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 <panfrost-job.h>
34 #include "util/macros.h"
35 #include "util/u_math.h"
37 #include "pan_pretty_print.h"
38 #include "midgard/disassemble.h"
39 #include "bifrost/disassemble.h"
41 #include "pan_encoder.h"
43 static void pandecode_swizzle(unsigned swizzle
, enum mali_format format
);
45 #define MEMORY_PROP(obj, p) {\
47 char *a = pointer_as_memory_reference(obj->p); \
48 pandecode_prop("%s = %s", #p, a); \
53 #define MEMORY_PROP_DIR(obj, p) {\
55 char *a = pointer_as_memory_reference(obj.p); \
56 pandecode_prop("%s = %s", #p, a); \
61 FILE *pandecode_dump_stream
;
63 /* Semantic logging type.
65 * Raw: for raw messages to be printed as is.
66 * Message: for helpful information to be commented out in replays.
67 * Property: for properties of a struct
69 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
72 enum pandecode_log_type
{
78 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
79 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
80 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
82 unsigned pandecode_indent
= 0;
85 pandecode_make_indent(void)
87 for (unsigned i
= 0; i
< pandecode_indent
; ++i
)
88 fprintf(pandecode_dump_stream
, " ");
92 pandecode_log_typed(enum pandecode_log_type type
, const char *format
, ...)
96 pandecode_make_indent();
98 if (type
== PANDECODE_MESSAGE
)
99 fprintf(pandecode_dump_stream
, "// ");
100 else if (type
== PANDECODE_PROPERTY
)
101 fprintf(pandecode_dump_stream
, ".");
103 va_start(ap
, format
);
104 vfprintf(pandecode_dump_stream
, format
, ap
);
107 if (type
== PANDECODE_PROPERTY
)
108 fprintf(pandecode_dump_stream
, ",\n");
112 pandecode_log_cont(const char *format
, ...)
116 va_start(ap
, format
);
117 vfprintf(pandecode_dump_stream
, format
, ap
);
121 /* To check for memory safety issues, validates that the given pointer in GPU
122 * memory is valid, containing at least sz bytes. The goal is to eliminate
123 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
124 * overruns) by statically validating pointers.
128 pandecode_validate_buffer(mali_ptr addr
, size_t sz
)
131 pandecode_msg("XXX: null pointer deref");
137 struct pandecode_mapped_memory
*bo
=
138 pandecode_find_mapped_gpu_mem_containing(addr
);
141 pandecode_msg("XXX: invalid memory dereference\n");
147 unsigned offset
= addr
- bo
->gpu_va
;
148 unsigned total
= offset
+ sz
;
150 if (total
> bo
->length
) {
151 pandecode_msg("XXX: buffer overrun. "
152 "Chunk of size %zu at offset %d in buffer of size %zu. "
153 "Overrun by %zu bytes. \n",
154 sz
, offset
, bo
->length
, total
- bo
->length
);
159 struct pandecode_flag_info
{
165 pandecode_log_decoded_flags(const struct pandecode_flag_info
*flag_info
,
168 bool decodable_flags_found
= false;
170 for (int i
= 0; flag_info
[i
].name
; i
++) {
171 if ((flags
& flag_info
[i
].flag
) != flag_info
[i
].flag
)
174 if (!decodable_flags_found
) {
175 decodable_flags_found
= true;
177 pandecode_log_cont(" | ");
180 pandecode_log_cont("%s", flag_info
[i
].name
);
182 flags
&= ~flag_info
[i
].flag
;
185 if (decodable_flags_found
) {
187 pandecode_log_cont(" | 0x%" PRIx64
, flags
);
189 pandecode_log_cont("0x%" PRIx64
, flags
);
193 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
194 static const struct pandecode_flag_info gl_enable_flag_info
[] = {
195 FLAG_INFO(OCCLUSION_QUERY
),
196 FLAG_INFO(OCCLUSION_PRECISE
),
197 FLAG_INFO(FRONT_CCW_TOP
),
198 FLAG_INFO(CULL_FACE_FRONT
),
199 FLAG_INFO(CULL_FACE_BACK
),
204 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
205 static const struct pandecode_flag_info clear_flag_info
[] = {
208 FLAG_INFO(SLOW_STENCIL
),
213 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
214 static const struct pandecode_flag_info mask_flag_info
[] = {
223 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
224 static const struct pandecode_flag_info u3_flag_info
[] = {
226 FLAG_INFO(CAN_DISCARD
),
227 FLAG_INFO(HAS_BLEND_SHADER
),
228 FLAG_INFO(DEPTH_WRITEMASK
),
232 static const struct pandecode_flag_info u4_flag_info
[] = {
234 FLAG_INFO(NO_DITHER
),
235 FLAG_INFO(DEPTH_RANGE_A
),
236 FLAG_INFO(DEPTH_RANGE_B
),
237 FLAG_INFO(STENCIL_TEST
),
238 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER
),
243 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
244 static const struct pandecode_flag_info mfbd_fmt_flag_info
[] = {
251 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
252 static const struct pandecode_flag_info mfbd_extra_flag_hi_info
[] = {
258 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
259 static const struct pandecode_flag_info mfbd_extra_flag_lo_info
[] = {
265 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
266 static const struct pandecode_flag_info shader_midgard1_flag_info
[] = {
268 FLAG_INFO(READS_TILEBUFFER
),
274 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
275 static const struct pandecode_flag_info mfbd_flag_info
[] = {
276 FLAG_INFO(DEPTH_WRITE
),
282 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
283 static const struct pandecode_flag_info sampler_flag_info
[] = {
284 FLAG_INFO(MAG_NEAREST
),
285 FLAG_INFO(MIN_NEAREST
),
286 FLAG_INFO(MIP_LINEAR_1
),
287 FLAG_INFO(MIP_LINEAR_2
),
288 FLAG_INFO(NORM_COORDS
),
293 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
294 static const struct pandecode_flag_info sfbd_unk1_info
[] = {
301 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
302 static const struct pandecode_flag_info sfbd_unk2_info
[] = {
309 extern char *replace_fragment
;
310 extern char *replace_vertex
;
313 pandecode_job_type(enum mali_job_type type
)
315 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
319 DEFINE_CASE(WRITE_VALUE
);
320 DEFINE_CASE(CACHE_FLUSH
);
321 DEFINE_CASE(COMPUTE
);
325 DEFINE_CASE(FRAGMENT
);
327 case JOB_NOT_STARTED
:
328 return "NOT_STARTED";
331 pandecode_log("Warning! Unknown job type %x\n", type
);
339 pandecode_draw_mode(enum mali_draw_mode mode
)
341 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
344 DEFINE_CASE(DRAW_NONE
);
347 DEFINE_CASE(TRIANGLES
);
348 DEFINE_CASE(TRIANGLE_STRIP
);
349 DEFINE_CASE(TRIANGLE_FAN
);
350 DEFINE_CASE(LINE_STRIP
);
351 DEFINE_CASE(LINE_LOOP
);
352 DEFINE_CASE(POLYGON
);
354 DEFINE_CASE(QUAD_STRIP
);
357 pandecode_msg("XXX: invalid draw mode %X\n", mode
);
364 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
366 pandecode_func(enum mali_func mode
)
373 DEFINE_CASE(GREATER
);
374 DEFINE_CASE(NOTEQUAL
);
379 pandecode_msg("XXX: invalid func %X\n", mode
);
385 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
387 pandecode_stencil_op(enum mali_stencil_op op
)
391 DEFINE_CASE(REPLACE
);
394 DEFINE_CASE(INCR_WRAP
);
395 DEFINE_CASE(DECR_WRAP
);
400 pandecode_msg("XXX: invalid stencil op %X\n", op
);
407 static char *pandecode_attr_mode_short(enum mali_attr_mode mode
)
410 /* TODO: Combine to just "instanced" once this can be done
411 * unambiguously in all known cases */
412 case MALI_ATTR_POT_DIVIDE
:
413 return "instanced_pot";
414 case MALI_ATTR_MODULO
:
415 return "instanced_mod";
416 case MALI_ATTR_NPOT_DIVIDE
:
417 return "instanced_npot";
418 case MALI_ATTR_IMAGE
:
421 pandecode_msg("XXX: invalid attribute mode %X\n", mode
);
427 pandecode_special_record(uint64_t v
, bool* attribute
)
430 case MALI_ATTR_VERTEXID
:
432 return "gl_VertexID";
433 case MALI_ATTR_INSTANCEID
:
435 return "gl_InstanceID";
436 case MALI_VARYING_FRAG_COORD
:
437 return "gl_FragCoord";
438 case MALI_VARYING_FRONT_FACING
:
439 return "gl_FrontFacing";
440 case MALI_VARYING_POINT_COORD
:
441 return "gl_PointCoord";
443 pandecode_msg("XXX: invalid special record %" PRIx64
"\n", v
);
448 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
450 pandecode_wrap_mode(enum mali_wrap_mode op
)
454 DEFINE_CASE(CLAMP_TO_EDGE
);
455 DEFINE_CASE(CLAMP_TO_BORDER
);
456 DEFINE_CASE(MIRRORED_REPEAT
);
459 pandecode_msg("XXX: invalid wrap mode %X\n", op
);
465 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
467 pandecode_block_format(enum mali_block_format fmt
)
471 DEFINE_CASE(UNKNOWN
);
476 unreachable("Invalid case");
481 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
483 pandecode_exception_access(unsigned access
)
487 DEFINE_CASE(EXECUTE
);
492 unreachable("Invalid case");
497 /* Midgard's tiler descriptor is embedded within the
501 pandecode_midgard_tiler_descriptor(
502 const struct midgard_tiler_descriptor
*t
,
508 pandecode_log(".tiler = {\n");
511 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
512 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
514 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
516 /* We know this name from the kernel, but we never see it nonzero */
519 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
521 MEMORY_PROP(t
, polygon_list
);
523 /* The body is offset from the base of the polygon list */
524 assert(t
->polygon_list_body
> t
->polygon_list
);
525 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
527 /* It needs to fit inside the reported size */
528 assert(t
->polygon_list_size
>= body_offset
);
530 /* Check that we fit */
531 struct pandecode_mapped_memory
*plist
=
532 pandecode_find_mapped_gpu_mem_containing(t
->polygon_list
);
534 assert(t
->polygon_list_size
<= plist
->length
);
536 /* Now that we've sanity checked, we'll try to calculate the sizes
537 * ourselves for comparison */
539 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
540 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
542 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
543 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
544 ref_header
, ref_size
);
545 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
546 pandecode_msg("body offset %d\n", body_offset
);
549 /* The tiler heap has a start and end specified -- it should be
550 * identical to what we have in the BO. The exception is if tiling is
553 MEMORY_PROP(t
, heap_start
);
554 assert(t
->heap_end
>= t
->heap_start
);
556 struct pandecode_mapped_memory
*heap
=
557 pandecode_find_mapped_gpu_mem_containing(t
->heap_start
);
559 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
561 /* Tiling is enabled with a special flag */
562 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
563 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
565 bool tiling_enabled
= hierarchy_mask
;
567 if (tiling_enabled
) {
568 /* When tiling is enabled, the heap should be a tight fit */
569 unsigned heap_offset
= t
->heap_start
- heap
->gpu_va
;
570 if ((heap_offset
+ heap_size
) != heap
->length
) {
571 pandecode_msg("XXX: heap size %u (expected %zu)\n",
572 heap_size
, heap
->length
- heap_offset
);
575 /* We should also have no other flags */
577 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
579 /* When tiling is disabled, we should have that flag and no others */
581 if (tiler_flags
!= MALI_TILER_DISABLED
) {
582 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
586 /* We should also have an empty heap */
588 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
592 /* Disabled tiling is used only for clear-only jobs, which are
593 * purely FRAGMENT, so we should never see this for
594 * non-FRAGMENT descriptors. */
597 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
600 /* We've never seen weights used in practice, but we know from the
601 * kernel these fields is there */
603 bool nonzero_weights
= false;
605 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
606 nonzero_weights
|= t
->weights
[w
] != 0x0;
609 if (nonzero_weights
) {
610 pandecode_log(".weights = {");
612 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
613 pandecode_log("%d, ", t
->weights
[w
]);
620 pandecode_log("}\n");
623 /* Information about the framebuffer passed back for
624 * additional analysis */
626 struct pandecode_fbd
{
634 pandecode_sfbd_format(struct mali_sfbd_format format
)
636 pandecode_log(".format = {\n");
639 pandecode_log(".unk1 = ");
640 pandecode_log_decoded_flags(sfbd_unk1_info
, format
.unk1
);
641 pandecode_log_cont(",\n");
643 /* TODO: Map formats so we can check swizzles and print nicely */
644 pandecode_log("swizzle");
645 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
646 pandecode_log_cont(",\n");
648 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
649 (format
.nr_channels
+ 1));
651 pandecode_log(".unk2 = ");
652 pandecode_log_decoded_flags(sfbd_unk2_info
, format
.unk2
);
653 pandecode_log_cont(",\n");
655 pandecode_prop("block = %s", pandecode_block_format(format
.block
));
657 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
660 pandecode_log("},\n");
663 static struct pandecode_fbd
664 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
, unsigned gpu_id
)
666 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
667 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
669 struct pandecode_fbd info
= {
674 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
677 pandecode_prop("unknown1 = 0x%" PRIx32
, s
->unknown1
);
678 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->unknown2
);
680 pandecode_sfbd_format(s
->format
);
682 info
.width
= s
->width
+ 1;
683 info
.height
= s
->height
+ 1;
685 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", info
.width
);
686 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", info
.height
);
688 MEMORY_PROP(s
, checksum
);
690 if (s
->checksum_stride
)
691 pandecode_prop("checksum_stride = %d", s
->checksum_stride
);
693 MEMORY_PROP(s
, framebuffer
);
694 pandecode_prop("stride = %d", s
->stride
);
696 /* Earlier in the actual commandstream -- right before width -- but we
697 * delay to flow nicer */
699 pandecode_log(".clear_flags = ");
700 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
701 pandecode_log_cont(",\n");
703 if (s
->depth_buffer
) {
704 MEMORY_PROP(s
, depth_buffer
);
705 pandecode_prop("depth_stride = %d", s
->depth_stride
);
708 if (s
->stencil_buffer
) {
709 MEMORY_PROP(s
, stencil_buffer
);
710 pandecode_prop("stencil_stride = %d", s
->stencil_stride
);
713 if (s
->depth_stride_zero
||
714 s
->stencil_stride_zero
||
715 s
->zero7
|| s
->zero8
) {
716 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
717 pandecode_prop("depth_stride_zero = 0x%x",
718 s
->depth_stride_zero
);
719 pandecode_prop("stencil_stride_zero = 0x%x",
720 s
->stencil_stride_zero
);
721 pandecode_prop("zero7 = 0x%" PRIx32
,
723 pandecode_prop("zero8 = 0x%" PRIx32
,
727 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
728 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
729 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
730 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
731 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
734 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
735 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
736 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
737 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
738 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
741 if (s
->clear_stencil
) {
742 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
745 MEMORY_PROP(s
, scratchpad
);
746 const struct midgard_tiler_descriptor t
= s
->tiler
;
748 bool has_hierarchy
= !(gpu_id
== 0x0720 || gpu_id
== 0x0820 || gpu_id
== 0x0830);
749 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
, has_hierarchy
);
752 pandecode_log("};\n");
754 pandecode_prop("zero0 = 0x%" PRIx64
, s
->zero0
);
755 pandecode_prop("zero1 = 0x%" PRIx64
, s
->zero1
);
756 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
757 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
758 pandecode_prop("zero5 = 0x%" PRIx32
, s
->zero5
);
760 pandecode_log_cont(".zero3 = {");
762 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
763 pandecode_log_cont("%X, ", s
->zero3
[i
]);
765 pandecode_log_cont("},\n");
767 pandecode_log_cont(".zero6 = {");
769 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
770 pandecode_log_cont("%X, ", s
->zero6
[i
]);
772 pandecode_log_cont("},\n");
778 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
780 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
781 const struct mali_compute_fbd
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
783 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
786 pandecode_log(".unknown1 = {");
788 for (int i
= 0; i
< ARRAY_SIZE(s
->unknown1
); ++i
)
789 pandecode_log_cont("%X, ", s
->unknown1
[i
]);
791 pandecode_log("},\n");
794 pandecode_log_cont("},\n");
797 /* Extracts the number of components associated with a Mali format */
800 pandecode_format_component_count(enum mali_format fmt
)
802 /* Mask out the format class */
803 unsigned top
= fmt
& 0b11100000;
806 case MALI_FORMAT_SNORM
:
807 case MALI_FORMAT_UINT
:
808 case MALI_FORMAT_UNORM
:
809 case MALI_FORMAT_SINT
:
810 return ((fmt
>> 3) & 3) + 1;
817 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
820 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
823 assert(MALI_CHANNEL_RED
== 0);
825 for (unsigned c
= 0; c
< 4; ++c
) {
826 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
828 if (chan
<= MALI_CHANNEL_ALPHA
)
835 /* Validates that a (format, swizzle) pair is valid, in the sense that the
836 * swizzle doesn't access any components that are undefined in the format.
837 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
841 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
843 unsigned nr_comp
= pandecode_format_component_count(fmt
);
844 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
845 unsigned valid_mask
= (1 << nr_comp
) - 1;
846 unsigned invalid_mask
= ~valid_mask
;
848 if (access_mask
& invalid_mask
) {
849 pandecode_msg("XXX: invalid components accessed\n");
853 /* Check for the default non-swizzling swizzle so we can suppress
854 * useless printing for the defaults */
856 unsigned default_swizzles
[4] = {
857 MALI_CHANNEL_RED
| (MALI_CHANNEL_ZERO
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
858 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
859 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ONE
<< 9),
860 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ALPHA
<< 9)
863 return (swizzle
== default_swizzles
[nr_comp
- 1]);
866 /* Maps MALI_RGBA32F to rgba32f, etc */
869 pandecode_format_short(enum mali_format fmt
, bool srgb
)
871 /* We want a type-like format, so cut off the initial MALI_ */
872 char *format
= pandecode_format(fmt
);
873 format
+= strlen("MALI_");
875 unsigned len
= strlen(format
);
876 char *lower_format
= calloc(1, len
+ 1);
878 for (unsigned i
= 0; i
< len
; ++i
)
879 lower_format
[i
] = tolower(format
[i
]);
881 /* Sanity check sRGB flag is applied to RGB, per the name */
882 if (srgb
&& lower_format
[0] != 'r')
883 pandecode_msg("XXX: sRGB applied to non-colour format\n");
885 /* Just prefix with an s, so you get formats like srgba8_unorm */
887 pandecode_log_cont("s");
889 pandecode_log_cont("%s", lower_format
);
894 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
896 /* First, do some validation */
897 bool trivial_swizzle
= pandecode_validate_format_swizzle(
903 /* Next, print the swizzle */
904 pandecode_log_cont(".");
906 static const char components
[] = "rgba01";
908 for (unsigned c
= 0; c
< 4; ++c
) {
909 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
911 if (chan
>= MALI_CHANNEL_RESERVED_0
) {
912 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
915 pandecode_log_cont("%c", components
[chan
]);
920 pandecode_rt_format(struct mali_rt_format format
)
922 pandecode_log(".format = {\n");
925 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
926 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
927 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
929 pandecode_prop("block = %s", pandecode_block_format(format
.block
));
931 /* TODO: Map formats so we can check swizzles and print nicely */
932 pandecode_log("swizzle");
933 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
934 pandecode_log_cont(",\n");
936 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
937 (format
.nr_channels
+ 1));
939 pandecode_log(".flags = ");
940 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
941 pandecode_log_cont(",\n");
943 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
944 * which is a faster hardware-based alternative to the wallpaper method
945 * to preserve framebuffer contents across frames. In practice, MFBD
946 * preload is buggy on Midgard, and so this is a chicken bit. If this
947 * bit isn't set, most likely something broke unrelated to preload */
949 if (!format
.no_preload
) {
950 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
951 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
955 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
958 pandecode_log("},\n");
962 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct bifrost_framebuffer
*fb
)
964 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
967 for (int i
= 0; i
< (fb
->rt_count_1
+ 1); i
++) {
968 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct bifrost_render_target
);
969 struct pandecode_mapped_memory
*mem
=
970 pandecode_find_mapped_gpu_mem_containing(rt_va
);
971 const struct bifrost_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
973 pandecode_log("{\n");
976 pandecode_rt_format(rt
->format
);
978 if (rt
->format
.block
== MALI_BLOCK_AFBC
) {
979 pandecode_log(".afbc = {\n");
982 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
983 pandecode_prop("metadata = %s", a
);
986 pandecode_prop("stride = %d", rt
->afbc
.stride
);
987 pandecode_prop("unk = 0x%" PRIx32
, rt
->afbc
.unk
);
990 pandecode_log("},\n");
991 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.unk
) {
992 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
998 MEMORY_PROP(rt
, framebuffer
);
999 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
1001 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
1002 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
1003 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
1004 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
1005 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
1008 if (rt
->zero1
|| rt
->zero2
|| rt
->zero3
) {
1009 pandecode_msg("XXX: render target zeros tripped\n");
1010 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
1011 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
1012 pandecode_prop("zero3 = 0x%" PRIx32
, rt
->zero3
);
1016 pandecode_log("},\n");
1020 pandecode_log("};\n");
1023 static struct pandecode_fbd
1024 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
, bool is_compute
)
1026 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1027 const struct bifrost_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
1029 struct pandecode_fbd info
;
1031 if (fb
->sample_locations
) {
1032 /* The blob stores all possible sample locations in a single buffer
1033 * allocated on startup, and just switches the pointer when switching
1034 * MSAA state. For now, we just put the data into the cmdstream, but we
1035 * should do something like what the blob does with a real driver.
1037 * There seem to be 32 slots for sample locations, followed by another
1038 * 16. The second 16 is just the center location followed by 15 zeros
1039 * in all the cases I've identified (maybe shader vs. depth/color
1043 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->sample_locations
);
1045 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->sample_locations
);
1047 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
1050 for (int i
= 0; i
< 32 + 16; i
++) {
1051 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
1055 pandecode_log("};\n");
1058 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1061 pandecode_prop("stack_shift = 0x%x", fb
->stack_shift
);
1062 pandecode_prop("unk0 = 0x%x", fb
->unk0
);
1064 if (fb
->sample_locations
)
1065 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
1067 /* Assume that unknown1 was emitted in the last job for
1069 MEMORY_PROP(fb
, unknown1
);
1071 info
.width
= fb
->width1
+ 1;
1072 info
.height
= fb
->height1
+ 1;
1073 info
.rt_count
= fb
->rt_count_1
+ 1;
1075 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
1076 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
1077 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
1078 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
1080 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
1081 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
1082 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
1083 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
1085 pandecode_log(".mfbd_flags = ");
1086 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
1087 pandecode_log_cont(",\n");
1089 if (fb
->clear_stencil
)
1090 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
1092 if (fb
->clear_depth
)
1093 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
1095 /* TODO: What is this? Let's not blow up.. */
1096 if (fb
->unknown2
!= 0x1F)
1097 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
1099 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
1100 MEMORY_PROP(fb
, scratchpad
);
1101 const struct midgard_tiler_descriptor t
= fb
->tiler
;
1103 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
, true);
1105 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1107 if (fb
->zero3
|| fb
->zero4
) {
1108 pandecode_msg("XXX: framebuffer zeros tripped\n");
1109 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
1110 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
1114 pandecode_log("};\n");
1116 gpu_va
+= sizeof(struct bifrost_framebuffer
);
1118 info
.has_extra
= (fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
;
1120 if (info
.has_extra
) {
1121 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1122 const struct bifrost_fb_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
1124 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1127 MEMORY_PROP(fbx
, checksum
);
1129 if (fbx
->checksum_stride
)
1130 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
1132 pandecode_log(".flags_hi = ");
1133 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info
, fbx
->flags_lo
);
1134 pandecode_log_cont(",\n");
1136 pandecode_log(".flags_lo = ");
1137 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info
, fbx
->flags_lo
);
1138 pandecode_log_cont(",\n");
1140 pandecode_prop("zs_block = %s\n", pandecode_block_format(fbx
->zs_block
));
1142 if (fbx
->zs_block
== MALI_BLOCK_AFBC
) {
1143 pandecode_log(".ds_afbc = {\n");
1146 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
1147 pandecode_prop("depth_stencil_afbc_stride = %d",
1148 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
1149 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
1151 if (fbx
->ds_afbc
.zero1
|| fbx
->ds_afbc
.padding
) {
1152 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1153 pandecode_prop("zero1 = 0x%" PRIx32
,
1154 fbx
->ds_afbc
.zero1
);
1155 pandecode_prop("padding = 0x%" PRIx64
,
1156 fbx
->ds_afbc
.padding
);
1160 pandecode_log("},\n");
1162 pandecode_log(".ds_linear = {\n");
1165 if (fbx
->ds_linear
.depth
) {
1166 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
1167 pandecode_prop("depth_stride = %d",
1168 fbx
->ds_linear
.depth_stride
);
1169 } else if (fbx
->ds_linear
.depth_stride
) {
1170 pandecode_msg("XXX: depth stride zero tripped %d\n", fbx
->ds_linear
.depth_stride
);
1173 if (fbx
->ds_linear
.stencil
) {
1174 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1175 pandecode_prop("stencil_stride = %d",
1176 fbx
->ds_linear
.stencil_stride
);
1177 } else if (fbx
->ds_linear
.stencil_stride
) {
1178 pandecode_msg("XXX: stencil stride zero tripped %d\n", fbx
->ds_linear
.stencil_stride
);
1181 if (fbx
->ds_linear
.depth_stride_zero
||
1182 fbx
->ds_linear
.stencil_stride_zero
||
1183 fbx
->ds_linear
.zero1
|| fbx
->ds_linear
.zero2
) {
1184 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1185 pandecode_prop("depth_stride_zero = 0x%x",
1186 fbx
->ds_linear
.depth_stride_zero
);
1187 pandecode_prop("stencil_stride_zero = 0x%x",
1188 fbx
->ds_linear
.stencil_stride_zero
);
1189 pandecode_prop("zero1 = 0x%" PRIx32
,
1190 fbx
->ds_linear
.zero1
);
1191 pandecode_prop("zero2 = 0x%" PRIx32
,
1192 fbx
->ds_linear
.zero2
);
1196 pandecode_log("},\n");
1199 if (fbx
->zero3
|| fbx
->zero4
) {
1200 pandecode_msg("XXX: fb_extra zeros tripped\n");
1201 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1202 pandecode_prop("zero4 = 0x%" PRIx64
, fbx
->zero4
);
1206 pandecode_log("};\n");
1208 gpu_va
+= sizeof(struct bifrost_fb_extra
);
1212 pandecode_render_target(gpu_va
, job_no
, fb
);
1217 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1221 pandecode_padded_vertices(unsigned shift
, unsigned k
)
1223 unsigned odd
= 2*k
+ 1;
1224 unsigned pot
= 1 << shift
;
1225 pandecode_msg("padded_num_vertices = %d\n", odd
* pot
);
1228 /* Given a magic divisor, recover what we were trying to divide by.
1230 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1231 * 0 <= m < 2^N, for N bits in m.
1233 * Let q represent the number we would like to divide by.
1235 * By definition of a magic divisor for N-bit unsigned integers (a number you
1236 * multiply by to magically get division), m is a number such that:
1238 * (m * x) & (2^N - 1) = floor(x/q).
1239 * for all x on Z where 0 <= x < 2^N
1241 * Ignore the case where any of the above values equals zero; it is irrelevant
1242 * for our purposes (instanced arrays).
1244 * Choose x = q. Then:
1246 * (m * x) & (2^N - 1) = floor(x/q).
1247 * (m * q) & (2^N - 1) = floor(q/q).
1249 * floor(q/q) = floor(1) = 1, therefore:
1251 * (m * q) & (2^N - 1) = 1
1253 * Recall the identity that the bitwise AND of one less than a power-of-two
1254 * equals the modulo with that power of two, i.e. for all x:
1256 * x & (2^N - 1) = x % N
1262 * By definition, a modular multiplicative inverse of a number m is the number
1263 * q such that with respect to a modulos M:
1267 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1272 pandecode_magic_divisor(uint32_t magic
, unsigned shift
, unsigned orig_divisor
, unsigned extra
)
1275 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1276 * shift) the most lame way possible... just repeatedly add.
1277 * Asymptoptically slow but nobody cares in practice, unless you have
1278 * massive numbers of vertices or high divisors. */
1280 unsigned inverse
= 0;
1282 /* Magic implicitly has the highest bit set */
1285 /* Depending on rounding direction */
1290 uint32_t product
= magic
* inverse
;
1302 pandecode_msg("dividing by %d (maybe off by two)\n", inverse
);
1304 /* Recall we're supposed to divide by (gl_level_divisor *
1305 * padded_num_vertices) */
1307 unsigned padded_num_vertices
= inverse
/ orig_divisor
;
1309 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices
);
1314 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1315 mali_ptr addr
, int job_no
, char *suffix
,
1316 int count
, bool varying
, enum mali_job_type job_type
)
1318 char *prefix
= varying
? "varying" : "attribute";
1322 pandecode_msg("warn: No %s records\n", prefix
);
1326 union mali_attr
*attr
= pandecode_fetch_gpu_mem(mem
, addr
, sizeof(union mali_attr
) * count
);
1328 for (int i
= 0; i
< count
; ++i
) {
1329 /* First, check for special records */
1330 if (attr
[i
].elements
< MALI_RECORD_SPECIAL
) {
1332 pandecode_msg("XXX: tripped size=%d\n", attr
[i
].size
);
1334 if (attr
[i
].stride
) {
1335 /* gl_InstanceID passes a magic divisor in the
1336 * stride field to divide by the padded vertex
1337 * count. No other records should do so, so
1338 * stride should otherwise be zero. Note that
1339 * stride in the usual attribute sense doesn't
1340 * apply to special records. */
1342 bool has_divisor
= attr
[i
].elements
== MALI_ATTR_INSTANCEID
;
1344 pandecode_log_cont("/* %smagic divisor = %X */ ",
1345 has_divisor
? "" : "XXX: ", attr
[i
].stride
);
1348 if (attr
[i
].shift
|| attr
[i
].extra_flags
) {
1349 /* Attributes use these fields for
1350 * instancing/padding/etc type issues, but
1353 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1354 varying
? "XXX: " : "",
1355 attr
[i
].shift
, attr
[i
].extra_flags
);
1358 /* Print the special record name */
1359 bool attribute
= false;
1360 pandecode_log("%s_%d = %s;\n", prefix
, i
, pandecode_special_record(attr
[i
].elements
, &attribute
));
1363 if (attribute
== varying
)
1364 pandecode_msg("XXX: mismatched special record\n");
1369 enum mali_attr_mode mode
= attr
[i
].elements
& 7;
1371 if (mode
== MALI_ATTR_UNUSED
)
1372 pandecode_msg("XXX: unused attribute record\n");
1374 /* For non-linear records, we need to print the type of record */
1375 if (mode
!= MALI_ATTR_LINEAR
)
1376 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode
));
1378 /* Print the name to link with attr_meta */
1379 pandecode_log_cont("%s_%d", prefix
, i
);
1381 /* Print the stride and size */
1382 pandecode_log_cont("<%u>[%u]", attr
[i
].stride
, attr
[i
].size
);
1384 /* TODO: Sanity check the quotient itself. It must be equal to
1385 * (or be greater than, if the driver added padding) the padded
1388 /* Finally, print the pointer */
1389 mali_ptr raw_elements
= attr
[i
].elements
& ~7;
1390 char *a
= pointer_as_memory_reference(raw_elements
);
1391 pandecode_log_cont(" = (%s);\n", a
);
1394 /* Check the pointer */
1395 pandecode_validate_buffer(raw_elements
, attr
[i
].size
);
1397 /* shift/extra_flags exist only for instanced */
1398 if (attr
[i
].shift
| attr
[i
].extra_flags
) {
1399 /* These are set to random values by the blob for
1400 * varyings, most likely a symptom of uninitialized
1401 * memory where the hardware masked the bug. As such we
1402 * put this at a warning, not an error. */
1404 if (mode
== MALI_ATTR_LINEAR
)
1405 pandecode_msg("warn: instancing fields set for linear\n");
1407 pandecode_prop("shift = %d", attr
[i
].shift
);
1408 pandecode_prop("extra_flags = %d", attr
[i
].extra_flags
);
1411 /* Decode further where possible */
1413 if (mode
== MALI_ATTR_MODULO
) {
1414 pandecode_padded_vertices(
1416 attr
[i
].extra_flags
);
1419 if (mode
== MALI_ATTR_NPOT_DIVIDE
) {
1421 pandecode_log("{\n");
1423 pandecode_prop("unk = 0x%x", attr
[i
].unk
);
1424 pandecode_prop("magic_divisor = 0x%08x", attr
[i
].magic_divisor
);
1425 if (attr
[i
].zero
!= 0)
1426 pandecode_prop("XXX: zero tripped (0x%x)\n", attr
[i
].zero
);
1427 pandecode_prop("divisor = %d", attr
[i
].divisor
);
1428 pandecode_magic_divisor(attr
[i
].magic_divisor
, attr
[i
- 1].shift
, attr
[i
].divisor
, attr
[i
- 1].extra_flags
);
1430 pandecode_log("}, \n");
1435 pandecode_log("\n");
1439 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1441 /* TODO: Decode flags */
1442 mali_ptr shader_ptr
= ptr
& ~15;
1444 char *a
= pointer_as_memory_reference(shader_ptr
);
1445 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1452 pandecode_stencil(const char *name
, const struct mali_stencil_test
*stencil
)
1454 unsigned any_nonzero
=
1455 stencil
->ref
| stencil
->mask
| stencil
->func
|
1456 stencil
->sfail
| stencil
->dpfail
| stencil
->dppass
;
1458 if (any_nonzero
== 0)
1461 const char *func
= pandecode_func(stencil
->func
);
1462 const char *sfail
= pandecode_stencil_op(stencil
->sfail
);
1463 const char *dpfail
= pandecode_stencil_op(stencil
->dpfail
);
1464 const char *dppass
= pandecode_stencil_op(stencil
->dppass
);
1467 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil
->zero
);
1469 pandecode_log(".stencil_%s = {\n", name
);
1471 pandecode_prop("ref = %d", stencil
->ref
);
1472 pandecode_prop("mask = 0x%02X", stencil
->mask
);
1473 pandecode_prop("func = %s", func
);
1474 pandecode_prop("sfail = %s", sfail
);
1475 pandecode_prop("dpfail = %s", dpfail
);
1476 pandecode_prop("dppass = %s", dppass
);
1478 pandecode_log("},\n");
1482 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1485 pandecode_msg("XXX: blend zero tripped: %X\n", blend
->zero1
);
1487 pandecode_log(".equation = {\n");
1490 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1491 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1493 pandecode_log(".color_mask = ");
1494 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1495 pandecode_log_cont(",\n");
1498 pandecode_log("},\n");
1501 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1504 decode_bifrost_constant(u16 constant
)
1506 float lo
= (float) (constant
& 0xFF);
1507 float hi
= (float) (constant
>> 8);
1509 return (hi
/ 255.0) + (lo
/ 65535.0);
1513 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1515 struct bifrost_blend_rt
*b
=
1516 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1518 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1521 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1522 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1523 b
->constant
, decode_bifrost_constant(b
->constant
));
1525 /* TODO figure out blend shader enable bit */
1526 pandecode_blend_equation(&b
->equation
);
1527 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1528 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1529 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1532 pandecode_log("},\n");
1538 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1540 /* constant/equation is in a union */
1544 pandecode_log(".blend = {\n");
1548 pandecode_shader_address("shader", blend
->shader
);
1550 pandecode_blend_equation(&blend
->equation
);
1551 pandecode_prop("constant = %f", blend
->constant
);
1555 pandecode_log("},\n");
1557 /* Return blend shader to disassemble if present */
1558 return is_shader
? (blend
->shader
& ~0xF) : 0;
1562 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1564 struct midgard_blend_rt
*b
=
1565 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1567 /* Flags determine presence of blend shader */
1568 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1570 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1573 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1575 union midgard_blend blend
= b
->blend
;
1576 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1579 pandecode_log("};\n");
1584 /* Attributes and varyings have descriptor records, which contain information
1585 * about their format and ordering with the attribute/varying buffers. We'll
1586 * want to validate that the combinations specified are self-consistent.
1590 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1593 char *prefix
= varying
? "varying" : "attribute";
1594 unsigned max_index
= 0;
1595 snprintf(base
, sizeof(base
), "%s_meta", prefix
);
1597 struct mali_attr_meta
*attr_meta
;
1598 mali_ptr p
= varying
? v
->varying_meta
: v
->attribute_meta
;
1600 struct pandecode_mapped_memory
*attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
);
1602 for (int i
= 0; i
< count
; ++i
, p
+= sizeof(struct mali_attr_meta
)) {
1603 attr_meta
= pandecode_fetch_gpu_mem(attr_mem
, p
,
1606 /* If the record is discard, it should be zero for everything else */
1608 if (attr_meta
->format
== MALI_VARYING_DISCARD
) {
1611 attr_meta
->unknown1
|
1612 attr_meta
->unknown3
|
1613 attr_meta
->src_offset
;
1616 pandecode_msg("XXX: expected empty record for varying discard\n");
1618 /* We want to look for a literal 0000 swizzle -- this
1619 * is not encoded with all zeroes, however */
1621 enum mali_channel z
= MALI_CHANNEL_ZERO
;
1622 unsigned zero_swizzle
= z
| (z
<< 3) | (z
<< 6) | (z
<< 9);
1623 bool good_swizzle
= attr_meta
->swizzle
== zero_swizzle
;
1626 pandecode_msg("XXX: expected zero swizzle for discard\n");
1629 pandecode_msg("XXX: cannot discard attribute\n");
1631 /* If we're all good, omit the record */
1632 if (!zero
&& varying
&& good_swizzle
) {
1633 pandecode_log("/* discarded varying */\n");
1638 if (attr_meta
->index
> max_index
)
1639 max_index
= attr_meta
->index
;
1641 if (attr_meta
->unknown1
!= 0x2) {
1642 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1643 pandecode_prop("unknown1 = 0x%" PRIx64
, (u64
) attr_meta
->unknown1
);
1646 if (attr_meta
->unknown3
) {
1647 pandecode_msg("XXX: unexpected unknown3 set\n");
1648 pandecode_prop("unknown3 = 0x%" PRIx64
, (u64
) attr_meta
->unknown3
);
1651 pandecode_format_short(attr_meta
->format
, false);
1652 pandecode_log_cont(" %s_%u", prefix
, attr_meta
->index
);
1654 if (attr_meta
->src_offset
)
1655 pandecode_log_cont("[%u]", attr_meta
->src_offset
);
1657 pandecode_swizzle(attr_meta
->swizzle
, attr_meta
->format
);
1659 pandecode_log_cont(";\n");
1662 pandecode_log("\n");
1664 return count
? (max_index
+ 1) : 0;
1667 /* return bits [lo, hi) of word */
1669 bits(u32 word
, u32 lo
, u32 hi
)
1672 return word
; // avoid undefined behavior with the shift
1674 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1678 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool graphics
)
1680 pandecode_log_cont("{\n");
1683 /* Decode invocation_count. See the comment before the definition of
1684 * invocation_count for an explanation.
1687 unsigned size_y_shift
= bits(p
->invocation_shifts
, 0, 5);
1688 unsigned size_z_shift
= bits(p
->invocation_shifts
, 5, 10);
1689 unsigned workgroups_x_shift
= bits(p
->invocation_shifts
, 10, 16);
1690 unsigned workgroups_y_shift
= bits(p
->invocation_shifts
, 16, 22);
1691 unsigned workgroups_z_shift
= bits(p
->invocation_shifts
, 22, 28);
1692 unsigned workgroups_x_shift_2
= bits(p
->invocation_shifts
, 28, 32);
1694 unsigned size_x
= bits(p
->invocation_count
, 0, size_y_shift
) + 1;
1695 unsigned size_y
= bits(p
->invocation_count
, size_y_shift
, size_z_shift
) + 1;
1696 unsigned size_z
= bits(p
->invocation_count
, size_z_shift
, workgroups_x_shift
) + 1;
1698 unsigned groups_x
= bits(p
->invocation_count
, workgroups_x_shift
, workgroups_y_shift
) + 1;
1699 unsigned groups_y
= bits(p
->invocation_count
, workgroups_y_shift
, workgroups_z_shift
) + 1;
1700 unsigned groups_z
= bits(p
->invocation_count
, workgroups_z_shift
, 32) + 1;
1702 /* Even though we have this decoded, we want to ensure that the
1703 * representation is "unique" so we don't lose anything by printing only
1704 * the final result. More specifically, we need to check that we were
1705 * passed something in canonical form, since the definition per the
1706 * hardware is inherently not unique. How? Well, take the resulting
1707 * decode and pack it ourselves! If it is bit exact with what we
1708 * decoded, we're good to go. */
1710 struct mali_vertex_tiler_prefix ref
;
1711 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, graphics
);
1714 (p
->invocation_count
== ref
.invocation_count
) &&
1715 (p
->invocation_shifts
== ref
.invocation_shifts
);
1718 pandecode_msg("XXX: non-canonical workgroups packing\n");
1719 pandecode_msg("expected: %X, %X",
1720 ref
.invocation_count
,
1721 ref
.invocation_shifts
);
1723 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1724 pandecode_prop("size_y_shift = %d", size_y_shift
);
1725 pandecode_prop("size_z_shift = %d", size_z_shift
);
1726 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift
);
1727 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift
);
1728 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift
);
1729 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2
);
1732 /* Regardless, print the decode */
1733 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1734 size_x
, size_y
, size_z
,
1735 groups_x
, groups_y
, groups_z
);
1738 if (p
->unknown_draw
)
1739 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1741 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1743 if (p
->draw_mode
!= MALI_DRAW_NONE
)
1744 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p
->draw_mode
));
1746 /* Index count only exists for tiler jobs anyway */
1749 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1752 unsigned index_raw_size
= (p
->unknown_draw
& MALI_DRAW_INDEXED_SIZE
);
1753 index_raw_size
>>= MALI_DRAW_INDEXED_SHIFT
;
1755 /* Validate an index buffer is present if we need one. TODO: verify
1756 * relationship between invocation_count and index_count */
1759 unsigned count
= p
->index_count
;
1762 unsigned size
= (index_raw_size
== 0x3) ? 4 : index_raw_size
;
1764 /* Ensure we got a size, and if so, validate the index buffer
1765 * is large enough to hold a full set of indices of the given
1768 if (!index_raw_size
)
1769 pandecode_msg("XXX: index size missing\n");
1771 pandecode_validate_buffer(p
->indices
, count
* size
);
1772 } else if (index_raw_size
)
1773 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size
);
1775 if (p
->offset_bias_correction
)
1776 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1778 /* TODO: Figure out what this is. It's not zero */
1779 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1782 pandecode_log("},\n");
1786 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1788 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1789 struct mali_uniform_buffer_meta
*PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1791 for (int i
= 0; i
< ubufs_count
; i
++) {
1792 unsigned size
= (ubufs
[i
].size
+ 1) * 16;
1793 mali_ptr addr
= ubufs
[i
].ptr
<< 2;
1795 pandecode_validate_buffer(addr
, size
);
1797 char *ptr
= pointer_as_memory_reference(ubufs
[i
].ptr
<< 2);
1798 pandecode_log("ubuf_%d[%u] = %s;\n", i
, size
, ptr
);
1802 pandecode_log("\n");
1806 pandecode_uniforms(mali_ptr uniforms
, unsigned uniform_count
)
1808 pandecode_validate_buffer(uniforms
, uniform_count
* 16);
1810 char *ptr
= pointer_as_memory_reference(uniforms
);
1811 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count
, ptr
);
1816 pandecode_scratchpad(uintptr_t pscratchpad
, int job_no
, char *suffix
)
1819 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(pscratchpad
);
1821 struct bifrost_scratchpad
*PANDECODE_PTR_VAR(scratchpad
, mem
, pscratchpad
);
1823 if (scratchpad
->zero
) {
1824 pandecode_msg("XXX: scratchpad zero tripped");
1825 pandecode_prop("zero = 0x%x\n", scratchpad
->zero
);
1828 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64
"_%d%s = {\n", pscratchpad
, job_no
, suffix
);
1831 pandecode_prop("flags = 0x%x", scratchpad
->flags
);
1832 MEMORY_PROP(scratchpad
, gpu_scratchpad
);
1835 pandecode_log("};\n");
1839 shader_type_for_job(unsigned type
)
1842 case JOB_TYPE_VERTEX
: return "VERTEX";
1843 case JOB_TYPE_TILER
: return "FRAGMENT";
1844 case JOB_TYPE_COMPUTE
: return "COMPUTE";
1850 static unsigned shader_id
= 0;
1852 static struct midgard_disasm_stats
1853 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1854 bool is_bifrost
, unsigned gpu_id
)
1856 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1857 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1859 /* Compute maximum possible size */
1860 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1862 /* Print some boilerplate to clearly denote the assembly (which doesn't
1863 * obey indentation rules), and actually do the disassembly! */
1865 pandecode_log_cont("\n\n");
1867 struct midgard_disasm_stats stats
;
1870 disassemble_bifrost(pandecode_dump_stream
, code
, sz
, false);
1872 /* TODO: Extend stats to Bifrost */
1873 stats
.texture_count
= -128;
1874 stats
.sampler_count
= -128;
1875 stats
.attribute_count
= -128;
1876 stats
.varying_count
= -128;
1877 stats
.uniform_count
= -128;
1878 stats
.uniform_buffer_count
= -128;
1879 stats
.work_count
= -128;
1881 stats
.instruction_count
= 0;
1882 stats
.bundle_count
= 0;
1883 stats
.quadword_count
= 0;
1884 stats
.helper_invocations
= false;
1886 stats
= disassemble_midgard(pandecode_dump_stream
,
1888 type
== JOB_TYPE_TILER
?
1889 MESA_SHADER_FRAGMENT
: MESA_SHADER_VERTEX
);
1892 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1893 * driver-internal purposes with the blob and interfere */
1895 bool should_shaderdb
= type
!= JOB_TYPE_COMPUTE
;
1897 if (should_shaderdb
) {
1898 unsigned nr_threads
=
1899 (stats
.work_count
<= 4) ? 4 :
1900 (stats
.work_count
<= 8) ? 2 :
1903 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
1904 "%u inst, %u bundles, %u quadwords, "
1905 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1907 shader_type_for_job(type
),
1908 stats
.instruction_count
, stats
.bundle_count
, stats
.quadword_count
,
1909 stats
.work_count
, nr_threads
);
1917 pandecode_texture(mali_ptr u
,
1918 struct pandecode_mapped_memory
*tmem
,
1919 unsigned job_no
, unsigned tex
)
1921 struct mali_texture_descriptor
*PANDECODE_PTR_VAR(t
, tmem
, u
);
1923 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64
"_%d_%d = {\n", u
, job_no
, tex
);
1926 struct mali_texture_format f
= t
->format
;
1928 /* See the definiton of enum mali_texture_type */
1930 bool is_cube
= f
.type
== MALI_TEX_CUBE
;
1931 unsigned dimension
= is_cube
? 2 : f
.type
;
1933 pandecode_make_indent();
1935 /* TODO: Are there others? */
1936 bool is_zs
= f
.format
== MALI_Z32_UNORM
;
1938 /* Recall Z/S switched the meaning of linear/tiled .. */
1939 if (is_zs
&& f
.layout
== MALI_TEXTURE_LINEAR
)
1940 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1942 /* Print the layout. Default is linear; a modifier can denote AFBC or
1943 * u-interleaved/tiled modes */
1945 if (f
.layout
== MALI_TEXTURE_AFBC
)
1946 pandecode_log_cont("afbc");
1947 else if (f
.layout
== MALI_TEXTURE_TILED
)
1948 pandecode_log_cont("tiled");
1949 else if (f
.layout
== MALI_TEXTURE_LINEAR
)
1950 pandecode_log_cont("linear");
1952 pandecode_msg("XXX: invalid texture layout 0x%X\n", f
.layout
);
1954 pandecode_swizzle(t
->swizzle
, f
.format
);
1955 pandecode_log_cont(" ");
1957 /* Distinguish cube/2D with modifier */
1960 pandecode_log_cont("cube ");
1962 pandecode_format_short(f
.format
, f
.srgb
);
1963 pandecode_swizzle(f
.swizzle
, f
.format
);
1965 /* All four width/height/depth/array_size dimensions are present
1966 * regardless of the type of texture, but it is an error to have
1967 * non-zero dimensions for unused dimensions. Verify this. array_size
1968 * can always be set, as can width. */
1970 if (t
->height
&& dimension
< 2)
1971 pandecode_msg("XXX: nonzero height for <2D texture\n");
1973 if (t
->depth
&& dimension
< 3)
1974 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1976 /* Print only the dimensions that are actually there */
1978 pandecode_log_cont(": %d", t
->width
+ 1);
1981 pandecode_log_cont("x%u", t
->height
+ 1);
1984 pandecode_log_cont("x%u", t
->depth
+ 1);
1987 pandecode_log_cont("[%u]", t
->array_size
+ 1);
1990 pandecode_log_cont(" mip %u", t
->levels
);
1992 pandecode_log_cont("\n");
1994 if (f
.unknown1
| f
.zero
) {
1995 pandecode_msg("XXX: texture format zero tripped\n");
1996 pandecode_prop("unknown1 = %" PRId32
, f
.unknown1
);
1997 pandecode_prop("zero = %" PRId32
, f
.zero
);
2001 pandecode_msg("XXX: expected unknown texture bit set\n");
2002 pandecode_prop("unknown2 = %" PRId32
, f
.unknown2
);
2005 if (t
->swizzle_zero
) {
2006 pandecode_msg("XXX: swizzle zero tripped\n");
2007 pandecode_prop("swizzle_zero = %d", t
->swizzle_zero
);
2010 if (t
->unknown3
| t
->unknown3A
| t
->unknown5
| t
->unknown6
| t
->unknown7
) {
2011 pandecode_msg("XXX: texture zero tripped\n");
2012 pandecode_prop("unknown3 = %" PRId16
, t
->unknown3
);
2013 pandecode_prop("unknown3A = %" PRId8
, t
->unknown3A
);
2014 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
2015 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
2016 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
2019 pandecode_log(".payload = {\n");
2022 /* A bunch of bitmap pointers follow.
2023 * We work out the correct number,
2024 * based on the mipmap/cubemap
2025 * properties, but dump extra
2026 * possibilities to futureproof */
2028 int bitmap_count
= t
->levels
+ 1;
2030 /* Miptree for each face */
2031 if (f
.type
== MALI_TEX_CUBE
)
2033 else if (f
.type
== MALI_TEX_3D
)
2034 bitmap_count
*= t
->depth
;
2036 /* Array of textures */
2037 bitmap_count
*= (t
->array_size
+ 1);
2039 /* Stride for each element */
2040 if (f
.manual_stride
)
2043 mali_ptr
*pointers_and_strides
= pandecode_fetch_gpu_mem(tmem
,
2044 u
+ sizeof(*t
), sizeof(mali_ptr
) * bitmap_count
);
2045 for (int i
= 0; i
< bitmap_count
; ++i
) {
2046 /* How we dump depends if this is a stride or a pointer */
2048 if (f
.manual_stride
&& (i
& 1)) {
2049 /* signed 32-bit snuck in as a 64-bit pointer */
2050 uint64_t stride_set
= pointers_and_strides
[i
];
2051 uint32_t clamped_stride
= stride_set
;
2052 int32_t stride
= clamped_stride
;
2053 assert(stride_set
== clamped_stride
);
2054 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
2056 char *a
= pointer_as_memory_reference(pointers_and_strides
[i
]);
2057 pandecode_log("%s, \n", a
);
2063 pandecode_log("},\n");
2066 pandecode_log("};\n");
2069 /* 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 */
2072 pandecode_shader_prop(const char *name
, unsigned claim
, signed truth
, bool fuzzy
)
2081 if ((truth
>= 0) && !fuzzy
) {
2082 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2083 (truth
< claim
) ? "warn" : "XXX",
2084 name
, truth
, claim
);
2085 } else if ((claim
> -truth
) && !fuzzy
) {
2086 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2087 name
, -truth
, claim
);
2088 } else if (fuzzy
&& (claim
< truth
))
2089 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2090 name
, truth
, claim
);
2092 pandecode_log(".%s = %" PRId16
, name
, claim
);
2095 pandecode_log_cont(" /* %u used */", truth
);
2097 pandecode_log_cont(",\n");
2101 pandecode_blend_shader_disassemble(mali_ptr shader
, int job_no
, int job_type
,
2102 bool is_bifrost
, unsigned gpu_id
)
2104 struct midgard_disasm_stats stats
=
2105 pandecode_shader_disassemble(shader
, job_no
, job_type
, is_bifrost
, gpu_id
);
2107 bool has_texture
= (stats
.texture_count
> 0);
2108 bool has_sampler
= (stats
.sampler_count
> 0);
2109 bool has_attribute
= (stats
.attribute_count
> 0);
2110 bool has_varying
= (stats
.varying_count
> 0);
2111 bool has_uniform
= (stats
.uniform_count
> 0);
2112 bool has_ubo
= (stats
.uniform_buffer_count
> 0);
2114 if (has_texture
|| has_sampler
)
2115 pandecode_msg("XXX: blend shader accessing textures\n");
2117 if (has_attribute
|| has_varying
)
2118 pandecode_msg("XXX: blend shader accessing interstage\n");
2120 if (has_uniform
|| has_ubo
)
2121 pandecode_msg("XXX: blend shader accessing uniforms\n");
2125 pandecode_vertex_tiler_postfix_pre(
2126 const struct mali_vertex_tiler_postfix
*p
,
2127 int job_no
, enum mali_job_type job_type
,
2128 char *suffix
, bool is_bifrost
, unsigned gpu_id
)
2130 struct pandecode_mapped_memory
*attr_mem
;
2132 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2133 * are the only things actually needed from the FBD, vertex/tiler jobs
2134 * no longer reference the FBD -- instead, this field points to some
2135 * info about the scratchpad.
2138 struct pandecode_fbd fbd_info
= {
2139 /* Default for Bifrost */
2144 pandecode_scratchpad(p
->framebuffer
& ~1, job_no
, suffix
);
2145 else if (p
->framebuffer
& MALI_MFBD
)
2146 fbd_info
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->framebuffer
) & FBD_MASK
, job_no
, false, job_type
== JOB_TYPE_COMPUTE
);
2147 else if (job_type
== JOB_TYPE_COMPUTE
)
2148 pandecode_compute_fbd((u64
) (uintptr_t) p
->framebuffer
, job_no
);
2150 fbd_info
= pandecode_sfbd((u64
) (uintptr_t) p
->framebuffer
, job_no
, false, gpu_id
);
2152 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
2153 int texture_count
= 0, sampler_count
= 0;
2156 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->shader
);
2157 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, p
->shader
);
2159 /* Disassemble ahead-of-time to get stats. Initialize with
2160 * stats for the missing-shader case so we get validation
2163 struct midgard_disasm_stats info
= {
2166 .attribute_count
= 0,
2170 .uniform_count
= -128,
2171 .uniform_buffer_count
= 0
2174 if (s
->shader
& ~0xF)
2175 info
= pandecode_shader_disassemble(s
->shader
& ~0xF, job_no
, job_type
, is_bifrost
, gpu_id
);
2177 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", p
->shader
, job_no
, suffix
);
2180 /* Save for dumps */
2181 attribute_count
= s
->attribute_count
;
2182 varying_count
= s
->varying_count
;
2183 texture_count
= s
->texture_count
;
2184 sampler_count
= s
->sampler_count
;
2187 uniform_count
= s
->bifrost2
.uniform_count
;
2188 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
2190 uniform_count
= s
->midgard1
.uniform_count
;
2191 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
2194 pandecode_shader_address("shader", s
->shader
);
2196 pandecode_shader_prop("texture_count", s
->texture_count
, info
.texture_count
, false);
2197 pandecode_shader_prop("sampler_count", s
->sampler_count
, info
.sampler_count
, false);
2198 pandecode_shader_prop("attribute_count", s
->attribute_count
, info
.attribute_count
, false);
2199 pandecode_shader_prop("varying_count", s
->varying_count
, info
.varying_count
, false);
2200 pandecode_shader_prop("uniform_buffer_count",
2201 uniform_buffer_count
,
2202 info
.uniform_buffer_count
, true);
2205 pandecode_shader_prop("uniform_count",
2207 info
.uniform_count
, false);
2209 pandecode_shader_prop("work_count",
2210 s
->midgard1
.work_count
, info
.work_count
, false);
2214 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32
, s
->bifrost1
.unk1
);
2216 bool helpers
= s
->midgard1
.flags
& MALI_HELPER_INVOCATIONS
;
2217 s
->midgard1
.flags
&= ~MALI_HELPER_INVOCATIONS
;
2219 if (helpers
!= info
.helper_invocations
) {
2220 pandecode_msg("XXX: expected helpers %u but got %u\n",
2221 info
.helper_invocations
, helpers
);
2224 pandecode_log(".midgard1.flags = ");
2225 pandecode_log_decoded_flags(shader_midgard1_flag_info
, s
->midgard1
.flags
);
2226 pandecode_log_cont(",\n");
2228 pandecode_prop("midgard1.unknown2 = 0x%" PRIx32
, s
->midgard1
.unknown2
);
2231 if (s
->depth_units
|| s
->depth_factor
) {
2232 pandecode_prop("depth_factor = %f", s
->depth_factor
);
2233 pandecode_prop("depth_units = %f", s
->depth_units
);
2236 if (s
->alpha_coverage
) {
2237 bool invert_alpha_coverage
= s
->alpha_coverage
& 0xFFF0;
2238 uint16_t inverted_coverage
= invert_alpha_coverage
? ~s
->alpha_coverage
: s
->alpha_coverage
;
2240 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2241 invert_alpha_coverage
? "~" : "",
2242 MALI_GET_ALPHA_COVERAGE(inverted_coverage
));
2245 if (s
->unknown2_3
|| s
->unknown2_4
) {
2246 pandecode_log(".unknown2_3 = ");
2248 int unknown2_3
= s
->unknown2_3
;
2249 int unknown2_4
= s
->unknown2_4
;
2251 /* We're not quite sure what these flags mean without the depth test, if anything */
2253 if (unknown2_3
& (MALI_DEPTH_WRITEMASK
| MALI_DEPTH_FUNC_MASK
)) {
2254 const char *func
= pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3
));
2255 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
2257 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
2260 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
2261 pandecode_log_cont(",\n");
2263 pandecode_log(".unknown2_4 = ");
2264 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
2265 pandecode_log_cont(",\n");
2268 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
2269 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
2270 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
2273 pandecode_stencil("front", &s
->stencil_front
);
2274 pandecode_stencil("back", &s
->stencil_back
);
2277 pandecode_log(".bifrost2 = {\n");
2280 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
2281 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
2282 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
2283 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
2286 pandecode_log("},\n");
2287 } else if (s
->midgard2
.unknown2_7
) {
2288 pandecode_log(".midgard2 = {\n");
2291 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
2293 pandecode_log("},\n");
2297 pandecode_prop("unknown2_8 = 0x%" PRIx32
, s
->unknown2_8
);
2300 /* TODO: Blend shaders routing/disasm */
2301 union midgard_blend blend
= s
->blend
;
2302 mali_ptr shader
= pandecode_midgard_blend(&blend
, s
->unknown2_3
& MALI_HAS_BLEND_SHADER
);
2304 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
2308 pandecode_log("};\n");
2310 /* MRT blend fields are used whenever MFBD is used, with
2311 * per-RT descriptors */
2313 if (job_type
== JOB_TYPE_TILER
&& p
->framebuffer
& MALI_MFBD
) {
2314 void* blend_base
= (void *) (s
+ 1);
2316 for (unsigned i
= 0; i
< fbd_info
.rt_count
; i
++) {
2317 mali_ptr shader
= 0;
2320 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
2322 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
2325 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
2330 pandecode_msg("XXX: missing shader descriptor\n");
2333 struct pandecode_mapped_memory
*fmem
= pandecode_find_mapped_gpu_mem_containing(p
->viewport
);
2334 struct mali_viewport
*PANDECODE_PTR_VAR(f
, fmem
, p
->viewport
);
2336 pandecode_log("struct mali_viewport viewport_%"PRIx64
"_%d%s = {\n", p
->viewport
, job_no
, suffix
);
2339 pandecode_prop("clip_minx = %f", f
->clip_minx
);
2340 pandecode_prop("clip_miny = %f", f
->clip_miny
);
2341 pandecode_prop("clip_minz = %f", f
->clip_minz
);
2342 pandecode_prop("clip_maxx = %f", f
->clip_maxx
);
2343 pandecode_prop("clip_maxy = %f", f
->clip_maxy
);
2344 pandecode_prop("clip_maxz = %f", f
->clip_maxz
);
2346 /* Only the higher coordinates are MALI_POSITIVE scaled */
2348 pandecode_prop("viewport0 = { %d, %d }",
2349 f
->viewport0
[0], f
->viewport0
[1]);
2351 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2352 f
->viewport1
[0] + 1, f
->viewport1
[1] + 1);
2355 pandecode_log("};\n");
2358 unsigned max_attr_index
= 0;
2360 if (p
->attribute_meta
)
2361 max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
2363 if (p
->attributes
) {
2364 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
2365 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false, job_type
);
2368 /* Varyings are encoded like attributes but not actually sent; we just
2369 * pass a zero buffer with the right stride/size set, (or whatever)
2370 * since the GPU will write to it itself */
2372 if (p
->varying_meta
) {
2373 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
2377 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
2379 /* Number of descriptors depends on whether there are
2380 * non-internal varyings */
2382 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true, job_type
);
2385 if (p
->uniform_buffers
) {
2386 if (uniform_buffer_count
)
2387 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
2389 pandecode_msg("warn: UBOs specified but not referenced\n");
2390 } else if (uniform_buffer_count
)
2391 pandecode_msg("XXX: UBOs referenced but not specified\n");
2393 /* We don't want to actually dump uniforms, but we do need to validate
2394 * that the counts we were given are sane */
2398 pandecode_uniforms(p
->uniforms
, uniform_count
);
2400 pandecode_msg("warn: Uniforms specified but not referenced\n");
2401 } else if (uniform_count
)
2402 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2404 if (p
->texture_trampoline
) {
2405 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(p
->texture_trampoline
);
2408 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
);
2410 pandecode_log("uint64_t texture_trampoline_%"PRIx64
"_%d[] = {\n", p
->texture_trampoline
, job_no
);
2413 for (int tex
= 0; tex
< texture_count
; ++tex
) {
2414 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
2415 char *a
= pointer_as_memory_reference(*u
);
2416 pandecode_log("%s,\n", a
);
2421 pandecode_log("};\n");
2423 /* Now, finally, descend down into the texture descriptor */
2424 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
2425 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
2426 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
2428 pandecode_texture(*u
, tmem
, job_no
, tex
);
2433 if (p
->sampler_descriptor
) {
2434 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->sampler_descriptor
);
2437 struct mali_sampler_descriptor
*s
;
2439 mali_ptr d
= p
->sampler_descriptor
;
2441 for (int i
= 0; i
< sampler_count
; ++i
) {
2442 s
= pandecode_fetch_gpu_mem(smem
, d
+ sizeof(*s
) * i
, sizeof(*s
));
2444 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", d
+ sizeof(*s
) * i
, job_no
, i
);
2447 pandecode_log(".filter_mode = ");
2448 pandecode_log_decoded_flags(sampler_flag_info
, s
->filter_mode
);
2449 pandecode_log_cont(",\n");
2451 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
2452 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
2455 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s
->lod_bias
));
2457 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s
->wrap_s
));
2458 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s
->wrap_t
));
2459 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s
->wrap_r
));
2461 pandecode_prop("compare_func = %s", pandecode_func(s
->compare_func
));
2463 if (s
->zero
|| s
->zero2
) {
2464 pandecode_msg("XXX: sampler zero tripped\n");
2465 pandecode_prop("zero = 0x%X, 0x%X\n", s
->zero
, s
->zero2
);
2468 pandecode_prop("seamless_cube_map = %d", s
->seamless_cube_map
);
2470 pandecode_prop("border_color = { %f, %f, %f, %f }",
2474 s
->border_color
[3]);
2477 pandecode_log("};\n");
2484 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2486 if (p
->shader
& 0xF)
2487 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p
->shader
& 0xF));
2489 if (!(p
->position_varying
|| p
->occlusion_counter
))
2492 pandecode_log(".postfix = {\n");
2495 MEMORY_PROP(p
, position_varying
);
2496 MEMORY_PROP(p
, occlusion_counter
);
2499 pandecode_log("},\n");
2503 pandecode_vertex_only_bfr(struct bifrost_vertex_only
*v
)
2505 pandecode_log_cont("{\n");
2508 pandecode_prop("unk2 = 0x%x", v
->unk2
);
2510 if (v
->zero0
|| v
->zero1
) {
2511 pandecode_msg("XXX: vertex only zero tripped");
2512 pandecode_prop("zero0 = 0x%" PRIx32
, v
->zero0
);
2513 pandecode_prop("zero1 = 0x%" PRIx64
, v
->zero1
);
2517 pandecode_log("}\n");
2521 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2524 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2525 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2527 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no
);
2531 pandecode_msg("XXX: tiler heap zero tripped\n");
2532 pandecode_prop("zero = 0x%x", h
->zero
);
2535 for (int i
= 0; i
< 12; i
++) {
2536 if (h
->zeros
[i
] != 0) {
2537 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2542 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2543 MEMORY_PROP(h
, tiler_heap_start
);
2544 MEMORY_PROP(h
, tiler_heap_free
);
2546 /* this might point to the beginning of another buffer, when it's
2547 * really the end of the tiler heap buffer, so we have to be careful
2548 * here. but for zero length, we need the same pointer.
2551 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2552 MEMORY_PROP(h
, tiler_heap_start
);
2554 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2555 pandecode_prop("tiler_heap_end = %s + 1", a
);
2560 pandecode_log("};\n");
2564 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2566 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2567 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2569 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2571 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no
);
2574 if (t
->zero0
|| t
->zero1
) {
2575 pandecode_msg("XXX: tiler meta zero tripped\n");
2576 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2577 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2580 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2581 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2583 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2584 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2586 for (int i
= 0; i
< 12; i
++) {
2587 if (t
->zeros
[i
] != 0) {
2588 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2594 pandecode_log("};\n");
2598 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2600 pandecode_log(".gl_enables = ");
2602 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2604 pandecode_log_cont(",\n");
2608 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2610 if (u
.pointer
== 0x0)
2613 pandecode_log(".primitive_size = {\n");
2617 pandecode_prop("constant = %f", u
.constant
);
2619 MEMORY_PROP((&u
), pointer
);
2623 pandecode_log("},\n");
2627 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2629 pandecode_log_cont("{\n");
2632 /* TODO: gl_PointSize on Bifrost */
2633 pandecode_primitive_size(t
->primitive_size
, true);
2635 pandecode_gl_enables(t
->gl_enables
, JOB_TYPE_TILER
);
2637 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2638 || t
->zero6
|| t
->zero7
|| t
->zero8
) {
2639 pandecode_msg("XXX: tiler only zero tripped\n");
2640 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2641 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2642 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2643 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2644 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2645 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2646 pandecode_prop("zero7 = 0x%" PRIx32
, t
->zero7
);
2647 pandecode_prop("zero8 = 0x%" PRIx64
, t
->zero8
);
2651 pandecode_log("},\n");
2655 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2656 const struct pandecode_mapped_memory
*mem
,
2657 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2659 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2661 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2663 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no
);
2666 pandecode_log(".prefix = ");
2667 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2669 pandecode_log(".vertex = ");
2670 pandecode_vertex_only_bfr(&v
->vertex
);
2672 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2675 pandecode_log("};\n");
2681 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2682 const struct pandecode_mapped_memory
*mem
,
2683 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2685 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2687 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2688 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2690 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no
);
2693 pandecode_log(".prefix = ");
2694 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2696 pandecode_log(".tiler = ");
2697 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2699 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2702 pandecode_log("};\n");
2708 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2709 const struct pandecode_mapped_memory
*mem
,
2710 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2712 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2714 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false, gpu_id
);
2716 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2719 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2720 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2722 bool is_graphics
= (h
->job_type
== JOB_TYPE_VERTEX
) || (h
->job_type
== JOB_TYPE_TILER
);
2724 pandecode_log(".prefix = ");
2725 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, is_graphics
);
2727 pandecode_gl_enables(v
->gl_enables
, h
->job_type
);
2729 if (v
->instance_shift
|| v
->instance_odd
) {
2730 pandecode_prop("instance_shift = 0x%d /* %d */",
2731 v
->instance_shift
, 1 << v
->instance_shift
);
2732 pandecode_prop("instance_odd = 0x%X /* %d */",
2733 v
->instance_odd
, (2 * v
->instance_odd
) + 1);
2735 pandecode_padded_vertices(v
->instance_shift
, v
->instance_odd
);
2738 if (v
->offset_start
)
2739 pandecode_prop("offset_start = %d", v
->offset_start
);
2742 pandecode_msg("XXX: midgard payload zero tripped\n");
2743 pandecode_prop("zero5 = 0x%" PRIx64
, v
->zero5
);
2746 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2749 pandecode_log("};\n");
2755 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2756 mali_ptr payload
, int job_no
,
2757 bool is_bifrost
, unsigned gpu_id
)
2759 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2761 bool is_mfbd
= s
->framebuffer
& MALI_MFBD
;
2763 /* Bifrost theoretically may retain support for SFBD on compute jobs,
2764 * but for graphics workloads with a FRAGMENT payload, use MFBD */
2766 if (!is_mfbd
&& is_bifrost
)
2767 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2769 struct pandecode_fbd info
;
2772 info
= pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true, false);
2774 info
= pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true, gpu_id
);
2776 /* Compute the tag for the tagged pointer. This contains the type of
2777 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2778 * additional structures follow the MFBD header (an extra payload or
2779 * not, as well as a count of render targets) */
2781 unsigned expected_tag
= is_mfbd
? MALI_MFBD
: 0;
2785 expected_tag
|= MALI_MFBD_TAG_EXTRA
;
2787 expected_tag
|= (MALI_POSITIVE(info
.rt_count
) << 2);
2790 if ((s
->min_tile_coord
| s
->max_tile_coord
) & ~(MALI_X_COORD_MASK
| MALI_Y_COORD_MASK
)) {
2791 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2792 pandecode_prop("min_tile_coord = 0x%X\n", s
->min_tile_coord
);
2793 pandecode_prop("max_tile_coord = 0x%X\n", s
->min_tile_coord
);
2796 /* Extract tile coordinates */
2798 unsigned min_x
= MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2799 unsigned min_y
= MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2801 unsigned max_x
= (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2802 unsigned max_y
= (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2804 /* For the max, we also want the floored (rather than ceiled) version for checking */
2806 unsigned max_x_f
= (MALI_TILE_COORD_X(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2807 unsigned max_y_f
= (MALI_TILE_COORD_Y(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2809 /* Validate the coordinates are well-ordered */
2812 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2813 else if (min_x
> max_x
)
2814 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2817 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2818 else if (min_y
> max_y
)
2819 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2821 /* Validate the coordinates fit inside the framebuffer. We use floor,
2822 * rather than ceil, for the max coordinates, since the tile
2823 * coordinates for something like an 800x600 framebuffer will actually
2824 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2826 if ((min_x
> info
.width
) || (max_x_f
> info
.width
))
2827 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2829 if ((min_y
> info
.height
) || (max_y_f
> info
.height
))
2830 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2832 /* After validation, we print */
2834 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x
, min_y
, max_x
, max_y
);
2836 /* The FBD is a tagged pointer */
2838 unsigned tag
= (s
->framebuffer
& ~FBD_MASK
);
2840 if (tag
!= expected_tag
)
2841 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag
, tag
);
2846 static int job_descriptor_number
= 0;
2849 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
, unsigned gpu_id
)
2851 struct mali_job_descriptor_header
*h
;
2853 int start_number
= 0;
2858 struct pandecode_mapped_memory
*mem
=
2859 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2863 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2865 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2866 * high 32-bits of the 64-bit pointer are reused to store
2869 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
2870 h
->job_type
!= JOB_TYPE_FRAGMENT
? 4 : 0;
2871 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
2873 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
, 256);
2875 int job_no
= job_descriptor_number
++;
2878 start_number
= job_no
;
2880 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2883 pandecode_prop("job_type = %s", pandecode_job_type(h
->job_type
));
2885 if (h
->job_descriptor_size
)
2886 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2888 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2889 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2890 h
->exception_status
,
2891 (h
->exception_status
>> 16) & 0xFFFF,
2892 pandecode_exception_access((h
->exception_status
>> 8) & 0x3),
2893 h
->exception_status
& 0xFF);
2895 if (h
->first_incomplete_task
)
2896 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2898 if (h
->fault_pointer
)
2899 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2902 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2904 pandecode_prop("job_index = %d", h
->job_index
);
2906 if (h
->unknown_flags
)
2907 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2909 if (h
->job_dependency_index_1
)
2910 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2912 if (h
->job_dependency_index_2
)
2913 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2916 pandecode_log("};\n");
2918 /* Do not touch the field yet -- decode the payload first, and
2919 * don't touch that either. This is essential for the uploads
2920 * to occur in sequence and therefore be dynamically allocated
2921 * correctly. Do note the size, however, for that related
2924 switch (h
->job_type
) {
2925 case JOB_TYPE_WRITE_VALUE
: {
2926 struct mali_payload_write_value
*s
= payload
;
2927 pandecode_log("struct mali_payload_write_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2929 MEMORY_PROP(s
, address
);
2931 if (s
->value_descriptor
!= MALI_WRITE_VALUE_ZERO
) {
2932 pandecode_msg("XXX: unknown value descriptor\n");
2933 pandecode_prop("value_descriptor = 0x%" PRIX32
, s
->value_descriptor
);
2937 pandecode_msg("XXX: set value tripped\n");
2938 pandecode_prop("reserved = 0x%" PRIX32
, s
->reserved
);
2941 pandecode_prop("immediate = 0x%" PRIX64
, s
->immediate
);
2943 pandecode_log("};\n");
2948 case JOB_TYPE_TILER
:
2949 case JOB_TYPE_VERTEX
:
2950 case JOB_TYPE_COMPUTE
:
2952 if (h
->job_type
== JOB_TYPE_TILER
)
2953 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2955 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2957 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2961 case JOB_TYPE_FRAGMENT
:
2962 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
, gpu_id
);
2969 /* Handle linkage */
2972 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no
- 1);
2973 pandecode_log_cont("next_job = job_%d_p;\n\n", job_no
);
2978 } while ((jc_gpu_va
= h
->next_job
));
2980 return start_number
;