391 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2395 case MALI_JOB_TYPE_FRAGMENT
:
2396 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
, gpu_id
);
2402 } while ((jc_gpu_va
= h
->next_job
));
2404 pandecode_map_read_write();