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 int pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
);
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 /* Semantic logging type.
63 * Raw: for raw messages to be printed as is.
64 * Message: for helpful information to be commented out in replays.
65 * Property: for properties of a struct
67 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
70 enum pandecode_log_type
{
76 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
77 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
78 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
80 unsigned pandecode_indent
= 0;
83 pandecode_make_indent(void)
85 for (unsigned i
= 0; i
< pandecode_indent
; ++i
)
90 pandecode_log_typed(enum pandecode_log_type type
, const char *format
, ...)
94 pandecode_make_indent();
96 if (type
== PANDECODE_MESSAGE
)
98 else if (type
== PANDECODE_PROPERTY
)
101 va_start(ap
, format
);
105 if (type
== PANDECODE_PROPERTY
)
110 pandecode_log_cont(const char *format
, ...)
114 va_start(ap
, format
);
119 /* To check for memory safety issues, validates that the given pointer in GPU
120 * memory is valid, containing at least sz bytes. The goal is to eliminate
121 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
122 * overruns) by statically validating pointers.
126 pandecode_validate_buffer(mali_ptr addr
, size_t sz
)
129 pandecode_msg("XXX: null pointer deref");
135 struct pandecode_mapped_memory
*bo
=
136 pandecode_find_mapped_gpu_mem_containing(addr
);
139 pandecode_msg("XXX: invalid memory dereference\n");
145 unsigned offset
= addr
- bo
->gpu_va
;
146 unsigned total
= offset
+ sz
;
148 if (total
> bo
->length
) {
149 pandecode_msg("XXX: buffer overrun."
150 "Chunk of size %d at offset %d in buffer of size %d. "
151 "Overrun by %d bytes.",
152 sz
, offset
, bo
->length
, total
- bo
->length
);
157 struct pandecode_flag_info
{
163 pandecode_log_decoded_flags(const struct pandecode_flag_info
*flag_info
,
166 bool decodable_flags_found
= false;
168 for (int i
= 0; flag_info
[i
].name
; i
++) {
169 if ((flags
& flag_info
[i
].flag
) != flag_info
[i
].flag
)
172 if (!decodable_flags_found
) {
173 decodable_flags_found
= true;
175 pandecode_log_cont(" | ");
178 pandecode_log_cont("%s", flag_info
[i
].name
);
180 flags
&= ~flag_info
[i
].flag
;
183 if (decodable_flags_found
) {
185 pandecode_log_cont(" | 0x%" PRIx64
, flags
);
187 pandecode_log_cont("0x%" PRIx64
, flags
);
191 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
192 static const struct pandecode_flag_info gl_enable_flag_info
[] = {
193 FLAG_INFO(OCCLUSION_QUERY
),
194 FLAG_INFO(OCCLUSION_PRECISE
),
195 FLAG_INFO(FRONT_CCW_TOP
),
196 FLAG_INFO(CULL_FACE_FRONT
),
197 FLAG_INFO(CULL_FACE_BACK
),
202 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
203 static const struct pandecode_flag_info clear_flag_info
[] = {
206 FLAG_INFO(SLOW_STENCIL
),
211 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
212 static const struct pandecode_flag_info mask_flag_info
[] = {
221 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
222 static const struct pandecode_flag_info u3_flag_info
[] = {
224 FLAG_INFO(CAN_DISCARD
),
225 FLAG_INFO(HAS_BLEND_SHADER
),
226 FLAG_INFO(DEPTH_TEST
),
230 static const struct pandecode_flag_info u4_flag_info
[] = {
232 FLAG_INFO(NO_DITHER
),
233 FLAG_INFO(DEPTH_RANGE_A
),
234 FLAG_INFO(DEPTH_RANGE_B
),
235 FLAG_INFO(STENCIL_TEST
),
236 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER
),
241 #define FLAG_INFO(flag) { MALI_FRAMEBUFFER_##flag, "MALI_FRAMEBUFFER_" #flag }
242 static const struct pandecode_flag_info fb_fmt_flag_info
[] = {
250 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
251 static const struct pandecode_flag_info mfbd_fmt_flag_info
[] = {
258 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
259 static const struct pandecode_flag_info mfbd_extra_flag_info
[] = {
267 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
268 static const struct pandecode_flag_info shader_midgard1_flag_info
[] = {
270 FLAG_INFO(HELPER_INVOCATIONS
),
271 FLAG_INFO(READS_TILEBUFFER
),
277 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
278 static const struct pandecode_flag_info mfbd_flag_info
[] = {
279 FLAG_INFO(DEPTH_WRITE
),
285 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
286 static const struct pandecode_flag_info sampler_flag_info
[] = {
287 FLAG_INFO(MAG_NEAREST
),
288 FLAG_INFO(MIN_NEAREST
),
289 FLAG_INFO(MIP_LINEAR_1
),
290 FLAG_INFO(MIP_LINEAR_2
),
291 FLAG_INFO(NORM_COORDS
),
296 extern char *replace_fragment
;
297 extern char *replace_vertex
;
300 pandecode_job_type(enum mali_job_type type
)
302 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
306 DEFINE_CASE(SET_VALUE
);
307 DEFINE_CASE(CACHE_FLUSH
);
308 DEFINE_CASE(COMPUTE
);
312 DEFINE_CASE(FRAGMENT
);
314 case JOB_NOT_STARTED
:
315 return "NOT_STARTED";
318 pandecode_log("Warning! Unknown job type %x\n", type
);
326 pandecode_draw_mode(enum mali_draw_mode mode
)
328 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
331 DEFINE_CASE(DRAW_NONE
);
334 DEFINE_CASE(TRIANGLES
);
335 DEFINE_CASE(TRIANGLE_STRIP
);
336 DEFINE_CASE(TRIANGLE_FAN
);
337 DEFINE_CASE(LINE_STRIP
);
338 DEFINE_CASE(LINE_LOOP
);
339 DEFINE_CASE(POLYGON
);
341 DEFINE_CASE(QUAD_STRIP
);
344 pandecode_msg("XXX: invalid draw mode %X\n", mode
);
351 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
353 pandecode_func(enum mali_func mode
)
360 DEFINE_CASE(GREATER
);
361 DEFINE_CASE(NOTEQUAL
);
366 pandecode_msg("XXX: invalid func %X\n", mode
);
372 /* Why is this duplicated? Who knows... */
373 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
375 pandecode_alt_func(enum mali_alt_func mode
)
382 DEFINE_CASE(GREATER
);
383 DEFINE_CASE(NOTEQUAL
);
388 pandecode_msg("XXX: invalid alt func %X\n", mode
);
394 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
396 pandecode_stencil_op(enum mali_stencil_op op
)
400 DEFINE_CASE(REPLACE
);
403 DEFINE_CASE(INCR_WRAP
);
404 DEFINE_CASE(DECR_WRAP
);
409 pandecode_msg("XXX: invalid stencil op %X\n", op
);
416 #define DEFINE_CASE(name) case MALI_ATTR_ ## name: return "MALI_ATTR_" #name
417 static char *pandecode_attr_mode(enum mali_attr_mode mode
)
422 DEFINE_CASE(POT_DIVIDE
);
424 DEFINE_CASE(NPOT_DIVIDE
);
426 DEFINE_CASE(INTERNAL
);
428 pandecode_msg("XXX: invalid attribute mode %X\n", mode
);
435 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
437 pandecode_wrap_mode(enum mali_wrap_mode op
)
441 DEFINE_CASE(CLAMP_TO_EDGE
);
442 DEFINE_CASE(CLAMP_TO_BORDER
);
443 DEFINE_CASE(MIRRORED_REPEAT
);
446 pandecode_msg("XXX: invalid wrap mode %X\n", op
);
452 #define DEFINE_CASE(name) case MALI_TEX_## name: return "MALI_TEX_" #name
454 pandecode_texture_type(enum mali_texture_type type
)
463 unreachable("Unknown case");
468 #define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
470 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt
)
474 DEFINE_CASE(UNKNOWN
);
479 unreachable("Invalid case");
484 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
486 pandecode_exception_access(enum mali_exception_access access
)
490 DEFINE_CASE(EXECUTE
);
495 unreachable("Invalid case");
500 /* Midgard's tiler descriptor is embedded within the
504 pandecode_midgard_tiler_descriptor(
505 const struct midgard_tiler_descriptor
*t
,
510 pandecode_log(".tiler = {\n");
513 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
514 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
516 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
518 /* We know this name from the kernel, but we never see it nonzero */
521 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
523 MEMORY_PROP(t
, polygon_list
);
525 /* The body is offset from the base of the polygon list */
526 assert(t
->polygon_list_body
> t
->polygon_list
);
527 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
529 /* It needs to fit inside the reported size */
530 assert(t
->polygon_list_size
>= body_offset
);
532 /* Check that we fit */
533 struct pandecode_mapped_memory
*plist
=
534 pandecode_find_mapped_gpu_mem_containing(t
->polygon_list
);
536 assert(t
->polygon_list_size
<= plist
->length
);
538 /* Now that we've sanity checked, we'll try to calculate the sizes
539 * ourselves for comparison */
541 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
);
542 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
);
544 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
545 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
546 ref_header
, ref_size
);
547 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
548 pandecode_msg("body offset %d\n", body_offset
);
551 /* The tiler heap has a start and end specified -- it should be
552 * identical to what we have in the BO. The exception is if tiling is
555 MEMORY_PROP(t
, heap_start
);
556 assert(t
->heap_end
>= t
->heap_start
);
558 struct pandecode_mapped_memory
*heap
=
559 pandecode_find_mapped_gpu_mem_containing(t
->heap_start
);
561 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
563 /* Tiling is enabled with a special flag */
564 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
565 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
567 bool tiling_enabled
= hierarchy_mask
;
569 if (tiling_enabled
) {
570 /* When tiling is enabled, the heap should be a tight fit */
571 unsigned heap_offset
= t
->heap_start
- heap
->gpu_va
;
572 if ((heap_offset
+ heap_size
) != heap
->length
) {
573 pandecode_msg("XXX: heap size %d (expected %d)\n",
574 heap_size
, heap
->length
- heap_offset
);
577 /* We should also have no other flags */
579 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
581 /* When tiling is disabled, we should have that flag and no others */
583 if (tiler_flags
!= MALI_TILER_DISABLED
) {
584 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
588 /* We should also have an empty heap */
590 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
594 /* Disabled tiling is used only for clear-only jobs, which are
595 * purely FRAGMENT, so we should never see this for
596 * non-FRAGMENT descriptors. */
599 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
602 /* We've never seen weights used in practice, but we know from the
603 * kernel these fields is there */
605 bool nonzero_weights
= false;
607 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
608 nonzero_weights
|= t
->weights
[w
] != 0x0;
611 if (nonzero_weights
) {
612 pandecode_log(".weights = {");
614 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
615 pandecode_log("%d, ", t
->weights
[w
]);
622 pandecode_log("}\n");
626 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
)
628 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
629 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
631 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
634 pandecode_prop("unknown1 = 0x%" PRIx32
, s
->unknown1
);
635 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->unknown2
);
637 pandecode_log(".format = ");
638 pandecode_log_decoded_flags(fb_fmt_flag_info
, s
->format
);
639 pandecode_log_cont(",\n");
641 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", s
->width
+ 1);
642 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", s
->height
+ 1);
644 MEMORY_PROP(s
, framebuffer
);
645 pandecode_prop("stride = %d", s
->stride
);
647 /* Earlier in the actual commandstream -- right before width -- but we
648 * delay to flow nicer */
650 pandecode_log(".clear_flags = ");
651 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
652 pandecode_log_cont(",\n");
654 if (s
->depth_buffer
| s
->depth_buffer_enable
) {
655 MEMORY_PROP(s
, depth_buffer
);
656 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s
->depth_buffer_enable
));
659 if (s
->stencil_buffer
| s
->stencil_buffer_enable
) {
660 MEMORY_PROP(s
, stencil_buffer
);
661 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s
->stencil_buffer_enable
));
664 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
665 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
666 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
667 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
668 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
671 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
672 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
673 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
674 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
675 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
678 if (s
->clear_stencil
) {
679 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
682 MEMORY_PROP(s
, unknown_address_0
);
683 const struct midgard_tiler_descriptor t
= s
->tiler
;
684 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
);
687 pandecode_log("};\n");
689 pandecode_prop("zero0 = 0x%" PRIx64
, s
->zero0
);
690 pandecode_prop("zero1 = 0x%" PRIx64
, s
->zero1
);
691 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
692 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
694 printf(".zero3 = {");
696 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
697 printf("%X, ", s
->zero3
[i
]);
701 printf(".zero6 = {");
703 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
704 printf("%X, ", s
->zero6
[i
]);
710 pandecode_u32_slide(unsigned name
, const u32
*slide
, unsigned count
)
712 pandecode_log(".unknown%d = {", name
);
714 for (int i
= 0; i
< count
; ++i
)
715 printf("%X, ", slide
[i
]);
717 pandecode_log("},\n");
720 #define SHORT_SLIDE(num) \
721 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
724 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
726 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
727 const struct mali_compute_fbd
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
729 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
738 /* Extracts the number of components associated with a Mali format */
741 pandecode_format_component_count(enum mali_format fmt
)
743 /* Mask out the format class */
744 unsigned top
= fmt
& 0b11100000;
747 case MALI_FORMAT_SNORM
:
748 case MALI_FORMAT_UINT
:
749 case MALI_FORMAT_UNORM
:
750 case MALI_FORMAT_SINT
:
751 return ((fmt
>> 3) & 3) + 1;
758 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
761 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
764 assert(MALI_CHANNEL_RED
== 0);
766 for (unsigned c
= 0; c
< 4; ++c
) {
767 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
769 if (chan
<= MALI_CHANNEL_ALPHA
)
776 /* Validates that a (format, swizzle) pair is valid, in the sense that the
777 * swizzle doesn't access any components that are undefined in the format.
778 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
782 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
784 unsigned nr_comp
= pandecode_format_component_count(fmt
);
785 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
786 unsigned valid_mask
= (1 << nr_comp
) - 1;
787 unsigned invalid_mask
= ~valid_mask
;
789 if (access_mask
& invalid_mask
) {
790 pandecode_msg("XXX: invalid components accessed\n");
794 /* Check for the default non-swizzling swizzle so we can suppress
795 * useless printing for the defaults */
797 unsigned default_swizzles
[4] = {
798 MALI_CHANNEL_RED
| (MALI_CHANNEL_ZERO
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
799 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
800 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ONE
<< 9),
801 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ALPHA
<< 9)
804 return (swizzle
== default_swizzles
[nr_comp
- 1]);
807 /* Maps MALI_RGBA32F to rgba32f, etc */
810 pandecode_format_short(enum mali_format fmt
, bool srgb
)
812 /* We want a type-like format, so cut off the initial MALI_ */
813 char *format
= pandecode_format(fmt
);
814 format
+= strlen("MALI_");
816 unsigned len
= strlen(format
);
817 char *lower_format
= calloc(1, len
+ 1);
819 for (unsigned i
= 0; i
< len
; ++i
)
820 lower_format
[i
] = tolower(format
[i
]);
822 /* Sanity check sRGB flag is applied to RGB, per the name */
823 if (srgb
&& lower_format
[0] != 'r')
824 pandecode_msg("XXX: sRGB applied to non-colour format\n");
826 /* Just prefix with an s, so you get formats like srgba8_unorm */
828 pandecode_log_cont("s");
830 pandecode_log_cont("%s", lower_format
);
835 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
837 /* First, do some validation */
838 bool trivial_swizzle
= pandecode_validate_format_swizzle(
844 /* Next, print the swizzle */
845 pandecode_log_cont(".");
847 static const char components
[] = "rgba01";
849 for (unsigned c
= 0; c
< 4; ++c
) {
850 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
852 if (chan
>= MALI_CHANNEL_RESERVED_0
) {
853 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
856 pandecode_log_cont("%c", components
[chan
]);
861 pandecode_rt_format(struct mali_rt_format format
)
863 pandecode_log(".format = {\n");
866 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
867 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
868 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
870 pandecode_prop("block = %s",
871 pandecode_mfbd_block_format(format
.block
));
873 /* TODO: Map formats so we can check swizzles and print nicely */
874 pandecode_log("swizzle");
875 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
876 pandecode_log_cont(",\n");
878 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
879 MALI_NEGATIVE(format
.nr_channels
));
881 pandecode_log(".flags = ");
882 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
883 pandecode_log_cont(",\n");
885 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
886 * which is a faster hardware-based alternative to the wallpaper method
887 * to preserve framebuffer contents across frames. In practice, MFBD
888 * preload is buggy on Midgard, and so this is a chicken bit. If this
889 * bit isn't set, most likely something broke unrelated to preload */
891 if (!format
.no_preload
) {
892 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
893 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
897 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
900 pandecode_log("},\n");
904 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct bifrost_framebuffer
*fb
)
906 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
909 for (int i
= 0; i
< MALI_NEGATIVE(fb
->rt_count_1
); i
++) {
910 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct bifrost_render_target
);
911 struct pandecode_mapped_memory
*mem
=
912 pandecode_find_mapped_gpu_mem_containing(rt_va
);
913 const struct bifrost_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
915 pandecode_log("{\n");
918 pandecode_rt_format(rt
->format
);
920 if (rt
->format
.block
== MALI_MFBD_BLOCK_AFBC
) {
921 pandecode_log(".afbc = {\n");
924 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
925 pandecode_prop("metadata = %s", a
);
928 pandecode_prop("stride = %d", rt
->afbc
.stride
);
929 pandecode_prop("unk = 0x%" PRIx32
, rt
->afbc
.unk
);
932 pandecode_log("},\n");
933 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.unk
) {
934 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
940 MEMORY_PROP(rt
, framebuffer
);
941 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
943 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
944 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
945 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
946 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
947 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
950 if (rt
->zero1
|| rt
->zero2
|| rt
->zero3
) {
951 pandecode_msg("XXX: render target zeros tripped\n");
952 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
953 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
954 pandecode_prop("zero3 = 0x%" PRIx32
, rt
->zero3
);
958 pandecode_log("},\n");
962 pandecode_log("};\n");
966 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
)
968 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
969 const struct bifrost_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
971 if (fb
->sample_locations
) {
972 /* The blob stores all possible sample locations in a single buffer
973 * allocated on startup, and just switches the pointer when switching
974 * MSAA state. For now, we just put the data into the cmdstream, but we
975 * should do something like what the blob does with a real driver.
977 * There seem to be 32 slots for sample locations, followed by another
978 * 16. The second 16 is just the center location followed by 15 zeros
979 * in all the cases I've identified (maybe shader vs. depth/color
983 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->sample_locations
);
985 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->sample_locations
);
987 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
990 for (int i
= 0; i
< 32 + 16; i
++) {
991 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
995 pandecode_log("};\n");
998 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1001 pandecode_prop("unk0 = 0x%x", fb
->unk0
);
1003 if (fb
->sample_locations
)
1004 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
1006 /* Assume that unknown1 was emitted in the last job for
1008 MEMORY_PROP(fb
, unknown1
);
1010 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
1011 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
1012 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
1013 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
1015 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
1016 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
1017 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
1018 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
1020 pandecode_log(".mfbd_flags = ");
1021 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
1022 pandecode_log_cont(",\n");
1024 if (fb
->clear_stencil
)
1025 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
1027 if (fb
->clear_depth
)
1028 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
1030 /* TODO: What is this? Let's not blow up.. */
1031 if (fb
->unknown2
!= 0x1F)
1032 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
1034 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
1035 MEMORY_PROP(fb
, scratchpad
);
1036 const struct midgard_tiler_descriptor t
= fb
->tiler
;
1037 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
);
1039 if (fb
->zero3
|| fb
->zero4
) {
1040 pandecode_msg("XXX: framebuffer zeros tripped\n");
1041 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
1042 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
1046 pandecode_log("};\n");
1048 gpu_va
+= sizeof(struct bifrost_framebuffer
);
1050 if ((fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
) {
1051 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1052 const struct bifrost_fb_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
1054 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1057 MEMORY_PROP(fbx
, checksum
);
1059 if (fbx
->checksum_stride
)
1060 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
1062 pandecode_log(".flags = ");
1063 pandecode_log_decoded_flags(mfbd_extra_flag_info
, fbx
->flags
);
1064 pandecode_log_cont(",\n");
1066 if (fbx
->flags
& MALI_EXTRA_AFBC_ZS
) {
1067 pandecode_log(".ds_afbc = {\n");
1070 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
1071 pandecode_prop("depth_stencil_afbc_stride = %d",
1072 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
1073 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
1075 if (fbx
->ds_afbc
.zero1
|| fbx
->ds_afbc
.padding
) {
1076 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1077 pandecode_prop("zero1 = 0x%" PRIx32
,
1078 fbx
->ds_afbc
.zero1
);
1079 pandecode_prop("padding = 0x%" PRIx64
,
1080 fbx
->ds_afbc
.padding
);
1084 pandecode_log("},\n");
1086 pandecode_log(".ds_linear = {\n");
1089 if (fbx
->ds_linear
.depth
) {
1090 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
1091 pandecode_prop("depth_stride = %d",
1092 fbx
->ds_linear
.depth_stride
);
1095 if (fbx
->ds_linear
.stencil
) {
1096 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1097 pandecode_prop("stencil_stride = %d",
1098 fbx
->ds_linear
.stencil_stride
);
1101 if (fbx
->ds_linear
.depth_stride_zero
||
1102 fbx
->ds_linear
.stencil_stride_zero
||
1103 fbx
->ds_linear
.zero1
|| fbx
->ds_linear
.zero2
) {
1104 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1105 pandecode_prop("depth_stride_zero = 0x%x",
1106 fbx
->ds_linear
.depth_stride_zero
);
1107 pandecode_prop("stencil_stride_zero = 0x%x",
1108 fbx
->ds_linear
.stencil_stride_zero
);
1109 pandecode_prop("zero1 = 0x%" PRIx32
,
1110 fbx
->ds_linear
.zero1
);
1111 pandecode_prop("zero2 = 0x%" PRIx32
,
1112 fbx
->ds_linear
.zero2
);
1116 pandecode_log("},\n");
1119 if (fbx
->zero3
|| fbx
->zero4
) {
1120 pandecode_msg("XXX: fb_extra zeros tripped\n");
1121 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1122 pandecode_prop("zero4 = 0x%" PRIx64
, fbx
->zero4
);
1126 pandecode_log("};\n");
1128 gpu_va
+= sizeof(struct bifrost_fb_extra
);
1132 pandecode_render_target(gpu_va
, job_no
, fb
);
1134 /* Passback the render target count */
1135 return MALI_NEGATIVE(fb
->rt_count_1
);
1138 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1142 pandecode_padded_vertices(unsigned shift
, unsigned k
)
1144 unsigned odd
= 2*k
+ 1;
1145 unsigned pot
= 1 << shift
;
1146 pandecode_msg("padded_num_vertices = %d\n", odd
* pot
);
1149 /* Given a magic divisor, recover what we were trying to divide by.
1151 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1152 * 0 <= m < 2^N, for N bits in m.
1154 * Let q represent the number we would like to divide by.
1156 * By definition of a magic divisor for N-bit unsigned integers (a number you
1157 * multiply by to magically get division), m is a number such that:
1159 * (m * x) & (2^N - 1) = floor(x/q).
1160 * for all x on Z where 0 <= x < 2^N
1162 * Ignore the case where any of the above values equals zero; it is irrelevant
1163 * for our purposes (instanced arrays).
1165 * Choose x = q. Then:
1167 * (m * x) & (2^N - 1) = floor(x/q).
1168 * (m * q) & (2^N - 1) = floor(q/q).
1170 * floor(q/q) = floor(1) = 1, therefore:
1172 * (m * q) & (2^N - 1) = 1
1174 * Recall the identity that the bitwise AND of one less than a power-of-two
1175 * equals the modulo with that power of two, i.e. for all x:
1177 * x & (2^N - 1) = x % N
1183 * By definition, a modular multiplicative inverse of a number m is the number
1184 * q such that with respect to a modulos M:
1188 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1193 pandecode_magic_divisor(uint32_t magic
, unsigned shift
, unsigned orig_divisor
, unsigned extra
)
1196 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1197 * shift) the most lame way possible... just repeatedly add.
1198 * Asymptoptically slow but nobody cares in practice, unless you have
1199 * massive numbers of vertices or high divisors. */
1201 unsigned inverse
= 0;
1203 /* Magic implicitly has the highest bit set */
1206 /* Depending on rounding direction */
1211 uint32_t product
= magic
* inverse
;
1223 pandecode_msg("dividing by %d (maybe off by two)\n", inverse
);
1225 /* Recall we're supposed to divide by (gl_level_divisor *
1226 * padded_num_vertices) */
1228 unsigned padded_num_vertices
= inverse
/ orig_divisor
;
1230 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices
);
1235 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1236 mali_ptr addr
, int job_no
, char *suffix
,
1237 int count
, bool varying
)
1239 char *prefix
= varying
? "varyings" : "attributes";
1242 pandecode_msg("no %s\n", prefix
);
1246 union mali_attr
*attr
= pandecode_fetch_gpu_mem(mem
, addr
, sizeof(union mali_attr
) * count
);
1249 snprintf(base
, sizeof(base
), "%s_data_%d%s", prefix
, job_no
, suffix
);
1251 for (int i
= 0; i
< count
; ++i
) {
1252 enum mali_attr_mode mode
= attr
[i
].elements
& 7;
1254 if (mode
== MALI_ATTR_UNUSED
)
1257 mali_ptr raw_elements
= attr
[i
].elements
& ~7;
1259 /* TODO: Do we maybe want to dump the attribute values
1260 * themselves given the specified format? Or is that too hard?
1263 char *a
= pointer_as_memory_reference(raw_elements
);
1264 pandecode_log("mali_ptr %s_%d_p = %s;\n", base
, i
, a
);
1268 pandecode_log("union mali_attr %s_%d[] = {\n", prefix
, job_no
);
1271 for (int i
= 0; i
< count
; ++i
) {
1272 pandecode_log("{\n");
1275 unsigned mode
= attr
[i
].elements
& 7;
1276 pandecode_prop("elements = (%s_%d_p) | %s", base
, i
, pandecode_attr_mode(mode
));
1277 pandecode_prop("shift = %d", attr
[i
].shift
);
1278 pandecode_prop("extra_flags = %d", attr
[i
].extra_flags
);
1279 pandecode_prop("stride = 0x%" PRIx32
, attr
[i
].stride
);
1280 pandecode_prop("size = 0x%" PRIx32
, attr
[i
].size
);
1282 /* Decode further where possible */
1284 if (mode
== MALI_ATTR_MODULO
) {
1285 pandecode_padded_vertices(
1287 attr
[i
].extra_flags
);
1291 pandecode_log("}, \n");
1293 if (mode
== MALI_ATTR_NPOT_DIVIDE
) {
1295 pandecode_log("{\n");
1297 pandecode_prop("unk = 0x%x", attr
[i
].unk
);
1298 pandecode_prop("magic_divisor = 0x%08x", attr
[i
].magic_divisor
);
1299 if (attr
[i
].zero
!= 0)
1300 pandecode_prop("XXX: zero tripped (0x%x)\n", attr
[i
].zero
);
1301 pandecode_prop("divisor = %d", attr
[i
].divisor
);
1302 pandecode_magic_divisor(attr
[i
].magic_divisor
, attr
[i
- 1].shift
, attr
[i
].divisor
, attr
[i
- 1].extra_flags
);
1304 pandecode_log("}, \n");
1310 pandecode_log("};\n");
1314 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1316 /* TODO: Decode flags */
1317 mali_ptr shader_ptr
= ptr
& ~15;
1319 char *a
= pointer_as_memory_reference(shader_ptr
);
1320 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1327 all_zero(unsigned *buffer
, unsigned count
)
1329 for (unsigned i
= 0; i
< count
; ++i
) {
1338 pandecode_stencil(const char *name
, const struct mali_stencil_test
*stencil
)
1340 if (all_zero((unsigned *) stencil
, sizeof(stencil
) / sizeof(unsigned)))
1343 const char *func
= pandecode_func(stencil
->func
);
1344 const char *sfail
= pandecode_stencil_op(stencil
->sfail
);
1345 const char *dpfail
= pandecode_stencil_op(stencil
->dpfail
);
1346 const char *dppass
= pandecode_stencil_op(stencil
->dppass
);
1349 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil
->zero
);
1351 pandecode_log(".stencil_%s = {\n", name
);
1353 pandecode_prop("ref = %d", stencil
->ref
);
1354 pandecode_prop("mask = 0x%02X", stencil
->mask
);
1355 pandecode_prop("func = %s", func
);
1356 pandecode_prop("sfail = %s", sfail
);
1357 pandecode_prop("dpfail = %s", dpfail
);
1358 pandecode_prop("dppass = %s", dppass
);
1360 pandecode_log("},\n");
1364 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1367 pandecode_msg("XXX: blend zero tripped: %X\n", blend
->zero1
);
1369 pandecode_log(".equation = {\n");
1372 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1373 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1375 pandecode_log(".color_mask = ");
1376 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1377 pandecode_log_cont(",\n");
1380 pandecode_log("},\n");
1383 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1386 decode_bifrost_constant(u16 constant
)
1388 float lo
= (float) (constant
& 0xFF);
1389 float hi
= (float) (constant
>> 8);
1391 return (hi
/ 255.0) + (lo
/ 65535.0);
1395 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1397 struct bifrost_blend_rt
*b
=
1398 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1400 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1403 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1404 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1405 b
->constant
, decode_bifrost_constant(b
->constant
));
1407 /* TODO figure out blend shader enable bit */
1408 pandecode_blend_equation(&b
->equation
);
1409 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1410 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1411 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1414 pandecode_log("},\n");
1420 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1422 if (all_zero((unsigned *) blend
, sizeof(blend
) / sizeof(unsigned)))
1425 pandecode_log(".blend = {\n");
1429 pandecode_shader_address("shader", blend
->shader
);
1431 pandecode_blend_equation(&blend
->equation
);
1432 pandecode_prop("constant = %f", blend
->constant
);
1436 pandecode_log("},\n");
1438 /* Return blend shader to disassemble if present */
1439 return is_shader
? (blend
->shader
& ~0xF) : 0;
1443 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1445 struct midgard_blend_rt
*b
=
1446 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1448 /* Flags determine presence of blend shader */
1449 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1451 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1454 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1456 union midgard_blend blend
= b
->blend
;
1457 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1460 pandecode_log("};\n");
1465 /* Attributes and varyings have descriptor records, which contain information
1466 * about their format and ordering with the attribute/varying buffers. We'll
1467 * want to validate that the combinations specified are self-consistent.
1471 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1474 char *prefix
= varying
? "varying" : "attribute";
1475 unsigned max_index
= 0;
1476 snprintf(base
, sizeof(base
), "%s_meta", prefix
);
1478 pandecode_log("struct mali_attr_meta %s_%d%s[] = {\n", base
, job_no
, suffix
);
1481 struct mali_attr_meta
*attr_meta
;
1482 mali_ptr p
= varying
? (v
->varying_meta
& ~0xF) : v
->attribute_meta
;
1484 struct pandecode_mapped_memory
*attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
);
1486 for (int i
= 0; i
< count
; ++i
, p
+= sizeof(struct mali_attr_meta
)) {
1487 attr_meta
= pandecode_fetch_gpu_mem(attr_mem
, p
,
1490 /* If the record is discard, it should be zero for everything else */
1492 if (attr_meta
->format
== MALI_VARYING_DISCARD
) {
1495 attr_meta
->unknown1
|
1496 attr_meta
->unknown3
|
1497 attr_meta
->src_offset
;
1500 pandecode_msg("XXX: expected empty record for varying discard\n");
1502 /* We want to look for a literal 0000 swizzle -- this
1503 * is not encoded with all zeroes, however */
1505 enum mali_channel z
= MALI_CHANNEL_ZERO
;
1506 unsigned zero_swizzle
= z
| (z
<< 3) | (z
<< 6) | (z
<< 9);
1507 bool good_swizzle
= attr_meta
->swizzle
== zero_swizzle
;
1510 pandecode_msg("XXX: expected zero swizzle for discard\n");
1513 pandecode_msg("XXX: cannot discard attribute\n");
1515 /* If we're all good, omit the record */
1516 if (!zero
&& varying
&& good_swizzle
) {
1517 pandecode_log("/* discarded varying */\n");
1522 if (attr_meta
->index
> max_index
)
1523 max_index
= attr_meta
->index
;
1525 if (attr_meta
->unknown1
!= 0x2) {
1526 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1527 pandecode_prop("unknown1 = 0x%" PRIx64
, (u64
) attr_meta
->unknown1
);
1530 if (attr_meta
->unknown3
) {
1531 pandecode_msg("XXX: unexpected unknown3 set\n");
1532 pandecode_prop("unknown3 = 0x%" PRIx64
, (u64
) attr_meta
->unknown3
);
1535 pandecode_make_indent();
1536 pandecode_format_short(attr_meta
->format
, false);
1537 pandecode_log_cont(" %s_%u", prefix
, attr_meta
->index
);
1539 if (attr_meta
->src_offset
)
1540 pandecode_log_cont("[%u]", attr_meta
->src_offset
);
1542 pandecode_swizzle(attr_meta
->swizzle
, attr_meta
->format
);
1544 pandecode_log_cont(";\n");
1548 pandecode_log("};\n");
1550 return count
? (max_index
+ 1) : 0;
1554 pandecode_indices(uintptr_t pindices
, uint32_t index_count
, int job_no
)
1556 struct pandecode_mapped_memory
*imem
= pandecode_find_mapped_gpu_mem_containing(pindices
);
1559 /* Indices are literally just a u32 array :) */
1561 uint32_t *PANDECODE_PTR_VAR(indices
, imem
, pindices
);
1563 pandecode_log("uint32_t indices_%d[] = {\n", job_no
);
1566 for (unsigned i
= 0; i
< (index_count
+ 1); i
+= 3)
1567 pandecode_log("%d, %d, %d,\n",
1573 pandecode_log("};\n");
1577 /* return bits [lo, hi) of word */
1579 bits(u32 word
, u32 lo
, u32 hi
)
1582 return word
; // avoid undefined behavior with the shift
1584 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1588 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool noninstanced
)
1590 pandecode_log_cont("{\n");
1593 /* Decode invocation_count. See the comment before the definition of
1594 * invocation_count for an explanation.
1597 unsigned size_x
= bits(p
->invocation_count
, 0, p
->size_y_shift
) + 1;
1598 unsigned size_y
= bits(p
->invocation_count
, p
->size_y_shift
, p
->size_z_shift
) + 1;
1599 unsigned size_z
= bits(p
->invocation_count
, p
->size_z_shift
, p
->workgroups_x_shift
) + 1;
1601 unsigned groups_x
= bits(p
->invocation_count
, p
->workgroups_x_shift
, p
->workgroups_y_shift
) + 1;
1602 unsigned groups_y
= bits(p
->invocation_count
, p
->workgroups_y_shift
, p
->workgroups_z_shift
) + 1;
1603 unsigned groups_z
= bits(p
->invocation_count
, p
->workgroups_z_shift
, 32) + 1;
1605 /* Even though we have this decoded, we want to ensure that the
1606 * representation is "unique" so we don't lose anything by printing only
1607 * the final result. More specifically, we need to check that we were
1608 * passed something in canonical form, since the definition per the
1609 * hardware is inherently not unique. How? Well, take the resulting
1610 * decode and pack it ourselves! If it is bit exact with what we
1611 * decoded, we're good to go. */
1613 struct mali_vertex_tiler_prefix ref
;
1614 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, noninstanced
);
1617 (p
->invocation_count
== ref
.invocation_count
) &&
1618 (p
->size_y_shift
== ref
.size_y_shift
) &&
1619 (p
->size_z_shift
== ref
.size_z_shift
) &&
1620 (p
->workgroups_x_shift
== ref
.workgroups_x_shift
) &&
1621 (p
->workgroups_y_shift
== ref
.workgroups_y_shift
) &&
1622 (p
->workgroups_z_shift
== ref
.workgroups_z_shift
) &&
1623 (p
->workgroups_x_shift_2
== ref
.workgroups_x_shift_2
);
1626 pandecode_msg("XXX: non-canonical workgroups packing\n");
1627 pandecode_msg("expected: %X, %d, %d, %d, %d, %d\n",
1628 ref
.invocation_count
,
1631 ref
.workgroups_x_shift
,
1632 ref
.workgroups_y_shift
,
1633 ref
.workgroups_z_shift
,
1634 ref
.workgroups_x_shift_2
);
1636 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1637 pandecode_prop("size_y_shift = %d", p
->size_y_shift
);
1638 pandecode_prop("size_z_shift = %d", p
->size_z_shift
);
1639 pandecode_prop("workgroups_x_shift = %d", p
->workgroups_x_shift
);
1640 pandecode_prop("workgroups_y_shift = %d", p
->workgroups_y_shift
);
1641 pandecode_prop("workgroups_z_shift = %d", p
->workgroups_z_shift
);
1642 pandecode_prop("workgroups_x_shift_2 = %d", p
->workgroups_x_shift_2
);
1645 /* Regardless, print the decode */
1646 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1647 size_x
, size_y
, size_z
,
1648 groups_x
, groups_y
, groups_z
);
1651 if (p
->unknown_draw
)
1652 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1654 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1656 if (p
->draw_mode
!= MALI_DRAW_NONE
)
1657 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p
->draw_mode
));
1659 /* Index count only exists for tiler jobs anyway */
1662 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1664 if (p
->offset_bias_correction
)
1665 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1667 /* TODO: Figure out what this is. It's not zero */
1668 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1671 pandecode_log("},\n");
1675 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1677 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1678 struct mali_uniform_buffer_meta
*PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1680 pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%"PRIx64
"_%d[] = {\n",
1684 for (int i
= 0; i
< ubufs_count
; i
++) {
1685 pandecode_log("{\n");
1688 unsigned size
= (ubufs
[i
].size
+ 1) * 16;
1689 mali_ptr addr
= ubufs
[i
].ptr
<< 2;
1691 pandecode_validate_buffer(addr
, size
);
1693 char *ptr
= pointer_as_memory_reference(ubufs
[i
].ptr
<< 2);
1694 pandecode_prop("size = %u", size
);
1695 pandecode_prop("ptr = (%s) >> 2", ptr
);
1697 pandecode_log("},\n");
1702 pandecode_log("};\n");
1706 pandecode_scratchpad(uintptr_t pscratchpad
, int job_no
, char *suffix
)
1709 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(pscratchpad
);
1711 struct bifrost_scratchpad
*PANDECODE_PTR_VAR(scratchpad
, mem
, pscratchpad
);
1713 if (scratchpad
->zero
) {
1714 pandecode_msg("XXX: scratchpad zero tripped");
1715 pandecode_prop("zero = 0x%x\n", scratchpad
->zero
);
1718 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64
"_%d%s = {\n", pscratchpad
, job_no
, suffix
);
1721 pandecode_prop("flags = 0x%x", scratchpad
->flags
);
1722 MEMORY_PROP(scratchpad
, gpu_scratchpad
);
1725 pandecode_log("};\n");
1728 static unsigned shader_id
= 0;
1731 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1732 bool is_bifrost
, unsigned nr_regs
)
1734 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1735 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1737 /* Compute maximum possible size */
1738 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1740 /* Print some boilerplate to clearly denote the assembly (which doesn't
1741 * obey indentation rules), and actually do the disassembly! */
1747 snprintf(prefix
, sizeof(prefix
) - 1, "shader%d - %s shader: ",
1749 (type
== JOB_TYPE_TILER
) ? "FRAGMENT" : "VERTEX");
1752 disassemble_bifrost(code
, sz
, false);
1754 disassemble_midgard(code
, sz
, true, nr_regs
, prefix
);
1761 pandecode_texture(mali_ptr u
,
1762 struct pandecode_mapped_memory
*tmem
,
1763 unsigned job_no
, unsigned tex
)
1765 struct mali_texture_descriptor
*PANDECODE_PTR_VAR(t
, tmem
, u
);
1767 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64
"_%d_%d = {\n", u
, job_no
, tex
);
1770 struct mali_texture_format f
= t
->format
;
1772 /* See the definiton of enum mali_texture_type */
1774 unsigned dimension
=
1775 (f
.type
== MALI_TEX_CUBE
) ? 2 : f
.type
;
1777 /* All four width/height/depth/array_size dimensions are present
1778 * regardless of the type of texture, but it is an error to have
1779 * non-zero dimensions for unused dimensions. Verify this. array_size
1780 * can always be set, as can width. */
1782 if (t
->height
&& dimension
< 2)
1783 pandecode_msg("XXX: nonzero height for <2D texture\n");
1785 if (t
->depth
&& dimension
< 3)
1786 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1788 /* Print only the dimensions that are actually there */
1790 pandecode_log("dim: %d", t
->width
+ 1);
1793 pandecode_log_cont("x%u", t
->height
+ 1);
1796 pandecode_log_cont("x%u", t
->depth
+ 1);
1799 pandecode_log_cont("[%u]", t
->array_size
+ 1);
1802 pandecode_log_cont(" mip %u", t
->levels
);
1804 pandecode_log_cont("\n");
1806 pandecode_log(".format = {\n");
1809 pandecode_log(".format = ");
1810 pandecode_format_short(f
.format
, f
.srgb
);
1811 pandecode_swizzle(f
.swizzle
, f
.format
);
1812 pandecode_log_cont(",\n");
1814 pandecode_prop("type = %s", pandecode_texture_type(f
.type
));
1815 pandecode_prop("usage2 = 0x%" PRIx32
, f
.usage2
);
1818 pandecode_msg("XXX: texture format zero tripped\n");
1819 pandecode_prop("unknown1 = %" PRId32
, f
.unknown1
);
1823 pandecode_log("},\n");
1825 if (t
->swizzle_zero
) {
1826 pandecode_msg("XXX: swizzle zero tripped\n");
1827 pandecode_prop("swizzle_zero = %d", t
->swizzle_zero
);
1830 if (t
->unknown3
| t
->unknown3A
| t
->unknown5
| t
->unknown6
| t
->unknown7
) {
1831 pandecode_msg("XXX: texture zero tripped\n");
1832 pandecode_prop("unknown3 = %" PRId16
, t
->unknown3
);
1833 pandecode_prop("unknown3A = %" PRId8
, t
->unknown3A
);
1834 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
1835 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
1836 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
1839 pandecode_log(".payload = {\n");
1842 /* A bunch of bitmap pointers follow.
1843 * We work out the correct number,
1844 * based on the mipmap/cubemap
1845 * properties, but dump extra
1846 * possibilities to futureproof */
1848 int bitmap_count
= MALI_NEGATIVE(t
->levels
);
1849 bool manual_stride
= f
.usage2
& MALI_TEX_MANUAL_STRIDE
;
1851 /* Miptree for each face */
1852 if (f
.type
== MALI_TEX_CUBE
)
1855 /* Array of textures */
1856 bitmap_count
*= MALI_NEGATIVE(t
->array_size
);
1858 /* Stride for each element */
1862 /* Sanity check the size */
1863 int max_count
= sizeof(t
->payload
) / sizeof(t
->payload
[0]);
1864 assert (bitmap_count
<= max_count
);
1866 for (int i
= 0; i
< bitmap_count
; ++i
) {
1867 /* How we dump depends if this is a stride or a pointer */
1869 if ((f
.usage2
& MALI_TEX_MANUAL_STRIDE
) && (i
& 1)) {
1870 /* signed 32-bit snuck in as a 64-bit pointer */
1871 uint64_t stride_set
= t
->payload
[i
];
1872 uint32_t clamped_stride
= stride_set
;
1873 int32_t stride
= clamped_stride
;
1874 assert(stride_set
== clamped_stride
);
1875 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
1877 char *a
= pointer_as_memory_reference(t
->payload
[i
]);
1878 pandecode_log("%s, \n", a
);
1884 pandecode_log("},\n");
1887 pandecode_log("};\n");
1891 pandecode_vertex_tiler_postfix_pre(
1892 const struct mali_vertex_tiler_postfix
*p
,
1893 int job_no
, enum mali_job_type job_type
,
1894 char *suffix
, bool is_bifrost
)
1896 mali_ptr shader_meta_ptr
= (u64
) (uintptr_t) (p
->_shader_upper
<< 4);
1897 struct pandecode_mapped_memory
*attr_mem
;
1899 unsigned rt_count
= 1;
1901 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1902 * are the only things actually needed from the FBD, vertex/tiler jobs
1903 * no longer reference the FBD -- instead, this field points to some
1904 * info about the scratchpad.
1907 pandecode_scratchpad(p
->framebuffer
& ~FBD_TYPE
, job_no
, suffix
);
1908 else if (p
->framebuffer
& MALI_MFBD
)
1909 rt_count
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->framebuffer
) & FBD_MASK
, job_no
, false);
1910 else if (job_type
== JOB_TYPE_COMPUTE
)
1911 pandecode_compute_fbd((u64
) (uintptr_t) p
->framebuffer
, job_no
);
1913 pandecode_sfbd((u64
) (uintptr_t) p
->framebuffer
, job_no
, false);
1915 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
1916 int texture_count
= 0, sampler_count
= 0;
1918 if (shader_meta_ptr
) {
1919 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr
);
1920 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, shader_meta_ptr
);
1922 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", shader_meta_ptr
, job_no
, suffix
);
1925 /* Save for dumps */
1926 attribute_count
= s
->attribute_count
;
1927 varying_count
= s
->varying_count
;
1928 texture_count
= s
->texture_count
;
1929 sampler_count
= s
->sampler_count
;
1932 uniform_count
= s
->bifrost2
.uniform_count
;
1933 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
1935 uniform_count
= s
->midgard1
.uniform_buffer_count
;
1936 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
1939 mali_ptr shader_ptr
= pandecode_shader_address("shader", s
->shader
);
1941 pandecode_prop("texture_count = %" PRId16
, s
->texture_count
);
1942 pandecode_prop("sampler_count = %" PRId16
, s
->sampler_count
);
1943 pandecode_prop("attribute_count = %" PRId16
, s
->attribute_count
);
1944 pandecode_prop("varying_count = %" PRId16
, s
->varying_count
);
1946 unsigned nr_registers
= 0;
1949 pandecode_log(".bifrost1 = {\n");
1952 pandecode_prop("uniform_buffer_count = %" PRId32
, s
->bifrost1
.uniform_buffer_count
);
1953 pandecode_prop("unk1 = 0x%" PRIx32
, s
->bifrost1
.unk1
);
1956 pandecode_log("},\n");
1958 pandecode_log(".midgard1 = {\n");
1961 pandecode_prop("uniform_count = %" PRId16
, s
->midgard1
.uniform_count
);
1962 pandecode_prop("uniform_buffer_count = %" PRId16
, s
->midgard1
.uniform_buffer_count
);
1963 pandecode_prop("work_count = %" PRId16
, s
->midgard1
.work_count
);
1964 nr_registers
= s
->midgard1
.work_count
;
1966 pandecode_log(".flags = ");
1967 pandecode_log_decoded_flags(shader_midgard1_flag_info
, s
->midgard1
.flags
);
1968 pandecode_log_cont(",\n");
1970 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->midgard1
.unknown2
);
1973 pandecode_log("},\n");
1976 if (s
->depth_units
|| s
->depth_factor
) {
1977 pandecode_prop("depth_factor = %f", s
->depth_factor
);
1978 pandecode_prop("depth_units = %f", s
->depth_units
);
1981 if (s
->alpha_coverage
) {
1982 bool invert_alpha_coverage
= s
->alpha_coverage
& 0xFFF0;
1983 uint16_t inverted_coverage
= invert_alpha_coverage
? ~s
->alpha_coverage
: s
->alpha_coverage
;
1985 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
1986 invert_alpha_coverage
? "~" : "",
1987 MALI_GET_ALPHA_COVERAGE(inverted_coverage
));
1990 if (s
->unknown2_3
|| s
->unknown2_4
) {
1991 pandecode_log(".unknown2_3 = ");
1993 int unknown2_3
= s
->unknown2_3
;
1994 int unknown2_4
= s
->unknown2_4
;
1996 /* We're not quite sure what these flags mean without the depth test, if anything */
1998 if (unknown2_3
& (MALI_DEPTH_TEST
| MALI_DEPTH_FUNC_MASK
)) {
1999 const char *func
= pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3
));
2000 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
2002 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
2005 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
2006 pandecode_log_cont(",\n");
2008 pandecode_log(".unknown2_4 = ");
2009 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
2010 pandecode_log_cont(",\n");
2013 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
2014 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
2015 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
2018 pandecode_stencil("front", &s
->stencil_front
);
2019 pandecode_stencil("back", &s
->stencil_back
);
2022 pandecode_log(".bifrost2 = {\n");
2025 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
2026 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
2027 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
2028 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
2031 pandecode_log("},\n");
2032 } else if (s
->midgard2
.unknown2_7
) {
2033 pandecode_log(".midgard2 = {\n");
2036 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
2038 pandecode_log("},\n");
2042 pandecode_prop("unknown2_8 = 0x%" PRIx32
, s
->unknown2_8
);
2045 /* TODO: Blend shaders routing/disasm */
2047 union midgard_blend blend
= s
->blend
;
2048 pandecode_midgard_blend(&blend
, false);
2052 pandecode_log("};\n");
2054 /* MRT blend fields are used whenever MFBD is used, with
2055 * per-RT descriptors */
2057 if (job_type
== JOB_TYPE_TILER
) {
2058 void* blend_base
= (void *) (s
+ 1);
2060 for (unsigned i
= 0; i
< rt_count
; i
++) {
2061 mali_ptr shader
= 0;
2064 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
2066 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
2069 pandecode_shader_disassemble(shader
, job_no
, job_type
, false, 0);
2073 if (shader_ptr
& ~0xF)
2074 pandecode_shader_disassemble(shader_ptr
, job_no
, job_type
, is_bifrost
, nr_registers
);
2076 pandecode_msg("<no shader>\n");
2079 struct pandecode_mapped_memory
*fmem
= pandecode_find_mapped_gpu_mem_containing(p
->viewport
);
2080 struct mali_viewport
*PANDECODE_PTR_VAR(f
, fmem
, p
->viewport
);
2082 pandecode_log("struct mali_viewport viewport_%"PRIx64
"_%d%s = {\n", p
->viewport
, job_no
, suffix
);
2085 pandecode_prop("clip_minx = %f", f
->clip_minx
);
2086 pandecode_prop("clip_miny = %f", f
->clip_miny
);
2087 pandecode_prop("clip_minz = %f", f
->clip_minz
);
2088 pandecode_prop("clip_maxx = %f", f
->clip_maxx
);
2089 pandecode_prop("clip_maxy = %f", f
->clip_maxy
);
2090 pandecode_prop("clip_maxz = %f", f
->clip_maxz
);
2092 /* Only the higher coordinates are MALI_POSITIVE scaled */
2094 pandecode_prop("viewport0 = { %d, %d }",
2095 f
->viewport0
[0], f
->viewport0
[1]);
2097 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2098 f
->viewport1
[0] + 1, f
->viewport1
[1] + 1);
2101 pandecode_log("};\n");
2104 if (p
->attribute_meta
) {
2105 unsigned max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
2107 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
2108 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false);
2111 /* Varyings are encoded like attributes but not actually sent; we just
2112 * pass a zero buffer with the right stride/size set, (or whatever)
2113 * since the GPU will write to it itself */
2115 if (p
->varying_meta
) {
2116 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
2120 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
2122 /* Number of descriptors depends on whether there are
2123 * non-internal varyings */
2125 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true);
2128 if (p
->uniform_buffers
) {
2129 if (uniform_buffer_count
)
2130 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
2132 pandecode_msg("XXX: UBOs specified but not referenced\n");
2133 } else if (uniform_buffer_count
)
2134 pandecode_msg("XXX: UBOs referenced but not specified\n");
2136 /* We don't want to actually dump uniforms, but we do need to validate
2137 * that the counts we were given are sane */
2141 pandecode_validate_buffer(p
->uniforms
, uniform_count
* 16);
2143 pandecode_msg("XXX: Uniforms specified but not referenced");
2144 } else if (uniform_count
)
2145 pandecode_msg("XXX: UBOs referenced but not specified\n");
2147 if (p
->texture_trampoline
) {
2148 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(p
->texture_trampoline
);
2151 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
);
2153 pandecode_log("uint64_t texture_trampoline_%"PRIx64
"_%d[] = {\n", p
->texture_trampoline
, job_no
);
2156 for (int tex
= 0; tex
< texture_count
; ++tex
) {
2157 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
2158 char *a
= pointer_as_memory_reference(*u
);
2159 pandecode_log("%s,\n", a
);
2164 pandecode_log("};\n");
2166 /* Now, finally, descend down into the texture descriptor */
2167 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
2168 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
2169 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
2171 pandecode_texture(*u
, tmem
, job_no
, tex
);
2176 if (p
->sampler_descriptor
) {
2177 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->sampler_descriptor
);
2180 struct mali_sampler_descriptor
*s
;
2182 mali_ptr d
= p
->sampler_descriptor
;
2184 for (int i
= 0; i
< sampler_count
; ++i
) {
2185 s
= pandecode_fetch_gpu_mem(smem
, d
+ sizeof(*s
) * i
, sizeof(*s
));
2187 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", d
+ sizeof(*s
) * i
, job_no
, i
);
2190 pandecode_log(".filter_mode = ");
2191 pandecode_log_decoded_flags(sampler_flag_info
, s
->filter_mode
);
2192 pandecode_log_cont(",\n");
2194 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
2195 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
2197 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s
->wrap_s
));
2198 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s
->wrap_t
));
2199 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s
->wrap_r
));
2201 pandecode_prop("compare_func = %s", pandecode_alt_func(s
->compare_func
));
2203 if (s
->zero
|| s
->zero2
) {
2204 pandecode_msg("XXX: sampler zero tripped\n");
2205 pandecode_prop("zero = 0x%X, 0x%X\n", s
->zero
, s
->zero2
);
2208 pandecode_prop("seamless_cube_map = %d", s
->seamless_cube_map
);
2210 pandecode_prop("border_color = { %f, %f, %f, %f }",
2214 s
->border_color
[3]);
2217 pandecode_log("};\n");
2224 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2226 if (!(p
->position_varying
|| p
->occlusion_counter
|| p
->flags
))
2229 pandecode_log(".postfix = {\n");
2232 MEMORY_PROP(p
, position_varying
);
2233 MEMORY_PROP(p
, occlusion_counter
);
2236 pandecode_prop("flags = %d", p
->flags
);
2239 pandecode_log("},\n");
2243 pandecode_vertex_only_bfr(struct bifrost_vertex_only
*v
)
2245 pandecode_log_cont("{\n");
2248 pandecode_prop("unk2 = 0x%x", v
->unk2
);
2250 if (v
->zero0
|| v
->zero1
) {
2251 pandecode_msg("XXX: vertex only zero tripped");
2252 pandecode_prop("zero0 = 0x%" PRIx32
, v
->zero0
);
2253 pandecode_prop("zero1 = 0x%" PRIx64
, v
->zero1
);
2257 pandecode_log("}\n");
2261 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2264 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2265 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2267 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no
);
2271 pandecode_msg("XXX: tiler heap zero tripped\n");
2272 pandecode_prop("zero = 0x%x", h
->zero
);
2275 for (int i
= 0; i
< 12; i
++) {
2276 if (h
->zeros
[i
] != 0) {
2277 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2282 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2283 MEMORY_PROP(h
, tiler_heap_start
);
2284 MEMORY_PROP(h
, tiler_heap_free
);
2286 /* this might point to the beginning of another buffer, when it's
2287 * really the end of the tiler heap buffer, so we have to be careful
2288 * here. but for zero length, we need the same pointer.
2291 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2292 MEMORY_PROP(h
, tiler_heap_start
);
2294 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2295 pandecode_prop("tiler_heap_end = %s + 1", a
);
2300 pandecode_log("};\n");
2304 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2306 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2307 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2309 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2311 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no
);
2314 if (t
->zero0
|| t
->zero1
) {
2315 pandecode_msg("XXX: tiler meta zero tripped\n");
2316 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2317 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2320 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2321 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2323 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2324 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2326 for (int i
= 0; i
< 12; i
++) {
2327 if (t
->zeros
[i
] != 0) {
2328 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2334 pandecode_log("};\n");
2338 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2340 pandecode_log(".gl_enables = ");
2342 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2344 pandecode_log_cont(",\n");
2348 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2350 if (u
.pointer
== 0x0)
2353 pandecode_log(".primitive_size = {\n");
2357 pandecode_prop("constant = %f", u
.constant
);
2359 MEMORY_PROP((&u
), pointer
);
2363 pandecode_log("},\n");
2367 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2369 pandecode_log_cont("{\n");
2372 /* TODO: gl_PointSize on Bifrost */
2373 pandecode_primitive_size(t
->primitive_size
, true);
2375 pandecode_gl_enables(t
->gl_enables
, JOB_TYPE_TILER
);
2377 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2378 || t
->zero6
|| t
->zero7
|| t
->zero8
) {
2379 pandecode_msg("XXX: tiler only zero tripped\n");
2380 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2381 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2382 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2383 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2384 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2385 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2386 pandecode_prop("zero7 = 0x%" PRIx32
, t
->zero7
);
2387 pandecode_prop("zero8 = 0x%" PRIx64
, t
->zero8
);
2391 pandecode_log("},\n");
2395 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2396 const struct pandecode_mapped_memory
*mem
,
2397 mali_ptr payload
, int job_no
)
2399 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2401 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true);
2403 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no
);
2406 pandecode_log(".prefix = ");
2407 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2409 pandecode_log(".vertex = ");
2410 pandecode_vertex_only_bfr(&v
->vertex
);
2412 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2415 pandecode_log("};\n");
2421 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2422 const struct pandecode_mapped_memory
*mem
,
2423 mali_ptr payload
, int job_no
)
2425 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2427 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true);
2429 pandecode_indices(t
->prefix
.indices
, t
->prefix
.index_count
, job_no
);
2430 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2432 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no
);
2435 pandecode_log(".prefix = ");
2436 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2438 pandecode_log(".tiler = ");
2439 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2441 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2444 pandecode_log("};\n");
2450 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2451 const struct pandecode_mapped_memory
*mem
,
2452 mali_ptr payload
, int job_no
)
2454 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2456 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false);
2458 pandecode_indices(v
->prefix
.indices
, v
->prefix
.index_count
, job_no
);
2460 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2463 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2464 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2466 bool instanced
= v
->instance_shift
|| v
->instance_odd
;
2467 bool is_graphics
= (h
->job_type
== JOB_TYPE_VERTEX
) || (h
->job_type
== JOB_TYPE_TILER
);
2469 pandecode_log(".prefix = ");
2470 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, !instanced
&& is_graphics
);
2472 pandecode_gl_enables(v
->gl_enables
, h
->job_type
);
2474 if (v
->instance_shift
|| v
->instance_odd
) {
2475 pandecode_prop("instance_shift = 0x%d /* %d */",
2476 v
->instance_shift
, 1 << v
->instance_shift
);
2477 pandecode_prop("instance_odd = 0x%X /* %d */",
2478 v
->instance_odd
, (2 * v
->instance_odd
) + 1);
2480 pandecode_padded_vertices(v
->instance_shift
, v
->instance_odd
);
2483 if (v
->offset_start
)
2484 pandecode_prop("offset_start = %d", v
->offset_start
);
2487 pandecode_msg("XXX: midgard payload zero tripped\n");
2488 pandecode_prop("zero5 = 0x%" PRIx64
, v
->zero5
);
2491 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2494 pandecode_log("};\n");
2500 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2501 mali_ptr payload
, int job_no
,
2504 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2506 bool fbd_dumped
= false;
2508 if (!is_bifrost
&& (s
->framebuffer
& FBD_TYPE
) == MALI_SFBD
) {
2509 /* Only SFBDs are understood, not MFBDs. We're speculating,
2510 * based on the versioning, kernel code, etc, that the
2511 * difference is between Single FrameBuffer Descriptor and
2512 * Multiple FrmaeBuffer Descriptor; the change apparently lines
2513 * up with multi-framebuffer support being added (T7xx onwards,
2514 * including Gxx). In any event, there's some field shuffling
2515 * that we haven't looked into yet. */
2517 pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true);
2519 } else if ((s
->framebuffer
& FBD_TYPE
) == MALI_MFBD
) {
2520 /* We don't know if Bifrost supports SFBD's at all, since the
2521 * driver never uses them. And the format is different from
2522 * Midgard anyways, due to the tiler heap and scratchpad being
2523 * moved out into separate structures, so it's not clear what a
2524 * Bifrost SFBD would even look like without getting an actual
2525 * trace, which appears impossible.
2528 pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true);
2532 uintptr_t p
= (uintptr_t) s
->framebuffer
& FBD_MASK
;
2533 pandecode_log("struct mali_payload_fragment payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2536 /* See the comments by the macro definitions for mathematical context
2537 * on why this is so weird */
2539 if (MALI_TILE_COORD_FLAGS(s
->max_tile_coord
) || MALI_TILE_COORD_FLAGS(s
->min_tile_coord
))
2540 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
2542 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
2543 MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
,
2544 MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
);
2546 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
2547 (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
,
2548 (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
);
2550 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2551 * we have to fallback on offsets. */
2553 const char *fbd_type
= s
->framebuffer
& MALI_MFBD
? "MALI_MFBD" : "MALI_SFBD";
2556 unsigned extra_flags
= (s
->framebuffer
& ~FBD_MASK
) & ~MALI_MFBD
;
2559 pandecode_prop("framebuffer = framebuffer_%d_p | %s | 0x%X", job_no
,
2560 fbd_type
, extra_flags
);
2562 pandecode_prop("framebuffer = %s | %s | 0x%X", pointer_as_memory_reference(p
),
2563 fbd_type
, extra_flags
);
2566 pandecode_log("};\n");
2571 static int job_descriptor_number
= 0;
2574 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
)
2576 struct mali_job_descriptor_header
*h
;
2578 int start_number
= 0;
2584 struct pandecode_mapped_memory
*mem
=
2585 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2589 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2591 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2592 * high 32-bits of the 64-bit pointer are reused to store
2595 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
2596 h
->job_type
!= JOB_TYPE_FRAGMENT
? 4 : 0;
2597 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
2599 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
,
2602 int job_no
= job_descriptor_number
++;
2605 start_number
= job_no
;
2607 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2610 pandecode_prop("job_type = %s", pandecode_job_type(h
->job_type
));
2612 /* Save for next job fixing */
2613 last_size
= h
->job_descriptor_size
;
2615 if (h
->job_descriptor_size
)
2616 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2618 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2619 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2620 h
->exception_status
,
2621 (h
->exception_status
>> 16) & 0xFFFF,
2622 pandecode_exception_access((h
->exception_status
>> 8) & 0x3),
2623 h
->exception_status
& 0xFF);
2625 if (h
->first_incomplete_task
)
2626 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2628 if (h
->fault_pointer
)
2629 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2632 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2634 pandecode_prop("job_index = %d", h
->job_index
);
2636 if (h
->unknown_flags
)
2637 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2639 if (h
->job_dependency_index_1
)
2640 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2642 if (h
->job_dependency_index_2
)
2643 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2646 pandecode_log("};\n");
2648 /* Do not touch the field yet -- decode the payload first, and
2649 * don't touch that either. This is essential for the uploads
2650 * to occur in sequence and therefore be dynamically allocated
2651 * correctly. Do note the size, however, for that related
2654 switch (h
->job_type
) {
2655 case JOB_TYPE_SET_VALUE
: {
2656 struct mali_payload_set_value
*s
= payload
;
2657 pandecode_log("struct mali_payload_set_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2659 MEMORY_PROP(s
, out
);
2660 pandecode_prop("unknown = 0x%" PRIX64
, s
->unknown
);
2662 pandecode_log("};\n");
2667 case JOB_TYPE_TILER
:
2668 case JOB_TYPE_VERTEX
:
2669 case JOB_TYPE_COMPUTE
:
2671 if (h
->job_type
== JOB_TYPE_TILER
)
2672 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
);
2674 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
);
2676 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
);
2680 case JOB_TYPE_FRAGMENT
:
2681 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
);
2688 /* Handle linkage */
2691 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no
- 1);
2694 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no
);
2696 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no
);
2701 } while ((jc_gpu_va
= h
->job_descriptor_size
? h
->next_job_64
: h
->next_job_32
));
2703 return start_number
;