2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
26 #include <panfrost-job.h>
34 #include "util/macros.h"
35 #include "util/u_math.h"
37 #include "pan_pretty_print.h"
38 #include "midgard/disassemble.h"
39 #include "bifrost/disassemble.h"
41 #include "pan_encoder.h"
43 static void pandecode_swizzle(unsigned swizzle
, enum mali_format format
);
45 #define MEMORY_PROP(obj, p) {\
47 char *a = pointer_as_memory_reference(obj->p); \
48 pandecode_prop("%s = %s", #p, a); \
53 #define MEMORY_PROP_DIR(obj, p) {\
55 char *a = pointer_as_memory_reference(obj.p); \
56 pandecode_prop("%s = %s", #p, a); \
61 FILE *pandecode_dump_stream
;
63 /* Semantic logging type.
65 * Raw: for raw messages to be printed as is.
66 * Message: for helpful information to be commented out in replays.
67 * Property: for properties of a struct
69 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
72 enum pandecode_log_type
{
78 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
79 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
80 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
82 unsigned pandecode_indent
= 0;
85 pandecode_make_indent(void)
87 for (unsigned i
= 0; i
< pandecode_indent
; ++i
)
88 fprintf(pandecode_dump_stream
, " ");
92 pandecode_log_typed(enum pandecode_log_type type
, const char *format
, ...)
96 pandecode_make_indent();
98 if (type
== PANDECODE_MESSAGE
)
99 fprintf(pandecode_dump_stream
, "// ");
100 else if (type
== PANDECODE_PROPERTY
)
101 fprintf(pandecode_dump_stream
, ".");
103 va_start(ap
, format
);
104 vfprintf(pandecode_dump_stream
, format
, ap
);
107 if (type
== PANDECODE_PROPERTY
)
108 fprintf(pandecode_dump_stream
, ",\n");
112 pandecode_log_cont(const char *format
, ...)
116 va_start(ap
, format
);
117 vfprintf(pandecode_dump_stream
, format
, ap
);
121 /* To check for memory safety issues, validates that the given pointer in GPU
122 * memory is valid, containing at least sz bytes. The goal is to eliminate
123 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
124 * overruns) by statically validating pointers.
128 pandecode_validate_buffer(mali_ptr addr
, size_t sz
)
131 pandecode_msg("XXX: null pointer deref");
137 struct pandecode_mapped_memory
*bo
=
138 pandecode_find_mapped_gpu_mem_containing(addr
);
141 pandecode_msg("XXX: invalid memory dereference\n");
147 unsigned offset
= addr
- bo
->gpu_va
;
148 unsigned total
= offset
+ sz
;
150 if (total
> bo
->length
) {
151 pandecode_msg("XXX: buffer overrun. "
152 "Chunk of size %zu at offset %d in buffer of size %zu. "
153 "Overrun by %zu bytes. \n",
154 sz
, offset
, bo
->length
, total
- bo
->length
);
159 struct pandecode_flag_info
{
165 pandecode_log_decoded_flags(const struct pandecode_flag_info
*flag_info
,
168 bool decodable_flags_found
= false;
170 for (int i
= 0; flag_info
[i
].name
; i
++) {
171 if ((flags
& flag_info
[i
].flag
) != flag_info
[i
].flag
)
174 if (!decodable_flags_found
) {
175 decodable_flags_found
= true;
177 pandecode_log_cont(" | ");
180 pandecode_log_cont("%s", flag_info
[i
].name
);
182 flags
&= ~flag_info
[i
].flag
;
185 if (decodable_flags_found
) {
187 pandecode_log_cont(" | 0x%" PRIx64
, flags
);
189 pandecode_log_cont("0x%" PRIx64
, flags
);
193 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
194 static const struct pandecode_flag_info gl_enable_flag_info
[] = {
195 FLAG_INFO(OCCLUSION_QUERY
),
196 FLAG_INFO(OCCLUSION_PRECISE
),
197 FLAG_INFO(FRONT_CCW_TOP
),
198 FLAG_INFO(CULL_FACE_FRONT
),
199 FLAG_INFO(CULL_FACE_BACK
),
204 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
205 static const struct pandecode_flag_info clear_flag_info
[] = {
208 FLAG_INFO(SLOW_STENCIL
),
213 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
214 static const struct pandecode_flag_info mask_flag_info
[] = {
223 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
224 static const struct pandecode_flag_info u3_flag_info
[] = {
226 FLAG_INFO(CAN_DISCARD
),
227 FLAG_INFO(HAS_BLEND_SHADER
),
228 FLAG_INFO(DEPTH_WRITEMASK
),
232 static const struct pandecode_flag_info u4_flag_info
[] = {
234 FLAG_INFO(NO_DITHER
),
235 FLAG_INFO(DEPTH_RANGE_A
),
236 FLAG_INFO(DEPTH_RANGE_B
),
237 FLAG_INFO(STENCIL_TEST
),
238 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER
),
243 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
244 static const struct pandecode_flag_info mfbd_fmt_flag_info
[] = {
251 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
252 static const struct pandecode_flag_info mfbd_extra_flag_hi_info
[] = {
258 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
259 static const struct pandecode_flag_info mfbd_extra_flag_lo_info
[] = {
265 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
266 static const struct pandecode_flag_info shader_midgard1_flag_lo_info
[] = {
269 FLAG_INFO(READS_TILEBUFFER
),
274 static const struct pandecode_flag_info shader_midgard1_flag_hi_info
[] = {
280 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
281 static const struct pandecode_flag_info mfbd_flag_info
[] = {
282 FLAG_INFO(DEPTH_WRITE
),
288 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
289 static const struct pandecode_flag_info sampler_flag_info
[] = {
290 FLAG_INFO(MAG_NEAREST
),
291 FLAG_INFO(MIN_NEAREST
),
292 FLAG_INFO(MIP_LINEAR_1
),
293 FLAG_INFO(MIP_LINEAR_2
),
294 FLAG_INFO(NORM_COORDS
),
299 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
300 static const struct pandecode_flag_info sfbd_unk1_info
[] = {
307 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
308 static const struct pandecode_flag_info sfbd_unk2_info
[] = {
315 extern char *replace_fragment
;
316 extern char *replace_vertex
;
319 pandecode_job_type(enum mali_job_type type
)
321 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
325 DEFINE_CASE(WRITE_VALUE
);
326 DEFINE_CASE(CACHE_FLUSH
);
327 DEFINE_CASE(COMPUTE
);
331 DEFINE_CASE(FRAGMENT
);
333 case JOB_NOT_STARTED
:
334 return "NOT_STARTED";
337 pandecode_log("Warning! Unknown job type %x\n", type
);
345 pandecode_draw_mode(enum mali_draw_mode mode
)
347 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
350 DEFINE_CASE(DRAW_NONE
);
353 DEFINE_CASE(TRIANGLES
);
354 DEFINE_CASE(TRIANGLE_STRIP
);
355 DEFINE_CASE(TRIANGLE_FAN
);
356 DEFINE_CASE(LINE_STRIP
);
357 DEFINE_CASE(LINE_LOOP
);
358 DEFINE_CASE(POLYGON
);
360 DEFINE_CASE(QUAD_STRIP
);
363 pandecode_msg("XXX: invalid draw mode %X\n", mode
);
370 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
372 pandecode_func(enum mali_func mode
)
379 DEFINE_CASE(GREATER
);
380 DEFINE_CASE(NOTEQUAL
);
385 pandecode_msg("XXX: invalid func %X\n", mode
);
391 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
393 pandecode_stencil_op(enum mali_stencil_op op
)
397 DEFINE_CASE(REPLACE
);
400 DEFINE_CASE(INCR_WRAP
);
401 DEFINE_CASE(DECR_WRAP
);
406 pandecode_msg("XXX: invalid stencil op %X\n", op
);
413 static char *pandecode_attr_mode_short(enum mali_attr_mode mode
)
416 /* TODO: Combine to just "instanced" once this can be done
417 * unambiguously in all known cases */
418 case MALI_ATTR_POT_DIVIDE
:
419 return "instanced_pot";
420 case MALI_ATTR_MODULO
:
421 return "instanced_mod";
422 case MALI_ATTR_NPOT_DIVIDE
:
423 return "instanced_npot";
424 case MALI_ATTR_IMAGE
:
427 pandecode_msg("XXX: invalid attribute mode %X\n", mode
);
433 pandecode_special_record(uint64_t v
, bool* attribute
)
436 case MALI_ATTR_VERTEXID
:
438 return "gl_VertexID";
439 case MALI_ATTR_INSTANCEID
:
441 return "gl_InstanceID";
442 case MALI_VARYING_FRAG_COORD
:
443 return "gl_FragCoord";
444 case MALI_VARYING_FRONT_FACING
:
445 return "gl_FrontFacing";
446 case MALI_VARYING_POINT_COORD
:
447 return "gl_PointCoord";
449 pandecode_msg("XXX: invalid special record %" PRIx64
"\n", v
);
454 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
456 pandecode_wrap_mode(enum mali_wrap_mode op
)
460 DEFINE_CASE(CLAMP_TO_EDGE
);
461 DEFINE_CASE(CLAMP_TO_BORDER
);
462 DEFINE_CASE(MIRRORED_REPEAT
);
465 pandecode_msg("XXX: invalid wrap mode %X\n", op
);
471 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
473 pandecode_block_format(enum mali_block_format fmt
)
477 DEFINE_CASE(UNKNOWN
);
482 unreachable("Invalid case");
487 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
489 pandecode_exception_access(unsigned access
)
493 DEFINE_CASE(EXECUTE
);
498 unreachable("Invalid case");
503 /* Midgard's tiler descriptor is embedded within the
507 pandecode_midgard_tiler_descriptor(
508 const struct midgard_tiler_descriptor
*t
,
514 pandecode_log(".tiler = {\n");
517 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
518 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
520 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
522 /* We know this name from the kernel, but we never see it nonzero */
525 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
527 MEMORY_PROP(t
, polygon_list
);
529 /* The body is offset from the base of the polygon list */
530 //assert(t->polygon_list_body > t->polygon_list);
531 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
533 /* It needs to fit inside the reported size */
534 //assert(t->polygon_list_size >= body_offset);
536 /* Check that we fit */
537 struct pandecode_mapped_memory
*plist
=
538 pandecode_find_mapped_gpu_mem_containing(t
->polygon_list
);
540 //assert(t->polygon_list_size <= plist->length);
542 /* Now that we've sanity checked, we'll try to calculate the sizes
543 * ourselves for comparison */
545 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
546 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
548 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
549 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
550 ref_header
, ref_size
);
551 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
552 pandecode_msg("body offset %d\n", body_offset
);
555 /* The tiler heap has a start and end specified -- it should be
556 * identical to what we have in the BO. The exception is if tiling is
559 MEMORY_PROP(t
, heap_start
);
560 assert(t
->heap_end
>= t
->heap_start
);
562 struct pandecode_mapped_memory
*heap
=
563 pandecode_find_mapped_gpu_mem_containing(t
->heap_start
);
565 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
567 /* Tiling is enabled with a special flag */
568 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
569 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
571 bool tiling_enabled
= hierarchy_mask
;
573 if (tiling_enabled
) {
574 /* When tiling is enabled, the heap should be a tight fit */
575 unsigned heap_offset
= t
->heap_start
- heap
->gpu_va
;
576 if ((heap_offset
+ heap_size
) != heap
->length
) {
577 pandecode_msg("XXX: heap size %u (expected %zu)\n",
578 heap_size
, heap
->length
- heap_offset
);
581 /* We should also have no other flags */
583 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
585 /* When tiling is disabled, we should have that flag and no others */
587 if (tiler_flags
!= MALI_TILER_DISABLED
) {
588 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
592 /* We should also have an empty heap */
594 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
598 /* Disabled tiling is used only for clear-only jobs, which are
599 * purely FRAGMENT, so we should never see this for
600 * non-FRAGMENT descriptors. */
603 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
606 /* We've never seen weights used in practice, but we know from the
607 * kernel these fields is there */
609 bool nonzero_weights
= false;
611 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
612 nonzero_weights
|= t
->weights
[w
] != 0x0;
615 if (nonzero_weights
) {
616 pandecode_log(".weights = { ");
618 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
619 pandecode_log_cont("%d, ", t
->weights
[w
]);
626 pandecode_log("}\n");
629 /* TODO: The Bifrost tiler is not understood at all yet */
632 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer
*fb
)
634 pandecode_log(".tiler = {\n");
637 MEMORY_PROP(fb
, tiler_meta
);
639 for (int i
= 0; i
< 16; i
++) {
640 if (fb
->zeros
[i
] != 0) {
641 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
646 pandecode_log("},\n");
649 pandecode_log("}\n");
653 /* Information about the framebuffer passed back for
654 * additional analysis */
656 struct pandecode_fbd
{
664 pandecode_sfbd_format(struct mali_sfbd_format format
)
666 pandecode_log(".format = {\n");
669 pandecode_log(".unk1 = ");
670 pandecode_log_decoded_flags(sfbd_unk1_info
, format
.unk1
);
671 pandecode_log_cont(",\n");
673 /* TODO: Map formats so we can check swizzles and print nicely */
674 pandecode_log("swizzle");
675 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
676 pandecode_log_cont(",\n");
678 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
679 (format
.nr_channels
+ 1));
681 pandecode_log(".unk2 = ");
682 pandecode_log_decoded_flags(sfbd_unk2_info
, format
.unk2
);
683 pandecode_log_cont(",\n");
685 pandecode_prop("block = %s", pandecode_block_format(format
.block
));
687 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
690 pandecode_log("},\n");
694 pandecode_shared_memory(const struct mali_shared_memory
*desc
, bool is_compute
)
696 pandecode_prop("stack_shift = 0x%x", desc
->stack_shift
);
699 pandecode_prop("unk0 = 0x%x", desc
->unk0
);
701 if (desc
->shared_workgroup_count
!= 0x1F) {
702 pandecode_prop("shared_workgroup_count = %d", desc
->shared_workgroup_count
);
704 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
707 if (desc
->shared_unk1
|| desc
->shared_shift
) {
708 pandecode_prop("shared_unk1 = %X", desc
->shared_unk1
);
709 pandecode_prop("shared_shift = %X", desc
->shared_shift
);
712 pandecode_msg("XXX: shared memory configured in noncompute shader");
715 if (desc
->shared_zero
) {
716 pandecode_msg("XXX: shared memory zero tripped\n");
717 pandecode_prop("shared_zero = 0x%" PRIx32
, desc
->shared_zero
);
720 if (desc
->shared_memory
&& !is_compute
)
721 pandecode_msg("XXX: shared memory used in noncompute shader\n");
723 MEMORY_PROP(desc
, scratchpad
);
724 MEMORY_PROP(desc
, shared_memory
);
725 MEMORY_PROP(desc
, unknown1
);
727 if (desc
->scratchpad
) {
728 struct pandecode_mapped_memory
*smem
=
729 pandecode_find_mapped_gpu_mem_containing(desc
->scratchpad
);
731 pandecode_msg("scratchpad size %u\n", smem
->length
);
736 static struct pandecode_fbd
737 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
, unsigned gpu_id
)
739 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
740 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
742 struct pandecode_fbd info
= {
747 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
750 pandecode_log(".shared_memory = {\n");
752 pandecode_shared_memory(&s
->shared_memory
, false);
754 pandecode_log("},\n");
756 pandecode_sfbd_format(s
->format
);
758 info
.width
= s
->width
+ 1;
759 info
.height
= s
->height
+ 1;
761 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", info
.width
);
762 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", info
.height
);
764 MEMORY_PROP(s
, checksum
);
766 if (s
->checksum_stride
)
767 pandecode_prop("checksum_stride = %d", s
->checksum_stride
);
769 MEMORY_PROP(s
, framebuffer
);
770 pandecode_prop("stride = %d", s
->stride
);
772 /* Earlier in the actual commandstream -- right before width -- but we
773 * delay to flow nicer */
775 pandecode_log(".clear_flags = ");
776 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
777 pandecode_log_cont(",\n");
779 if (s
->depth_buffer
) {
780 MEMORY_PROP(s
, depth_buffer
);
781 pandecode_prop("depth_stride = %d", s
->depth_stride
);
784 if (s
->stencil_buffer
) {
785 MEMORY_PROP(s
, stencil_buffer
);
786 pandecode_prop("stencil_stride = %d", s
->stencil_stride
);
789 if (s
->depth_stride_zero
||
790 s
->stencil_stride_zero
||
791 s
->zero7
|| s
->zero8
) {
792 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
793 pandecode_prop("depth_stride_zero = 0x%x",
794 s
->depth_stride_zero
);
795 pandecode_prop("stencil_stride_zero = 0x%x",
796 s
->stencil_stride_zero
);
797 pandecode_prop("zero7 = 0x%" PRIx32
,
799 pandecode_prop("zero8 = 0x%" PRIx32
,
803 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
804 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
805 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
806 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
807 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
810 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
811 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
812 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
813 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
814 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
817 if (s
->clear_stencil
) {
818 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
821 const struct midgard_tiler_descriptor t
= s
->tiler
;
823 bool has_hierarchy
= !(gpu_id
== 0x0720 || gpu_id
== 0x0820 || gpu_id
== 0x0830);
824 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
, has_hierarchy
);
827 pandecode_log("};\n");
829 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
830 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
831 pandecode_prop("zero5 = 0x%" PRIx32
, s
->zero5
);
833 pandecode_log_cont(".zero3 = {");
835 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
836 pandecode_log_cont("%X, ", s
->zero3
[i
]);
838 pandecode_log_cont("},\n");
840 pandecode_log_cont(".zero6 = {");
842 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
843 pandecode_log_cont("%X, ", s
->zero6
[i
]);
845 pandecode_log_cont("},\n");
851 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
853 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
854 const struct mali_shared_memory
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
856 pandecode_log("struct mali_shared_memory shared_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
858 pandecode_shared_memory(s
, true);
860 pandecode_log("},\n");
863 /* Extracts the number of components associated with a Mali format */
866 pandecode_format_component_count(enum mali_format fmt
)
868 /* Mask out the format class */
869 unsigned top
= fmt
& 0b11100000;
872 case MALI_FORMAT_SNORM
:
873 case MALI_FORMAT_UINT
:
874 case MALI_FORMAT_UNORM
:
875 case MALI_FORMAT_SINT
:
876 return ((fmt
>> 3) & 3) + 1;
883 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
886 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
889 assert(MALI_CHANNEL_RED
== 0);
891 for (unsigned c
= 0; c
< 4; ++c
) {
892 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
894 if (chan
<= MALI_CHANNEL_ALPHA
)
901 /* Validates that a (format, swizzle) pair is valid, in the sense that the
902 * swizzle doesn't access any components that are undefined in the format.
903 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
907 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
909 unsigned nr_comp
= pandecode_format_component_count(fmt
);
910 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
911 unsigned valid_mask
= (1 << nr_comp
) - 1;
912 unsigned invalid_mask
= ~valid_mask
;
914 if (access_mask
& invalid_mask
) {
915 pandecode_msg("XXX: invalid components accessed\n");
919 /* Check for the default non-swizzling swizzle so we can suppress
920 * useless printing for the defaults */
922 unsigned default_swizzles
[4] = {
923 MALI_CHANNEL_RED
| (MALI_CHANNEL_ZERO
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
924 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
925 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ONE
<< 9),
926 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ALPHA
<< 9)
929 return (swizzle
== default_swizzles
[nr_comp
- 1]);
932 /* Maps MALI_RGBA32F to rgba32f, etc */
935 pandecode_format_short(enum mali_format fmt
, bool srgb
)
937 /* We want a type-like format, so cut off the initial MALI_ */
938 char *format
= pandecode_format(fmt
);
939 format
+= strlen("MALI_");
941 unsigned len
= strlen(format
);
942 char *lower_format
= calloc(1, len
+ 1);
944 for (unsigned i
= 0; i
< len
; ++i
)
945 lower_format
[i
] = tolower(format
[i
]);
947 /* Sanity check sRGB flag is applied to RGB, per the name */
948 if (srgb
&& lower_format
[0] != 'r')
949 pandecode_msg("XXX: sRGB applied to non-colour format\n");
951 /* Just prefix with an s, so you get formats like srgba8_unorm */
953 pandecode_log_cont("s");
955 pandecode_log_cont("%s", lower_format
);
960 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
962 /* First, do some validation */
963 bool trivial_swizzle
= pandecode_validate_format_swizzle(
969 /* Next, print the swizzle */
970 pandecode_log_cont(".");
972 static const char components
[] = "rgba01";
974 for (unsigned c
= 0; c
< 4; ++c
) {
975 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
977 if (chan
>= MALI_CHANNEL_RESERVED_0
) {
978 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
981 pandecode_log_cont("%c", components
[chan
]);
986 pandecode_rt_format(struct mali_rt_format format
)
988 pandecode_log(".format = {\n");
991 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
992 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
993 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
995 pandecode_prop("block = %s", pandecode_block_format(format
.block
));
997 /* TODO: Map formats so we can check swizzles and print nicely */
998 pandecode_log("swizzle");
999 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
1000 pandecode_log_cont(",\n");
1002 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
1003 (format
.nr_channels
+ 1));
1005 pandecode_log(".flags = ");
1006 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
1007 pandecode_log_cont(",\n");
1009 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
1010 * which is a faster hardware-based alternative to the wallpaper method
1011 * to preserve framebuffer contents across frames. In practice, MFBD
1012 * preload is buggy on Midgard, and so this is a chicken bit. If this
1013 * bit isn't set, most likely something broke unrelated to preload */
1015 if (!format
.no_preload
) {
1016 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
1017 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
1021 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
1024 pandecode_log("},\n");
1028 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct mali_framebuffer
*fb
)
1030 pandecode_log("struct mali_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
1033 for (int i
= 0; i
< (fb
->rt_count_1
+ 1); i
++) {
1034 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct mali_render_target
);
1035 struct pandecode_mapped_memory
*mem
=
1036 pandecode_find_mapped_gpu_mem_containing(rt_va
);
1037 const struct mali_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
1039 pandecode_log("{\n");
1042 pandecode_rt_format(rt
->format
);
1044 if (rt
->format
.block
== MALI_BLOCK_AFBC
) {
1045 pandecode_log(".afbc = {\n");
1048 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
1049 pandecode_prop("metadata = %s", a
);
1052 pandecode_prop("stride = %d", rt
->afbc
.stride
);
1053 pandecode_prop("unk = 0x%" PRIx32
, rt
->afbc
.unk
);
1056 pandecode_log("},\n");
1057 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.unk
) {
1058 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1064 MEMORY_PROP(rt
, framebuffer
);
1065 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
1067 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
1068 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
1069 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
1070 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
1071 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
1074 if (rt
->zero1
|| rt
->zero2
|| rt
->zero3
) {
1075 pandecode_msg("XXX: render target zeros tripped\n");
1076 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
1077 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
1078 pandecode_prop("zero3 = 0x%" PRIx32
, rt
->zero3
);
1082 pandecode_log("},\n");
1086 pandecode_log("};\n");
1089 static struct pandecode_fbd
1090 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
, bool is_compute
, bool is_bifrost
)
1092 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1093 const struct mali_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
1095 struct pandecode_fbd info
;
1097 if (is_bifrost
&& fb
->msaa
.sample_locations
) {
1098 /* The blob stores all possible sample locations in a single buffer
1099 * allocated on startup, and just switches the pointer when switching
1100 * MSAA state. For now, we just put the data into the cmdstream, but we
1101 * should do something like what the blob does with a real driver.
1103 * There seem to be 32 slots for sample locations, followed by another
1104 * 16. The second 16 is just the center location followed by 15 zeros
1105 * in all the cases I've identified (maybe shader vs. depth/color
1109 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->msaa
.sample_locations
);
1111 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->msaa
.sample_locations
);
1113 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
1116 for (int i
= 0; i
< 32 + 16; i
++) {
1117 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
1121 pandecode_log("};\n");
1124 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1128 pandecode_log(".msaa = {\n");
1131 if (fb
->msaa
.sample_locations
)
1132 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
1134 pandecode_msg("XXX: sample_locations missing\n");
1136 if (fb
->msaa
.zero1
|| fb
->msaa
.zero2
|| fb
->msaa
.zero4
) {
1137 pandecode_msg("XXX: multisampling zero tripped\n");
1138 pandecode_prop("zero1 = %" PRIx64
, fb
->msaa
.zero1
);
1139 pandecode_prop("zero2 = %" PRIx64
, fb
->msaa
.zero2
);
1140 pandecode_prop("zero4 = %" PRIx64
, fb
->msaa
.zero4
);
1144 pandecode_log("},\n");
1146 pandecode_log(".shared_memory = {\n");
1148 pandecode_shared_memory(&fb
->shared_memory
, is_compute
);
1150 pandecode_log("},\n");
1153 info
.width
= fb
->width1
+ 1;
1154 info
.height
= fb
->height1
+ 1;
1155 info
.rt_count
= fb
->rt_count_1
+ 1;
1157 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
1158 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
1159 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
1160 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
1162 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
1163 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
1164 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
1165 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
1167 pandecode_log(".mfbd_flags = ");
1168 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
1169 pandecode_log_cont(",\n");
1171 if (fb
->clear_stencil
)
1172 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
1174 if (fb
->clear_depth
)
1175 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
1179 pandecode_bifrost_tiler_descriptor(fb
);
1181 pandecode_midgard_tiler_descriptor(&fb
->tiler
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
, true);
1183 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1185 if (fb
->zero3
|| fb
->zero4
) {
1186 pandecode_msg("XXX: framebuffer zeros tripped\n");
1187 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
1188 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
1192 pandecode_log("};\n");
1194 gpu_va
+= sizeof(struct mali_framebuffer
);
1196 info
.has_extra
= (fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
;
1198 if (info
.has_extra
) {
1199 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1200 const struct mali_framebuffer_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
1202 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1205 MEMORY_PROP(fbx
, checksum
);
1207 if (fbx
->checksum_stride
)
1208 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
1210 pandecode_log(".flags_hi = ");
1211 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info
, fbx
->flags_lo
);
1212 pandecode_log_cont(",\n");
1214 pandecode_log(".flags_lo = ");
1215 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info
, fbx
->flags_lo
);
1216 pandecode_log_cont(",\n");
1218 pandecode_prop("zs_block = %s", pandecode_block_format(fbx
->zs_block
));
1220 if (fbx
->zs_block
== MALI_BLOCK_AFBC
) {
1221 pandecode_log(".ds_afbc = {\n");
1224 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
1225 pandecode_prop("depth_stencil_afbc_stride = %d",
1226 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
1227 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
1229 if (fbx
->ds_afbc
.zero1
|| fbx
->ds_afbc
.padding
) {
1230 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1231 pandecode_prop("zero1 = 0x%" PRIx32
,
1232 fbx
->ds_afbc
.zero1
);
1233 pandecode_prop("padding = 0x%" PRIx64
,
1234 fbx
->ds_afbc
.padding
);
1238 pandecode_log("},\n");
1240 pandecode_log(".ds_linear = {\n");
1243 if (fbx
->ds_linear
.depth
) {
1244 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
1245 pandecode_prop("depth_stride = %d",
1246 fbx
->ds_linear
.depth_stride
);
1247 } else if (fbx
->ds_linear
.depth_stride
) {
1248 pandecode_msg("XXX: depth stride zero tripped %d\n", fbx
->ds_linear
.depth_stride
);
1251 if (fbx
->ds_linear
.stencil
) {
1252 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1253 pandecode_prop("stencil_stride = %d",
1254 fbx
->ds_linear
.stencil_stride
);
1255 } else if (fbx
->ds_linear
.stencil_stride
) {
1256 pandecode_msg("XXX: stencil stride zero tripped %d\n", fbx
->ds_linear
.stencil_stride
);
1259 if (fbx
->ds_linear
.depth_stride_zero
||
1260 fbx
->ds_linear
.stencil_stride_zero
||
1261 fbx
->ds_linear
.zero1
|| fbx
->ds_linear
.zero2
) {
1262 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1263 pandecode_prop("depth_stride_zero = 0x%x",
1264 fbx
->ds_linear
.depth_stride_zero
);
1265 pandecode_prop("stencil_stride_zero = 0x%x",
1266 fbx
->ds_linear
.stencil_stride_zero
);
1267 pandecode_prop("zero1 = 0x%" PRIx32
,
1268 fbx
->ds_linear
.zero1
);
1269 pandecode_prop("zero2 = 0x%" PRIx32
,
1270 fbx
->ds_linear
.zero2
);
1274 pandecode_log("},\n");
1277 if (fbx
->zero3
|| fbx
->zero4
) {
1278 pandecode_msg("XXX: fb_extra zeros tripped\n");
1279 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1280 pandecode_prop("zero4 = 0x%" PRIx64
, fbx
->zero4
);
1284 pandecode_log("};\n");
1286 gpu_va
+= sizeof(struct mali_framebuffer_extra
);
1290 pandecode_render_target(gpu_va
, job_no
, fb
);
1295 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1299 pandecode_padded_vertices(unsigned shift
, unsigned k
)
1301 unsigned odd
= 2*k
+ 1;
1302 unsigned pot
= 1 << shift
;
1303 pandecode_msg("padded_num_vertices = %d\n", odd
* pot
);
1306 /* Given a magic divisor, recover what we were trying to divide by.
1308 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1309 * 0 <= m < 2^N, for N bits in m.
1311 * Let q represent the number we would like to divide by.
1313 * By definition of a magic divisor for N-bit unsigned integers (a number you
1314 * multiply by to magically get division), m is a number such that:
1316 * (m * x) & (2^N - 1) = floor(x/q).
1317 * for all x on Z where 0 <= x < 2^N
1319 * Ignore the case where any of the above values equals zero; it is irrelevant
1320 * for our purposes (instanced arrays).
1322 * Choose x = q. Then:
1324 * (m * x) & (2^N - 1) = floor(x/q).
1325 * (m * q) & (2^N - 1) = floor(q/q).
1327 * floor(q/q) = floor(1) = 1, therefore:
1329 * (m * q) & (2^N - 1) = 1
1331 * Recall the identity that the bitwise AND of one less than a power-of-two
1332 * equals the modulo with that power of two, i.e. for all x:
1334 * x & (2^N - 1) = x % N
1340 * By definition, a modular multiplicative inverse of a number m is the number
1341 * q such that with respect to a modulos M:
1345 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1350 pandecode_magic_divisor(uint32_t magic
, unsigned shift
, unsigned orig_divisor
, unsigned extra
)
1353 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1354 * shift) the most lame way possible... just repeatedly add.
1355 * Asymptoptically slow but nobody cares in practice, unless you have
1356 * massive numbers of vertices or high divisors. */
1358 unsigned inverse
= 0;
1360 /* Magic implicitly has the highest bit set */
1363 /* Depending on rounding direction */
1368 uint32_t product
= magic
* inverse
;
1380 pandecode_msg("dividing by %d (maybe off by two)\n", inverse
);
1382 /* Recall we're supposed to divide by (gl_level_divisor *
1383 * padded_num_vertices) */
1385 unsigned padded_num_vertices
= inverse
/ orig_divisor
;
1387 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices
);
1392 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1393 mali_ptr addr
, int job_no
, char *suffix
,
1394 int count
, bool varying
, enum mali_job_type job_type
)
1396 char *prefix
= varying
? "varying" : "attribute";
1400 pandecode_msg("warn: No %s records\n", prefix
);
1404 union mali_attr
*attr
= pandecode_fetch_gpu_mem(mem
, addr
, sizeof(union mali_attr
) * count
);
1406 for (int i
= 0; i
< count
; ++i
) {
1407 /* First, check for special records */
1408 if (attr
[i
].elements
< MALI_RECORD_SPECIAL
) {
1410 pandecode_msg("XXX: tripped size=%d\n", attr
[i
].size
);
1412 if (attr
[i
].stride
) {
1413 /* gl_InstanceID passes a magic divisor in the
1414 * stride field to divide by the padded vertex
1415 * count. No other records should do so, so
1416 * stride should otherwise be zero. Note that
1417 * stride in the usual attribute sense doesn't
1418 * apply to special records. */
1420 bool has_divisor
= attr
[i
].elements
== MALI_ATTR_INSTANCEID
;
1422 pandecode_log_cont("/* %smagic divisor = %X */ ",
1423 has_divisor
? "" : "XXX: ", attr
[i
].stride
);
1426 if (attr
[i
].shift
|| attr
[i
].extra_flags
) {
1427 /* Attributes use these fields for
1428 * instancing/padding/etc type issues, but
1431 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1432 varying
? "XXX: " : "",
1433 attr
[i
].shift
, attr
[i
].extra_flags
);
1436 /* Print the special record name */
1437 bool attribute
= false;
1438 pandecode_log("%s_%d = %s;\n", prefix
, i
, pandecode_special_record(attr
[i
].elements
, &attribute
));
1441 if (attribute
== varying
)
1442 pandecode_msg("XXX: mismatched special record\n");
1447 enum mali_attr_mode mode
= attr
[i
].elements
& 7;
1449 if (mode
== MALI_ATTR_UNUSED
)
1450 pandecode_msg("XXX: unused attribute record\n");
1452 /* For non-linear records, we need to print the type of record */
1453 if (mode
!= MALI_ATTR_LINEAR
)
1454 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode
));
1456 /* Print the name to link with attr_meta */
1457 pandecode_log_cont("%s_%d", prefix
, i
);
1459 /* Print the stride and size */
1460 pandecode_log_cont("<%u>[%u]", attr
[i
].stride
, attr
[i
].size
);
1462 /* TODO: Sanity check the quotient itself. It must be equal to
1463 * (or be greater than, if the driver added padding) the padded
1466 /* Finally, print the pointer */
1467 mali_ptr raw_elements
= attr
[i
].elements
& ~7;
1468 char *a
= pointer_as_memory_reference(raw_elements
);
1469 pandecode_log_cont(" = (%s);\n", a
);
1472 /* Check the pointer */
1473 pandecode_validate_buffer(raw_elements
, attr
[i
].size
);
1475 /* shift/extra_flags exist only for instanced */
1476 if (attr
[i
].shift
| attr
[i
].extra_flags
) {
1477 /* These are set to random values by the blob for
1478 * varyings, most likely a symptom of uninitialized
1479 * memory where the hardware masked the bug. As such we
1480 * put this at a warning, not an error. */
1482 if (mode
== MALI_ATTR_LINEAR
)
1483 pandecode_msg("warn: instancing fields set for linear\n");
1485 pandecode_prop("shift = %d", attr
[i
].shift
);
1486 pandecode_prop("extra_flags = %d", attr
[i
].extra_flags
);
1489 /* Decode further where possible */
1491 if (mode
== MALI_ATTR_MODULO
) {
1492 pandecode_padded_vertices(
1494 attr
[i
].extra_flags
);
1497 if (mode
== MALI_ATTR_NPOT_DIVIDE
) {
1499 pandecode_log("{\n");
1501 pandecode_prop("unk = 0x%x", attr
[i
].unk
);
1502 pandecode_prop("magic_divisor = 0x%08x", attr
[i
].magic_divisor
);
1503 if (attr
[i
].zero
!= 0)
1504 pandecode_prop("XXX: zero tripped (0x%x)\n", attr
[i
].zero
);
1505 pandecode_prop("divisor = %d", attr
[i
].divisor
);
1506 pandecode_magic_divisor(attr
[i
].magic_divisor
, attr
[i
- 1].shift
, attr
[i
].divisor
, attr
[i
- 1].extra_flags
);
1508 pandecode_log("}, \n");
1513 pandecode_log("\n");
1517 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1519 /* TODO: Decode flags */
1520 mali_ptr shader_ptr
= ptr
& ~15;
1522 char *a
= pointer_as_memory_reference(shader_ptr
);
1523 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1530 pandecode_stencil(const char *name
, const struct mali_stencil_test
*stencil
)
1532 unsigned any_nonzero
=
1533 stencil
->ref
| stencil
->mask
| stencil
->func
|
1534 stencil
->sfail
| stencil
->dpfail
| stencil
->dppass
;
1536 if (any_nonzero
== 0)
1539 const char *func
= pandecode_func(stencil
->func
);
1540 const char *sfail
= pandecode_stencil_op(stencil
->sfail
);
1541 const char *dpfail
= pandecode_stencil_op(stencil
->dpfail
);
1542 const char *dppass
= pandecode_stencil_op(stencil
->dppass
);
1545 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil
->zero
);
1547 pandecode_log(".stencil_%s = {\n", name
);
1549 pandecode_prop("ref = %d", stencil
->ref
);
1550 pandecode_prop("mask = 0x%02X", stencil
->mask
);
1551 pandecode_prop("func = %s", func
);
1552 pandecode_prop("sfail = %s", sfail
);
1553 pandecode_prop("dpfail = %s", dpfail
);
1554 pandecode_prop("dppass = %s", dppass
);
1556 pandecode_log("},\n");
1560 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1563 pandecode_msg("XXX: blend zero tripped: %X\n", blend
->zero1
);
1565 pandecode_log(".equation = {\n");
1568 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1569 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1571 pandecode_log(".color_mask = ");
1572 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1573 pandecode_log_cont(",\n");
1576 pandecode_log("},\n");
1579 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1582 decode_bifrost_constant(u16 constant
)
1584 float lo
= (float) (constant
& 0xFF);
1585 float hi
= (float) (constant
>> 8);
1587 return (hi
/ 255.0) + (lo
/ 65535.0);
1591 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1593 struct bifrost_blend_rt
*b
=
1594 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1596 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1599 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1600 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1601 b
->constant
, decode_bifrost_constant(b
->constant
));
1603 /* TODO figure out blend shader enable bit */
1604 pandecode_blend_equation(&b
->equation
);
1605 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1606 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1607 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1610 pandecode_log("},\n");
1616 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1618 /* constant/equation is in a union */
1622 pandecode_log(".blend = {\n");
1626 pandecode_shader_address("shader", blend
->shader
);
1628 pandecode_blend_equation(&blend
->equation
);
1629 pandecode_prop("constant = %f", blend
->constant
);
1633 pandecode_log("},\n");
1635 /* Return blend shader to disassemble if present */
1636 return is_shader
? (blend
->shader
& ~0xF) : 0;
1640 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1642 struct midgard_blend_rt
*b
=
1643 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1645 /* Flags determine presence of blend shader */
1646 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1648 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1651 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1653 union midgard_blend blend
= b
->blend
;
1654 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1657 pandecode_log("};\n");
1662 /* Attributes and varyings have descriptor records, which contain information
1663 * about their format and ordering with the attribute/varying buffers. We'll
1664 * want to validate that the combinations specified are self-consistent.
1668 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1671 char *prefix
= varying
? "varying" : "attribute";
1672 unsigned max_index
= 0;
1673 snprintf(base
, sizeof(base
), "%s_meta", prefix
);
1675 struct mali_attr_meta
*attr_meta
;
1676 mali_ptr p
= varying
? v
->varying_meta
: v
->attribute_meta
;
1678 struct pandecode_mapped_memory
*attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
);
1680 for (int i
= 0; i
< count
; ++i
, p
+= sizeof(struct mali_attr_meta
)) {
1681 attr_meta
= pandecode_fetch_gpu_mem(attr_mem
, p
,
1684 /* If the record is discard, it should be zero for everything else */
1686 if (attr_meta
->format
== MALI_VARYING_DISCARD
) {
1689 attr_meta
->unknown1
|
1690 attr_meta
->unknown3
|
1691 attr_meta
->src_offset
;
1694 pandecode_msg("XXX: expected empty record for varying discard\n");
1696 /* We want to look for a literal 0000 swizzle -- this
1697 * is not encoded with all zeroes, however */
1699 enum mali_channel z
= MALI_CHANNEL_ZERO
;
1700 unsigned zero_swizzle
= z
| (z
<< 3) | (z
<< 6) | (z
<< 9);
1701 bool good_swizzle
= attr_meta
->swizzle
== zero_swizzle
;
1704 pandecode_msg("XXX: expected zero swizzle for discard\n");
1707 pandecode_msg("XXX: cannot discard attribute\n");
1709 /* If we're all good, omit the record */
1710 if (!zero
&& varying
&& good_swizzle
) {
1711 pandecode_log("/* discarded varying */\n");
1716 if (attr_meta
->index
> max_index
)
1717 max_index
= attr_meta
->index
;
1719 if (attr_meta
->unknown1
!= 0x2) {
1720 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1721 pandecode_prop("unknown1 = 0x%" PRIx64
, (u64
) attr_meta
->unknown1
);
1724 if (attr_meta
->unknown3
) {
1725 pandecode_msg("XXX: unexpected unknown3 set\n");
1726 pandecode_prop("unknown3 = 0x%" PRIx64
, (u64
) attr_meta
->unknown3
);
1729 pandecode_format_short(attr_meta
->format
, false);
1730 pandecode_log_cont(" %s_%u", prefix
, attr_meta
->index
);
1732 if (attr_meta
->src_offset
)
1733 pandecode_log_cont("[%u]", attr_meta
->src_offset
);
1735 pandecode_swizzle(attr_meta
->swizzle
, attr_meta
->format
);
1737 pandecode_log_cont(";\n");
1740 pandecode_log("\n");
1742 return count
? (max_index
+ 1) : 0;
1745 /* return bits [lo, hi) of word */
1747 bits(u32 word
, u32 lo
, u32 hi
)
1750 return word
; // avoid undefined behavior with the shift
1752 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1756 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool graphics
)
1758 pandecode_log(".prefix = {\n");
1761 /* Decode invocation_count. See the comment before the definition of
1762 * invocation_count for an explanation.
1765 unsigned size_y_shift
= bits(p
->invocation_shifts
, 0, 5);
1766 unsigned size_z_shift
= bits(p
->invocation_shifts
, 5, 10);
1767 unsigned workgroups_x_shift
= bits(p
->invocation_shifts
, 10, 16);
1768 unsigned workgroups_y_shift
= bits(p
->invocation_shifts
, 16, 22);
1769 unsigned workgroups_z_shift
= bits(p
->invocation_shifts
, 22, 28);
1770 unsigned workgroups_x_shift_2
= bits(p
->invocation_shifts
, 28, 32);
1772 unsigned size_x
= bits(p
->invocation_count
, 0, size_y_shift
) + 1;
1773 unsigned size_y
= bits(p
->invocation_count
, size_y_shift
, size_z_shift
) + 1;
1774 unsigned size_z
= bits(p
->invocation_count
, size_z_shift
, workgroups_x_shift
) + 1;
1776 unsigned groups_x
= bits(p
->invocation_count
, workgroups_x_shift
, workgroups_y_shift
) + 1;
1777 unsigned groups_y
= bits(p
->invocation_count
, workgroups_y_shift
, workgroups_z_shift
) + 1;
1778 unsigned groups_z
= bits(p
->invocation_count
, workgroups_z_shift
, 32) + 1;
1780 /* Even though we have this decoded, we want to ensure that the
1781 * representation is "unique" so we don't lose anything by printing only
1782 * the final result. More specifically, we need to check that we were
1783 * passed something in canonical form, since the definition per the
1784 * hardware is inherently not unique. How? Well, take the resulting
1785 * decode and pack it ourselves! If it is bit exact with what we
1786 * decoded, we're good to go. */
1788 struct mali_vertex_tiler_prefix ref
;
1789 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, graphics
);
1792 (p
->invocation_count
== ref
.invocation_count
) &&
1793 (p
->invocation_shifts
== ref
.invocation_shifts
);
1796 pandecode_msg("XXX: non-canonical workgroups packing\n");
1797 pandecode_msg("expected: %X, %X",
1798 ref
.invocation_count
,
1799 ref
.invocation_shifts
);
1801 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1802 pandecode_prop("size_y_shift = %d", size_y_shift
);
1803 pandecode_prop("size_z_shift = %d", size_z_shift
);
1804 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift
);
1805 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift
);
1806 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift
);
1807 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2
);
1810 /* Regardless, print the decode */
1811 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1812 size_x
, size_y
, size_z
,
1813 groups_x
, groups_y
, groups_z
);
1816 if (p
->unknown_draw
)
1817 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1819 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1821 if (p
->draw_mode
!= MALI_DRAW_NONE
)
1822 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p
->draw_mode
));
1824 /* Index count only exists for tiler jobs anyway */
1827 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1830 unsigned index_raw_size
= (p
->unknown_draw
& MALI_DRAW_INDEXED_SIZE
);
1831 index_raw_size
>>= MALI_DRAW_INDEXED_SHIFT
;
1833 /* Validate an index buffer is present if we need one. TODO: verify
1834 * relationship between invocation_count and index_count */
1837 unsigned count
= p
->index_count
;
1840 unsigned size
= (index_raw_size
== 0x3) ? 4 : index_raw_size
;
1842 /* Ensure we got a size, and if so, validate the index buffer
1843 * is large enough to hold a full set of indices of the given
1846 if (!index_raw_size
)
1847 pandecode_msg("XXX: index size missing\n");
1849 pandecode_validate_buffer(p
->indices
, count
* size
);
1850 } else if (index_raw_size
)
1851 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size
);
1853 if (p
->offset_bias_correction
)
1854 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1856 /* TODO: Figure out what this is. It's not zero */
1857 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1860 pandecode_log("},\n");
1864 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1866 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1867 uint64_t *PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1869 for (int i
= 0; i
< ubufs_count
; i
++) {
1870 unsigned size
= (ubufs
[i
] & ((1 << 10) - 1)) * 16;
1871 mali_ptr addr
= (ubufs
[i
] >> 10) << 2;
1873 pandecode_validate_buffer(addr
, size
);
1875 char *ptr
= pointer_as_memory_reference(addr
);
1876 pandecode_log("ubuf_%d[%u] = %s;\n", i
, size
, ptr
);
1880 pandecode_log("\n");
1884 pandecode_uniforms(mali_ptr uniforms
, unsigned uniform_count
)
1886 pandecode_validate_buffer(uniforms
, uniform_count
* 16);
1888 char *ptr
= pointer_as_memory_reference(uniforms
);
1889 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count
, ptr
);
1894 shader_type_for_job(unsigned type
)
1897 case JOB_TYPE_VERTEX
: return "VERTEX";
1898 case JOB_TYPE_TILER
: return "FRAGMENT";
1899 case JOB_TYPE_COMPUTE
: return "COMPUTE";
1905 static unsigned shader_id
= 0;
1907 static struct midgard_disasm_stats
1908 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1909 bool is_bifrost
, unsigned gpu_id
)
1911 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1912 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1914 /* Compute maximum possible size */
1915 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1917 /* Print some boilerplate to clearly denote the assembly (which doesn't
1918 * obey indentation rules), and actually do the disassembly! */
1920 pandecode_log_cont("\n\n");
1922 struct midgard_disasm_stats stats
;
1925 disassemble_bifrost(pandecode_dump_stream
, code
, sz
, true);
1927 /* TODO: Extend stats to Bifrost */
1928 stats
.texture_count
= -128;
1929 stats
.sampler_count
= -128;
1930 stats
.attribute_count
= -128;
1931 stats
.varying_count
= -128;
1932 stats
.uniform_count
= -128;
1933 stats
.uniform_buffer_count
= -128;
1934 stats
.work_count
= -128;
1936 stats
.instruction_count
= 0;
1937 stats
.bundle_count
= 0;
1938 stats
.quadword_count
= 0;
1939 stats
.helper_invocations
= false;
1941 stats
= disassemble_midgard(pandecode_dump_stream
,
1943 type
== JOB_TYPE_TILER
?
1944 MESA_SHADER_FRAGMENT
: MESA_SHADER_VERTEX
);
1947 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1948 * driver-internal purposes with the blob and interfere */
1950 bool should_shaderdb
= type
!= JOB_TYPE_COMPUTE
;
1952 if (should_shaderdb
) {
1953 unsigned nr_threads
=
1954 (stats
.work_count
<= 4) ? 4 :
1955 (stats
.work_count
<= 8) ? 2 :
1958 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
1959 "%u inst, %u bundles, %u quadwords, "
1960 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1962 shader_type_for_job(type
),
1963 stats
.instruction_count
, stats
.bundle_count
, stats
.quadword_count
,
1964 stats
.work_count
, nr_threads
);
1972 pandecode_texture(mali_ptr u
,
1973 struct pandecode_mapped_memory
*tmem
,
1974 unsigned job_no
, unsigned tex
)
1976 struct mali_texture_descriptor
*PANDECODE_PTR_VAR(t
, tmem
, u
);
1978 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64
"_%d_%d = {\n", u
, job_no
, tex
);
1981 struct mali_texture_format f
= t
->format
;
1983 /* See the definiton of enum mali_texture_type */
1985 bool is_cube
= f
.type
== MALI_TEX_CUBE
;
1986 unsigned dimension
= is_cube
? 2 : f
.type
;
1988 pandecode_make_indent();
1990 /* TODO: Are there others? */
1991 bool is_zs
= f
.format
== MALI_Z32_UNORM
;
1993 /* Recall Z/S switched the meaning of linear/tiled .. */
1994 if (is_zs
&& f
.layout
== MALI_TEXTURE_LINEAR
)
1995 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1997 /* Print the layout. Default is linear; a modifier can denote AFBC or
1998 * u-interleaved/tiled modes */
2000 if (f
.layout
== MALI_TEXTURE_AFBC
)
2001 pandecode_log_cont("afbc");
2002 else if (f
.layout
== MALI_TEXTURE_TILED
)
2003 pandecode_log_cont("tiled");
2004 else if (f
.layout
== MALI_TEXTURE_LINEAR
)
2005 pandecode_log_cont("linear");
2007 pandecode_msg("XXX: invalid texture layout 0x%X\n", f
.layout
);
2009 pandecode_swizzle(t
->swizzle
, f
.format
);
2010 pandecode_log_cont(" ");
2012 /* Distinguish cube/2D with modifier */
2015 pandecode_log_cont("cube ");
2017 pandecode_format_short(f
.format
, f
.srgb
);
2018 pandecode_swizzle(f
.swizzle
, f
.format
);
2020 /* All four width/height/depth/array_size dimensions are present
2021 * regardless of the type of texture, but it is an error to have
2022 * non-zero dimensions for unused dimensions. Verify this. array_size
2023 * can always be set, as can width. */
2025 if (t
->height
&& dimension
< 2)
2026 pandecode_msg("XXX: nonzero height for <2D texture\n");
2028 if (t
->depth
&& dimension
< 3)
2029 pandecode_msg("XXX: nonzero depth for <2D texture\n");
2031 /* Print only the dimensions that are actually there */
2033 pandecode_log_cont(": %d", t
->width
+ 1);
2036 pandecode_log_cont("x%u", t
->height
+ 1);
2039 pandecode_log_cont("x%u", t
->depth
+ 1);
2042 pandecode_log_cont("[%u]", t
->array_size
+ 1);
2045 pandecode_log_cont(" mip %u", t
->levels
);
2047 pandecode_log_cont("\n");
2049 if (f
.unknown1
| f
.zero
) {
2050 pandecode_msg("XXX: texture format zero tripped\n");
2051 pandecode_prop("unknown1 = %" PRId32
, f
.unknown1
);
2052 pandecode_prop("zero = %" PRId32
, f
.zero
);
2056 pandecode_msg("XXX: expected unknown texture bit set\n");
2057 pandecode_prop("unknown2 = %" PRId32
, f
.unknown2
);
2060 if (t
->swizzle_zero
) {
2061 pandecode_msg("XXX: swizzle zero tripped\n");
2062 pandecode_prop("swizzle_zero = %d", t
->swizzle_zero
);
2065 if (t
->unknown3
| t
->unknown3A
| t
->unknown5
| t
->unknown6
| t
->unknown7
) {
2066 pandecode_msg("XXX: texture zero tripped\n");
2067 pandecode_prop("unknown3 = %" PRId16
, t
->unknown3
);
2068 pandecode_prop("unknown3A = %" PRId8
, t
->unknown3A
);
2069 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
2070 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
2071 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
2074 pandecode_log(".payload = {\n");
2077 /* A bunch of bitmap pointers follow.
2078 * We work out the correct number,
2079 * based on the mipmap/cubemap
2080 * properties, but dump extra
2081 * possibilities to futureproof */
2083 int bitmap_count
= t
->levels
+ 1;
2085 /* Miptree for each face */
2086 if (f
.type
== MALI_TEX_CUBE
)
2088 else if (f
.type
== MALI_TEX_3D
&& f
.layout
== MALI_TEXTURE_LINEAR
)
2089 bitmap_count
*= (t
->depth
+ 1);
2091 /* Array of textures */
2092 bitmap_count
*= (t
->array_size
+ 1);
2094 /* Stride for each element */
2095 if (f
.manual_stride
)
2098 mali_ptr
*pointers_and_strides
= pandecode_fetch_gpu_mem(tmem
,
2099 u
+ sizeof(*t
), sizeof(mali_ptr
) * bitmap_count
);
2100 for (int i
= 0; i
< bitmap_count
; ++i
) {
2101 /* How we dump depends if this is a stride or a pointer */
2103 if (f
.manual_stride
&& (i
& 1)) {
2104 /* signed 32-bit snuck in as a 64-bit pointer */
2105 uint64_t stride_set
= pointers_and_strides
[i
];
2106 uint32_t clamped_stride
= stride_set
;
2107 int32_t stride
= clamped_stride
;
2108 assert(stride_set
== clamped_stride
);
2109 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
2111 char *a
= pointer_as_memory_reference(pointers_and_strides
[i
]);
2112 pandecode_log("%s, \n", a
);
2118 pandecode_log("},\n");
2121 pandecode_log("};\n");
2124 /* 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 */
2127 pandecode_shader_prop(const char *name
, unsigned claim
, signed truth
, bool fuzzy
)
2133 if (fuzzy
&& (truth
< 0))
2134 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name
, claim
, truth
);
2136 if ((truth
>= 0) && !fuzzy
) {
2137 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2138 (truth
< claim
) ? "warn" : "XXX",
2139 name
, truth
, claim
);
2140 } else if ((claim
> -truth
) && !fuzzy
) {
2141 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2142 name
, -truth
, claim
);
2143 } else if (fuzzy
&& (claim
< truth
))
2144 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2145 name
, truth
, claim
);
2147 pandecode_log(".%s = %" PRId16
, name
, claim
);
2150 pandecode_log_cont(" /* %u used */", truth
);
2152 pandecode_log_cont(",\n");
2156 pandecode_blend_shader_disassemble(mali_ptr shader
, int job_no
, int job_type
,
2157 bool is_bifrost
, unsigned gpu_id
)
2159 struct midgard_disasm_stats stats
=
2160 pandecode_shader_disassemble(shader
, job_no
, job_type
, is_bifrost
, gpu_id
);
2162 bool has_texture
= (stats
.texture_count
> 0);
2163 bool has_sampler
= (stats
.sampler_count
> 0);
2164 bool has_attribute
= (stats
.attribute_count
> 0);
2165 bool has_varying
= (stats
.varying_count
> 0);
2166 bool has_uniform
= (stats
.uniform_count
> 0);
2167 bool has_ubo
= (stats
.uniform_buffer_count
> 0);
2169 if (has_texture
|| has_sampler
)
2170 pandecode_msg("XXX: blend shader accessing textures\n");
2172 if (has_attribute
|| has_varying
)
2173 pandecode_msg("XXX: blend shader accessing interstage\n");
2175 if (has_uniform
|| has_ubo
)
2176 pandecode_msg("XXX: blend shader accessing uniforms\n");
2180 pandecode_vertex_tiler_postfix_pre(
2181 const struct mali_vertex_tiler_postfix
*p
,
2182 int job_no
, enum mali_job_type job_type
,
2183 char *suffix
, bool is_bifrost
, unsigned gpu_id
)
2185 struct pandecode_mapped_memory
*attr_mem
;
2187 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2188 * are the only things actually needed from the FBD, vertex/tiler jobs
2189 * no longer reference the FBD -- instead, this field points to some
2190 * info about the scratchpad.
2193 struct pandecode_fbd fbd_info
= {
2194 /* Default for Bifrost */
2199 pandecode_log_cont("\t/* %X %/\n", p
->shared_memory
& 1);
2200 pandecode_compute_fbd(p
->shared_memory
& ~1, job_no
);
2201 } else if (p
->shared_memory
& MALI_MFBD
)
2202 fbd_info
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->shared_memory
) & FBD_MASK
, job_no
, false, job_type
== JOB_TYPE_COMPUTE
, false);
2203 else if (job_type
== JOB_TYPE_COMPUTE
)
2204 pandecode_compute_fbd((u64
) (uintptr_t) p
->shared_memory
, job_no
);
2206 fbd_info
= pandecode_sfbd((u64
) (uintptr_t) p
->shared_memory
, job_no
, false, gpu_id
);
2208 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
2209 int texture_count
= 0, sampler_count
= 0;
2212 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->shader
);
2213 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, p
->shader
);
2215 /* Disassemble ahead-of-time to get stats. Initialize with
2216 * stats for the missing-shader case so we get validation
2219 struct midgard_disasm_stats info
= {
2222 .attribute_count
= 0,
2226 .uniform_count
= -128,
2227 .uniform_buffer_count
= 0
2230 if (s
->shader
& ~0xF)
2231 info
= pandecode_shader_disassemble(s
->shader
& ~0xF, job_no
, job_type
, is_bifrost
, gpu_id
);
2233 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", p
->shader
, job_no
, suffix
);
2236 /* Save for dumps */
2237 attribute_count
= s
->attribute_count
;
2238 varying_count
= s
->varying_count
;
2239 texture_count
= s
->texture_count
;
2240 sampler_count
= s
->sampler_count
;
2243 uniform_count
= s
->bifrost2
.uniform_count
;
2244 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
2246 uniform_count
= s
->midgard1
.uniform_count
;
2247 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
2250 pandecode_shader_address("shader", s
->shader
);
2252 pandecode_shader_prop("texture_count", s
->texture_count
, info
.texture_count
, false);
2253 pandecode_shader_prop("sampler_count", s
->sampler_count
, info
.sampler_count
, false);
2254 pandecode_shader_prop("attribute_count", s
->attribute_count
, info
.attribute_count
, false);
2255 pandecode_shader_prop("varying_count", s
->varying_count
, info
.varying_count
, false);
2256 pandecode_shader_prop("uniform_buffer_count",
2257 uniform_buffer_count
,
2258 info
.uniform_buffer_count
, true);
2261 pandecode_shader_prop("uniform_count",
2263 info
.uniform_count
, false);
2265 pandecode_shader_prop("work_count",
2266 s
->midgard1
.work_count
, info
.work_count
, false);
2270 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32
, s
->bifrost1
.unk1
);
2272 bool helpers
= s
->midgard1
.flags_lo
& MALI_HELPER_INVOCATIONS
;
2273 s
->midgard1
.flags_lo
&= ~MALI_HELPER_INVOCATIONS
;
2275 if (helpers
!= info
.helper_invocations
) {
2276 pandecode_msg("XXX: expected helpers %u but got %u\n",
2277 info
.helper_invocations
, helpers
);
2280 pandecode_log(".midgard1.flags_lo = ");
2281 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info
, s
->midgard1
.flags_lo
);
2282 pandecode_log_cont(",\n");
2284 pandecode_log(".midgard1.flags_hi = ");
2285 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info
, s
->midgard1
.flags_hi
);
2286 pandecode_log_cont(",\n");
2289 if (s
->depth_units
|| s
->depth_factor
) {
2290 pandecode_prop("depth_factor = %f", s
->depth_factor
);
2291 pandecode_prop("depth_units = %f", s
->depth_units
);
2294 if (s
->alpha_coverage
) {
2295 bool invert_alpha_coverage
= s
->alpha_coverage
& 0xFFF0;
2296 uint16_t inverted_coverage
= invert_alpha_coverage
? ~s
->alpha_coverage
: s
->alpha_coverage
;
2298 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2299 invert_alpha_coverage
? "~" : "",
2300 MALI_GET_ALPHA_COVERAGE(inverted_coverage
));
2303 if (s
->unknown2_3
|| s
->unknown2_4
) {
2304 pandecode_log(".unknown2_3 = ");
2306 int unknown2_3
= s
->unknown2_3
;
2307 int unknown2_4
= s
->unknown2_4
;
2309 /* We're not quite sure what these flags mean without the depth test, if anything */
2311 if (unknown2_3
& (MALI_DEPTH_WRITEMASK
| MALI_DEPTH_FUNC_MASK
)) {
2312 const char *func
= pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3
));
2313 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
2315 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
2318 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
2319 pandecode_log_cont(",\n");
2321 pandecode_log(".unknown2_4 = ");
2322 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
2323 pandecode_log_cont(",\n");
2326 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
2327 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
2328 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
2331 pandecode_stencil("front", &s
->stencil_front
);
2332 pandecode_stencil("back", &s
->stencil_back
);
2335 pandecode_log(".bifrost2 = {\n");
2338 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
2339 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
2340 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
2341 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
2344 pandecode_log("},\n");
2345 } else if (s
->midgard2
.unknown2_7
) {
2346 pandecode_log(".midgard2 = {\n");
2349 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
2351 pandecode_log("},\n");
2355 pandecode_msg("XXX: shader padding tripped\n");
2356 pandecode_prop("padding = 0x%" PRIx32
, s
->padding
);
2360 /* TODO: Blend shaders routing/disasm */
2361 union midgard_blend blend
= s
->blend
;
2362 mali_ptr shader
= pandecode_midgard_blend(&blend
, s
->unknown2_3
& MALI_HAS_BLEND_SHADER
);
2364 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
2368 pandecode_log("};\n");
2370 /* MRT blend fields are used whenever MFBD is used, with
2371 * per-RT descriptors */
2373 if (job_type
== JOB_TYPE_TILER
&& (is_bifrost
|| p
->shared_memory
& MALI_MFBD
)) {
2374 void* blend_base
= (void *) (s
+ 1);
2376 for (unsigned i
= 0; i
< fbd_info
.rt_count
; i
++) {
2377 mali_ptr shader
= 0;
2380 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
2382 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
2385 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
2390 pandecode_msg("XXX: missing shader descriptor\n");
2393 struct pandecode_mapped_memory
*fmem
= pandecode_find_mapped_gpu_mem_containing(p
->viewport
);
2394 struct mali_viewport
*PANDECODE_PTR_VAR(f
, fmem
, p
->viewport
);
2396 pandecode_log("struct mali_viewport viewport_%"PRIx64
"_%d%s = {\n", p
->viewport
, job_no
, suffix
);
2399 pandecode_prop("clip_minx = %f", f
->clip_minx
);
2400 pandecode_prop("clip_miny = %f", f
->clip_miny
);
2401 pandecode_prop("clip_minz = %f", f
->clip_minz
);
2402 pandecode_prop("clip_maxx = %f", f
->clip_maxx
);
2403 pandecode_prop("clip_maxy = %f", f
->clip_maxy
);
2404 pandecode_prop("clip_maxz = %f", f
->clip_maxz
);
2406 /* Only the higher coordinates are MALI_POSITIVE scaled */
2408 pandecode_prop("viewport0 = { %d, %d }",
2409 f
->viewport0
[0], f
->viewport0
[1]);
2411 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2412 f
->viewport1
[0] + 1, f
->viewport1
[1] + 1);
2415 pandecode_log("};\n");
2418 unsigned max_attr_index
= 0;
2420 if (p
->attribute_meta
)
2421 max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
2423 if (p
->attributes
) {
2424 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
2425 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false, job_type
);
2428 /* Varyings are encoded like attributes but not actually sent; we just
2429 * pass a zero buffer with the right stride/size set, (or whatever)
2430 * since the GPU will write to it itself */
2432 if (p
->varying_meta
) {
2433 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
2437 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
2439 /* Number of descriptors depends on whether there are
2440 * non-internal varyings */
2442 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true, job_type
);
2445 if (p
->uniform_buffers
) {
2446 if (uniform_buffer_count
)
2447 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
2449 pandecode_msg("warn: UBOs specified but not referenced\n");
2450 } else if (uniform_buffer_count
)
2451 pandecode_msg("XXX: UBOs referenced but not specified\n");
2453 /* We don't want to actually dump uniforms, but we do need to validate
2454 * that the counts we were given are sane */
2458 pandecode_uniforms(p
->uniforms
, uniform_count
);
2460 pandecode_msg("warn: Uniforms specified but not referenced\n");
2461 } else if (uniform_count
)
2462 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2464 if (p
->texture_trampoline
) {
2465 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(p
->texture_trampoline
);
2468 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
);
2470 pandecode_log("uint64_t texture_trampoline_%"PRIx64
"_%d[] = {\n", p
->texture_trampoline
, job_no
);
2473 for (int tex
= 0; tex
< texture_count
; ++tex
) {
2474 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
2475 char *a
= pointer_as_memory_reference(*u
);
2476 pandecode_log("%s,\n", a
);
2481 pandecode_log("};\n");
2483 /* Now, finally, descend down into the texture descriptor */
2484 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
2485 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
2486 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
2488 pandecode_texture(*u
, tmem
, job_no
, tex
);
2493 if (p
->sampler_descriptor
) {
2494 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->sampler_descriptor
);
2497 struct mali_sampler_descriptor
*s
;
2499 mali_ptr d
= p
->sampler_descriptor
;
2501 for (int i
= 0; i
< sampler_count
; ++i
) {
2502 s
= pandecode_fetch_gpu_mem(smem
, d
+ sizeof(*s
) * i
, sizeof(*s
));
2504 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", d
+ sizeof(*s
) * i
, job_no
, i
);
2507 pandecode_log(".filter_mode = ");
2508 pandecode_log_decoded_flags(sampler_flag_info
, s
->filter_mode
);
2509 pandecode_log_cont(",\n");
2511 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
2512 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
2515 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s
->lod_bias
));
2517 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s
->wrap_s
));
2518 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s
->wrap_t
));
2519 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s
->wrap_r
));
2521 pandecode_prop("compare_func = %s", pandecode_func(s
->compare_func
));
2523 if (s
->zero
|| s
->zero2
) {
2524 pandecode_msg("XXX: sampler zero tripped\n");
2525 pandecode_prop("zero = 0x%X, 0x%X\n", s
->zero
, s
->zero2
);
2528 pandecode_prop("seamless_cube_map = %d", s
->seamless_cube_map
);
2530 pandecode_prop("border_color = { %f, %f, %f, %f }",
2534 s
->border_color
[3]);
2537 pandecode_log("};\n");
2544 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2546 pandecode_log(".gl_enables = ");
2548 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2550 pandecode_log_cont(",\n");
2554 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2556 if (p
->shader
& 0xF)
2557 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p
->shader
& 0xF));
2559 pandecode_log(".postfix = {\n");
2562 pandecode_gl_enables(p
->gl_enables
, JOB_TYPE_TILER
);
2563 pandecode_prop("instance_shift = 0x%x", p
->instance_shift
);
2564 pandecode_prop("instance_odd = 0x%x", p
->instance_odd
);
2567 pandecode_msg("XXX: vertex only zero tripped");
2568 pandecode_prop("zero4 = 0x%" PRIx32
, p
->zero4
);
2571 pandecode_prop("offset_start = 0x%x", p
->offset_start
);
2574 pandecode_msg("XXX: vertex only zero tripped");
2575 pandecode_prop("zero5 = 0x%" PRIx32
, p
->zero5
);
2578 MEMORY_PROP(p
, position_varying
);
2579 MEMORY_PROP(p
, occlusion_counter
);
2582 pandecode_log("},\n");
2586 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2588 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2589 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2591 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2595 pandecode_msg("XXX: tiler heap zero tripped\n");
2596 pandecode_prop("zero = 0x%x", h
->zero
);
2599 for (int i
= 0; i
< 12; i
++) {
2600 if (h
->zeros
[i
] != 0) {
2601 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2606 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2607 MEMORY_PROP(h
, tiler_heap_start
);
2608 MEMORY_PROP(h
, tiler_heap_free
);
2610 /* this might point to the beginning of another buffer, when it's
2611 * really the end of the tiler heap buffer, so we have to be careful
2612 * here. but for zero length, we need the same pointer.
2615 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2616 MEMORY_PROP(h
, tiler_heap_start
);
2618 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2619 pandecode_prop("tiler_heap_end = %s + 1", a
);
2624 pandecode_log("};\n");
2628 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2630 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2631 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2633 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2635 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2638 if (t
->zero0
|| t
->zero1
) {
2639 pandecode_msg("XXX: tiler meta zero tripped\n");
2640 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2641 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2644 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2645 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2647 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2648 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2650 for (int i
= 0; i
< 12; i
++) {
2651 if (t
->zeros
[i
] != 0) {
2652 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2658 pandecode_log("};\n");
2662 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2664 if (u
.pointer
== 0x0)
2667 pandecode_log(".primitive_size = {\n");
2671 pandecode_prop("constant = %f", u
.constant
);
2673 MEMORY_PROP((&u
), pointer
);
2677 pandecode_log("},\n");
2681 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2683 pandecode_log_cont("{\n");
2686 /* TODO: gl_PointSize on Bifrost */
2687 pandecode_primitive_size(t
->primitive_size
, true);
2689 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2691 pandecode_msg("XXX: tiler only zero tripped\n");
2692 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2693 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2694 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2695 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2696 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2697 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2701 pandecode_log("},\n");
2705 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2706 const struct pandecode_mapped_memory
*mem
,
2707 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2709 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2711 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2713 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2716 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2717 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2720 pandecode_log("};\n");
2726 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2727 const struct pandecode_mapped_memory
*mem
,
2728 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2730 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2732 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2733 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2735 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2738 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2740 pandecode_log(".tiler = ");
2741 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2743 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2746 pandecode_log("};\n");
2752 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2753 const struct pandecode_mapped_memory
*mem
,
2754 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2756 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2757 bool is_graphics
= (h
->job_type
== JOB_TYPE_VERTEX
) || (h
->job_type
== JOB_TYPE_TILER
);
2759 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false, gpu_id
);
2761 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2764 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, is_graphics
);
2765 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2767 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2768 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2771 pandecode_log("};\n");
2777 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2778 mali_ptr payload
, int job_no
,
2779 bool is_bifrost
, unsigned gpu_id
)
2781 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2783 bool is_mfbd
= s
->framebuffer
& MALI_MFBD
;
2785 if (!is_mfbd
&& is_bifrost
)
2786 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2788 struct pandecode_fbd info
;
2791 info
= pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true, false, is_bifrost
);
2793 info
= pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true, gpu_id
);
2795 /* Compute the tag for the tagged pointer. This contains the type of
2796 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2797 * additional structures follow the MFBD header (an extra payload or
2798 * not, as well as a count of render targets) */
2800 unsigned expected_tag
= is_mfbd
? MALI_MFBD
: 0;
2804 expected_tag
|= MALI_MFBD_TAG_EXTRA
;
2806 expected_tag
|= (MALI_POSITIVE(info
.rt_count
) << 2);
2809 if ((s
->min_tile_coord
| s
->max_tile_coord
) & ~(MALI_X_COORD_MASK
| MALI_Y_COORD_MASK
)) {
2810 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2811 pandecode_prop("min_tile_coord = 0x%X\n", s
->min_tile_coord
);
2812 pandecode_prop("max_tile_coord = 0x%X\n", s
->min_tile_coord
);
2815 /* Extract tile coordinates */
2817 unsigned min_x
= MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2818 unsigned min_y
= MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2820 unsigned max_x
= (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2821 unsigned max_y
= (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2823 /* For the max, we also want the floored (rather than ceiled) version for checking */
2825 unsigned max_x_f
= (MALI_TILE_COORD_X(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2826 unsigned max_y_f
= (MALI_TILE_COORD_Y(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2828 /* Validate the coordinates are well-ordered */
2831 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2832 else if (min_x
> max_x
)
2833 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2836 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2837 else if (min_y
> max_y
)
2838 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2840 /* Validate the coordinates fit inside the framebuffer. We use floor,
2841 * rather than ceil, for the max coordinates, since the tile
2842 * coordinates for something like an 800x600 framebuffer will actually
2843 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2845 if ((min_x
> info
.width
) || (max_x_f
> info
.width
))
2846 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2848 if ((min_y
> info
.height
) || (max_y_f
> info
.height
))
2849 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2851 /* After validation, we print */
2853 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x
, min_y
, max_x
, max_y
);
2855 /* The FBD is a tagged pointer */
2857 unsigned tag
= (s
->framebuffer
& ~FBD_MASK
);
2859 if (tag
!= expected_tag
)
2860 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag
, tag
);
2865 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
2866 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
2867 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
2868 * might wish to combine this with the bifrost parameter) because some details
2869 * are model-specific even within a particular architecture. Minimal traces
2870 * *only* examine the job descriptors, skipping printing entirely if there is
2871 * no faults, and only descends into the payload if there are faults. This is
2872 * useful for looking for faults without the overhead of invasive traces. */
2875 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
, unsigned gpu_id
, bool minimal
)
2877 struct mali_job_descriptor_header
*h
;
2878 unsigned job_descriptor_number
= 0;
2881 struct pandecode_mapped_memory
*mem
=
2882 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2886 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2888 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2889 * high 32-bits of the 64-bit pointer are reused to store
2892 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
2893 h
->job_type
!= JOB_TYPE_FRAGMENT
? 4 : 0;
2894 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
2896 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
, 256);
2898 int job_no
= job_descriptor_number
++;
2900 /* If the job is good to go, skip it in minimal mode */
2901 if (minimal
&& (h
->exception_status
== 0x0 || h
->exception_status
== 0x1))
2904 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2907 pandecode_prop("job_type = %s", pandecode_job_type(h
->job_type
));
2909 if (h
->job_descriptor_size
)
2910 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2912 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2913 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2914 h
->exception_status
,
2915 (h
->exception_status
>> 16) & 0xFFFF,
2916 pandecode_exception_access((h
->exception_status
>> 8) & 0x3),
2917 h
->exception_status
& 0xFF);
2919 if (h
->first_incomplete_task
)
2920 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2922 if (h
->fault_pointer
)
2923 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2926 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2928 pandecode_prop("job_index = %d", h
->job_index
);
2930 if (h
->unknown_flags
)
2931 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2933 if (h
->job_dependency_index_1
)
2934 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2936 if (h
->job_dependency_index_2
)
2937 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2940 pandecode_log("};\n");
2942 switch (h
->job_type
) {
2943 case JOB_TYPE_WRITE_VALUE
: {
2944 struct mali_payload_write_value
*s
= payload
;
2945 pandecode_log("struct mali_payload_write_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2947 MEMORY_PROP(s
, address
);
2949 if (s
->value_descriptor
!= MALI_WRITE_VALUE_ZERO
) {
2950 pandecode_msg("XXX: unknown value descriptor\n");
2951 pandecode_prop("value_descriptor = 0x%" PRIX32
, s
->value_descriptor
);
2955 pandecode_msg("XXX: set value tripped\n");
2956 pandecode_prop("reserved = 0x%" PRIX32
, s
->reserved
);
2959 pandecode_prop("immediate = 0x%" PRIX64
, s
->immediate
);
2961 pandecode_log("};\n");
2966 case JOB_TYPE_TILER
:
2967 case JOB_TYPE_VERTEX
:
2968 case JOB_TYPE_COMPUTE
:
2970 if (h
->job_type
== JOB_TYPE_TILER
)
2971 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2973 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2975 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2979 case JOB_TYPE_FRAGMENT
:
2980 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
, gpu_id
);
2986 } while ((jc_gpu_va
= h
->next_job
));