2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
26 #include <midgard_pack.h>
34 #include "util/macros.h"
35 #include "util/u_math.h"
37 #include "midgard/disassemble.h"
38 #include "bifrost/disassemble.h"
40 #include "pan_encoder.h"
42 static void pandecode_swizzle(unsigned swizzle
, enum mali_format format
);
44 #define MEMORY_PROP(obj, p) {\
46 char *a = pointer_as_memory_reference(obj->p); \
47 pandecode_prop("%s = %s", #p, a); \
52 #define MEMORY_PROP_DIR(obj, p) {\
54 char *a = pointer_as_memory_reference(obj.p); \
55 pandecode_prop("%s = %s", #p, a); \
60 #define DUMP_CL(title, T, cl, indent) {\
61 fprintf(pandecode_dump_stream, "%s\n", title); \
62 struct MALI_ ## T temp; \
63 MALI_ ## T ## _unpack((const uint8_t *) cl, &temp); \
64 MALI_ ## T ## _print(pandecode_dump_stream, &temp, 0); \
67 #define DUMP_ADDR(title, T, addr, indent) {\
68 struct pandecode_mapped_memory *mapped_mem = pandecode_find_mapped_gpu_mem_containing(addr); \
69 const uint8_t *cl = pandecode_fetch_gpu_mem(mapped_mem, addr, MALI_ ## T ## _LENGTH); \
70 DUMP_CL(title, T, cl, indent); \
73 FILE *pandecode_dump_stream
;
75 /* Semantic logging type.
77 * Raw: for raw messages to be printed as is.
78 * Message: for helpful information to be commented out in replays.
79 * Property: for properties of a struct
81 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
84 enum pandecode_log_type
{
90 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
91 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
92 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
94 unsigned pandecode_indent
= 0;
97 pandecode_make_indent(void)
99 for (unsigned i
= 0; i
< pandecode_indent
; ++i
)
100 fprintf(pandecode_dump_stream
, " ");
104 pandecode_log_typed(enum pandecode_log_type type
, const char *format
, ...)
108 pandecode_make_indent();
110 if (type
== PANDECODE_MESSAGE
)
111 fprintf(pandecode_dump_stream
, "// ");
112 else if (type
== PANDECODE_PROPERTY
)
113 fprintf(pandecode_dump_stream
, ".");
115 va_start(ap
, format
);
116 vfprintf(pandecode_dump_stream
, format
, ap
);
119 if (type
== PANDECODE_PROPERTY
)
120 fprintf(pandecode_dump_stream
, ",\n");
124 pandecode_log_cont(const char *format
, ...)
128 va_start(ap
, format
);
129 vfprintf(pandecode_dump_stream
, format
, ap
);
133 /* To check for memory safety issues, validates that the given pointer in GPU
134 * memory is valid, containing at least sz bytes. The goal is to eliminate
135 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
136 * overruns) by statically validating pointers.
140 pandecode_validate_buffer(mali_ptr addr
, size_t sz
)
143 pandecode_msg("XXX: null pointer deref");
149 struct pandecode_mapped_memory
*bo
=
150 pandecode_find_mapped_gpu_mem_containing(addr
);
153 pandecode_msg("XXX: invalid memory dereference\n");
159 unsigned offset
= addr
- bo
->gpu_va
;
160 unsigned total
= offset
+ sz
;
162 if (total
> bo
->length
) {
163 pandecode_msg("XXX: buffer overrun. "
164 "Chunk of size %zu at offset %d in buffer of size %zu. "
165 "Overrun by %zu bytes. \n",
166 sz
, offset
, bo
->length
, total
- bo
->length
);
171 struct pandecode_flag_info
{
177 pandecode_log_decoded_flags(const struct pandecode_flag_info
*flag_info
,
180 bool decodable_flags_found
= false;
182 for (int i
= 0; flag_info
[i
].name
; i
++) {
183 if ((flags
& flag_info
[i
].flag
) != flag_info
[i
].flag
)
186 if (!decodable_flags_found
) {
187 decodable_flags_found
= true;
189 pandecode_log_cont(" | ");
192 pandecode_log_cont("%s", flag_info
[i
].name
);
194 flags
&= ~flag_info
[i
].flag
;
197 if (decodable_flags_found
) {
199 pandecode_log_cont(" | 0x%" PRIx64
, flags
);
201 pandecode_log_cont("0x%" PRIx64
, flags
);
205 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
206 static const struct pandecode_flag_info gl_enable_flag_info
[] = {
207 FLAG_INFO(OCCLUSION_QUERY
),
208 FLAG_INFO(OCCLUSION_PRECISE
),
209 FLAG_INFO(FRONT_CCW_TOP
),
210 FLAG_INFO(CULL_FACE_FRONT
),
211 FLAG_INFO(CULL_FACE_BACK
),
216 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
217 static const struct pandecode_flag_info clear_flag_info
[] = {
220 FLAG_INFO(SLOW_STENCIL
),
225 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
226 static const struct pandecode_flag_info mask_flag_info
[] = {
235 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
236 static const struct pandecode_flag_info u3_flag_info
[] = {
238 FLAG_INFO(PER_SAMPLE
),
239 FLAG_INFO(CAN_DISCARD
),
240 FLAG_INFO(HAS_BLEND_SHADER
),
241 FLAG_INFO(DEPTH_WRITEMASK
),
242 FLAG_INFO(DEPTH_CLIP_NEAR
),
243 FLAG_INFO(DEPTH_CLIP_FAR
),
247 static const struct pandecode_flag_info u4_flag_info
[] = {
249 FLAG_INFO(NO_DITHER
),
250 FLAG_INFO(DEPTH_RANGE_A
),
251 FLAG_INFO(DEPTH_RANGE_B
),
252 FLAG_INFO(STENCIL_TEST
),
253 FLAG_INFO(ALPHA_TO_COVERAGE
),
258 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
259 static const struct pandecode_flag_info mfbd_fmt_flag_info
[] = {
265 #define FLAG_INFO(flag) { MALI_AFBC_##flag, "MALI_AFBC_" #flag }
266 static const struct pandecode_flag_info afbc_fmt_flag_info
[] = {
272 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
273 static const struct pandecode_flag_info mfbd_extra_flag_hi_info
[] = {
279 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
280 static const struct pandecode_flag_info mfbd_extra_flag_lo_info
[] = {
286 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
287 static const struct pandecode_flag_info shader_midgard1_flag_lo_info
[] = {
290 FLAG_INFO(READS_TILEBUFFER
),
291 FLAG_INFO(WRITES_GLOBAL
),
296 static const struct pandecode_flag_info shader_midgard1_flag_hi_info
[] = {
298 FLAG_INFO(SUPPRESS_INF_NAN
),
303 #define FLAG_INFO(flag) { MALI_BIFROST_##flag, "MALI_BIFROST_" #flag }
304 static const struct pandecode_flag_info shader_bifrost_info
[] = {
305 FLAG_INFO(FULL_THREAD
),
307 FLAG_INFO(FIRST_ATEST
),
313 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
314 static const struct pandecode_flag_info mfbd_flag_info
[] = {
315 FLAG_INFO(DEPTH_WRITE
),
321 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
322 static const struct pandecode_flag_info sampler_flag_info
[] = {
323 FLAG_INFO(MAG_NEAREST
),
324 FLAG_INFO(MIN_NEAREST
),
325 FLAG_INFO(MIP_LINEAR_1
),
326 FLAG_INFO(MIP_LINEAR_2
),
327 FLAG_INFO(NORM_COORDS
),
332 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
333 static const struct pandecode_flag_info sfbd_unk1_info
[] = {
340 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
341 static const struct pandecode_flag_info sfbd_unk2_info
[] = {
348 #define DEFINE_CASE(name) case MALI_## name: return "MALI_" #name
349 static char *pandecode_format(enum mali_format format
)
351 static char unk_format_str
[10];
354 DEFINE_CASE(ETC2_RGB8
);
355 DEFINE_CASE(ETC2_R11_UNORM
);
356 DEFINE_CASE(ETC2_RGBA8
);
357 DEFINE_CASE(ETC2_RG11_UNORM
);
358 DEFINE_CASE(ETC2_R11_SNORM
);
359 DEFINE_CASE(ETC2_RG11_SNORM
);
360 DEFINE_CASE(ETC2_RGB8A1
);
362 DEFINE_CASE(BC1_UNORM
);
363 DEFINE_CASE(BC2_UNORM
);
364 DEFINE_CASE(BC3_UNORM
);
365 DEFINE_CASE(BC4_UNORM
);
366 DEFINE_CASE(BC4_SNORM
);
367 DEFINE_CASE(BC5_UNORM
);
368 DEFINE_CASE(BC5_SNORM
);
369 DEFINE_CASE(BC6H_UF16
);
370 DEFINE_CASE(BC6H_SF16
);
371 DEFINE_CASE(BC7_UNORM
);
372 DEFINE_CASE(ASTC_3D_LDR
);
373 DEFINE_CASE(ASTC_3D_HDR
);
374 DEFINE_CASE(ASTC_2D_LDR
);
375 DEFINE_CASE(ASTC_2D_HDR
);
377 DEFINE_CASE(RGB5_X1_UNORM
);
378 DEFINE_CASE(RGB5_A1_UNORM
);
379 DEFINE_CASE(RGB10_A2_UNORM
);
380 DEFINE_CASE(RGB10_A2_SNORM
);
381 DEFINE_CASE(RGB10_A2UI
);
382 DEFINE_CASE(RGB10_A2I
);
383 DEFINE_CASE(RGB332_UNORM
);
384 DEFINE_CASE(RGB233_UNORM
);
385 DEFINE_CASE(Z24X8_UNORM
);
386 DEFINE_CASE(R32_FIXED
);
387 DEFINE_CASE(RG32_FIXED
);
388 DEFINE_CASE(RGB32_FIXED
);
389 DEFINE_CASE(RGBA32_FIXED
);
390 DEFINE_CASE(R11F_G11F_B10F
);
391 DEFINE_CASE(R9F_G9F_B9F_E5F
);
392 DEFINE_CASE(VARYING_POS
);
393 DEFINE_CASE(VARYING_DISCARD
);
395 DEFINE_CASE(R8_SNORM
);
396 DEFINE_CASE(R16_SNORM
);
397 DEFINE_CASE(R32_SNORM
);
398 DEFINE_CASE(RG8_SNORM
);
399 DEFINE_CASE(RG16_SNORM
);
400 DEFINE_CASE(RG32_SNORM
);
401 DEFINE_CASE(RGB8_SNORM
);
402 DEFINE_CASE(RGB16_SNORM
);
403 DEFINE_CASE(RGB32_SNORM
);
404 DEFINE_CASE(RGBA8_SNORM
);
405 DEFINE_CASE(RGBA16_SNORM
);
406 DEFINE_CASE(RGBA32_SNORM
);
415 DEFINE_CASE(RGB16UI
);
416 DEFINE_CASE(RGB32UI
);
417 DEFINE_CASE(RGBA8UI
);
418 DEFINE_CASE(RGBA16UI
);
419 DEFINE_CASE(RGBA32UI
);
421 DEFINE_CASE(R8_UNORM
);
422 DEFINE_CASE(R16_UNORM
);
423 DEFINE_CASE(R32_UNORM
);
425 DEFINE_CASE(RG8_UNORM
);
426 DEFINE_CASE(RG16_UNORM
);
427 DEFINE_CASE(RG32_UNORM
);
429 DEFINE_CASE(RGB8_UNORM
);
430 DEFINE_CASE(RGB16_UNORM
);
431 DEFINE_CASE(RGB32_UNORM
);
433 DEFINE_CASE(RGBA4_UNORM
);
434 DEFINE_CASE(RGBA8_UNORM
);
435 DEFINE_CASE(RGBA16_UNORM
);
436 DEFINE_CASE(RGBA32_UNORM
);
437 DEFINE_CASE(RGBA32F
);
452 DEFINE_CASE(RGBA16I
);
453 DEFINE_CASE(RGBA32I
);
454 DEFINE_CASE(RGBA16F
);
457 DEFINE_CASE(RGBA8_2
);
458 DEFINE_CASE(RGB10_A2_2
);
460 snprintf(unk_format_str
, sizeof(unk_format_str
), "MALI_0x%02x", format
);
461 return unk_format_str
;
468 pandecode_draw_mode(enum mali_draw_mode mode
)
470 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
473 DEFINE_CASE(DRAW_NONE
);
476 DEFINE_CASE(TRIANGLES
);
477 DEFINE_CASE(TRIANGLE_STRIP
);
478 DEFINE_CASE(TRIANGLE_FAN
);
479 DEFINE_CASE(LINE_STRIP
);
480 DEFINE_CASE(LINE_LOOP
);
481 DEFINE_CASE(POLYGON
);
483 DEFINE_CASE(QUAD_STRIP
);
486 pandecode_msg("XXX: invalid draw mode %X\n", mode
);
493 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
495 pandecode_func(enum mali_func mode
)
502 DEFINE_CASE(GREATER
);
503 DEFINE_CASE(NOTEQUAL
);
508 pandecode_msg("XXX: invalid func %X\n", mode
);
514 #define DEFINE_CASE(name) case MALI_MSAA_ ## name: return "MALI_MSAA_" #name
516 pandecode_msaa_mode(enum mali_msaa_mode mode
)
520 DEFINE_CASE(AVERAGE
);
521 DEFINE_CASE(MULTIPLE
);
522 DEFINE_CASE(LAYERED
);
524 unreachable("Impossible");
530 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
532 pandecode_stencil_op(enum mali_stencil_op op
)
536 DEFINE_CASE(REPLACE
);
539 DEFINE_CASE(INCR_WRAP
);
540 DEFINE_CASE(DECR_WRAP
);
545 pandecode_msg("XXX: invalid stencil op %X\n", op
);
552 static char *pandecode_attr_mode_short(enum mali_attr_mode mode
)
555 /* TODO: Combine to just "instanced" once this can be done
556 * unambiguously in all known cases */
557 case MALI_ATTR_POT_DIVIDE
:
558 return "instanced_pot";
559 case MALI_ATTR_MODULO
:
560 return "instanced_mod";
561 case MALI_ATTR_NPOT_DIVIDE
:
562 return "instanced_npot";
563 case MALI_ATTR_IMAGE
:
566 pandecode_msg("XXX: invalid attribute mode %X\n", mode
);
572 pandecode_special_record(uint64_t v
, bool* attribute
)
575 case MALI_ATTR_VERTEXID
:
577 return "gl_VertexID";
578 case MALI_ATTR_INSTANCEID
:
580 return "gl_InstanceID";
581 case MALI_VARYING_FRAG_COORD
:
582 return "gl_FragCoord";
583 case MALI_VARYING_FRONT_FACING
:
584 return "gl_FrontFacing";
585 case MALI_VARYING_POINT_COORD
:
586 return "gl_PointCoord";
588 pandecode_msg("XXX: invalid special record %" PRIx64
"\n", v
);
593 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
595 pandecode_wrap_mode(enum mali_wrap_mode op
)
599 DEFINE_CASE(CLAMP_TO_EDGE
);
601 DEFINE_CASE(CLAMP_TO_BORDER
);
602 DEFINE_CASE(MIRRORED_REPEAT
);
603 DEFINE_CASE(MIRRORED_CLAMP_TO_EDGE
);
604 DEFINE_CASE(MIRRORED_CLAMP
);
605 DEFINE_CASE(MIRRORED_CLAMP_TO_BORDER
);
608 pandecode_msg("XXX: invalid wrap mode %X\n", op
);
614 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
616 pandecode_block_format(enum mali_block_format fmt
)
620 DEFINE_CASE(UNKNOWN
);
625 unreachable("Invalid case");
630 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
632 pandecode_exception_access(unsigned access
)
636 DEFINE_CASE(EXECUTE
);
641 unreachable("Invalid case");
646 /* Midgard's tiler descriptor is embedded within the
650 pandecode_midgard_tiler_descriptor(
651 const struct midgard_tiler_descriptor
*t
,
657 pandecode_log(".tiler = {\n");
660 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
661 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
663 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
665 /* We know this name from the kernel, but we never see it nonzero */
668 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
670 MEMORY_PROP(t
, polygon_list
);
672 /* The body is offset from the base of the polygon list */
673 //assert(t->polygon_list_body > t->polygon_list);
674 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
676 /* It needs to fit inside the reported size */
677 //assert(t->polygon_list_size >= body_offset);
679 /* Now that we've sanity checked, we'll try to calculate the sizes
680 * ourselves for comparison */
682 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
683 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
685 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
686 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
687 ref_header
, ref_size
);
688 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
689 pandecode_msg("body offset %d\n", body_offset
);
692 /* The tiler heap has a start and end specified -- it should be
693 * identical to what we have in the BO. The exception is if tiling is
696 MEMORY_PROP(t
, heap_start
);
697 assert(t
->heap_end
>= t
->heap_start
);
699 struct pandecode_mapped_memory
*heap
=
700 pandecode_find_mapped_gpu_mem_containing(t
->heap_start
);
702 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
704 /* Tiling is enabled with a special flag */
705 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
706 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
708 bool tiling_enabled
= hierarchy_mask
;
710 if (tiling_enabled
) {
711 /* When tiling is enabled, the heap should be a tight fit */
712 unsigned heap_offset
= t
->heap_start
- heap
->gpu_va
;
713 if ((heap_offset
+ heap_size
) != heap
->length
) {
714 pandecode_msg("XXX: heap size %u (expected %zu)\n",
715 heap_size
, heap
->length
- heap_offset
);
718 /* We should also have no other flags */
720 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
722 /* When tiling is disabled, we should have that flag and no others */
724 if (tiler_flags
!= MALI_TILER_DISABLED
) {
725 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
729 /* We should also have an empty heap */
731 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
735 /* Disabled tiling is used only for clear-only jobs, which are
736 * purely FRAGMENT, so we should never see this for
737 * non-FRAGMENT descriptors. */
740 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
743 /* We've never seen weights used in practice, but we know from the
744 * kernel these fields is there */
746 bool nonzero_weights
= false;
748 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
749 nonzero_weights
|= t
->weights
[w
] != 0x0;
752 if (nonzero_weights
) {
753 pandecode_log(".weights = { ");
755 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
756 pandecode_log_cont("%d, ", t
->weights
[w
]);
763 pandecode_log("}\n");
766 /* TODO: The Bifrost tiler is not understood at all yet */
769 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer
*fb
)
771 pandecode_log(".tiler = {\n");
774 MEMORY_PROP(fb
, tiler_meta
);
776 for (int i
= 0; i
< 16; i
++) {
777 if (fb
->zeros
[i
] != 0) {
778 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
783 pandecode_log("},\n");
786 pandecode_log("}\n");
790 /* Information about the framebuffer passed back for
791 * additional analysis */
793 struct pandecode_fbd
{
801 pandecode_sfbd_format(struct mali_sfbd_format format
)
803 pandecode_log(".format = {\n");
806 pandecode_log(".unk1 = ");
807 pandecode_log_decoded_flags(sfbd_unk1_info
, format
.unk1
);
808 pandecode_log_cont(",\n");
810 /* TODO: Map formats so we can check swizzles and print nicely */
811 pandecode_log("swizzle");
812 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
813 pandecode_log_cont(",\n");
815 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
816 (format
.nr_channels
+ 1));
818 pandecode_log(".unk2 = ");
819 pandecode_log_decoded_flags(sfbd_unk2_info
, format
.unk2
);
820 pandecode_log_cont(",\n");
822 pandecode_prop("block = %s", pandecode_block_format(format
.block
));
824 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
827 pandecode_log("},\n");
831 pandecode_shared_memory(const struct mali_shared_memory
*desc
, bool is_compute
)
833 pandecode_prop("stack_shift = 0x%x", desc
->stack_shift
);
836 pandecode_prop("unk0 = 0x%x", desc
->unk0
);
838 if (desc
->shared_workgroup_count
!= 0x1F) {
839 pandecode_prop("shared_workgroup_count = %d", desc
->shared_workgroup_count
);
841 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
844 if (desc
->shared_unk1
|| desc
->shared_shift
) {
845 pandecode_prop("shared_unk1 = %X", desc
->shared_unk1
);
846 pandecode_prop("shared_shift = %X", desc
->shared_shift
);
849 pandecode_msg("XXX: shared memory configured in noncompute shader");
852 if (desc
->shared_zero
) {
853 pandecode_msg("XXX: shared memory zero tripped\n");
854 pandecode_prop("shared_zero = 0x%" PRIx32
, desc
->shared_zero
);
857 if (desc
->shared_memory
&& !is_compute
)
858 pandecode_msg("XXX: shared memory used in noncompute shader\n");
860 MEMORY_PROP(desc
, scratchpad
);
861 MEMORY_PROP(desc
, shared_memory
);
862 MEMORY_PROP(desc
, unknown1
);
864 if (desc
->scratchpad
) {
865 struct pandecode_mapped_memory
*smem
=
866 pandecode_find_mapped_gpu_mem_containing(desc
->scratchpad
);
868 pandecode_msg("scratchpad size %u\n", smem
->length
);
873 static struct pandecode_fbd
874 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
, unsigned gpu_id
)
876 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
877 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
879 struct pandecode_fbd info
= {
884 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
887 pandecode_log(".shared_memory = {\n");
889 pandecode_shared_memory(&s
->shared_memory
, false);
891 pandecode_log("},\n");
893 pandecode_sfbd_format(s
->format
);
895 info
.width
= s
->width
+ 1;
896 info
.height
= s
->height
+ 1;
898 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", info
.width
);
899 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", info
.height
);
901 MEMORY_PROP(s
, checksum
);
903 if (s
->checksum_stride
)
904 pandecode_prop("checksum_stride = %d", s
->checksum_stride
);
906 MEMORY_PROP(s
, framebuffer
);
907 pandecode_prop("stride = %d", s
->stride
);
909 /* Earlier in the actual commandstream -- right before width -- but we
910 * delay to flow nicer */
912 pandecode_log(".clear_flags = ");
913 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
914 pandecode_log_cont(",\n");
916 if (s
->depth_buffer
) {
917 MEMORY_PROP(s
, depth_buffer
);
918 pandecode_prop("depth_stride = %d", s
->depth_stride
);
921 if (s
->stencil_buffer
) {
922 MEMORY_PROP(s
, stencil_buffer
);
923 pandecode_prop("stencil_stride = %d", s
->stencil_stride
);
926 if (s
->depth_stride_zero
||
927 s
->stencil_stride_zero
||
928 s
->zero7
|| s
->zero8
) {
929 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
930 pandecode_prop("depth_stride_zero = 0x%x",
931 s
->depth_stride_zero
);
932 pandecode_prop("stencil_stride_zero = 0x%x",
933 s
->stencil_stride_zero
);
934 pandecode_prop("zero7 = 0x%" PRIx32
,
936 pandecode_prop("zero8 = 0x%" PRIx32
,
940 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
941 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
942 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
943 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
944 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
947 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
948 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
949 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
950 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
951 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
954 if (s
->clear_stencil
) {
955 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
958 const struct midgard_tiler_descriptor t
= s
->tiler
;
960 bool has_hierarchy
= !(gpu_id
== 0x0720 || gpu_id
== 0x0820 || gpu_id
== 0x0830);
961 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
, has_hierarchy
);
964 pandecode_log("};\n");
966 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
967 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
968 pandecode_prop("zero5 = 0x%" PRIx32
, s
->zero5
);
970 pandecode_log_cont(".zero3 = {");
972 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
973 pandecode_log_cont("%X, ", s
->zero3
[i
]);
975 pandecode_log_cont("},\n");
977 pandecode_log_cont(".zero6 = {");
979 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
980 pandecode_log_cont("%X, ", s
->zero6
[i
]);
982 pandecode_log_cont("},\n");
988 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
990 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
991 const struct mali_shared_memory
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
993 pandecode_log("struct mali_shared_memory shared_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
995 pandecode_shared_memory(s
, true);
997 pandecode_log("},\n");
1000 /* Extracts the number of components associated with a Mali format */
1003 pandecode_format_component_count(enum mali_format fmt
)
1005 /* Mask out the format class */
1006 unsigned top
= fmt
& 0b11100000;
1009 case MALI_FORMAT_SNORM
:
1010 case MALI_FORMAT_UINT
:
1011 case MALI_FORMAT_UNORM
:
1012 case MALI_FORMAT_SINT
:
1013 return ((fmt
>> 3) & 3) + 1;
1015 /* TODO: Validate */
1020 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
1023 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
1026 assert(MALI_CHANNEL_RED
== 0);
1028 for (unsigned c
= 0; c
< 4; ++c
) {
1029 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
1031 if (chan
<= MALI_CHANNEL_ALPHA
)
1032 mask
|= (1 << chan
);
1038 /* Validates that a (format, swizzle) pair is valid, in the sense that the
1039 * swizzle doesn't access any components that are undefined in the format.
1040 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
1044 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
1046 unsigned nr_comp
= pandecode_format_component_count(fmt
);
1047 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
1048 unsigned valid_mask
= (1 << nr_comp
) - 1;
1049 unsigned invalid_mask
= ~valid_mask
;
1051 if (access_mask
& invalid_mask
) {
1052 pandecode_msg("XXX: invalid components accessed\n");
1056 /* Check for the default non-swizzling swizzle so we can suppress
1057 * useless printing for the defaults */
1059 unsigned default_swizzles
[4] = {
1060 MALI_CHANNEL_RED
| (MALI_CHANNEL_ZERO
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
1061 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
1062 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ONE
<< 9),
1063 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ALPHA
<< 9)
1066 return (swizzle
== default_swizzles
[nr_comp
- 1]);
1069 /* Maps MALI_RGBA32F to rgba32f, etc */
1072 pandecode_format_short(enum mali_format fmt
, bool srgb
)
1074 /* We want a type-like format, so cut off the initial MALI_ */
1075 char *format
= pandecode_format(fmt
);
1076 format
+= strlen("MALI_");
1078 unsigned len
= strlen(format
);
1079 char *lower_format
= calloc(1, len
+ 1);
1081 for (unsigned i
= 0; i
< len
; ++i
)
1082 lower_format
[i
] = tolower(format
[i
]);
1084 /* Sanity check sRGB flag is applied to RGB, per the name */
1085 if (srgb
&& lower_format
[0] != 'r')
1086 pandecode_msg("XXX: sRGB applied to non-colour format\n");
1088 /* Just prefix with an s, so you get formats like srgba8_unorm */
1090 pandecode_log_cont("s");
1092 pandecode_log_cont("%s", lower_format
);
1097 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
1099 /* First, do some validation */
1100 bool trivial_swizzle
= pandecode_validate_format_swizzle(
1103 if (trivial_swizzle
)
1106 /* Next, print the swizzle */
1107 pandecode_log_cont(".");
1109 static const char components
[] = "rgba01";
1111 for (unsigned c
= 0; c
< 4; ++c
) {
1112 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
1114 if (chan
>= MALI_CHANNEL_RESERVED_0
) {
1115 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
1118 pandecode_log_cont("%c", components
[chan
]);
1123 pandecode_rt_format(struct mali_rt_format format
)
1125 pandecode_log(".format = {\n");
1128 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
1129 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
1130 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
1131 pandecode_prop("unk4 = 0x%" PRIx32
, format
.unk4
);
1133 pandecode_prop("block = %s", pandecode_block_format(format
.block
));
1135 /* TODO: Map formats so we can check swizzles and print nicely */
1136 pandecode_log("swizzle");
1137 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
1138 pandecode_log_cont(",\n");
1140 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
1141 (format
.nr_channels
+ 1));
1143 pandecode_log(".flags = ");
1144 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
1145 pandecode_log_cont(",\n");
1147 pandecode_prop("msaa = %s", pandecode_msaa_mode(format
.msaa
));
1149 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
1150 * which is a faster hardware-based alternative to the wallpaper method
1151 * to preserve framebuffer contents across frames. In practice, MFBD
1152 * preload is buggy on Midgard, and so this is a chicken bit. If this
1153 * bit isn't set, most likely something broke unrelated to preload */
1155 if (!format
.no_preload
) {
1156 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
1157 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
1161 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
1164 pandecode_log("},\n");
1168 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct mali_framebuffer
*fb
)
1170 pandecode_log("struct mali_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
1173 for (int i
= 0; i
< (fb
->rt_count_1
+ 1); i
++) {
1174 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct mali_render_target
);
1175 struct pandecode_mapped_memory
*mem
=
1176 pandecode_find_mapped_gpu_mem_containing(rt_va
);
1177 const struct mali_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
1179 pandecode_log("{\n");
1182 pandecode_rt_format(rt
->format
);
1184 if (rt
->format
.block
== MALI_BLOCK_AFBC
) {
1185 pandecode_log(".afbc = {\n");
1188 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
1189 pandecode_prop("metadata = %s", a
);
1192 pandecode_prop("stride = %d", rt
->afbc
.stride
);
1194 pandecode_log(".flags = ");
1195 pandecode_log_decoded_flags(afbc_fmt_flag_info
, rt
->afbc
.flags
);
1196 pandecode_log_cont(",\n");
1199 pandecode_log("},\n");
1200 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.flags
) {
1201 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1207 MEMORY_PROP(rt
, framebuffer
);
1208 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
1210 if (rt
->layer_stride
)
1211 pandecode_prop("layer_stride = %d", rt
->layer_stride
);
1213 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
1214 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
1215 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
1216 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
1217 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
1220 if (rt
->zero1
|| rt
->zero2
) {
1221 pandecode_msg("XXX: render target zeros tripped\n");
1222 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
1223 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
1227 pandecode_log("},\n");
1231 pandecode_log("};\n");
1234 static struct pandecode_fbd
1235 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
, bool is_compute
, bool is_bifrost
)
1237 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1238 const struct mali_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
1240 struct pandecode_fbd info
;
1242 if (is_bifrost
&& fb
->msaa
.sample_locations
) {
1243 /* The blob stores all possible sample locations in a single buffer
1244 * allocated on startup, and just switches the pointer when switching
1245 * MSAA state. For now, we just put the data into the cmdstream, but we
1246 * should do something like what the blob does with a real driver.
1248 * There seem to be 32 slots for sample locations, followed by another
1249 * 16. The second 16 is just the center location followed by 15 zeros
1250 * in all the cases I've identified (maybe shader vs. depth/color
1254 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->msaa
.sample_locations
);
1256 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->msaa
.sample_locations
);
1258 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
1261 for (int i
= 0; i
< 32 + 16; i
++) {
1262 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
1266 pandecode_log("};\n");
1269 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1273 pandecode_log(".msaa = {\n");
1276 if (fb
->msaa
.sample_locations
)
1277 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
1279 pandecode_msg("XXX: sample_locations missing\n");
1281 if (fb
->msaa
.zero1
|| fb
->msaa
.zero2
|| fb
->msaa
.zero4
) {
1282 pandecode_msg("XXX: multisampling zero tripped\n");
1283 pandecode_prop("zero1 = %" PRIx64
, fb
->msaa
.zero1
);
1284 pandecode_prop("zero2 = %" PRIx64
, fb
->msaa
.zero2
);
1285 pandecode_prop("zero4 = %" PRIx64
, fb
->msaa
.zero4
);
1289 pandecode_log("},\n");
1291 pandecode_log(".shared_memory = {\n");
1293 pandecode_shared_memory(&fb
->shared_memory
, is_compute
);
1295 pandecode_log("},\n");
1298 info
.width
= fb
->width1
+ 1;
1299 info
.height
= fb
->height1
+ 1;
1300 info
.rt_count
= fb
->rt_count_1
+ 1;
1302 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
1303 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
1304 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
1305 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
1307 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
1308 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
1309 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
1310 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
1312 pandecode_log(".mfbd_flags = ");
1313 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
1314 pandecode_log_cont(",\n");
1316 if (fb
->clear_stencil
)
1317 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
1319 if (fb
->clear_depth
)
1320 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
1324 pandecode_bifrost_tiler_descriptor(fb
);
1326 const struct midgard_tiler_descriptor t
= fb
->tiler
;
1327 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
, true);
1330 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1332 if (fb
->zero3
|| fb
->zero4
) {
1333 pandecode_msg("XXX: framebuffer zeros tripped\n");
1334 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
1335 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
1339 pandecode_log("};\n");
1341 gpu_va
+= sizeof(struct mali_framebuffer
);
1343 info
.has_extra
= (fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
;
1345 if (info
.has_extra
) {
1346 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1347 const struct mali_framebuffer_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
1349 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1352 MEMORY_PROP(fbx
, checksum
);
1354 if (fbx
->checksum_stride
)
1355 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
1357 pandecode_log(".flags_hi = ");
1358 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info
, fbx
->flags_hi
);
1359 pandecode_log_cont(",\n");
1361 pandecode_log(".flags_lo = ");
1362 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info
, fbx
->flags_lo
);
1363 pandecode_log_cont(",\n");
1365 pandecode_prop("zs_block = %s", pandecode_block_format(fbx
->zs_block
));
1366 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx
->zs_samples
+ 1);
1368 if (fbx
->zs_block
== MALI_BLOCK_AFBC
) {
1369 pandecode_log(".ds_afbc = {\n");
1372 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
1373 pandecode_prop("depth_stencil_afbc_stride = %d",
1374 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
1375 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
1377 pandecode_log(".flags = ");
1378 pandecode_log_decoded_flags(afbc_fmt_flag_info
, fbx
->ds_afbc
.flags
);
1379 pandecode_log_cont(",\n");
1381 if (fbx
->ds_afbc
.padding
) {
1382 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1383 pandecode_prop("padding = 0x%" PRIx64
, fbx
->ds_afbc
.padding
);
1387 pandecode_log("},\n");
1389 pandecode_log(".ds_linear = {\n");
1392 if (fbx
->ds_linear
.depth
) {
1393 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
1394 pandecode_prop("depth_stride = %d",
1395 fbx
->ds_linear
.depth_stride
);
1396 pandecode_prop("depth_layer_stride = %d",
1397 fbx
->ds_linear
.depth_layer_stride
);
1398 } else if (fbx
->ds_linear
.depth_stride
|| fbx
->ds_linear
.depth_layer_stride
) {
1399 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx
->ds_linear
.depth_stride
, fbx
->ds_linear
.depth_layer_stride
);
1402 if (fbx
->ds_linear
.stencil
) {
1403 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1404 pandecode_prop("stencil_stride = %d",
1405 fbx
->ds_linear
.stencil_stride
);
1406 pandecode_prop("stencil_layer_stride = %d",
1407 fbx
->ds_linear
.stencil_layer_stride
);
1408 } else if (fbx
->ds_linear
.stencil_stride
|| fbx
->ds_linear
.stencil_layer_stride
) {
1409 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx
->ds_linear
.stencil_stride
, fbx
->ds_linear
.stencil_layer_stride
);
1412 if (fbx
->ds_linear
.depth_stride_zero
||
1413 fbx
->ds_linear
.stencil_stride_zero
) {
1414 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1415 pandecode_prop("depth_stride_zero = 0x%x",
1416 fbx
->ds_linear
.depth_stride_zero
);
1417 pandecode_prop("stencil_stride_zero = 0x%x",
1418 fbx
->ds_linear
.stencil_stride_zero
);
1422 pandecode_log("},\n");
1425 if (fbx
->clear_color_1
| fbx
->clear_color_2
) {
1426 pandecode_prop("clear_color_1 = 0x%" PRIx32
, fbx
->clear_color_1
);
1427 pandecode_prop("clear_color_2 = 0x%" PRIx32
, fbx
->clear_color_2
);
1431 pandecode_msg("XXX: fb_extra zeros tripped\n");
1432 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1436 pandecode_log("};\n");
1438 gpu_va
+= sizeof(struct mali_framebuffer_extra
);
1442 pandecode_render_target(gpu_va
, job_no
, fb
);
1447 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1451 pandecode_padded_vertices(unsigned shift
, unsigned k
)
1453 unsigned odd
= 2*k
+ 1;
1454 unsigned pot
= 1 << shift
;
1455 pandecode_msg("padded_num_vertices = %d\n", odd
* pot
);
1458 /* Given a magic divisor, recover what we were trying to divide by.
1460 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1461 * 0 <= m < 2^N, for N bits in m.
1463 * Let q represent the number we would like to divide by.
1465 * By definition of a magic divisor for N-bit unsigned integers (a number you
1466 * multiply by to magically get division), m is a number such that:
1468 * (m * x) & (2^N - 1) = floor(x/q).
1469 * for all x on Z where 0 <= x < 2^N
1471 * Ignore the case where any of the above values equals zero; it is irrelevant
1472 * for our purposes (instanced arrays).
1474 * Choose x = q. Then:
1476 * (m * x) & (2^N - 1) = floor(x/q).
1477 * (m * q) & (2^N - 1) = floor(q/q).
1479 * floor(q/q) = floor(1) = 1, therefore:
1481 * (m * q) & (2^N - 1) = 1
1483 * Recall the identity that the bitwise AND of one less than a power-of-two
1484 * equals the modulo with that power of two, i.e. for all x:
1486 * x & (2^N - 1) = x % N
1492 * By definition, a modular multiplicative inverse of a number m is the number
1493 * q such that with respect to a modulos M:
1497 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1502 pandecode_magic_divisor(uint32_t magic
, unsigned shift
, unsigned orig_divisor
, unsigned extra
)
1505 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1506 * shift) the most lame way possible... just repeatedly add.
1507 * Asymptoptically slow but nobody cares in practice, unless you have
1508 * massive numbers of vertices or high divisors. */
1510 unsigned inverse
= 0;
1512 /* Magic implicitly has the highest bit set */
1515 /* Depending on rounding direction */
1520 uint32_t product
= magic
* inverse
;
1532 pandecode_msg("dividing by %d (maybe off by two)\n", inverse
);
1534 /* Recall we're supposed to divide by (gl_level_divisor *
1535 * padded_num_vertices) */
1537 unsigned padded_num_vertices
= inverse
/ orig_divisor
;
1539 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices
);
1544 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1545 mali_ptr addr
, int job_no
, char *suffix
,
1546 int count
, bool varying
, enum mali_job_type job_type
)
1548 char *prefix
= varying
? "varying" : "attribute";
1552 pandecode_msg("warn: No %s records\n", prefix
);
1556 union mali_attr
*attr
= pandecode_fetch_gpu_mem(mem
, addr
, sizeof(union mali_attr
) * count
);
1558 for (int i
= 0; i
< count
; ++i
) {
1559 /* First, check for special records */
1560 if (attr
[i
].elements
< MALI_RECORD_SPECIAL
) {
1562 pandecode_msg("XXX: tripped size=%d\n", attr
[i
].size
);
1564 if (attr
[i
].stride
) {
1565 /* gl_InstanceID passes a magic divisor in the
1566 * stride field to divide by the padded vertex
1567 * count. No other records should do so, so
1568 * stride should otherwise be zero. Note that
1569 * stride in the usual attribute sense doesn't
1570 * apply to special records. */
1572 bool has_divisor
= attr
[i
].elements
== MALI_ATTR_INSTANCEID
;
1574 pandecode_log_cont("/* %smagic divisor = %X */ ",
1575 has_divisor
? "" : "XXX: ", attr
[i
].stride
);
1578 if (attr
[i
].shift
|| attr
[i
].extra_flags
) {
1579 /* Attributes use these fields for
1580 * instancing/padding/etc type issues, but
1583 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1584 varying
? "XXX: " : "",
1585 attr
[i
].shift
, attr
[i
].extra_flags
);
1588 /* Print the special record name */
1589 bool attribute
= false;
1590 pandecode_log("%s_%d = %s;\n", prefix
, i
, pandecode_special_record(attr
[i
].elements
, &attribute
));
1593 if (attribute
== varying
)
1594 pandecode_msg("XXX: mismatched special record\n");
1599 enum mali_attr_mode mode
= attr
[i
].elements
& 7;
1601 if (mode
== MALI_ATTR_UNUSED
)
1602 pandecode_msg("XXX: unused attribute record\n");
1604 /* For non-linear records, we need to print the type of record */
1605 if (mode
!= MALI_ATTR_LINEAR
)
1606 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode
));
1608 /* Print the name to link with attr_meta */
1609 pandecode_log_cont("%s_%d", prefix
, i
);
1611 /* Print the stride and size */
1612 pandecode_log_cont("<%u>[%u]", attr
[i
].stride
, attr
[i
].size
);
1614 /* TODO: Sanity check the quotient itself. It must be equal to
1615 * (or be greater than, if the driver added padding) the padded
1618 /* Finally, print the pointer */
1619 mali_ptr raw_elements
= attr
[i
].elements
& ~7;
1620 char *a
= pointer_as_memory_reference(raw_elements
);
1621 pandecode_log_cont(" = (%s);\n", a
);
1624 /* Check the pointer */
1625 pandecode_validate_buffer(raw_elements
, attr
[i
].size
);
1627 /* shift/extra_flags exist only for instanced */
1628 if (attr
[i
].shift
| attr
[i
].extra_flags
) {
1629 /* These are set to random values by the blob for
1630 * varyings, most likely a symptom of uninitialized
1631 * memory where the hardware masked the bug. As such we
1632 * put this at a warning, not an error. */
1634 if (mode
== MALI_ATTR_LINEAR
)
1635 pandecode_msg("warn: instancing fields set for linear\n");
1637 pandecode_prop("shift = %d", attr
[i
].shift
);
1638 pandecode_prop("extra_flags = %d", attr
[i
].extra_flags
);
1641 /* Decode further where possible */
1643 if (mode
== MALI_ATTR_MODULO
) {
1644 pandecode_padded_vertices(
1646 attr
[i
].extra_flags
);
1649 if (mode
== MALI_ATTR_NPOT_DIVIDE
) {
1651 pandecode_log("{\n");
1653 pandecode_prop("unk = 0x%x", attr
[i
].unk
);
1654 pandecode_prop("magic_divisor = 0x%08x", attr
[i
].magic_divisor
);
1655 if (attr
[i
].zero
!= 0)
1656 pandecode_prop("XXX: zero tripped (0x%x)\n", attr
[i
].zero
);
1657 pandecode_prop("divisor = %d", attr
[i
].divisor
);
1658 pandecode_magic_divisor(attr
[i
].magic_divisor
, attr
[i
- 1].shift
, attr
[i
].divisor
, attr
[i
- 1].extra_flags
);
1660 pandecode_log("}, \n");
1665 pandecode_log("\n");
1669 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1671 /* TODO: Decode flags */
1672 mali_ptr shader_ptr
= ptr
& ~15;
1674 char *a
= pointer_as_memory_reference(shader_ptr
);
1675 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1682 pandecode_stencil(const char *name
, const struct mali_stencil_test
*stencil
)
1684 unsigned any_nonzero
=
1685 stencil
->ref
| stencil
->mask
| stencil
->func
|
1686 stencil
->sfail
| stencil
->dpfail
| stencil
->dppass
;
1688 if (any_nonzero
== 0)
1691 const char *func
= pandecode_func(stencil
->func
);
1692 const char *sfail
= pandecode_stencil_op(stencil
->sfail
);
1693 const char *dpfail
= pandecode_stencil_op(stencil
->dpfail
);
1694 const char *dppass
= pandecode_stencil_op(stencil
->dppass
);
1697 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil
->zero
);
1699 pandecode_log(".stencil_%s = {\n", name
);
1701 pandecode_prop("ref = %d", stencil
->ref
);
1702 pandecode_prop("mask = 0x%02X", stencil
->mask
);
1703 pandecode_prop("func = %s", func
);
1704 pandecode_prop("sfail = %s", sfail
);
1705 pandecode_prop("dpfail = %s", dpfail
);
1706 pandecode_prop("dppass = %s", dppass
);
1708 pandecode_log("},\n");
1712 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1715 pandecode_msg("XXX: blend zero tripped: %X\n", blend
->zero1
);
1717 pandecode_log(".equation = {\n");
1720 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1721 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1723 pandecode_log(".color_mask = ");
1724 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1725 pandecode_log_cont(",\n");
1728 pandecode_log("},\n");
1731 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1734 decode_bifrost_constant(u16 constant
)
1736 float lo
= (float) (constant
& 0xFF);
1737 float hi
= (float) (constant
>> 8);
1739 return (hi
/ 255.0) + (lo
/ 65535.0);
1743 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1745 struct bifrost_blend_rt
*b
=
1746 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1748 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1751 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1752 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1753 b
->constant
, decode_bifrost_constant(b
->constant
));
1755 /* TODO figure out blend shader enable bit */
1756 pandecode_blend_equation(&b
->equation
);
1758 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1759 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1761 pandecode_log(".format = ");
1762 pandecode_format_short(b
->format
, false);
1763 pandecode_swizzle(b
->swizzle
, b
->format
);
1764 pandecode_log_cont(",\n");
1766 pandecode_prop("swizzle = 0x%" PRIx32
, b
->swizzle
);
1767 pandecode_prop("format = 0x%" PRIx32
, b
->format
);
1770 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1771 pandecode_prop("zero1 = 0x%" PRIx32
, b
->zero1
);
1774 pandecode_log(".shader_type = ");
1775 switch(b
->shader_type
) {
1776 case BIFROST_BLEND_F16
:
1777 pandecode_log_cont("BIFROST_BLEND_F16");
1779 case BIFROST_BLEND_F32
:
1780 pandecode_log_cont("BIFROST_BLEND_F32");
1782 case BIFROST_BLEND_I32
:
1783 pandecode_log_cont("BIFROST_BLEND_I32");
1785 case BIFROST_BLEND_U32
:
1786 pandecode_log_cont("BIFROST_BLEND_U32");
1788 case BIFROST_BLEND_I16
:
1789 pandecode_log_cont("BIFROST_BLEND_I16");
1791 case BIFROST_BLEND_U16
:
1792 pandecode_log_cont("BIFROST_BLEND_U16");
1795 pandecode_log_cont(",\n");
1798 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1799 pandecode_prop("zero2 = 0x%" PRIx32
, b
->zero2
);
1802 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1805 pandecode_log("},\n");
1811 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1813 /* constant/equation is in a union */
1817 pandecode_log(".blend = {\n");
1821 pandecode_shader_address("shader", blend
->shader
);
1823 pandecode_blend_equation(&blend
->equation
);
1824 pandecode_prop("constant = %f", blend
->constant
);
1828 pandecode_log("},\n");
1830 /* Return blend shader to disassemble if present */
1831 return is_shader
? (blend
->shader
& ~0xF) : 0;
1835 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1837 struct midgard_blend_rt
*b
=
1838 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1840 /* Flags determine presence of blend shader */
1841 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1843 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1846 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1848 union midgard_blend blend
= b
->blend
;
1849 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1852 pandecode_log("};\n");
1857 /* Attributes and varyings have descriptor records, which contain information
1858 * about their format and ordering with the attribute/varying buffers. We'll
1859 * want to validate that the combinations specified are self-consistent.
1863 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1866 char *prefix
= varying
? "varying" : "attribute";
1867 unsigned max_index
= 0;
1868 snprintf(base
, sizeof(base
), "%s_meta", prefix
);
1870 struct mali_attr_meta
*attr_meta
;
1871 mali_ptr p
= varying
? v
->varying_meta
: v
->attribute_meta
;
1873 struct pandecode_mapped_memory
*attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
);
1875 for (int i
= 0; i
< count
; ++i
, p
+= sizeof(struct mali_attr_meta
)) {
1876 attr_meta
= pandecode_fetch_gpu_mem(attr_mem
, p
,
1879 /* If the record is discard, it should be zero for everything else */
1881 if (attr_meta
->format
== MALI_VARYING_DISCARD
) {
1884 attr_meta
->unknown1
|
1885 attr_meta
->unknown3
|
1886 attr_meta
->src_offset
;
1889 pandecode_msg("XXX: expected empty record for varying discard\n");
1891 /* We want to look for a literal 0000 swizzle -- this
1892 * is not encoded with all zeroes, however */
1894 enum mali_channel z
= MALI_CHANNEL_ZERO
;
1895 unsigned zero_swizzle
= z
| (z
<< 3) | (z
<< 6) | (z
<< 9);
1896 bool good_swizzle
= attr_meta
->swizzle
== zero_swizzle
;
1899 pandecode_msg("XXX: expected zero swizzle for discard\n");
1902 pandecode_msg("XXX: cannot discard attribute\n");
1904 /* If we're all good, omit the record */
1905 if (!zero
&& varying
&& good_swizzle
) {
1906 pandecode_log("/* discarded varying */\n");
1911 if (attr_meta
->index
> max_index
)
1912 max_index
= attr_meta
->index
;
1914 if (attr_meta
->unknown1
!= 0x2) {
1915 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1916 pandecode_prop("unknown1 = 0x%" PRIx64
, (u64
) attr_meta
->unknown1
);
1919 if (attr_meta
->unknown3
) {
1920 pandecode_msg("XXX: unexpected unknown3 set\n");
1921 pandecode_prop("unknown3 = 0x%" PRIx64
, (u64
) attr_meta
->unknown3
);
1924 pandecode_format_short(attr_meta
->format
, false);
1925 pandecode_log_cont(" %s_%u", prefix
, attr_meta
->index
);
1927 if (attr_meta
->src_offset
)
1928 pandecode_log_cont("[%u]", attr_meta
->src_offset
);
1930 pandecode_swizzle(attr_meta
->swizzle
, attr_meta
->format
);
1932 pandecode_log_cont(";\n");
1935 pandecode_log("\n");
1937 return count
? (max_index
+ 1) : 0;
1940 /* return bits [lo, hi) of word */
1942 bits(u32 word
, u32 lo
, u32 hi
)
1945 return word
; // avoid undefined behavior with the shift
1947 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1951 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool graphics
)
1953 pandecode_log(".prefix = {\n");
1956 /* Decode invocation_count. See the comment before the definition of
1957 * invocation_count for an explanation.
1960 unsigned size_y_shift
= bits(p
->invocation_shifts
, 0, 5);
1961 unsigned size_z_shift
= bits(p
->invocation_shifts
, 5, 10);
1962 unsigned workgroups_x_shift
= bits(p
->invocation_shifts
, 10, 16);
1963 unsigned workgroups_y_shift
= bits(p
->invocation_shifts
, 16, 22);
1964 unsigned workgroups_z_shift
= bits(p
->invocation_shifts
, 22, 28);
1965 unsigned workgroups_x_shift_2
= bits(p
->invocation_shifts
, 28, 32);
1967 unsigned size_x
= bits(p
->invocation_count
, 0, size_y_shift
) + 1;
1968 unsigned size_y
= bits(p
->invocation_count
, size_y_shift
, size_z_shift
) + 1;
1969 unsigned size_z
= bits(p
->invocation_count
, size_z_shift
, workgroups_x_shift
) + 1;
1971 unsigned groups_x
= bits(p
->invocation_count
, workgroups_x_shift
, workgroups_y_shift
) + 1;
1972 unsigned groups_y
= bits(p
->invocation_count
, workgroups_y_shift
, workgroups_z_shift
) + 1;
1973 unsigned groups_z
= bits(p
->invocation_count
, workgroups_z_shift
, 32) + 1;
1975 /* Even though we have this decoded, we want to ensure that the
1976 * representation is "unique" so we don't lose anything by printing only
1977 * the final result. More specifically, we need to check that we were
1978 * passed something in canonical form, since the definition per the
1979 * hardware is inherently not unique. How? Well, take the resulting
1980 * decode and pack it ourselves! If it is bit exact with what we
1981 * decoded, we're good to go. */
1983 struct mali_vertex_tiler_prefix ref
;
1984 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, graphics
);
1987 (p
->invocation_count
== ref
.invocation_count
) &&
1988 (p
->invocation_shifts
== ref
.invocation_shifts
);
1991 pandecode_msg("XXX: non-canonical workgroups packing\n");
1992 pandecode_msg("expected: %X, %X",
1993 ref
.invocation_count
,
1994 ref
.invocation_shifts
);
1996 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1997 pandecode_prop("size_y_shift = %d", size_y_shift
);
1998 pandecode_prop("size_z_shift = %d", size_z_shift
);
1999 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift
);
2000 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift
);
2001 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift
);
2002 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2
);
2005 /* Regardless, print the decode */
2006 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
2007 size_x
, size_y
, size_z
,
2008 groups_x
, groups_y
, groups_z
);
2011 if (p
->unknown_draw
)
2012 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
2014 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
2016 if (p
->draw_mode
!= MALI_DRAW_NONE
)
2017 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p
->draw_mode
));
2019 /* Index count only exists for tiler jobs anyway */
2022 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
2025 unsigned index_raw_size
= (p
->unknown_draw
& MALI_DRAW_INDEXED_SIZE
);
2026 index_raw_size
>>= MALI_DRAW_INDEXED_SHIFT
;
2028 /* Validate an index buffer is present if we need one. TODO: verify
2029 * relationship between invocation_count and index_count */
2032 unsigned count
= p
->index_count
;
2035 unsigned size
= (index_raw_size
== 0x3) ? 4 : index_raw_size
;
2037 /* Ensure we got a size, and if so, validate the index buffer
2038 * is large enough to hold a full set of indices of the given
2041 if (!index_raw_size
)
2042 pandecode_msg("XXX: index size missing\n");
2044 pandecode_validate_buffer(p
->indices
, count
* size
);
2045 } else if (index_raw_size
)
2046 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size
);
2048 if (p
->offset_bias_correction
)
2049 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
2051 /* TODO: Figure out what this is. It's not zero */
2052 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
2055 pandecode_log("},\n");
2059 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
2061 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
2062 uint64_t *PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
2064 for (int i
= 0; i
< ubufs_count
; i
++) {
2065 unsigned size
= (ubufs
[i
] & ((1 << 10) - 1)) * 16;
2066 mali_ptr addr
= (ubufs
[i
] >> 10) << 2;
2068 pandecode_validate_buffer(addr
, size
);
2070 char *ptr
= pointer_as_memory_reference(addr
);
2071 pandecode_log("ubuf_%d[%u] = %s;\n", i
, size
, ptr
);
2075 pandecode_log("\n");
2079 pandecode_uniforms(mali_ptr uniforms
, unsigned uniform_count
)
2081 pandecode_validate_buffer(uniforms
, uniform_count
* 16);
2083 char *ptr
= pointer_as_memory_reference(uniforms
);
2084 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count
, ptr
);
2089 shader_type_for_job(unsigned type
)
2092 case MALI_JOB_TYPE_VERTEX
: return "VERTEX";
2093 case MALI_JOB_TYPE_TILER
: return "FRAGMENT";
2094 case MALI_JOB_TYPE_COMPUTE
: return "COMPUTE";
2100 static unsigned shader_id
= 0;
2102 static struct midgard_disasm_stats
2103 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
2104 bool is_bifrost
, unsigned gpu_id
)
2106 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
2107 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
2109 /* Compute maximum possible size */
2110 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
2112 /* Print some boilerplate to clearly denote the assembly (which doesn't
2113 * obey indentation rules), and actually do the disassembly! */
2115 pandecode_log_cont("\n\n");
2117 struct midgard_disasm_stats stats
;
2120 disassemble_bifrost(pandecode_dump_stream
, code
, sz
, true);
2122 /* TODO: Extend stats to Bifrost */
2123 stats
.texture_count
= -128;
2124 stats
.sampler_count
= -128;
2125 stats
.attribute_count
= -128;
2126 stats
.varying_count
= -128;
2127 stats
.uniform_count
= -128;
2128 stats
.uniform_buffer_count
= -128;
2129 stats
.work_count
= -128;
2131 stats
.instruction_count
= 0;
2132 stats
.bundle_count
= 0;
2133 stats
.quadword_count
= 0;
2134 stats
.helper_invocations
= false;
2136 stats
= disassemble_midgard(pandecode_dump_stream
,
2138 type
== MALI_JOB_TYPE_TILER
?
2139 MESA_SHADER_FRAGMENT
: MESA_SHADER_VERTEX
);
2142 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
2143 * driver-internal purposes with the blob and interfere */
2145 bool should_shaderdb
= type
!= MALI_JOB_TYPE_COMPUTE
;
2147 if (should_shaderdb
) {
2148 unsigned nr_threads
=
2149 (stats
.work_count
<= 4) ? 4 :
2150 (stats
.work_count
<= 8) ? 2 :
2153 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
2154 "%u inst, %u bundles, %u quadwords, "
2155 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
2157 shader_type_for_job(type
),
2158 stats
.instruction_count
, stats
.bundle_count
, stats
.quadword_count
,
2159 stats
.work_count
, nr_threads
);
2167 pandecode_texture_payload(mali_ptr payload
,
2168 enum mali_texture_type type
,
2169 enum mali_texture_layout layout
,
2173 uint16_t array_size
,
2174 struct pandecode_mapped_memory
*tmem
)
2176 pandecode_log(".payload = {\n");
2179 /* A bunch of bitmap pointers follow.
2180 * We work out the correct number,
2181 * based on the mipmap/cubemap
2182 * properties, but dump extra
2183 * possibilities to futureproof */
2185 int bitmap_count
= levels
+ 1;
2187 /* Miptree for each face */
2188 if (type
== MALI_TEX_CUBE
)
2191 /* Array of layers */
2192 bitmap_count
*= (depth
+ 1);
2194 /* Array of textures */
2195 bitmap_count
*= (array_size
+ 1);
2197 /* Stride for each element */
2201 mali_ptr
*pointers_and_strides
= pandecode_fetch_gpu_mem(tmem
,
2202 payload
, sizeof(mali_ptr
) * bitmap_count
);
2203 for (int i
= 0; i
< bitmap_count
; ++i
) {
2204 /* How we dump depends if this is a stride or a pointer */
2206 if (manual_stride
&& (i
& 1)) {
2207 /* signed 32-bit snuck in as a 64-bit pointer */
2208 uint64_t stride_set
= pointers_and_strides
[i
];
2209 uint32_t clamped_stride
= stride_set
;
2210 int32_t stride
= clamped_stride
;
2211 assert(stride_set
== clamped_stride
);
2212 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
2214 char *a
= pointer_as_memory_reference(pointers_and_strides
[i
]);
2215 pandecode_log("%s, \n", a
);
2221 pandecode_log("},\n");
2225 pandecode_texture(mali_ptr u
,
2226 struct pandecode_mapped_memory
*tmem
,
2227 unsigned job_no
, unsigned tex
)
2229 struct mali_texture_descriptor
*PANDECODE_PTR_VAR(t
, tmem
, u
);
2231 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64
"_%d_%d = {\n", u
, job_no
, tex
);
2234 pandecode_prop("width = %" PRId32
, t
->width
);
2235 pandecode_prop("height = %" PRId32
, t
->height
);
2236 pandecode_prop("depth = %" PRId32
, t
->depth
);
2237 pandecode_prop("array_size = %" PRId32
, t
->array_size
);
2239 pandecode_log("\n");
2240 pandecode_prop("f.swizzle = 0x%" PRIx32
, t
->format
.swizzle
);
2241 pandecode_prop("f.format = 0x%" PRIx32
, t
->format
.format
);
2242 pandecode_prop("f.srgb = 0x%" PRIx32
, t
->format
.srgb
);
2243 pandecode_prop("f.unknown1 = 0x%" PRIx32
, t
->format
.unknown1
);
2244 pandecode_prop("f.type = %" PRId32
, t
->format
.type
);
2245 pandecode_prop("f.layout = %" PRId32
, t
->format
.layout
);
2246 pandecode_prop("f.unknown2 = 0x%" PRIx32
, t
->format
.unknown2
);
2247 pandecode_prop("f.manual_stride = %" PRId32
, t
->format
.manual_stride
);
2248 pandecode_prop("f.zero = 0x%" PRIx32
, t
->format
.zero
);
2249 pandecode_log("\n");
2251 pandecode_prop("unknown3 = 0x%" PRIx32
, t
->unknown3
);
2252 pandecode_prop("unknown3A = 0x%" PRIx32
, t
->unknown3A
);
2253 pandecode_prop("levels = %" PRId32
, t
->levels
);
2254 pandecode_prop("swizzle = 0x%" PRIx32
, t
->swizzle
);
2255 pandecode_prop("swizzle_zero = 0x%" PRIx32
, t
->swizzle_zero
);
2257 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
2258 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
2259 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
2260 pandecode_log("\n");
2262 struct mali_texture_format f
= t
->format
;
2264 /* See the definiton of enum mali_texture_type */
2266 bool is_cube
= f
.type
== MALI_TEX_CUBE
;
2267 unsigned dimension
= is_cube
? 2 : f
.type
;
2269 pandecode_make_indent();
2271 /* Print the layout. Default is linear; a modifier can denote AFBC or
2272 * u-interleaved/tiled modes */
2274 if (f
.layout
== MALI_TEXTURE_AFBC
)
2275 pandecode_log_cont("afbc");
2276 else if (f
.layout
== MALI_TEXTURE_TILED
)
2277 pandecode_log_cont("tiled");
2278 else if (f
.layout
== MALI_TEXTURE_LINEAR
)
2279 pandecode_log_cont("linear");
2281 pandecode_msg("XXX: invalid texture layout 0x%X\n", f
.layout
);
2283 pandecode_swizzle(t
->swizzle
, f
.format
);
2284 pandecode_log_cont(" ");
2286 /* Distinguish cube/2D with modifier */
2289 pandecode_log_cont("cube ");
2291 pandecode_format_short(f
.format
, f
.srgb
);
2292 pandecode_swizzle(f
.swizzle
, f
.format
);
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. Depth used for MSAA. */
2299 if (t
->height
&& dimension
< 2)
2300 pandecode_msg("XXX: nonzero height for <2D texture\n");
2302 /* Print only the dimensions that are actually there */
2304 pandecode_log_cont(": %d", t
->width
+ 1);
2306 if (t
->height
|| t
->depth
)
2307 pandecode_log_cont("x%u", t
->height
+ 1);
2310 pandecode_log_cont("x%u", t
->depth
+ 1);
2313 pandecode_log_cont("[%u]", t
->array_size
+ 1);
2316 pandecode_log_cont(" mip %u", t
->levels
);
2318 pandecode_log_cont("\n");
2320 if (f
.unknown1
| f
.zero
) {
2321 pandecode_msg("XXX: texture format zero tripped\n");
2322 pandecode_prop("unknown1 = %" PRId32
, f
.unknown1
);
2323 pandecode_prop("zero = %" PRId32
, f
.zero
);
2327 pandecode_msg("XXX: expected unknown texture bit set\n");
2328 pandecode_prop("unknown2 = %" PRId32
, f
.unknown2
);
2331 if (t
->swizzle_zero
) {
2332 pandecode_msg("XXX: swizzle zero tripped\n");
2333 pandecode_prop("swizzle_zero = %d", t
->swizzle_zero
);
2336 if (t
->unknown3
| t
->unknown3A
| t
->unknown5
| t
->unknown6
| t
->unknown7
) {
2337 pandecode_msg("XXX: texture zero tripped\n");
2338 pandecode_prop("unknown3 = %" PRId16
, t
->unknown3
);
2339 pandecode_prop("unknown3A = %" PRId8
, t
->unknown3A
);
2340 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
2341 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
2342 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
2345 pandecode_texture_payload(u
+ sizeof(*t
), f
.type
, f
.layout
, f
.manual_stride
, t
->levels
, t
->depth
, t
->array_size
, tmem
);
2348 pandecode_log("};\n");
2352 pandecode_bifrost_texture(
2353 const struct bifrost_texture_descriptor
*t
,
2357 pandecode_log("struct bifrost_texture_descriptor texture_descriptor_%d_%d = {\n", job_no
, tex
);
2360 pandecode_prop("format_unk = 0x%" PRIx32
, t
->format_unk
);
2361 pandecode_prop("type = %" PRId32
, t
->type
);
2364 pandecode_msg("XXX: zero tripped\n");
2365 pandecode_prop("zero = 0x%" PRIx32
, t
->zero
);
2368 pandecode_prop("format_swizzle = 0x%" PRIx32
, t
->format_swizzle
);
2369 pandecode_prop("format = 0x%" PRIx32
, t
->format
);
2370 pandecode_prop("srgb = 0x%" PRIx32
, t
->srgb
);
2371 pandecode_prop("format_unk3 = 0x%" PRIx32
, t
->format_unk3
);
2372 pandecode_prop("width = %" PRId32
, t
->width
);
2373 pandecode_prop("height = %" PRId32
, t
->height
);
2374 pandecode_prop("swizzle = 0x%" PRIx32
, t
->swizzle
);
2375 pandecode_prop("levels = %" PRId32
, t
->levels
);
2376 pandecode_prop("unk1 = 0x%" PRIx32
, t
->unk1
);
2377 pandecode_prop("levels_unk = %" PRId32
, t
->levels_unk
);
2378 pandecode_prop("level_2 = %" PRId32
, t
->level_2
);
2379 pandecode_prop("payload = 0x%" PRIx64
, t
->payload
);
2380 pandecode_prop("array_size = %" PRId32
, t
->array_size
);
2381 pandecode_prop("unk4 = 0x%" PRIx32
, t
->unk4
);
2382 pandecode_prop("depth = %" PRId32
, t
->depth
);
2383 pandecode_prop("unk5 = 0x%" PRIx32
, t
->unk5
);
2384 pandecode_log("\n");
2386 /* See the definiton of enum mali_texture_type */
2388 bool is_cube
= t
->type
== MALI_TEX_CUBE
;
2389 unsigned dimension
= is_cube
? 2 : t
->type
;
2391 /* Print the layout. Default is linear; a modifier can denote AFBC or
2392 * u-interleaved/tiled modes */
2394 if (t
->layout
== MALI_TEXTURE_AFBC
)
2395 pandecode_log_cont("afbc");
2396 else if (t
->layout
== MALI_TEXTURE_TILED
)
2397 pandecode_log_cont("tiled");
2398 else if (t
->layout
== MALI_TEXTURE_LINEAR
)
2399 pandecode_log_cont("linear");
2401 pandecode_msg("XXX: invalid texture layout 0x%X\n", t
->layout
);
2403 pandecode_swizzle(t
->swizzle
, t
->format
);
2404 pandecode_log_cont(" ");
2406 /* Distinguish cube/2D with modifier */
2409 pandecode_log_cont("cube ");
2411 pandecode_format_short(t
->format
, t
->srgb
);
2413 /* All four width/height/depth/array_size dimensions are present
2414 * regardless of the type of texture, but it is an error to have
2415 * non-zero dimensions for unused dimensions. Verify this. array_size
2416 * can always be set, as can width. */
2418 if (t
->height
&& dimension
< 2)
2419 pandecode_msg("XXX: nonzero height for <2D texture\n");
2421 if (t
->depth
&& dimension
< 3)
2422 pandecode_msg("XXX: nonzero depth for <2D texture\n");
2424 /* Print only the dimensions that are actually there */
2426 pandecode_log_cont(": %d", t
->width
+ 1);
2429 pandecode_log_cont("x%u", t
->height
+ 1);
2432 pandecode_log_cont("x%u", t
->depth
+ 1);
2435 pandecode_log_cont("[%u]", t
->array_size
+ 1);
2438 pandecode_log_cont(" mip %u", t
->levels
);
2440 pandecode_log_cont("\n");
2442 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(t
->payload
);
2444 pandecode_texture_payload(t
->payload
, t
->type
, t
->layout
,
2445 true, t
->levels
, t
->depth
,
2446 t
->array_size
, tmem
);
2450 pandecode_log("};\n");
2453 /* 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 */
2456 pandecode_shader_prop(const char *name
, unsigned claim
, signed truth
, bool fuzzy
)
2462 if (fuzzy
&& (truth
< 0))
2463 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name
, claim
, truth
);
2465 if ((truth
>= 0) && !fuzzy
) {
2466 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2467 (truth
< claim
) ? "warn" : "XXX",
2468 name
, truth
, claim
);
2469 } else if ((claim
> -truth
) && !fuzzy
) {
2470 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2471 name
, -truth
, claim
);
2472 } else if (fuzzy
&& (claim
< truth
))
2473 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2474 name
, truth
, claim
);
2476 pandecode_log(".%s = %" PRId16
, name
, claim
);
2479 pandecode_log_cont(" /* %u used */", truth
);
2481 pandecode_log_cont(",\n");
2485 pandecode_blend_shader_disassemble(mali_ptr shader
, int job_no
, int job_type
,
2486 bool is_bifrost
, unsigned gpu_id
)
2488 struct midgard_disasm_stats stats
=
2489 pandecode_shader_disassemble(shader
, job_no
, job_type
, is_bifrost
, gpu_id
);
2491 bool has_texture
= (stats
.texture_count
> 0);
2492 bool has_sampler
= (stats
.sampler_count
> 0);
2493 bool has_attribute
= (stats
.attribute_count
> 0);
2494 bool has_varying
= (stats
.varying_count
> 0);
2495 bool has_uniform
= (stats
.uniform_count
> 0);
2496 bool has_ubo
= (stats
.uniform_buffer_count
> 0);
2498 if (has_texture
|| has_sampler
)
2499 pandecode_msg("XXX: blend shader accessing textures\n");
2501 if (has_attribute
|| has_varying
)
2502 pandecode_msg("XXX: blend shader accessing interstage\n");
2504 if (has_uniform
|| has_ubo
)
2505 pandecode_msg("XXX: blend shader accessing uniforms\n");
2509 pandecode_textures(mali_ptr textures
, unsigned texture_count
, int job_no
, bool is_bifrost
)
2511 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(textures
);
2517 const struct bifrost_texture_descriptor
*PANDECODE_PTR_VAR(t
, mmem
, textures
);
2519 pandecode_log("uint64_t textures_%"PRIx64
"_%d[] = {\n", textures
, job_no
);
2522 for (unsigned tex
= 0; tex
< texture_count
; ++tex
)
2523 pandecode_bifrost_texture(&t
[tex
], job_no
, tex
);
2526 pandecode_log("};\n");
2528 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
);
2530 pandecode_log("uint64_t textures_%"PRIx64
"_%d[] = {\n", textures
, job_no
);
2533 for (int tex
= 0; tex
< texture_count
; ++tex
) {
2534 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
2535 char *a
= pointer_as_memory_reference(*u
);
2536 pandecode_log("%s,\n", a
);
2541 pandecode_log("};\n");
2543 /* Now, finally, descend down into the texture descriptor */
2544 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
2545 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
2546 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
2548 pandecode_texture(*u
, tmem
, job_no
, tex
);
2554 pandecode_samplers(mali_ptr samplers
, unsigned sampler_count
, int job_no
, bool is_bifrost
)
2556 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(samplers
);
2562 struct bifrost_sampler_descriptor
*s
;
2564 for (int i
= 0; i
< sampler_count
; ++i
) {
2565 s
= pandecode_fetch_gpu_mem(smem
, samplers
+ sizeof(*s
) * i
, sizeof(*s
));
2567 pandecode_log("struct bifrost_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", samplers
+ sizeof(*s
) * i
, job_no
, i
);
2571 pandecode_msg("XXX: unk1 tripped\n");
2572 pandecode_prop("unk1 = 0x%x", s
->unk1
);
2575 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s
->wrap_s
));
2576 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s
->wrap_t
));
2577 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s
->wrap_r
));
2579 if (s
->unk8
!= 0x8) {
2580 pandecode_msg("XXX: unk8 tripped\n");
2581 pandecode_prop("unk8 = 0x%x", s
->unk8
);
2584 pandecode_prop("unk2 = 0x%x", s
->unk2
);
2585 pandecode_prop("unk3 = 0x%x", s
->unk3
);
2586 pandecode_prop("min_filter = %s", s
->min_filter
? "nearest" : "linear");
2587 pandecode_prop("norm_coords = 0x%x", s
->norm_coords
& 0x1);
2588 pandecode_prop("zero1 = 0x%x", s
->zero1
& 0x1);
2589 pandecode_prop("mip_filter = %s", s
->mip_filter
? "linear" : "nearest");
2590 pandecode_prop("mag_filter = %s", s
->mag_filter
? "linear" : "nearest");
2592 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
2593 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
2595 if (s
->zero1
|| s
->zero2
|| s
->zero3
|| s
->zero4
) {
2596 pandecode_msg("XXX: sampler zero tripped\n");
2597 pandecode_prop("zero = 0x%" PRIx8
", 0x%" PRIx64
", 0x%" PRIx64
", 0x%" PRIx64
"\n", s
->zero1
, s
->zero2
, s
->zero3
, s
->zero4
);
2601 pandecode_log("};\n");
2604 struct mali_sampler_descriptor
*s
;
2606 for (int i
= 0; i
< sampler_count
; ++i
) {
2607 s
= pandecode_fetch_gpu_mem(smem
, samplers
+ sizeof(*s
) * i
, sizeof(*s
));
2609 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", samplers
+ sizeof(*s
) * i
, job_no
, i
);
2612 pandecode_log(".filter_mode = ");
2613 pandecode_log_decoded_flags(sampler_flag_info
, s
->filter_mode
);
2614 pandecode_log_cont(",\n");
2616 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
2617 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
2620 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s
->lod_bias
));
2622 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s
->wrap_s
));
2623 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s
->wrap_t
));
2624 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s
->wrap_r
));
2626 pandecode_prop("compare_func = %s", pandecode_func(s
->compare_func
));
2628 if (s
->zero
|| s
->zero2
) {
2629 pandecode_msg("XXX: sampler zero tripped\n");
2630 pandecode_prop("zero = 0x%X, 0x%X\n", s
->zero
, s
->zero2
);
2633 pandecode_prop("seamless_cube_map = %d", s
->seamless_cube_map
);
2635 pandecode_prop("border_color = { %f, %f, %f, %f }",
2639 s
->border_color
[3]);
2642 pandecode_log("};\n");
2648 pandecode_vertex_tiler_postfix_pre(
2649 const struct mali_vertex_tiler_postfix
*p
,
2650 int job_no
, enum mali_job_type job_type
,
2651 char *suffix
, bool is_bifrost
, unsigned gpu_id
)
2653 struct pandecode_mapped_memory
*attr_mem
;
2655 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2656 * are the only things actually needed from the FBD, vertex/tiler jobs
2657 * no longer reference the FBD -- instead, this field points to some
2658 * info about the scratchpad.
2661 struct pandecode_fbd fbd_info
= {
2662 /* Default for Bifrost */
2667 pandecode_log_cont("\t/* %X %/\n", p
->shared_memory
& 1);
2668 pandecode_compute_fbd(p
->shared_memory
& ~1, job_no
);
2669 } else if (p
->shared_memory
& MALI_MFBD
)
2670 fbd_info
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->shared_memory
) & FBD_MASK
, job_no
, false, job_type
== MALI_JOB_TYPE_COMPUTE
, false);
2671 else if (job_type
== MALI_JOB_TYPE_COMPUTE
)
2672 pandecode_compute_fbd((u64
) (uintptr_t) p
->shared_memory
, job_no
);
2674 fbd_info
= pandecode_sfbd((u64
) (uintptr_t) p
->shared_memory
, job_no
, false, gpu_id
);
2676 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
2677 int texture_count
= 0, sampler_count
= 0;
2680 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->shader
);
2681 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, p
->shader
);
2683 /* Disassemble ahead-of-time to get stats. Initialize with
2684 * stats for the missing-shader case so we get validation
2687 struct midgard_disasm_stats info
= {
2690 .attribute_count
= 0,
2694 .uniform_count
= -128,
2695 .uniform_buffer_count
= 0
2698 if (s
->shader
& ~0xF)
2699 info
= pandecode_shader_disassemble(s
->shader
& ~0xF, job_no
, job_type
, is_bifrost
, gpu_id
);
2701 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", p
->shader
, job_no
, suffix
);
2704 /* Save for dumps */
2705 attribute_count
= s
->attribute_count
;
2706 varying_count
= s
->varying_count
;
2707 texture_count
= s
->texture_count
;
2708 sampler_count
= s
->sampler_count
;
2711 uniform_count
= s
->bifrost2
.uniform_count
;
2712 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
2714 uniform_count
= s
->midgard1
.uniform_count
;
2715 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
2718 pandecode_shader_address("shader", s
->shader
);
2720 pandecode_shader_prop("texture_count", s
->texture_count
, info
.texture_count
, false);
2721 pandecode_shader_prop("sampler_count", s
->sampler_count
, info
.sampler_count
, false);
2722 pandecode_shader_prop("attribute_count", s
->attribute_count
, info
.attribute_count
, false);
2723 pandecode_shader_prop("varying_count", s
->varying_count
, info
.varying_count
, false);
2724 pandecode_shader_prop("uniform_buffer_count",
2725 uniform_buffer_count
,
2726 info
.uniform_buffer_count
, true);
2729 pandecode_shader_prop("uniform_count",
2731 info
.uniform_count
, false);
2733 pandecode_shader_prop("work_count",
2734 s
->midgard1
.work_count
, info
.work_count
, false);
2738 pandecode_log("bifrost1.unk1 = ");
2739 pandecode_log_decoded_flags(shader_bifrost_info
, s
->bifrost1
.unk1
);
2740 pandecode_log_cont(",\n");
2742 bool helpers
= s
->midgard1
.flags_lo
& MALI_HELPER_INVOCATIONS
;
2744 if (helpers
!= info
.helper_invocations
) {
2745 pandecode_msg("XXX: expected helpers %u but got %u\n",
2746 info
.helper_invocations
, helpers
);
2749 pandecode_log(".midgard1.flags_lo = ");
2750 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info
,
2751 s
->midgard1
.flags_lo
& ~MALI_HELPER_INVOCATIONS
);
2752 pandecode_log_cont(",\n");
2754 pandecode_log(".midgard1.flags_hi = ");
2755 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info
, s
->midgard1
.flags_hi
);
2756 pandecode_log_cont(",\n");
2759 if (s
->depth_units
|| s
->depth_factor
) {
2760 pandecode_prop("depth_factor = %f", s
->depth_factor
);
2761 pandecode_prop("depth_units = %f", s
->depth_units
);
2764 if (s
->coverage_mask
)
2765 pandecode_prop("coverage_mask = 0x%X", s
->coverage_mask
);
2768 pandecode_prop(".unknown2_2 = %X", s
->unknown2_2
);
2770 if (s
->unknown2_3
|| s
->unknown2_4
) {
2771 pandecode_log(".unknown2_3 = ");
2773 int unknown2_3
= s
->unknown2_3
;
2774 int unknown2_4
= s
->unknown2_4
;
2776 /* We're not quite sure what these flags mean without the depth test, if anything */
2778 if (unknown2_3
& (MALI_DEPTH_WRITEMASK
| MALI_DEPTH_FUNC_MASK
)) {
2779 const char *func
= pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3
));
2780 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
2782 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
2785 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
2786 pandecode_log_cont(",\n");
2788 pandecode_log(".unknown2_4 = ");
2789 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
2790 pandecode_log_cont(",\n");
2793 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
2794 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
2795 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
2798 pandecode_stencil("front", &s
->stencil_front
);
2799 pandecode_stencil("back", &s
->stencil_back
);
2802 pandecode_log(".bifrost2 = {\n");
2805 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
2806 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
2807 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
2808 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
2811 pandecode_log("},\n");
2812 } else if (s
->midgard2
.unknown2_7
) {
2813 pandecode_log(".midgard2 = {\n");
2816 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
2818 pandecode_log("},\n");
2822 pandecode_msg("XXX: shader padding tripped\n");
2823 pandecode_prop("padding = 0x%" PRIx32
, s
->padding
);
2827 /* TODO: Blend shaders routing/disasm */
2828 union midgard_blend blend
= s
->blend
;
2829 mali_ptr shader
= pandecode_midgard_blend(&blend
, s
->unknown2_3
& MALI_HAS_BLEND_SHADER
);
2831 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
2833 pandecode_msg("mdg_blend = %" PRIx64
"\n", s
->blend
.shader
);
2837 pandecode_log("};\n");
2839 /* MRT blend fields are used whenever MFBD is used, with
2840 * per-RT descriptors */
2842 if (job_type
== MALI_JOB_TYPE_TILER
&& (is_bifrost
|| p
->shared_memory
& MALI_MFBD
)) {
2843 void* blend_base
= (void *) (s
+ 1);
2845 for (unsigned i
= 0; i
< fbd_info
.rt_count
; i
++) {
2846 mali_ptr shader
= 0;
2849 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
2851 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
2854 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
2859 pandecode_msg("XXX: missing shader descriptor\n");
2862 struct pandecode_mapped_memory
*fmem
= pandecode_find_mapped_gpu_mem_containing(p
->viewport
);
2863 struct mali_viewport
*PANDECODE_PTR_VAR(f
, fmem
, p
->viewport
);
2865 pandecode_log("struct mali_viewport viewport_%"PRIx64
"_%d%s = {\n", p
->viewport
, job_no
, suffix
);
2868 pandecode_prop("clip_minx = %f", f
->clip_minx
);
2869 pandecode_prop("clip_miny = %f", f
->clip_miny
);
2870 pandecode_prop("clip_minz = %f", f
->clip_minz
);
2871 pandecode_prop("clip_maxx = %f", f
->clip_maxx
);
2872 pandecode_prop("clip_maxy = %f", f
->clip_maxy
);
2873 pandecode_prop("clip_maxz = %f", f
->clip_maxz
);
2875 /* Only the higher coordinates are MALI_POSITIVE scaled */
2877 pandecode_prop("viewport0 = { %d, %d }",
2878 f
->viewport0
[0], f
->viewport0
[1]);
2880 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2881 f
->viewport1
[0] + 1, f
->viewport1
[1] + 1);
2884 pandecode_log("};\n");
2887 unsigned max_attr_index
= 0;
2889 if (p
->attribute_meta
)
2890 max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
2892 if (p
->attributes
) {
2893 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
2894 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false, job_type
);
2897 /* Varyings are encoded like attributes but not actually sent; we just
2898 * pass a zero buffer with the right stride/size set, (or whatever)
2899 * since the GPU will write to it itself */
2901 if (p
->varying_meta
) {
2902 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
2906 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
2908 /* Number of descriptors depends on whether there are
2909 * non-internal varyings */
2911 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true, job_type
);
2914 if (p
->uniform_buffers
) {
2915 if (uniform_buffer_count
)
2916 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
2918 pandecode_msg("warn: UBOs specified but not referenced\n");
2919 } else if (uniform_buffer_count
)
2920 pandecode_msg("XXX: UBOs referenced but not specified\n");
2922 /* We don't want to actually dump uniforms, but we do need to validate
2923 * that the counts we were given are sane */
2927 pandecode_uniforms(p
->uniforms
, uniform_count
);
2929 pandecode_msg("warn: Uniforms specified but not referenced\n");
2930 } else if (uniform_count
)
2931 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2934 pandecode_textures(p
->textures
, texture_count
, job_no
, is_bifrost
);
2936 if (p
->sampler_descriptor
)
2937 pandecode_samplers(p
->sampler_descriptor
, sampler_count
, job_no
, is_bifrost
);
2941 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2943 pandecode_log(".gl_enables = ");
2945 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2947 pandecode_log_cont(",\n");
2951 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2953 if (p
->shader
& 0xF)
2954 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p
->shader
& 0xF));
2956 pandecode_log(".postfix = {\n");
2959 pandecode_gl_enables(p
->gl_enables
, MALI_JOB_TYPE_TILER
);
2960 pandecode_prop("instance_shift = 0x%x", p
->instance_shift
);
2961 pandecode_prop("instance_odd = 0x%x", p
->instance_odd
);
2964 pandecode_msg("XXX: vertex only zero tripped");
2965 pandecode_prop("zero4 = 0x%" PRIx32
, p
->zero4
);
2968 pandecode_prop("offset_start = 0x%x", p
->offset_start
);
2971 pandecode_msg("XXX: vertex only zero tripped");
2972 pandecode_prop("zero5 = 0x%" PRIx64
, p
->zero5
);
2975 MEMORY_PROP(p
, position_varying
);
2976 MEMORY_PROP(p
, occlusion_counter
);
2979 pandecode_log("},\n");
2983 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2985 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2986 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2988 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2992 pandecode_msg("XXX: tiler heap zero tripped\n");
2993 pandecode_prop("zero = 0x%x", h
->zero
);
2996 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2997 MEMORY_PROP(h
, tiler_heap_start
);
2998 MEMORY_PROP(h
, tiler_heap_free
);
3000 /* this might point to the beginning of another buffer, when it's
3001 * really the end of the tiler heap buffer, so we have to be careful
3002 * here. but for zero length, we need the same pointer.
3005 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
3006 MEMORY_PROP(h
, tiler_heap_start
);
3008 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
3009 pandecode_prop("tiler_heap_end = %s + 1", a
);
3013 for (int i
= 0; i
< 10; i
++) {
3014 if (h
->zeros
[i
] != 0) {
3015 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
3020 if (h
->unk1
!= 0x1) {
3021 pandecode_msg("XXX: tiler heap unk1 tripped\n");
3022 pandecode_prop("unk1 = 0x%x", h
->unk1
);
3025 if (h
->unk7e007e
!= 0x7e007e) {
3026 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
3027 pandecode_prop("unk7e007e = 0x%x", h
->unk7e007e
);
3031 pandecode_log("};\n");
3035 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
3037 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
3038 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
3040 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
3042 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
3045 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32
, t
->tiler_heap_next_start
);
3046 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32
, t
->used_hierarchy_mask
);
3048 if (t
->hierarchy_mask
!= 0xa &&
3049 t
->hierarchy_mask
!= 0x14 &&
3050 t
->hierarchy_mask
!= 0x28 &&
3051 t
->hierarchy_mask
!= 0x50 &&
3052 t
->hierarchy_mask
!= 0xa0)
3053 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
3055 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
3057 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
3059 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
3060 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
3063 pandecode_msg("XXX: tiler meta zero tripped\n");
3064 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
3067 for (int i
= 0; i
< 12; i
++) {
3068 if (t
->zeros
[i
] != 0) {
3069 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
3075 pandecode_log("};\n");
3079 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
3081 if (u
.pointer
== 0x0)
3084 pandecode_log(".primitive_size = {\n");
3088 pandecode_prop("constant = %f", u
.constant
);
3090 MEMORY_PROP((&u
), pointer
);
3094 pandecode_log("},\n");
3098 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
3100 pandecode_log_cont("{\n");
3103 /* TODO: gl_PointSize on Bifrost */
3104 pandecode_primitive_size(t
->primitive_size
, true);
3106 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
3108 pandecode_msg("XXX: tiler only zero tripped\n");
3109 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
3110 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
3111 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
3112 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
3113 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
3114 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
3118 pandecode_log("},\n");
3122 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
3123 const struct pandecode_mapped_memory
*mem
,
3124 mali_ptr payload
, int job_no
, unsigned gpu_id
)
3126 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
3128 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
3130 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
3133 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
3134 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
3137 pandecode_log("};\n");
3143 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
3144 const struct pandecode_mapped_memory
*mem
,
3145 mali_ptr payload
, int job_no
, unsigned gpu_id
)
3147 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
3149 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
3150 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
3152 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
3155 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
3157 pandecode_log(".tiler = ");
3158 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
3160 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
3163 pandecode_log("};\n");
3169 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
3170 const struct pandecode_mapped_memory
*mem
,
3171 mali_ptr payload
, int job_no
, unsigned gpu_id
)
3173 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
3174 bool is_graphics
= (h
->job_type
== MALI_JOB_TYPE_VERTEX
) || (h
->job_type
== MALI_JOB_TYPE_TILER
);
3176 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false, gpu_id
);
3178 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
3181 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, is_graphics
);
3182 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
3184 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
3185 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
3188 pandecode_log("};\n");
3194 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
3195 mali_ptr payload
, int job_no
,
3196 bool is_bifrost
, unsigned gpu_id
)
3198 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
3200 bool is_mfbd
= s
->framebuffer
& MALI_MFBD
;
3202 if (!is_mfbd
&& is_bifrost
)
3203 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
3205 struct pandecode_fbd info
;
3208 info
= pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true, false, is_bifrost
);
3210 info
= pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true, gpu_id
);
3212 /* Compute the tag for the tagged pointer. This contains the type of
3213 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
3214 * additional structures follow the MFBD header (an extra payload or
3215 * not, as well as a count of render targets) */
3217 unsigned expected_tag
= is_mfbd
? MALI_MFBD
: 0;
3221 expected_tag
|= MALI_MFBD_TAG_EXTRA
;
3223 expected_tag
|= (MALI_POSITIVE(info
.rt_count
) << 2);
3226 if ((s
->min_tile_coord
| s
->max_tile_coord
) & ~(MALI_X_COORD_MASK
| MALI_Y_COORD_MASK
)) {
3227 pandecode_msg("XXX: unexpected tile coordinate bits\n");
3228 pandecode_prop("min_tile_coord = 0x%X\n", s
->min_tile_coord
);
3229 pandecode_prop("max_tile_coord = 0x%X\n", s
->max_tile_coord
);
3232 /* Extract tile coordinates */
3234 unsigned min_x
= MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
3235 unsigned min_y
= MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
3237 unsigned max_x
= (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
3238 unsigned max_y
= (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
3240 /* For the max, we also want the floored (rather than ceiled) version for checking */
3242 unsigned max_x_f
= (MALI_TILE_COORD_X(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
3243 unsigned max_y_f
= (MALI_TILE_COORD_Y(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
3245 /* Validate the coordinates are well-ordered */
3248 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
3249 else if (min_x
> max_x
)
3250 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
3253 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
3254 else if (min_y
> max_y
)
3255 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
3257 /* Validate the coordinates fit inside the framebuffer. We use floor,
3258 * rather than ceil, for the max coordinates, since the tile
3259 * coordinates for something like an 800x600 framebuffer will actually
3260 * resolve to 800x608, which would otherwise trigger a Y-overflow */
3262 if ((min_x
> info
.width
) || (max_x_f
> info
.width
))
3263 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
3265 if ((min_y
> info
.height
) || (max_y_f
> info
.height
))
3266 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
3268 /* After validation, we print */
3270 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x
, min_y
, max_x
, max_y
);
3272 /* The FBD is a tagged pointer */
3274 unsigned tag
= (s
->framebuffer
& ~FBD_MASK
);
3276 if (tag
!= expected_tag
)
3277 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag
, tag
);
3282 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
3283 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
3284 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
3285 * might wish to combine this with the bifrost parameter) because some details
3286 * are model-specific even within a particular architecture. Minimal traces
3287 * *only* examine the job descriptors, skipping printing entirely if there is
3288 * no faults, and only descends into the payload if there are faults. This is
3289 * useful for looking for faults without the overhead of invasive traces. */
3292 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
, unsigned gpu_id
, bool minimal
)
3294 pandecode_dump_file_open();
3296 struct mali_job_descriptor_header
*h
;
3297 unsigned job_descriptor_number
= 0;
3300 struct pandecode_mapped_memory
*mem
=
3301 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
3305 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
3307 /* On Midgard, for 32-bit jobs except for fragment jobs, the
3308 * high 32-bits of the 64-bit pointer are reused to store
3311 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
3312 h
->job_type
!= MALI_JOB_TYPE_FRAGMENT
? 4 : 0;
3313 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
3315 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
, 256);
3317 int job_no
= job_descriptor_number
++;
3319 /* If the job is good to go, skip it in minimal mode */
3320 if (minimal
&& (h
->exception_status
== 0x0 || h
->exception_status
== 0x1))
3323 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
3326 pandecode_prop("job_type = %s", mali_job_type_as_str(h
->job_type
));
3328 if (h
->job_descriptor_size
)
3329 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
3331 if (h
->exception_status
&& h
->exception_status
!= 0x1)
3332 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
3333 h
->exception_status
,
3334 (h
->exception_status
>> 16) & 0xFFFF,
3335 pandecode_exception_access((h
->exception_status
>> 8) & 0x3),
3336 h
->exception_status
& 0xFF);
3338 if (h
->first_incomplete_task
)
3339 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
3341 if (h
->fault_pointer
)
3342 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
3345 pandecode_prop("job_barrier = %d", h
->job_barrier
);
3347 pandecode_prop("job_index = %d", h
->job_index
);
3349 if (h
->unknown_flags
)
3350 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
3352 if (h
->job_dependency_index_1
)
3353 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
3355 if (h
->job_dependency_index_2
)
3356 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
3359 pandecode_log("};\n");
3361 switch (h
->job_type
) {
3362 case MALI_JOB_TYPE_WRITE_VALUE
: {
3363 struct mali_payload_write_value
*s
= payload
;
3364 pandecode_log("struct mali_payload_write_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
3366 MEMORY_PROP(s
, address
);
3368 if (s
->value_descriptor
!= MALI_WRITE_VALUE_ZERO
) {
3369 pandecode_msg("XXX: unknown value descriptor\n");
3370 pandecode_prop("value_descriptor = 0x%" PRIX32
, s
->value_descriptor
);
3374 pandecode_msg("XXX: set value tripped\n");
3375 pandecode_prop("reserved = 0x%" PRIX32
, s
->reserved
);
3378 pandecode_prop("immediate = 0x%" PRIX64
, s
->immediate
);
3380 pandecode_log("};\n");
3385 case MALI_JOB_TYPE_TILER
:
3386 case MALI_JOB_TYPE_VERTEX
:
3387 case MALI_JOB_TYPE_COMPUTE
:
3389 if (h
->job_type
== MALI_JOB_TYPE_TILER
)
3390 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
3392 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
3394 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
, gpu_id
);
3398 case MALI_JOB_TYPE_FRAGMENT
:
3399 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
, gpu_id
);
3405 } while ((jc_gpu_va
= h
->next_job
));
3407 pandecode_map_read_write();