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
),
229 FLAG_INFO(DEPTH_CLIP_NEAR
),
230 FLAG_INFO(DEPTH_CLIP_FAR
),
234 static const struct pandecode_flag_info u4_flag_info
[] = {
236 FLAG_INFO(NO_DITHER
),
237 FLAG_INFO(DEPTH_RANGE_A
),
238 FLAG_INFO(DEPTH_RANGE_B
),
239 FLAG_INFO(STENCIL_TEST
),
240 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER
),
245 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
246 static const struct pandecode_flag_info mfbd_fmt_flag_info
[] = {
254 #define FLAG_INFO(flag) { MALI_AFBC_##flag, "MALI_AFBC_" #flag }
255 static const struct pandecode_flag_info afbc_fmt_flag_info
[] = {
261 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
262 static const struct pandecode_flag_info mfbd_extra_flag_hi_info
[] = {
268 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
269 static const struct pandecode_flag_info mfbd_extra_flag_lo_info
[] = {
275 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
276 static const struct pandecode_flag_info shader_midgard1_flag_lo_info
[] = {
279 FLAG_INFO(READS_TILEBUFFER
),
280 FLAG_INFO(WRITES_GLOBAL
),
285 static const struct pandecode_flag_info shader_midgard1_flag_hi_info
[] = {
287 FLAG_INFO(SUPPRESS_INF_NAN
),
292 #define FLAG_INFO(flag) { MALI_BIFROST_##flag, "MALI_BIFROST_" #flag }
293 static const struct pandecode_flag_info shader_bifrost_info
[] = {
294 FLAG_INFO(FULL_THREAD
),
296 FLAG_INFO(FIRST_ATEST
),
302 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
303 static const struct pandecode_flag_info mfbd_flag_info
[] = {
304 FLAG_INFO(DEPTH_WRITE
),
310 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
311 static const struct pandecode_flag_info sampler_flag_info
[] = {
312 FLAG_INFO(MAG_NEAREST
),
313 FLAG_INFO(MIN_NEAREST
),
314 FLAG_INFO(MIP_LINEAR_1
),
315 FLAG_INFO(MIP_LINEAR_2
),
316 FLAG_INFO(NORM_COORDS
),
321 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
322 static const struct pandecode_flag_info sfbd_unk1_info
[] = {
329 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
330 static const struct pandecode_flag_info sfbd_unk2_info
[] = {
337 extern char *replace_fragment
;
338 extern char *replace_vertex
;
341 pandecode_job_type(enum mali_job_type type
)
343 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
347 DEFINE_CASE(WRITE_VALUE
);
348 DEFINE_CASE(CACHE_FLUSH
);
349 DEFINE_CASE(COMPUTE
);
353 DEFINE_CASE(FRAGMENT
);
355 case JOB_NOT_STARTED
:
356 return "NOT_STARTED";
359 pandecode_log("Warning! Unknown job type %x\n", type
);
367 pandecode_draw_mode(enum mali_draw_mode mode
)
369 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
372 DEFINE_CASE(DRAW_NONE
);
375 DEFINE_CASE(TRIANGLES
);
376 DEFINE_CASE(TRIANGLE_STRIP
);
377 DEFINE_CASE(TRIANGLE_FAN
);
378 DEFINE_CASE(LINE_STRIP
);
379 DEFINE_CASE(LINE_LOOP
);
380 DEFINE_CASE(POLYGON
);
382 DEFINE_CASE(QUAD_STRIP
);
385 pandecode_msg("XXX: invalid draw mode %X\n", mode
);
392 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
394 pandecode_func(enum mali_func mode
)
401 DEFINE_CASE(GREATER
);
402 DEFINE_CASE(NOTEQUAL
);
407 pandecode_msg("XXX: invalid func %X\n", mode
);
413 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
415 pandecode_stencil_op(enum mali_stencil_op op
)
419 DEFINE_CASE(REPLACE
);
422 DEFINE_CASE(INCR_WRAP
);
423 DEFINE_CASE(DECR_WRAP
);
428 pandecode_msg("XXX: invalid stencil op %X\n", op
);
435 static char *pandecode_attr_mode_short(enum mali_attr_mode mode
)
438 /* TODO: Combine to just "instanced" once this can be done
439 * unambiguously in all known cases */
440 case MALI_ATTR_POT_DIVIDE
:
441 return "instanced_pot";
442 case MALI_ATTR_MODULO
:
443 return "instanced_mod";
444 case MALI_ATTR_NPOT_DIVIDE
:
445 return "instanced_npot";
446 case MALI_ATTR_IMAGE
:
449 pandecode_msg("XXX: invalid attribute mode %X\n", mode
);
455 pandecode_special_record(uint64_t v
, bool* attribute
)
458 case MALI_ATTR_VERTEXID
:
460 return "gl_VertexID";
461 case MALI_ATTR_INSTANCEID
:
463 return "gl_InstanceID";
464 case MALI_VARYING_FRAG_COORD
:
465 return "gl_FragCoord";
466 case MALI_VARYING_FRONT_FACING
:
467 return "gl_FrontFacing";
468 case MALI_VARYING_POINT_COORD
:
469 return "gl_PointCoord";
471 pandecode_msg("XXX: invalid special record %" PRIx64
"\n", v
);
476 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
478 pandecode_wrap_mode(enum mali_wrap_mode op
)
482 DEFINE_CASE(CLAMP_TO_EDGE
);
484 DEFINE_CASE(CLAMP_TO_BORDER
);
485 DEFINE_CASE(MIRRORED_REPEAT
);
486 DEFINE_CASE(MIRRORED_CLAMP_TO_EDGE
);
487 DEFINE_CASE(MIRRORED_CLAMP
);
488 DEFINE_CASE(MIRRORED_CLAMP_TO_BORDER
);
491 pandecode_msg("XXX: invalid wrap mode %X\n", op
);
497 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
499 pandecode_block_format(enum mali_block_format fmt
)
503 DEFINE_CASE(UNKNOWN
);
508 unreachable("Invalid case");
513 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
515 pandecode_exception_access(unsigned access
)
519 DEFINE_CASE(EXECUTE
);
524 unreachable("Invalid case");
529 /* Midgard's tiler descriptor is embedded within the
533 pandecode_midgard_tiler_descriptor(
534 const struct midgard_tiler_descriptor
*t
,
540 pandecode_log(".tiler = {\n");
543 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
544 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
546 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
548 /* We know this name from the kernel, but we never see it nonzero */
551 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
553 MEMORY_PROP(t
, polygon_list
);
555 /* The body is offset from the base of the polygon list */
556 //assert(t->polygon_list_body > t->polygon_list);
557 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
559 /* It needs to fit inside the reported size */
560 //assert(t->polygon_list_size >= body_offset);
562 /* Now that we've sanity checked, we'll try to calculate the sizes
563 * ourselves for comparison */
565 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
566 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
568 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
569 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
570 ref_header
, ref_size
);
571 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
572 pandecode_msg("body offset %d\n", body_offset
);
575 /* The tiler heap has a start and end specified -- it should be
576 * identical to what we have in the BO. The exception is if tiling is
579 MEMORY_PROP(t
, heap_start
);
580 assert(t
->heap_end
>= t
->heap_start
);
582 struct pandecode_mapped_memory
*heap
=
583 pandecode_find_mapped_gpu_mem_containing(t
->heap_start
);
585 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
587 /* Tiling is enabled with a special flag */
588 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
589 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
591 bool tiling_enabled
= hierarchy_mask
;
593 if (tiling_enabled
) {
594 /* When tiling is enabled, the heap should be a tight fit */
595 unsigned heap_offset
= t
->heap_start
- heap
->gpu_va
;
596 if ((heap_offset
+ heap_size
) != heap
->length
) {
597 pandecode_msg("XXX: heap size %u (expected %zu)\n",
598 heap_size
, heap
->length
- heap_offset
);
601 /* We should also have no other flags */
603 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
605 /* When tiling is disabled, we should have that flag and no others */
607 if (tiler_flags
!= MALI_TILER_DISABLED
) {
608 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
612 /* We should also have an empty heap */
614 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
618 /* Disabled tiling is used only for clear-only jobs, which are
619 * purely FRAGMENT, so we should never see this for
620 * non-FRAGMENT descriptors. */
623 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
626 /* We've never seen weights used in practice, but we know from the
627 * kernel these fields is there */
629 bool nonzero_weights
= false;
631 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
632 nonzero_weights
|= t
->weights
[w
] != 0x0;
635 if (nonzero_weights
) {
636 pandecode_log(".weights = { ");
638 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
639 pandecode_log_cont("%d, ", t
->weights
[w
]);
646 pandecode_log("}\n");
649 /* TODO: The Bifrost tiler is not understood at all yet */
652 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer
*fb
)
654 pandecode_log(".tiler = {\n");
657 MEMORY_PROP(fb
, tiler_meta
);
659 for (int i
= 0; i
< 16; i
++) {
660 if (fb
->zeros
[i
] != 0) {
661 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
666 pandecode_log("},\n");
669 pandecode_log("}\n");
673 /* Information about the framebuffer passed back for
674 * additional analysis */
676 struct pandecode_fbd
{
684 pandecode_sfbd_format(struct mali_sfbd_format format
)
686 pandecode_log(".format = {\n");
689 pandecode_log(".unk1 = ");
690 pandecode_log_decoded_flags(sfbd_unk1_info
, format
.unk1
);
691 pandecode_log_cont(",\n");
693 /* TODO: Map formats so we can check swizzles and print nicely */
694 pandecode_log("swizzle");
695 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
696 pandecode_log_cont(",\n");
698 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
699 (format
.nr_channels
+ 1));
701 pandecode_log(".unk2 = ");
702 pandecode_log_decoded_flags(sfbd_unk2_info
, format
.unk2
);
703 pandecode_log_cont(",\n");
705 pandecode_prop("block = %s", pandecode_block_format(format
.block
));
707 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
710 pandecode_log("},\n");
714 pandecode_shared_memory(const struct mali_shared_memory
*desc
, bool is_compute
)
716 pandecode_prop("stack_shift = 0x%x", desc
->stack_shift
);
719 pandecode_prop("unk0 = 0x%x", desc
->unk0
);
721 if (desc
->shared_workgroup_count
!= 0x1F) {
722 pandecode_prop("shared_workgroup_count = %d", desc
->shared_workgroup_count
);
724 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
727 if (desc
->shared_unk1
|| desc
->shared_shift
) {
728 pandecode_prop("shared_unk1 = %X", desc
->shared_unk1
);
729 pandecode_prop("shared_shift = %X", desc
->shared_shift
);
732 pandecode_msg("XXX: shared memory configured in noncompute shader");
735 if (desc
->shared_zero
) {
736 pandecode_msg("XXX: shared memory zero tripped\n");
737 pandecode_prop("shared_zero = 0x%" PRIx32
, desc
->shared_zero
);
740 if (desc
->shared_memory
&& !is_compute
)
741 pandecode_msg("XXX: shared memory used in noncompute shader\n");
743 MEMORY_PROP(desc
, scratchpad
);
744 MEMORY_PROP(desc
, shared_memory
);
745 MEMORY_PROP(desc
, unknown1
);
747 if (desc
->scratchpad
) {
748 struct pandecode_mapped_memory
*smem
=
749 pandecode_find_mapped_gpu_mem_containing(desc
->scratchpad
);
751 pandecode_msg("scratchpad size %u\n", smem
->length
);
756 static struct pandecode_fbd
757 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
, unsigned gpu_id
)
759 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
760 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
762 struct pandecode_fbd info
= {
767 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
770 pandecode_log(".shared_memory = {\n");
772 pandecode_shared_memory(&s
->shared_memory
, false);
774 pandecode_log("},\n");
776 pandecode_sfbd_format(s
->format
);
778 info
.width
= s
->width
+ 1;
779 info
.height
= s
->height
+ 1;
781 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", info
.width
);
782 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", info
.height
);
784 MEMORY_PROP(s
, checksum
);
786 if (s
->checksum_stride
)
787 pandecode_prop("checksum_stride = %d", s
->checksum_stride
);
789 MEMORY_PROP(s
, framebuffer
);
790 pandecode_prop("stride = %d", s
->stride
);
792 /* Earlier in the actual commandstream -- right before width -- but we
793 * delay to flow nicer */
795 pandecode_log(".clear_flags = ");
796 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
797 pandecode_log_cont(",\n");
799 if (s
->depth_buffer
) {
800 MEMORY_PROP(s
, depth_buffer
);
801 pandecode_prop("depth_stride = %d", s
->depth_stride
);
804 if (s
->stencil_buffer
) {
805 MEMORY_PROP(s
, stencil_buffer
);
806 pandecode_prop("stencil_stride = %d", s
->stencil_stride
);
809 if (s
->depth_stride_zero
||
810 s
->stencil_stride_zero
||
811 s
->zero7
|| s
->zero8
) {
812 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
813 pandecode_prop("depth_stride_zero = 0x%x",
814 s
->depth_stride_zero
);
815 pandecode_prop("stencil_stride_zero = 0x%x",
816 s
->stencil_stride_zero
);
817 pandecode_prop("zero7 = 0x%" PRIx32
,
819 pandecode_prop("zero8 = 0x%" PRIx32
,
823 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
824 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
825 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
826 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
827 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
830 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
831 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
832 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
833 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
834 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
837 if (s
->clear_stencil
) {
838 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
841 const struct midgard_tiler_descriptor t
= s
->tiler
;
843 bool has_hierarchy
= !(gpu_id
== 0x0720 || gpu_id
== 0x0820 || gpu_id
== 0x0830);
844 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
, has_hierarchy
);
847 pandecode_log("};\n");
849 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
850 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
851 pandecode_prop("zero5 = 0x%" PRIx32
, s
->zero5
);
853 pandecode_log_cont(".zero3 = {");
855 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
856 pandecode_log_cont("%X, ", s
->zero3
[i
]);
858 pandecode_log_cont("},\n");
860 pandecode_log_cont(".zero6 = {");
862 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
863 pandecode_log_cont("%X, ", s
->zero6
[i
]);
865 pandecode_log_cont("},\n");
871 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
873 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
874 const struct mali_shared_memory
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
876 pandecode_log("struct mali_shared_memory shared_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
878 pandecode_shared_memory(s
, true);
880 pandecode_log("},\n");
883 /* Extracts the number of components associated with a Mali format */
886 pandecode_format_component_count(enum mali_format fmt
)
888 /* Mask out the format class */
889 unsigned top
= fmt
& 0b11100000;
892 case MALI_FORMAT_SNORM
:
893 case MALI_FORMAT_UINT
:
894 case MALI_FORMAT_UNORM
:
895 case MALI_FORMAT_SINT
:
896 return ((fmt
>> 3) & 3) + 1;
903 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
906 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
909 assert(MALI_CHANNEL_RED
== 0);
911 for (unsigned c
= 0; c
< 4; ++c
) {
912 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
914 if (chan
<= MALI_CHANNEL_ALPHA
)
921 /* Validates that a (format, swizzle) pair is valid, in the sense that the
922 * swizzle doesn't access any components that are undefined in the format.
923 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
927 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
929 unsigned nr_comp
= pandecode_format_component_count(fmt
);
930 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
931 unsigned valid_mask
= (1 << nr_comp
) - 1;
932 unsigned invalid_mask
= ~valid_mask
;
934 if (access_mask
& invalid_mask
) {
935 pandecode_msg("XXX: invalid components accessed\n");
939 /* Check for the default non-swizzling swizzle so we can suppress
940 * useless printing for the defaults */
942 unsigned default_swizzles
[4] = {
943 MALI_CHANNEL_RED
| (MALI_CHANNEL_ZERO
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
944 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
945 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ONE
<< 9),
946 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ALPHA
<< 9)
949 return (swizzle
== default_swizzles
[nr_comp
- 1]);
952 /* Maps MALI_RGBA32F to rgba32f, etc */
955 pandecode_format_short(enum mali_format fmt
, bool srgb
)
957 /* We want a type-like format, so cut off the initial MALI_ */
958 char *format
= pandecode_format(fmt
);
959 format
+= strlen("MALI_");
961 unsigned len
= strlen(format
);
962 char *lower_format
= calloc(1, len
+ 1);
964 for (unsigned i
= 0; i
< len
; ++i
)
965 lower_format
[i
] = tolower(format
[i
]);
967 /* Sanity check sRGB flag is applied to RGB, per the name */
968 if (srgb
&& lower_format
[0] != 'r')
969 pandecode_msg("XXX: sRGB applied to non-colour format\n");
971 /* Just prefix with an s, so you get formats like srgba8_unorm */
973 pandecode_log_cont("s");
975 pandecode_log_cont("%s", lower_format
);
980 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
982 /* First, do some validation */
983 bool trivial_swizzle
= pandecode_validate_format_swizzle(
989 /* Next, print the swizzle */
990 pandecode_log_cont(".");
992 static const char components
[] = "rgba01";
994 for (unsigned c
= 0; c
< 4; ++c
) {
995 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
997 if (chan
>= MALI_CHANNEL_RESERVED_0
) {
998 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
1001 pandecode_log_cont("%c", components
[chan
]);
1006 pandecode_rt_format(struct mali_rt_format format
)
1008 pandecode_log(".format = {\n");
1011 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
1012 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
1013 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
1014 pandecode_prop("unk4 = 0x%" PRIx32
, format
.unk4
);
1016 pandecode_prop("block = %s", pandecode_block_format(format
.block
));
1018 /* TODO: Map formats so we can check swizzles and print nicely */
1019 pandecode_log("swizzle");
1020 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
1021 pandecode_log_cont(",\n");
1023 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
1024 (format
.nr_channels
+ 1));
1026 pandecode_log(".flags = ");
1027 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
1028 pandecode_log_cont(",\n");
1030 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
1031 * which is a faster hardware-based alternative to the wallpaper method
1032 * to preserve framebuffer contents across frames. In practice, MFBD
1033 * preload is buggy on Midgard, and so this is a chicken bit. If this
1034 * bit isn't set, most likely something broke unrelated to preload */
1036 if (!format
.no_preload
) {
1037 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
1038 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
1042 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
1045 pandecode_log("},\n");
1049 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct mali_framebuffer
*fb
)
1051 pandecode_log("struct mali_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
1054 for (int i
= 0; i
< (fb
->rt_count_1
+ 1); i
++) {
1055 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct mali_render_target
);
1056 struct pandecode_mapped_memory
*mem
=
1057 pandecode_find_mapped_gpu_mem_containing(rt_va
);
1058 const struct mali_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
1060 pandecode_log("{\n");
1063 pandecode_rt_format(rt
->format
);
1065 if (rt
->format
.block
== MALI_BLOCK_AFBC
) {
1066 pandecode_log(".afbc = {\n");
1069 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
1070 pandecode_prop("metadata = %s", a
);
1073 pandecode_prop("stride = %d", rt
->afbc
.stride
);
1075 pandecode_log(".flags = ");
1076 pandecode_log_decoded_flags(afbc_fmt_flag_info
, rt
->afbc
.flags
);
1077 pandecode_log_cont(",\n");
1080 pandecode_log("},\n");
1081 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.flags
) {
1082 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1088 MEMORY_PROP(rt
, framebuffer
);
1089 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
1091 if (rt
->layer_stride
)
1092 pandecode_prop("layer_stride = %d", rt
->layer_stride
);
1094 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
1095 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
1096 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
1097 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
1098 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
1101 if (rt
->zero1
|| rt
->zero2
) {
1102 pandecode_msg("XXX: render target zeros tripped\n");
1103 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
1104 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
1108 pandecode_log("},\n");
1112 pandecode_log("};\n");
1115 static struct pandecode_fbd
1116 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
, bool is_compute
, bool is_bifrost
)
1118 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1119 const struct mali_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
1121 struct pandecode_fbd info
;
1123 if (is_bifrost
&& fb
->msaa
.sample_locations
) {
1124 /* The blob stores all possible sample locations in a single buffer
1125 * allocated on startup, and just switches the pointer when switching
1126 * MSAA state. For now, we just put the data into the cmdstream, but we
1127 * should do something like what the blob does with a real driver.
1129 * There seem to be 32 slots for sample locations, followed by another
1130 * 16. The second 16 is just the center location followed by 15 zeros
1131 * in all the cases I've identified (maybe shader vs. depth/color
1135 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->msaa
.sample_locations
);
1137 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->msaa
.sample_locations
);
1139 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
1142 for (int i
= 0; i
< 32 + 16; i
++) {
1143 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
1147 pandecode_log("};\n");
1150 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1154 pandecode_log(".msaa = {\n");
1157 if (fb
->msaa
.sample_locations
)
1158 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
1160 pandecode_msg("XXX: sample_locations missing\n");
1162 if (fb
->msaa
.zero1
|| fb
->msaa
.zero2
|| fb
->msaa
.zero4
) {
1163 pandecode_msg("XXX: multisampling zero tripped\n");
1164 pandecode_prop("zero1 = %" PRIx64
, fb
->msaa
.zero1
);
1165 pandecode_prop("zero2 = %" PRIx64
, fb
->msaa
.zero2
);
1166 pandecode_prop("zero4 = %" PRIx64
, fb
->msaa
.zero4
);
1170 pandecode_log("},\n");
1172 pandecode_log(".shared_memory = {\n");
1174 pandecode_shared_memory(&fb
->shared_memory
, is_compute
);
1176 pandecode_log("},\n");
1179 info
.width
= fb
->width1
+ 1;
1180 info
.height
= fb
->height1
+ 1;
1181 info
.rt_count
= fb
->rt_count_1
+ 1;
1183 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
1184 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
1185 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
1186 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
1188 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
1189 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
1190 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
1191 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
1193 pandecode_log(".mfbd_flags = ");
1194 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
1195 pandecode_log_cont(",\n");
1197 if (fb
->clear_stencil
)
1198 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
1200 if (fb
->clear_depth
)
1201 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
1205 pandecode_bifrost_tiler_descriptor(fb
);
1207 const struct midgard_tiler_descriptor t
= fb
->tiler
;
1208 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
, true);
1211 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1213 if (fb
->zero3
|| fb
->zero4
) {
1214 pandecode_msg("XXX: framebuffer zeros tripped\n");
1215 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
1216 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
1220 pandecode_log("};\n");
1222 gpu_va
+= sizeof(struct mali_framebuffer
);
1224 info
.has_extra
= (fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
;
1226 if (info
.has_extra
) {
1227 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1228 const struct mali_framebuffer_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
1230 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1233 MEMORY_PROP(fbx
, checksum
);
1235 if (fbx
->checksum_stride
)
1236 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
1238 pandecode_log(".flags_hi = ");
1239 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info
, fbx
->flags_hi
);
1240 pandecode_log_cont(",\n");
1242 pandecode_log(".flags_lo = ");
1243 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info
, fbx
->flags_lo
);
1244 pandecode_log_cont(",\n");
1246 pandecode_prop("zs_block = %s", pandecode_block_format(fbx
->zs_block
));
1248 if (fbx
->zs_block
== MALI_BLOCK_AFBC
) {
1249 pandecode_log(".ds_afbc = {\n");
1252 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
1253 pandecode_prop("depth_stencil_afbc_stride = %d",
1254 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
1255 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
1257 pandecode_log(".flags = ");
1258 pandecode_log_decoded_flags(afbc_fmt_flag_info
, fbx
->ds_afbc
.flags
);
1259 pandecode_log_cont(",\n");
1261 if (fbx
->ds_afbc
.padding
) {
1262 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1263 pandecode_prop("padding = 0x%" PRIx64
, fbx
->ds_afbc
.padding
);
1267 pandecode_log("},\n");
1269 pandecode_log(".ds_linear = {\n");
1272 if (fbx
->ds_linear
.depth
) {
1273 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
1274 pandecode_prop("depth_stride = %d",
1275 fbx
->ds_linear
.depth_stride
);
1276 } else if (fbx
->ds_linear
.depth_stride
) {
1277 pandecode_msg("XXX: depth stride zero tripped %d\n", fbx
->ds_linear
.depth_stride
);
1280 if (fbx
->ds_linear
.stencil
) {
1281 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1282 pandecode_prop("stencil_stride = %d",
1283 fbx
->ds_linear
.stencil_stride
);
1284 } else if (fbx
->ds_linear
.stencil_stride
) {
1285 pandecode_msg("XXX: stencil stride zero tripped %d\n", fbx
->ds_linear
.stencil_stride
);
1288 if (fbx
->ds_linear
.depth_stride_zero
||
1289 fbx
->ds_linear
.stencil_stride_zero
||
1290 fbx
->ds_linear
.zero1
|| fbx
->ds_linear
.zero2
) {
1291 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1292 pandecode_prop("depth_stride_zero = 0x%x",
1293 fbx
->ds_linear
.depth_stride_zero
);
1294 pandecode_prop("stencil_stride_zero = 0x%x",
1295 fbx
->ds_linear
.stencil_stride_zero
);
1296 pandecode_prop("zero1 = 0x%" PRIx32
,
1297 fbx
->ds_linear
.zero1
);
1298 pandecode_prop("zero2 = 0x%" PRIx32
,
1299 fbx
->ds_linear
.zero2
);
1303 pandecode_log("},\n");
1306 if (fbx
->clear_color_1
| fbx
->clear_color_2
) {
1307 pandecode_prop("clear_color_1 = 0x%" PRIx32
, fbx
->clear_color_1
);
1308 pandecode_prop("clear_color_2 = 0x%" PRIx32
, fbx
->clear_color_2
);
1312 pandecode_msg("XXX: fb_extra zeros tripped\n");
1313 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1317 pandecode_log("};\n");
1319 gpu_va
+= sizeof(struct mali_framebuffer_extra
);
1323 pandecode_render_target(gpu_va
, job_no
, fb
);
1328 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1332 pandecode_padded_vertices(unsigned shift
, unsigned k
)
1334 unsigned odd
= 2*k
+ 1;
1335 unsigned pot
= 1 << shift
;
1336 pandecode_msg("padded_num_vertices = %d\n", odd
* pot
);
1339 /* Given a magic divisor, recover what we were trying to divide by.
1341 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1342 * 0 <= m < 2^N, for N bits in m.
1344 * Let q represent the number we would like to divide by.
1346 * By definition of a magic divisor for N-bit unsigned integers (a number you
1347 * multiply by to magically get division), m is a number such that:
1349 * (m * x) & (2^N - 1) = floor(x/q).
1350 * for all x on Z where 0 <= x < 2^N
1352 * Ignore the case where any of the above values equals zero; it is irrelevant
1353 * for our purposes (instanced arrays).
1355 * Choose x = q. Then:
1357 * (m * x) & (2^N - 1) = floor(x/q).
1358 * (m * q) & (2^N - 1) = floor(q/q).
1360 * floor(q/q) = floor(1) = 1, therefore:
1362 * (m * q) & (2^N - 1) = 1
1364 * Recall the identity that the bitwise AND of one less than a power-of-two
1365 * equals the modulo with that power of two, i.e. for all x:
1367 * x & (2^N - 1) = x % N
1373 * By definition, a modular multiplicative inverse of a number m is the number
1374 * q such that with respect to a modulos M:
1378 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1383 pandecode_magic_divisor(uint32_t magic
, unsigned shift
, unsigned orig_divisor
, unsigned extra
)
1386 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1387 * shift) the most lame way possible... just repeatedly add.
1388 * Asymptoptically slow but nobody cares in practice, unless you have
1389 * massive numbers of vertices or high divisors. */
1391 unsigned inverse
= 0;
1393 /* Magic implicitly has the highest bit set */
1396 /* Depending on rounding direction */
1401 uint32_t product
= magic
* inverse
;
1413 pandecode_msg("dividing by %d (maybe off by two)\n", inverse
);
1415 /* Recall we're supposed to divide by (gl_level_divisor *
1416 * padded_num_vertices) */
1418 unsigned padded_num_vertices
= inverse
/ orig_divisor
;
1420 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices
);
1425 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1426 mali_ptr addr
, int job_no
, char *suffix
,
1427 int count
, bool varying
, enum mali_job_type job_type
)
1429 char *prefix
= varying
? "varying" : "attribute";
1433 pandecode_msg("warn: No %s records\n", prefix
);
1437 union mali_attr
*attr
= pandecode_fetch_gpu_mem(mem
, addr
, sizeof(union mali_attr
) * count
);
1439 for (int i
= 0; i
< count
; ++i
) {
1440 /* First, check for special records */
1441 if (attr
[i
].elements
< MALI_RECORD_SPECIAL
) {
1443 pandecode_msg("XXX: tripped size=%d\n", attr
[i
].size
);
1445 if (attr
[i
].stride
) {
1446 /* gl_InstanceID passes a magic divisor in the
1447 * stride field to divide by the padded vertex
1448 * count. No other records should do so, so
1449 * stride should otherwise be zero. Note that
1450 * stride in the usual attribute sense doesn't
1451 * apply to special records. */
1453 bool has_divisor
= attr
[i
].elements
== MALI_ATTR_INSTANCEID
;
1455 pandecode_log_cont("/* %smagic divisor = %X */ ",
1456 has_divisor
? "" : "XXX: ", attr
[i
].stride
);
1459 if (attr
[i
].shift
|| attr
[i
].extra_flags
) {
1460 /* Attributes use these fields for
1461 * instancing/padding/etc type issues, but
1464 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1465 varying
? "XXX: " : "",
1466 attr
[i
].shift
, attr
[i
].extra_flags
);
1469 /* Print the special record name */
1470 bool attribute
= false;
1471 pandecode_log("%s_%d = %s;\n", prefix
, i
, pandecode_special_record(attr
[i
].elements
, &attribute
));
1474 if (attribute
== varying
)
1475 pandecode_msg("XXX: mismatched special record\n");
1480 enum mali_attr_mode mode
= attr
[i
].elements
& 7;
1482 if (mode
== MALI_ATTR_UNUSED
)
1483 pandecode_msg("XXX: unused attribute record\n");
1485 /* For non-linear records, we need to print the type of record */
1486 if (mode
!= MALI_ATTR_LINEAR
)
1487 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode
));
1489 /* Print the name to link with attr_meta */
1490 pandecode_log_cont("%s_%d", prefix
, i
);
1492 /* Print the stride and size */
1493 pandecode_log_cont("<%u>[%u]", attr
[i
].stride
, attr
[i
].size
);
1495 /* TODO: Sanity check the quotient itself. It must be equal to
1496 * (or be greater than, if the driver added padding) the padded
1499 /* Finally, print the pointer */
1500 mali_ptr raw_elements
= attr
[i
].elements
& ~7;
1501 char *a
= pointer_as_memory_reference(raw_elements
);
1502 pandecode_log_cont(" = (%s);\n", a
);
1505 /* Check the pointer */
1506 pandecode_validate_buffer(raw_elements
, attr
[i
].size
);
1508 /* shift/extra_flags exist only for instanced */
1509 if (attr
[i
].shift
| attr
[i
].extra_flags
) {
1510 /* These are set to random values by the blob for
1511 * varyings, most likely a symptom of uninitialized
1512 * memory where the hardware masked the bug. As such we
1513 * put this at a warning, not an error. */
1515 if (mode
== MALI_ATTR_LINEAR
)
1516 pandecode_msg("warn: instancing fields set for linear\n");
1518 pandecode_prop("shift = %d", attr
[i
].shift
);
1519 pandecode_prop("extra_flags = %d", attr
[i
].extra_flags
);
1522 /* Decode further where possible */
1524 if (mode
== MALI_ATTR_MODULO
) {
1525 pandecode_padded_vertices(
1527 attr
[i
].extra_flags
);
1530 if (mode
== MALI_ATTR_NPOT_DIVIDE
) {
1532 pandecode_log("{\n");
1534 pandecode_prop("unk = 0x%x", attr
[i
].unk
);
1535 pandecode_prop("magic_divisor = 0x%08x", attr
[i
].magic_divisor
);
1536 if (attr
[i
].zero
!= 0)
1537 pandecode_prop("XXX: zero tripped (0x%x)\n", attr
[i
].zero
);
1538 pandecode_prop("divisor = %d", attr
[i
].divisor
);
1539 pandecode_magic_divisor(attr
[i
].magic_divisor
, attr
[i
- 1].shift
, attr
[i
].divisor
, attr
[i
- 1].extra_flags
);
1541 pandecode_log("}, \n");
1546 pandecode_log("\n");
1550 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1552 /* TODO: Decode flags */
1553 mali_ptr shader_ptr
= ptr
& ~15;
1555 char *a
= pointer_as_memory_reference(shader_ptr
);
1556 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1563 pandecode_stencil(const char *name
, const struct mali_stencil_test
*stencil
)
1565 unsigned any_nonzero
=
1566 stencil
->ref
| stencil
->mask
| stencil
->func
|
1567 stencil
->sfail
| stencil
->dpfail
| stencil
->dppass
;
1569 if (any_nonzero
== 0)
1572 const char *func
= pandecode_func(stencil
->func
);
1573 const char *sfail
= pandecode_stencil_op(stencil
->sfail
);
1574 const char *dpfail
= pandecode_stencil_op(stencil
->dpfail
);
1575 const char *dppass
= pandecode_stencil_op(stencil
->dppass
);
1578 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil
->zero
);
1580 pandecode_log(".stencil_%s = {\n", name
);
1582 pandecode_prop("ref = %d", stencil
->ref
);
1583 pandecode_prop("mask = 0x%02X", stencil
->mask
);
1584 pandecode_prop("func = %s", func
);
1585 pandecode_prop("sfail = %s", sfail
);
1586 pandecode_prop("dpfail = %s", dpfail
);
1587 pandecode_prop("dppass = %s", dppass
);
1589 pandecode_log("},\n");
1593 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1596 pandecode_msg("XXX: blend zero tripped: %X\n", blend
->zero1
);
1598 pandecode_log(".equation = {\n");
1601 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1602 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1604 pandecode_log(".color_mask = ");
1605 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1606 pandecode_log_cont(",\n");
1609 pandecode_log("},\n");
1612 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1615 decode_bifrost_constant(u16 constant
)
1617 float lo
= (float) (constant
& 0xFF);
1618 float hi
= (float) (constant
>> 8);
1620 return (hi
/ 255.0) + (lo
/ 65535.0);
1624 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1626 struct bifrost_blend_rt
*b
=
1627 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1629 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1632 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1633 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1634 b
->constant
, decode_bifrost_constant(b
->constant
));
1636 /* TODO figure out blend shader enable bit */
1637 pandecode_blend_equation(&b
->equation
);
1639 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1640 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1642 pandecode_log(".format = ");
1643 pandecode_format_short(b
->format
, false);
1644 pandecode_swizzle(b
->swizzle
, b
->format
);
1645 pandecode_log_cont(",\n");
1647 pandecode_prop("swizzle = 0x%" PRIx32
, b
->swizzle
);
1648 pandecode_prop("format = 0x%" PRIx32
, b
->format
);
1651 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1652 pandecode_prop("zero1 = 0x%" PRIx32
, b
->zero1
);
1655 pandecode_log(".shader_type = ");
1656 switch(b
->shader_type
) {
1657 case BIFROST_BLEND_F16
:
1658 pandecode_log_cont("BIFROST_BLEND_F16");
1660 case BIFROST_BLEND_F32
:
1661 pandecode_log_cont("BIFROST_BLEND_F32");
1663 case BIFROST_BLEND_I32
:
1664 pandecode_log_cont("BIFROST_BLEND_I32");
1666 case BIFROST_BLEND_U32
:
1667 pandecode_log_cont("BIFROST_BLEND_U32");
1669 case BIFROST_BLEND_I16
:
1670 pandecode_log_cont("BIFROST_BLEND_I16");
1672 case BIFROST_BLEND_U16
:
1673 pandecode_log_cont("BIFROST_BLEND_U16");
1676 pandecode_log_cont(",\n");
1679 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1680 pandecode_prop("zero2 = 0x%" PRIx32
, b
->zero2
);
1683 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1686 pandecode_log("},\n");
1692 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1694 /* constant/equation is in a union */
1698 pandecode_log(".blend = {\n");
1702 pandecode_shader_address("shader", blend
->shader
);
1704 pandecode_blend_equation(&blend
->equation
);
1705 pandecode_prop("constant = %f", blend
->constant
);
1709 pandecode_log("},\n");
1711 /* Return blend shader to disassemble if present */
1712 return is_shader
? (blend
->shader
& ~0xF) : 0;
1716 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1718 struct midgard_blend_rt
*b
=
1719 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1721 /* Flags determine presence of blend shader */
1722 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1724 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1727 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1729 union midgard_blend blend
= b
->blend
;
1730 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1733 pandecode_log("};\n");
1738 /* Attributes and varyings have descriptor records, which contain information
1739 * about their format and ordering with the attribute/varying buffers. We'll
1740 * want to validate that the combinations specified are self-consistent.
1744 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1747 char *prefix
= varying
? "varying" : "attribute";
1748 unsigned max_index
= 0;
1749 snprintf(base
, sizeof(base
), "%s_meta", prefix
);
1751 struct mali_attr_meta
*attr_meta
;
1752 mali_ptr p
= varying
? v
->varying_meta
: v
->attribute_meta
;
1754 struct pandecode_mapped_memory
*attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
);
1756 for (int i
= 0; i
< count
; ++i
, p
+= sizeof(struct mali_attr_meta
)) {
1757 attr_meta
= pandecode_fetch_gpu_mem(attr_mem
, p
,
1760 /* If the record is discard, it should be zero for everything else */
1762 if (attr_meta
->format
== MALI_VARYING_DISCARD
) {
1765 attr_meta
->unknown1
|
1766 attr_meta
->unknown3
|
1767 attr_meta
->src_offset
;
1770 pandecode_msg("XXX: expected empty record for varying discard\n");
1772 /* We want to look for a literal 0000 swizzle -- this
1773 * is not encoded with all zeroes, however */
1775 enum mali_channel z
= MALI_CHANNEL_ZERO
;
1776 unsigned zero_swizzle
= z
| (z
<< 3) | (z
<< 6) | (z
<< 9);
1777 bool good_swizzle
= attr_meta
->swizzle
== zero_swizzle
;
1780 pandecode_msg("XXX: expected zero swizzle for discard\n");
1783 pandecode_msg("XXX: cannot discard attribute\n");
1785 /* If we're all good, omit the record */
1786 if (!zero
&& varying
&& good_swizzle
) {
1787 pandecode_log("/* discarded varying */\n");
1792 if (attr_meta
->index
> max_index
)
1793 max_index
= attr_meta
->index
;
1795 if (attr_meta
->unknown1
!= 0x2) {
1796 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1797 pandecode_prop("unknown1 = 0x%" PRIx64
, (u64
) attr_meta
->unknown1
);
1800 if (attr_meta
->unknown3
) {
1801 pandecode_msg("XXX: unexpected unknown3 set\n");
1802 pandecode_prop("unknown3 = 0x%" PRIx64
, (u64
) attr_meta
->unknown3
);
1805 pandecode_format_short(attr_meta
->format
, false);
1806 pandecode_log_cont(" %s_%u", prefix
, attr_meta
->index
);
1808 if (attr_meta
->src_offset
)
1809 pandecode_log_cont("[%u]", attr_meta
->src_offset
);
1811 pandecode_swizzle(attr_meta
->swizzle
, attr_meta
->format
);
1813 pandecode_log_cont(";\n");
1816 pandecode_log("\n");
1818 return count
? (max_index
+ 1) : 0;
1821 /* return bits [lo, hi) of word */
1823 bits(u32 word
, u32 lo
, u32 hi
)
1826 return word
; // avoid undefined behavior with the shift
1828 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1832 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool graphics
)
1834 pandecode_log(".prefix = {\n");
1837 /* Decode invocation_count. See the comment before the definition of
1838 * invocation_count for an explanation.
1841 unsigned size_y_shift
= bits(p
->invocation_shifts
, 0, 5);
1842 unsigned size_z_shift
= bits(p
->invocation_shifts
, 5, 10);
1843 unsigned workgroups_x_shift
= bits(p
->invocation_shifts
, 10, 16);
1844 unsigned workgroups_y_shift
= bits(p
->invocation_shifts
, 16, 22);
1845 unsigned workgroups_z_shift
= bits(p
->invocation_shifts
, 22, 28);
1846 unsigned workgroups_x_shift_2
= bits(p
->invocation_shifts
, 28, 32);
1848 unsigned size_x
= bits(p
->invocation_count
, 0, size_y_shift
) + 1;
1849 unsigned size_y
= bits(p
->invocation_count
, size_y_shift
, size_z_shift
) + 1;
1850 unsigned size_z
= bits(p
->invocation_count
, size_z_shift
, workgroups_x_shift
) + 1;
1852 unsigned groups_x
= bits(p
->invocation_count
, workgroups_x_shift
, workgroups_y_shift
) + 1;
1853 unsigned groups_y
= bits(p
->invocation_count
, workgroups_y_shift
, workgroups_z_shift
) + 1;
1854 unsigned groups_z
= bits(p
->invocation_count
, workgroups_z_shift
, 32) + 1;
1856 /* Even though we have this decoded, we want to ensure that the
1857 * representation is "unique" so we don't lose anything by printing only
1858 * the final result. More specifically, we need to check that we were
1859 * passed something in canonical form, since the definition per the
1860 * hardware is inherently not unique. How? Well, take the resulting
1861 * decode and pack it ourselves! If it is bit exact with what we
1862 * decoded, we're good to go. */
1864 struct mali_vertex_tiler_prefix ref
;
1865 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, graphics
);
1868 (p
->invocation_count
== ref
.invocation_count
) &&
1869 (p
->invocation_shifts
== ref
.invocation_shifts
);
1872 pandecode_msg("XXX: non-canonical workgroups packing\n");
1873 pandecode_msg("expected: %X, %X",
1874 ref
.invocation_count
,
1875 ref
.invocation_shifts
);
1877 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1878 pandecode_prop("size_y_shift = %d", size_y_shift
);
1879 pandecode_prop("size_z_shift = %d", size_z_shift
);
1880 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift
);
1881 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift
);
1882 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift
);
1883 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2
);
1886 /* Regardless, print the decode */
1887 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1888 size_x
, size_y
, size_z
,
1889 groups_x
, groups_y
, groups_z
);
1892 if (p
->unknown_draw
)
1893 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1895 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1897 if (p
->draw_mode
!= MALI_DRAW_NONE
)
1898 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p
->draw_mode
));
1900 /* Index count only exists for tiler jobs anyway */
1903 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1906 unsigned index_raw_size
= (p
->unknown_draw
& MALI_DRAW_INDEXED_SIZE
);
1907 index_raw_size
>>= MALI_DRAW_INDEXED_SHIFT
;
1909 /* Validate an index buffer is present if we need one. TODO: verify
1910 * relationship between invocation_count and index_count */
1913 unsigned count
= p
->index_count
;
1916 unsigned size
= (index_raw_size
== 0x3) ? 4 : index_raw_size
;
1918 /* Ensure we got a size, and if so, validate the index buffer
1919 * is large enough to hold a full set of indices of the given
1922 if (!index_raw_size
)
1923 pandecode_msg("XXX: index size missing\n");
1925 pandecode_validate_buffer(p
->indices
, count
* size
);
1926 } else if (index_raw_size
)
1927 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size
);
1929 if (p
->offset_bias_correction
)
1930 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1932 /* TODO: Figure out what this is. It's not zero */
1933 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1936 pandecode_log("},\n");
1940 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1942 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1943 uint64_t *PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1945 for (int i
= 0; i
< ubufs_count
; i
++) {
1946 unsigned size
= (ubufs
[i
] & ((1 << 10) - 1)) * 16;
1947 mali_ptr addr
= (ubufs
[i
] >> 10) << 2;
1949 pandecode_validate_buffer(addr
, size
);
1951 char *ptr
= pointer_as_memory_reference(addr
);
1952 pandecode_log("ubuf_%d[%u] = %s;\n", i
, size
, ptr
);
1956 pandecode_log("\n");
1960 pandecode_uniforms(mali_ptr uniforms
, unsigned uniform_count
)
1962 pandecode_validate_buffer(uniforms
, uniform_count
* 16);
1964 char *ptr
= pointer_as_memory_reference(uniforms
);
1965 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count
, ptr
);
1970 shader_type_for_job(unsigned type
)
1973 case JOB_TYPE_VERTEX
: return "VERTEX";
1974 case JOB_TYPE_TILER
: return "FRAGMENT";
1975 case JOB_TYPE_COMPUTE
: return "COMPUTE";
1981 static unsigned shader_id
= 0;
1983 static struct midgard_disasm_stats
1984 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1985 bool is_bifrost
, unsigned gpu_id
)
1987 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1988 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1990 /* Compute maximum possible size */
1991 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1993 /* Print some boilerplate to clearly denote the assembly (which doesn't
1994 * obey indentation rules), and actually do the disassembly! */
1996 pandecode_log_cont("\n\n");
1998 struct midgard_disasm_stats stats
;
2001 disassemble_bifrost(pandecode_dump_stream
, code
, sz
, true);
2003 /* TODO: Extend stats to Bifrost */
2004 stats
.texture_count
= -128;
2005 stats
.sampler_count
= -128;
2006 stats
.attribute_count
= -128;
2007 stats
.varying_count
= -128;
2008 stats
.uniform_count
= -128;
2009 stats
.uniform_buffer_count
= -128;
2010 stats
.work_count
= -128;
2012 stats
.instruction_count
= 0;
2013 stats
.bundle_count
= 0;
2014 stats
.quadword_count
= 0;
2015 stats
.helper_invocations
= false;
2017 stats
= disassemble_midgard(pandecode_dump_stream
,
2019 type
== JOB_TYPE_TILER
?
2020 MESA_SHADER_FRAGMENT
: MESA_SHADER_VERTEX
);
2023 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
2024 * driver-internal purposes with the blob and interfere */
2026 bool should_shaderdb
= type
!= JOB_TYPE_COMPUTE
;
2028 if (should_shaderdb
) {
2029 unsigned nr_threads
=
2030 (stats
.work_count
<= 4) ? 4 :
2031 (stats
.work_count
<= 8) ? 2 :
2034 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
2035 "%u inst, %u bundles, %u quadwords, "
2036 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
2038 shader_type_for_job(type
),
2039 stats
.instruction_count
, stats
.bundle_count
, stats
.quadword_count
,
2040 stats
.work_count
, nr_threads
);
2048 pandecode_texture_payload(mali_ptr payload
,
2049 enum mali_texture_type type
,
2050 enum mali_texture_layout layout
,
2054 uint16_t array_size
,
2055 struct pandecode_mapped_memory
*tmem
)
2057 pandecode_log(".payload = {\n");
2060 /* A bunch of bitmap pointers follow.
2061 * We work out the correct number,
2062 * based on the mipmap/cubemap
2063 * properties, but dump extra
2064 * possibilities to futureproof */
2066 int bitmap_count
= levels
+ 1;
2068 /* Miptree for each face */
2069 if (type
== MALI_TEX_CUBE
)
2072 /* Array of layers */
2073 bitmap_count
*= (depth
+ 1);
2075 /* Array of textures */
2076 bitmap_count
*= (array_size
+ 1);
2078 /* Stride for each element */
2082 mali_ptr
*pointers_and_strides
= pandecode_fetch_gpu_mem(tmem
,
2083 payload
, sizeof(mali_ptr
) * bitmap_count
);
2084 for (int i
= 0; i
< bitmap_count
; ++i
) {
2085 /* How we dump depends if this is a stride or a pointer */
2087 if (manual_stride
&& (i
& 1)) {
2088 /* signed 32-bit snuck in as a 64-bit pointer */
2089 uint64_t stride_set
= pointers_and_strides
[i
];
2090 uint32_t clamped_stride
= stride_set
;
2091 int32_t stride
= clamped_stride
;
2092 assert(stride_set
== clamped_stride
);
2093 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
2095 char *a
= pointer_as_memory_reference(pointers_and_strides
[i
]);
2096 pandecode_log("%s, \n", a
);
2102 pandecode_log("},\n");
2106 pandecode_texture(mali_ptr u
,
2107 struct pandecode_mapped_memory
*tmem
,
2108 unsigned job_no
, unsigned tex
)
2110 struct mali_texture_descriptor
*PANDECODE_PTR_VAR(t
, tmem
, u
);
2112 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64
"_%d_%d = {\n", u
, job_no
, tex
);
2115 pandecode_prop("width = %" PRId32
, t
->width
);
2116 pandecode_prop("height = %" PRId32
, t
->height
);
2117 pandecode_prop("depth = %" PRId32
, t
->depth
);
2118 pandecode_prop("array_size = %" PRId32
, t
->array_size
);
2120 pandecode_log("\n");
2121 pandecode_prop("f.swizzle = 0x%" PRIx32
, t
->format
.swizzle
);
2122 pandecode_prop("f.format = 0x%" PRIx32
, t
->format
.format
);
2123 pandecode_prop("f.srgb = 0x%" PRIx32
, t
->format
.srgb
);
2124 pandecode_prop("f.unknown1 = 0x%" PRIx32
, t
->format
.unknown1
);
2125 pandecode_prop("f.type = %" PRId32
, t
->format
.type
);
2126 pandecode_prop("f.layout = %" PRId32
, t
->format
.layout
);
2127 pandecode_prop("f.unknown2 = 0x%" PRIx32
, t
->format
.unknown2
);
2128 pandecode_prop("f.manual_stride = %" PRId32
, t
->format
.manual_stride
);
2129 pandecode_prop("f.zero = 0x%" PRIx32
, t
->format
.zero
);
2130 pandecode_log("\n");
2132 pandecode_prop("unknown3 = 0x%" PRIx32
, t
->unknown3
);
2133 pandecode_prop("unknown3A = 0x%" PRIx32
, t
->unknown3A
);
2134 pandecode_prop("levels = %" PRId32
, t
->levels
);
2135 pandecode_prop("swizzle = 0x%" PRIx32
, t
->swizzle
);
2136 pandecode_prop("swizzle_zero = 0x%" PRIx32
, t
->swizzle_zero
);
2138 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
2139 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
2140 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
2141 pandecode_log("\n");
2143 struct mali_texture_format f
= t
->format
;
2145 /* See the definiton of enum mali_texture_type */
2147 bool is_cube
= f
.type
== MALI_TEX_CUBE
;
2148 unsigned dimension
= is_cube
? 2 : f
.type
;
2150 pandecode_make_indent();
2152 /* Print the layout. Default is linear; a modifier can denote AFBC or
2153 * u-interleaved/tiled modes */
2155 if (f
.layout
== MALI_TEXTURE_AFBC
)
2156 pandecode_log_cont("afbc");
2157 else if (f
.layout
== MALI_TEXTURE_TILED
)
2158 pandecode_log_cont("tiled");
2159 else if (f
.layout
== MALI_TEXTURE_LINEAR
)
2160 pandecode_log_cont("linear");
2162 pandecode_msg("XXX: invalid texture layout 0x%X\n", f
.layout
);
2164 pandecode_swizzle(t
->swizzle
, f
.format
);
2165 pandecode_log_cont(" ");
2167 /* Distinguish cube/2D with modifier */
2170 pandecode_log_cont("cube ");
2172 pandecode_format_short(f
.format
, f
.srgb
);
2173 pandecode_swizzle(f
.swizzle
, f
.format
);
2175 /* All four width/height/depth/array_size dimensions are present
2176 * regardless of the type of texture, but it is an error to have
2177 * non-zero dimensions for unused dimensions. Verify this. array_size
2178 * can always be set, as can width. Depth used for MSAA. */
2180 if (t
->height
&& dimension
< 2)
2181 pandecode_msg("XXX: nonzero height for <2D texture\n");
2183 /* Print only the dimensions that are actually there */
2185 pandecode_log_cont(": %d", t
->width
+ 1);
2187 if (t
->height
|| t
->depth
)
2188 pandecode_log_cont("x%u", t
->height
+ 1);
2191 pandecode_log_cont("x%u", t
->depth
+ 1);
2194 pandecode_log_cont("[%u]", t
->array_size
+ 1);
2197 pandecode_log_cont(" mip %u", t
->levels
);
2199 pandecode_log_cont("\n");
2201 if (f
.unknown1
| f
.zero
) {
2202 pandecode_msg("XXX: texture format zero tripped\n");
2203 pandecode_prop("unknown1 = %" PRId32
, f
.unknown1
);
2204 pandecode_prop("zero = %" PRId32
, f
.zero
);
2208 pandecode_msg("XXX: expected unknown texture bit set\n");
2209 pandecode_prop("unknown2 = %" PRId32
, f
.unknown2
);
2212 if (t
->swizzle_zero
) {
2213 pandecode_msg("XXX: swizzle zero tripped\n");
2214 pandecode_prop("swizzle_zero = %d", t
->swizzle_zero
);
2217 if (t
->unknown3
| t
->unknown3A
| t
->unknown5
| t
->unknown6
| t
->unknown7
) {
2218 pandecode_msg("XXX: texture zero tripped\n");
2219 pandecode_prop("unknown3 = %" PRId16
, t
->unknown3
);
2220 pandecode_prop("unknown3A = %" PRId8
, t
->unknown3A
);
2221 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
2222 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
2223 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
2226 pandecode_texture_payload(u
+ sizeof(*t
), f
.type
, f
.layout
, f
.manual_stride
, t
->levels
, t
->depth
, t
->array_size
, tmem
);
2229 pandecode_log("};\n");
2233 pandecode_bifrost_texture(
2234 const struct bifrost_texture_descriptor
*t
,
2238 pandecode_log("struct bifrost_texture_descriptor texture_descriptor_%d_%d = {\n", job_no
, tex
);
2241 pandecode_prop("format_unk = 0x%" PRIx32
, t
->format_unk
);
2242 pandecode_prop("type = %" PRId32
, t
->type
);
2245 pandecode_msg("XXX: zero tripped\n");
2246 pandecode_prop("zero = 0x%" PRIx32
, t
->zero
);
2249 pandecode_prop("format_swizzle = 0x%" PRIx32
, t
->format_swizzle
);
2250 pandecode_prop("format = 0x%" PRIx32
, t
->format
);
2251 pandecode_prop("srgb = 0x%" PRIx32
, t
->srgb
);
2252 pandecode_prop("format_unk3 = 0x%" PRIx32
, t
->format_unk3
);
2253 pandecode_prop("width = %" PRId32
, t
->width
);
2254 pandecode_prop("height = %" PRId32
, t
->height
);
2255 pandecode_prop("swizzle = 0x%" PRIx32
, t
->swizzle
);
2256 pandecode_prop("levels = %" PRId32
, t
->levels
);
2257 pandecode_prop("unk1 = 0x%" PRIx32
, t
->unk1
);
2258 pandecode_prop("levels_unk = %" PRId32
, t
->levels_unk
);
2259 pandecode_prop("level_2 = %" PRId32
, t
->level_2
);
2260 pandecode_prop("payload = 0x%" PRIx64
, t
->payload
);
2261 pandecode_prop("array_size = %" PRId32
, t
->array_size
);
2262 pandecode_prop("unk4 = 0x%" PRIx32
, t
->unk4
);
2263 pandecode_prop("depth = %" PRId32
, t
->depth
);
2264 pandecode_prop("unk5 = 0x%" PRIx32
, t
->unk5
);
2265 pandecode_log("\n");
2267 /* See the definiton of enum mali_texture_type */
2269 bool is_cube
= t
->type
== MALI_TEX_CUBE
;
2270 unsigned dimension
= is_cube
? 2 : t
->type
;
2272 /* Print the layout. Default is linear; a modifier can denote AFBC or
2273 * u-interleaved/tiled modes */
2275 if (t
->layout
== MALI_TEXTURE_AFBC
)
2276 pandecode_log_cont("afbc");
2277 else if (t
->layout
== MALI_TEXTURE_TILED
)
2278 pandecode_log_cont("tiled");
2279 else if (t
->layout
== MALI_TEXTURE_LINEAR
)
2280 pandecode_log_cont("linear");
2282 pandecode_msg("XXX: invalid texture layout 0x%X\n", t
->layout
);
2284 pandecode_swizzle(t
->swizzle
, t
->format
);
2285 pandecode_log_cont(" ");
2287 /* Distinguish cube/2D with modifier */
2290 pandecode_log_cont("cube ");
2292 pandecode_format_short(t
->format
, t
->srgb
);
2294 /* All four width/height/depth/array_size dimensions are present
2295 * regardless of the type of texture, but it is an error to have
2296 * non-zero dimensions for unused dimensions. Verify this. array_size
2297 * can always be set, as can width. */
2299 if (t
->height
&& dimension
< 2)
2300 pandecode_msg("XXX: nonzero height for <2D texture\n");
2302 if (t
->depth
&& dimension
< 3)
2303 pandecode_msg("XXX: nonzero depth for <2D texture\n");
2305 /* Print only the dimensions that are actually there */
2307 pandecode_log_cont(": %d", t
->width
+ 1);
2310 pandecode_log_cont("x%u", t
->height
+ 1);
2313 pandecode_log_cont("x%u", t
->depth
+ 1);
2316 pandecode_log_cont("[%u]", t
->array_size
+ 1);
2319 pandecode_log_cont(" mip %u", t
->levels
);
2321 pandecode_log_cont("\n");
2323 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(t
->payload
);
2325 pandecode_texture_payload(t
->payload
, t
->type
, t
->layout
,
2326 true, t
->levels
, t
->depth
,
2327 t
->array_size
, tmem
);
2331 pandecode_log("};\n");
2334 /* 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 */
2337 pandecode_shader_prop(const char *name
, unsigned claim
, signed truth
, bool fuzzy
)
2343 if (fuzzy
&& (truth
< 0))
2344 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name
, claim
, truth
);
2346 if ((truth
>= 0) && !fuzzy
) {
2347 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2348 (truth
< claim
) ? "warn" : "XXX",
2349 name
, truth
, claim
);
2350 } else if ((claim
> -truth
) && !fuzzy
) {
2351 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2352 name
, -truth
, claim
);
2353 } else if (fuzzy
&& (claim
< truth
))
2354 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2355 name
, truth
, claim
);
2357 pandecode_log(".%s = %" PRId16
, name
, claim
);
2360 pandecode_log_cont(" /* %u used */", truth
);
2362 pandecode_log_cont(",\n");
2366 pandecode_blend_shader_disassemble(mali_ptr shader
, int job_no
, int job_type
,
2367 bool is_bifrost
, unsigned gpu_id
)
2369 struct midgard_disasm_stats stats
=
2370 pandecode_shader_disassemble(shader
, job_no
, job_type
, is_bifrost
, gpu_id
);
2372 bool has_texture
= (stats
.texture_count
> 0);
2373 bool has_sampler
= (stats
.sampler_count
> 0);
2374 bool has_attribute
= (stats
.attribute_count
> 0);
2375 bool has_varying
= (stats
.varying_count
> 0);
2376 bool has_uniform
= (stats
.uniform_count
> 0);
2377 bool has_ubo
= (stats
.uniform_buffer_count
> 0);
2379 if (has_texture
|| has_sampler
)
2380 pandecode_msg("XXX: blend shader accessing textures\n");
2382 if (has_attribute
|| has_varying
)
2383 pandecode_msg("XXX: blend shader accessing interstage\n");
2385 if (has_uniform
|| has_ubo
)
2386 pandecode_msg("XXX: blend shader accessing uniforms\n");
2390 pandecode_textures(mali_ptr textures
, unsigned texture_count
, int job_no
, bool is_bifrost
)
2392 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(textures
);
2398 const struct bifrost_texture_descriptor
*PANDECODE_PTR_VAR(t
, mmem
, textures
);
2400 pandecode_log("uint64_t textures_%"PRIx64
"_%d[] = {\n", textures
, job_no
);
2403 for (unsigned tex
= 0; tex
< texture_count
; ++tex
)
2404 pandecode_bifrost_texture(&t
[tex
], job_no
, tex
);
2407 pandecode_log("};\n");
2409 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
);
2411 pandecode_log("uint64_t textures_%"PRIx64
"_%d[] = {\n", textures
, job_no
);
2414 for (int tex
= 0; tex
< texture_count
; ++tex
) {
2415 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
2416 char *a
= pointer_as_memory_reference(*u
);
2417 pandecode_log("%s,\n", a
);
2422 pandecode_log("};\n");
2424 /* Now, finally, descend down into the texture descriptor */
2425 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
2426 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
2427 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
2429 pandecode_texture(*u
, tmem
, job_no
, tex
);
2435 pandecode_samplers(mali_ptr samplers
, unsigned sampler_count
, int job_no
, bool is_bifrost
)
2437 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(samplers
);
2443 struct bifrost_sampler_descriptor
*s
;
2445 for (int i
= 0; i
< sampler_count
; ++i
) {
2446 s
= pandecode_fetch_gpu_mem(smem
, samplers
+ sizeof(*s
) * i
, sizeof(*s
));
2448 pandecode_log("struct bifrost_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", samplers
+ sizeof(*s
) * i
, job_no
, i
);
2452 pandecode_msg("XXX: unk1 tripped\n");
2453 pandecode_prop("unk1 = 0x%x", s
->unk1
);
2456 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s
->wrap_s
));
2457 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s
->wrap_t
));
2458 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s
->wrap_r
));
2460 if (s
->unk8
!= 0x8) {
2461 pandecode_msg("XXX: unk8 tripped\n");
2462 pandecode_prop("unk8 = 0x%x", s
->unk8
);
2465 pandecode_prop("unk2 = 0x%x", s
->unk2
);
2466 pandecode_prop("unk3 = 0x%x", s
->unk3
);
2467 pandecode_prop("min_filter = %s", s
->min_filter
? "nearest" : "linear");
2468 pandecode_prop("norm_coords = 0x%x", s
->norm_coords
& 0x1);
2469 pandecode_prop("zero1 = 0x%x", s
->zero1
& 0x1);
2470 pandecode_prop("mip_filter = %s", s
->mip_filter
? "linear" : "nearest");
2471 pandecode_prop("mag_filter = %s", s
->mag_filter
? "linear" : "nearest");
2473 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
2474 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
2476 if (s
->zero1
|| s
->zero2
|| s
->zero3
|| s
->zero4
) {
2477 pandecode_msg("XXX: sampler zero tripped\n");
2478 pandecode_prop("zero = 0x%" PRIx8
", 0x%" PRIx64
", 0x%" PRIx64
", 0x%" PRIx64
"\n", s
->zero1
, s
->zero2
, s
->zero3
, s
->zero4
);
2482 pandecode_log("};\n");
2485 struct mali_sampler_descriptor
*s
;
2487 for (int i
= 0; i
< sampler_count
; ++i
) {
2488 s
= pandecode_fetch_gpu_mem(smem
, samplers
+ sizeof(*s
) * i
, sizeof(*s
));
2490 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", samplers
+ sizeof(*s
) * i
, job_no
, i
);
2493 pandecode_log(".filter_mode = ");
2494 pandecode_log_decoded_flags(sampler_flag_info
, s
->filter_mode
);
2495 pandecode_log_cont(",\n");
2497 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
2498 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
2501 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s
->lod_bias
));
2503 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s
->wrap_s
));
2504 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s
->wrap_t
));
2505 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s
->wrap_r
));
2507 pandecode_prop("compare_func = %s", pandecode_func(s
->compare_func
));
2509 if (s
->zero
|| s
->zero2
) {
2510 pandecode_msg("XXX: sampler zero tripped\n");
2511 pandecode_prop("zero = 0x%X, 0x%X\n", s
->zero
, s
->zero2
);
2514 pandecode_prop("seamless_cube_map = %d", s
->seamless_cube_map
);
2516 pandecode_prop("border_color = { %f, %f, %f, %f }",
2520 s
->border_color
[3]);
2523 pandecode_log("};\n");
2529 pandecode_vertex_tiler_postfix_pre(
2530 const struct mali_vertex_tiler_postfix
*p
,
2531 int job_no
, enum mali_job_type job_type
,
2532 char *suffix
, bool is_bifrost
, unsigned gpu_id
)
2534 struct pandecode_mapped_memory
*attr_mem
;
2536 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2537 * are the only things actually needed from the FBD, vertex/tiler jobs
2538 * no longer reference the FBD -- instead, this field points to some
2539 * info about the scratchpad.
2542 struct pandecode_fbd fbd_info
= {
2543 /* Default for Bifrost */
2548 pandecode_log_cont("\t/* %X %/\n", p
->shared_memory
& 1);
2549 pandecode_compute_fbd(p
->shared_memory
& ~1, job_no
);
2550 } else if (p
->shared_memory
& MALI_MFBD
)
2551 fbd_info
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->shared_memory
) & FBD_MASK
, job_no
, false, job_type
== JOB_TYPE_COMPUTE
, false);
2552 else if (job_type
== JOB_TYPE_COMPUTE
)
2553 pandecode_compute_fbd((u64
) (uintptr_t) p
->shared_memory
, job_no
);
2555 fbd_info
= pandecode_sfbd((u64
) (uintptr_t) p
->shared_memory
, job_no
, false, gpu_id
);
2557 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
2558 int texture_count
= 0, sampler_count
= 0;
2561 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->shader
);
2562 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, p
->shader
);
2564 /* Disassemble ahead-of-time to get stats. Initialize with
2565 * stats for the missing-shader case so we get validation
2568 struct midgard_disasm_stats info
= {
2571 .attribute_count
= 0,
2575 .uniform_count
= -128,
2576 .uniform_buffer_count
= 0
2579 if (s
->shader
& ~0xF)
2580 info
= pandecode_shader_disassemble(s
->shader
& ~0xF, job_no
, job_type
, is_bifrost
, gpu_id
);
2582 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", p
->shader
, job_no
, suffix
);
2585 /* Save for dumps */
2586 attribute_count
= s
->attribute_count
;
2587 varying_count
= s
->varying_count
;
2588 texture_count
= s
->texture_count
;
2589 sampler_count
= s
->sampler_count
;
2592 uniform_count
= s
->bifrost2
.uniform_count
;
2593 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
2595 uniform_count
= s
->midgard1
.uniform_count
;
2596 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
2599 pandecode_shader_address("shader", s
->shader
);
2601 pandecode_shader_prop("texture_count", s
->texture_count
, info
.texture_count
, false);
2602 pandecode_shader_prop("sampler_count", s
->sampler_count
, info
.sampler_count
, false);
2603 pandecode_shader_prop("attribute_count", s
->attribute_count
, info
.attribute_count
, false);
2604 pandecode_shader_prop("varying_count", s
->varying_count
, info
.varying_count
, false);
2605 pandecode_shader_prop("uniform_buffer_count",
2606 uniform_buffer_count
,
2607 info
.uniform_buffer_count
, true);
2610 pandecode_shader_prop("uniform_count",
2612 info
.uniform_count
, false);
2614 pandecode_shader_prop("work_count",
2615 s
->midgard1
.work_count
, info
.work_count
, false);
2619 pandecode_log("bifrost1.unk1 = ");
2620 pandecode_log_decoded_flags(shader_bifrost_info
, s
->bifrost1
.unk1
);
2621 pandecode_log_cont(",\n");
2623 bool helpers
= s
->midgard1
.flags_lo
& MALI_HELPER_INVOCATIONS
;
2625 if (helpers
!= info
.helper_invocations
) {
2626 pandecode_msg("XXX: expected helpers %u but got %u\n",
2627 info
.helper_invocations
, helpers
);
2630 pandecode_log(".midgard1.flags_lo = ");
2631 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info
,
2632 s
->midgard1
.flags_lo
& ~MALI_HELPER_INVOCATIONS
);
2633 pandecode_log_cont(",\n");
2635 pandecode_log(".midgard1.flags_hi = ");
2636 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info
, s
->midgard1
.flags_hi
);
2637 pandecode_log_cont(",\n");
2640 if (s
->depth_units
|| s
->depth_factor
) {
2641 pandecode_prop("depth_factor = %f", s
->depth_factor
);
2642 pandecode_prop("depth_units = %f", s
->depth_units
);
2645 if (s
->alpha_coverage
) {
2646 bool invert_alpha_coverage
= s
->alpha_coverage
& 0xFFF0;
2647 uint16_t inverted_coverage
= invert_alpha_coverage
? ~s
->alpha_coverage
: s
->alpha_coverage
;
2649 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2650 invert_alpha_coverage
? "~" : "",
2651 MALI_GET_ALPHA_COVERAGE(inverted_coverage
));
2655 pandecode_prop(".unknown2_2 = %X", s
->unknown2_2
);
2657 if (s
->unknown2_3
|| s
->unknown2_4
) {
2658 pandecode_log(".unknown2_3 = ");
2660 int unknown2_3
= s
->unknown2_3
;
2661 int unknown2_4
= s
->unknown2_4
;
2663 /* We're not quite sure what these flags mean without the depth test, if anything */
2665 if (unknown2_3
& (MALI_DEPTH_WRITEMASK
| MALI_DEPTH_FUNC_MASK
)) {
2666 const char *func
= pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3
));
2667 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
2669 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
2672 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
2673 pandecode_log_cont(",\n");
2675 pandecode_log(".unknown2_4 = ");
2676 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
2677 pandecode_log_cont(",\n");
2680 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
2681 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
2682 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
2685 pandecode_stencil("front", &s
->stencil_front
);
2686 pandecode_stencil("back", &s
->stencil_back
);
2689 pandecode_log(".bifrost2 = {\n");
2692 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
2693 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
2694 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
2695 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
2698 pandecode_log("},\n");
2699 } else if (s
->midgard2
.unknown2_7
) {
2700 pandecode_log(".midgard2 = {\n");
2703 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
2705 pandecode_log("},\n");
2709 pandecode_msg("XXX: shader padding tripped\n");
2710 pandecode_prop("padding = 0x%" PRIx32
, s
->padding
);
2714 /* TODO: Blend shaders routing/disasm */
2715 union midgard_blend blend
= s
->blend
;
2716 mali_ptr shader
= pandecode_midgard_blend(&blend
, s
->unknown2_3
& MALI_HAS_BLEND_SHADER
);
2718 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
2720 pandecode_msg("mdg_blend = %" PRIx64
"\n", s
->blend
.shader
);
2724 pandecode_log("};\n");
2726 /* MRT blend fields are used whenever MFBD is used, with
2727 * per-RT descriptors */
2729 if (job_type
== JOB_TYPE_TILER
&& (is_bifrost
|| p
->shared_memory
& MALI_MFBD
)) {
2730 void* blend_base
= (void *) (s
+ 1);
2732 for (unsigned i
= 0; i
< fbd_info
.rt_count
; i
++) {
2733 mali_ptr shader
= 0;
2736 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
2738 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
2741 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
2746 pandecode_msg("XXX: missing shader descriptor\n");
2749 struct pandecode_mapped_memory
*fmem
= pandecode_find_mapped_gpu_mem_containing(p
->viewport
);
2750 struct mali_viewport
*PANDECODE_PTR_VAR(f
, fmem
, p
->viewport
);
2752 pandecode_log("struct mali_viewport viewport_%"PRIx64
"_%d%s = {\n", p
->viewport
, job_no
, suffix
);
2755 pandecode_prop("clip_minx = %f", f
->clip_minx
);
2756 pandecode_prop("clip_miny = %f", f
->clip_miny
);
2757 pandecode_prop("clip_minz = %f", f
->clip_minz
);
2758 pandecode_prop("clip_maxx = %f", f
->clip_maxx
);
2759 pandecode_prop("clip_maxy = %f", f
->clip_maxy
);
2760 pandecode_prop("clip_maxz = %f", f
->clip_maxz
);
2762 /* Only the higher coordinates are MALI_POSITIVE scaled */
2764 pandecode_prop("viewport0 = { %d, %d }",
2765 f
->viewport0
[0], f
->viewport0
[1]);
2767 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2768 f
->viewport1
[0] + 1, f
->viewport1
[1] + 1);
2771 pandecode_log("};\n");
2774 unsigned max_attr_index
= 0;
2776 if (p
->attribute_meta
)
2777 max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
2779 if (p
->attributes
) {
2780 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
2781 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false, job_type
);
2784 /* Varyings are encoded like attributes but not actually sent; we just
2785 * pass a zero buffer with the right stride/size set, (or whatever)
2786 * since the GPU will write to it itself */
2788 if (p
->varying_meta
) {
2789 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
2793 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
2795 /* Number of descriptors depends on whether there are
2796 * non-internal varyings */
2798 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true, job_type
);
2801 if (p
->uniform_buffers
) {
2802 if (uniform_buffer_count
)
2803 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
2805 pandecode_msg("warn: UBOs specified but not referenced\n");
2806 } else if (uniform_buffer_count
)
2807 pandecode_msg("XXX: UBOs referenced but not specified\n");
2809 /* We don't want to actually dump uniforms, but we do need to validate
2810 * that the counts we were given are sane */
2814 pandecode_uniforms(p
->uniforms
, uniform_count
);
2816 pandecode_msg("warn: Uniforms specified but not referenced\n");
2817 } else if (uniform_count
)
2818 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2821 pandecode_textures(p
->textures
, texture_count
, job_no
, is_bifrost
);
2823 if (p
->sampler_descriptor
)
2824 pandecode_samplers(p
->sampler_descriptor
, sampler_count
, job_no
, is_bifrost
);
2828 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2830 pandecode_log(".gl_enables = ");
2832 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2834 pandecode_log_cont(",\n");
2838 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2840 if (p
->shader
& 0xF)
2841 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p
->shader
& 0xF));
2843 pandecode_log(".postfix = {\n");
2846 pandecode_gl_enables(p
->gl_enables
, JOB_TYPE_TILER
);
2847 pandecode_prop("instance_shift = 0x%x", p
->instance_shift
);
2848 pandecode_prop("instance_odd = 0x%x", p
->instance_odd
);
2851 pandecode_msg("XXX: vertex only zero tripped");
2852 pandecode_prop("zero4 = 0x%" PRIx32
, p
->zero4
);
2855 pandecode_prop("offset_start = 0x%x", p
->offset_start
);
2858 pandecode_msg("XXX: vertex only zero tripped");
2859 pandecode_prop("zero5 = 0x%" PRIx64
, p
->zero5
);
2862 MEMORY_PROP(p
, position_varying
);
2863 MEMORY_PROP(p
, occlusion_counter
);
2866 pandecode_log("},\n");
2870 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2872 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2873 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2875 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2879 pandecode_msg("XXX: tiler heap zero tripped\n");
2880 pandecode_prop("zero = 0x%x", h
->zero
);
2883 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2884 MEMORY_PROP(h
, tiler_heap_start
);
2885 MEMORY_PROP(h
, tiler_heap_free
);
2887 /* this might point to the beginning of another buffer, when it's
2888 * really the end of the tiler heap buffer, so we have to be careful
2889 * here. but for zero length, we need the same pointer.
2892 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2893 MEMORY_PROP(h
, tiler_heap_start
);
2895 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2896 pandecode_prop("tiler_heap_end = %s + 1", a
);
2900 for (int i
= 0; i
< 10; i
++) {
2901 if (h
->zeros
[i
] != 0) {
2902 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2907 if (h
->unk1
!= 0x1) {
2908 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2909 pandecode_prop("unk1 = 0x%x", h
->unk1
);
2912 if (h
->unk7e007e
!= 0x7e007e) {
2913 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
2914 pandecode_prop("unk7e007e = 0x%x", h
->unk7e007e
);
2918 pandecode_log("};\n");
2922 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2924 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2925 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2927 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2929 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2932 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32
, t
->tiler_heap_next_start
);
2933 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32
, t
->used_hierarchy_mask
);
2935 if (t
->hierarchy_mask
!= 0xa &&
2936 t
->hierarchy_mask
!= 0x14 &&
2937 t
->hierarchy_mask
!= 0x28 &&
2938 t
->hierarchy_mask
!= 0x50 &&
2939 t
->hierarchy_mask
!= 0xa0)
2940 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
2942 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2944 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2946 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2947 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2950 pandecode_msg("XXX: tiler meta zero tripped\n");
2951 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2954 for (int i
= 0; i
< 12; i
++) {
2955 if (t
->zeros
[i
] != 0) {
2956 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2962 pandecode_log("};\n");
2966 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2968 if (u
.pointer
== 0x0)
2971 pandecode_log(".primitive_size = {\n");
2975 pandecode_prop("constant = %f", u
.constant
);
2977 MEMORY_PROP((&u
), pointer
);
2981 pandecode_log("},\n");
2985 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2987 pandecode_log_cont("{\n");
2990 /* TODO: gl_PointSize on Bifrost */
2991 pandecode_primitive_size(t
->primitive_size
, true);
2993 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2995 pandecode_msg("XXX: tiler only zero tripped\n");
2996 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2997 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2998 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2999 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
3000 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
3001 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
3005 pandecode_log("},\n");
3009 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
3010 const struct pandecode_mapped_memory
*mem
,
3011 mali_ptr payload
, int job_no
, unsigned gpu_id
)
3013 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
3015 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
3017 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
3020 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
3021 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
3024 pandecode_log("};\n");
3030 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
3031 const struct pandecode_mapped_memory
*mem
,
3032 mali_ptr payload
, int job_no
, unsigned gpu_id
)
3034 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
3036 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
3037 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
3039 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
3042 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
3044 pandecode_log(".tiler = ");
3045 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
3047 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
3050 pandecode_log("};\n");
3056 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
3057 const struct pandecode_mapped_memory
*mem
,
3058 mali_ptr payload
, int job_no
, unsigned gpu_id
)
3060 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
3061 bool is_graphics
= (h
->job_type
== JOB_TYPE_VERTEX
) || (h
->job_type
== JOB_TYPE_TILER
);
3063 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false, gpu_id
);
3065 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
3068 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, is_graphics
);
3069 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
3071 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
3072 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
3075 pandecode_log("};\n");
3081 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
3082 mali_ptr payload
, int job_no
,
3083 bool is_bifrost
, unsigned gpu_id
)
3085 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
3087 bool is_mfbd
= s
->framebuffer
& MALI_MFBD
;
3089 if (!is_mfbd
&& is_bifrost
)
3090 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
3092 struct pandecode_fbd info
;
3095 info
= pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true, false, is_bifrost
);
3097 info
= pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true, gpu_id
);
3099 /* Compute the tag for the tagged pointer. This contains the type of
3100 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
3101 * additional structures follow the MFBD header (an extra payload or
3102 * not, as well as a count of render targets) */
3104 unsigned expected_tag
= is_mfbd
? MALI_MFBD
: 0;
3108 expected_tag
|= MALI_MFBD_TAG_EXTRA
;
3110 expected_tag
|= (MALI_POSITIVE(info
.rt_count
) << 2);
3113 if ((s
->min_tile_coord
| s
->max_tile_coord
) & ~(MALI_X_COORD_MASK
| MALI_Y_COORD_MASK
)) {
3114 pandecode_msg("XXX: unexpected tile coordinate bits\n");
3115 pandecode_prop("min_tile_coord = 0x%X\n", s
->min_tile_coord
);
3116 pandecode_prop("max_tile_coord = 0x%X\n", s
->max_tile_coord
);
3119 /* Extract tile coordinates */
3121 unsigned min_x
= MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
3122 unsigned min_y
= MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
3124 unsigned max_x
= (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
3125 unsigned max_y
= (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
3127 /* For the max, we also want the floored (rather than ceiled) version for checking */
3129 unsigned max_x_f
= (MALI_TILE_COORD_X(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
3130 unsigned max_y_f
= (MALI_TILE_COORD_Y(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
3132 /* Validate the coordinates are well-ordered */
3135 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
3136 else if (min_x
> max_x
)
3137 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
3140 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
3141 else if (min_y
> max_y
)
3142 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
3144 /* Validate the coordinates fit inside the framebuffer. We use floor,
3145 * rather than ceil, for the max coordinates, since the tile
3146 * coordinates for something like an 800x600 framebuffer will actually
3147 * resolve to 800x608, which would otherwise trigger a Y-overflow */
3149 if ((min_x
> info
.width
) || (max_x_f
> info
.width
))
3150 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
3152 if ((min_y
> info
.height
) || (max_y_f
> info
.height
))
3153 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
3155 /* After validation, we print */
3157 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x
, min_y
, max_x
, max_y
);
3159 /* The FBD is a tagged pointer */
3161 unsigned tag
= (s
->framebuffer
& ~FBD_MASK
);
3163 if (tag
!= expected_tag
)
3164 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag
, tag
);
3169 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
3170 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
3171 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
3172 * might wish to combine this with the bifrost parameter) because some details
3173 * are model-specific even within a particular architecture. Minimal traces
3174 * *only* examine the job descriptors, skipping printing entirely if there is
3175 * no faults, and only descends into the payload if there are faults. This is
3176 * useful for looking for faults without the overhead of invasive traces. */
3179 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
, unsigned gpu_id
, bool minimal
)
3181 struct mali_job_descriptor_header
*h
;
3182 unsigned job_descriptor_number
= 0;
3185 struct pandecode_mapped_memory
*mem
=
3186 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
3190 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
3192 /* On Midgard, for 32-bit jobs except for fragment jobs, the
3193 * high 32-bits of the 64-bit pointer are reused to store
3196 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
3197 h
->job_type
!= JOB_TYPE_FRAGMENT
? 4 : 0;
3198 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
3200 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
, 256);
3202 int job_no
= job_descriptor_number
++;
3204 /* If the job is good to go, skip it in minimal mode */
3205 if (minimal
&& (h
->exception_status
== 0x0 || h
->exception_status
== 0x1))
3208 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
3211 pandecode_prop("job_type = %s", pandecode_job_type(h
->job_type
));
3213 if (h
->job_descriptor_size
)
3214 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
3216 if (h
->exception_status
&& h
->exception_status
!= 0x1)
3217 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
3218 h
->exception_status
,
3219 (h
->exception_status
>> 16) & 0xFFFF,
3220 pandecode_exception_access((h
->exception_status
>> 8) & 0x3),
3221 h
->exception_status
& 0xFF);
3223 if (h
->first_incomplete_task
)
3224 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
3226 if (h
->fault_pointer
)
3227 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
3230 pandecode_prop("job_barrier = %d", h
->job_barrier
);
3232 pandecode_prop("job_index = %d", h
->job_index
);
3234 if (h
->unknown_flags
)
3235 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
3237 if (h
->job_dependency_index_1
)
3238 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
3240 if (h
->job_dependency_index_2
)
3241 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
3244 pandecode_log("};\n");
3246 switch (h
->job_type
) {
3247 case JOB_TYPE_WRITE_VALUE
: {
3248 struct mali_payload_write_value
*s
= payload
;
3249 pandecode_log("struct mali_payload_write_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
3251 MEMORY_PROP(s
, address
);
3253 if (s
->value_descriptor
!= MALI_WRITE_VALUE_ZERO
) {
3254 pandecode_msg("XXX: unknown value descriptor\n");
3255 pandecode_prop("value_descriptor = 0x%" PRIX32
, s
->value_descriptor
);
3259 pandecode_msg("XXX: set value tripped\n");
3260 pandecode_prop("reserved = 0x%" PRIX32
, s
->reserved
);
3263 pandecode_prop("immediate = 0x%" PRIX64
, s
->immediate
);
3265 pandecode_log("};\n");
3270 case JOB_TYPE_TILER
:
3271 case JOB_TYPE_VERTEX
:
3272 case JOB_TYPE_COMPUTE
:
3274 if (h
->job_type
== JOB_TYPE_TILER
)
3275 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
3277 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
3279 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
, gpu_id
);
3283 case JOB_TYPE_FRAGMENT
:
3284 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
, gpu_id
);
3290 } while ((jc_gpu_va
= h
->next_job
));