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_short(enum mali_attr_mode mode
)
420 /* TODO: Combine to just "instanced" once this can be done
421 * unambiguously in all known cases */
422 case MALI_ATTR_POT_DIVIDE
:
423 return "instanced_pot";
424 case MALI_ATTR_MODULO
:
425 return "instanced_mod";
426 case MALI_ATTR_NPOT_DIVIDE
:
427 return "instanced_npot";
428 case MALI_ATTR_IMAGE
:
430 case MALI_ATTR_INTERNAL
:
433 pandecode_msg("XXX: invalid attribute mode %X\n", mode
);
440 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
442 pandecode_wrap_mode(enum mali_wrap_mode op
)
446 DEFINE_CASE(CLAMP_TO_EDGE
);
447 DEFINE_CASE(CLAMP_TO_BORDER
);
448 DEFINE_CASE(MIRRORED_REPEAT
);
451 pandecode_msg("XXX: invalid wrap mode %X\n", op
);
457 #define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
459 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt
)
463 DEFINE_CASE(UNKNOWN
);
468 unreachable("Invalid case");
473 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
475 pandecode_exception_access(enum mali_exception_access access
)
479 DEFINE_CASE(EXECUTE
);
484 unreachable("Invalid case");
489 /* Midgard's tiler descriptor is embedded within the
493 pandecode_midgard_tiler_descriptor(
494 const struct midgard_tiler_descriptor
*t
,
499 pandecode_log(".tiler = {\n");
502 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
503 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
505 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
507 /* We know this name from the kernel, but we never see it nonzero */
510 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
512 MEMORY_PROP(t
, polygon_list
);
514 /* The body is offset from the base of the polygon list */
515 assert(t
->polygon_list_body
> t
->polygon_list
);
516 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
518 /* It needs to fit inside the reported size */
519 assert(t
->polygon_list_size
>= body_offset
);
521 /* Check that we fit */
522 struct pandecode_mapped_memory
*plist
=
523 pandecode_find_mapped_gpu_mem_containing(t
->polygon_list
);
525 assert(t
->polygon_list_size
<= plist
->length
);
527 /* Now that we've sanity checked, we'll try to calculate the sizes
528 * ourselves for comparison */
530 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
);
531 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
);
533 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
534 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
535 ref_header
, ref_size
);
536 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
537 pandecode_msg("body offset %d\n", body_offset
);
540 /* The tiler heap has a start and end specified -- it should be
541 * identical to what we have in the BO. The exception is if tiling is
544 MEMORY_PROP(t
, heap_start
);
545 assert(t
->heap_end
>= t
->heap_start
);
547 struct pandecode_mapped_memory
*heap
=
548 pandecode_find_mapped_gpu_mem_containing(t
->heap_start
);
550 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
552 /* Tiling is enabled with a special flag */
553 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
554 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
556 bool tiling_enabled
= hierarchy_mask
;
558 if (tiling_enabled
) {
559 /* When tiling is enabled, the heap should be a tight fit */
560 unsigned heap_offset
= t
->heap_start
- heap
->gpu_va
;
561 if ((heap_offset
+ heap_size
) != heap
->length
) {
562 pandecode_msg("XXX: heap size %d (expected %d)\n",
563 heap_size
, heap
->length
- heap_offset
);
566 /* We should also have no other flags */
568 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
570 /* When tiling is disabled, we should have that flag and no others */
572 if (tiler_flags
!= MALI_TILER_DISABLED
) {
573 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
577 /* We should also have an empty heap */
579 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
583 /* Disabled tiling is used only for clear-only jobs, which are
584 * purely FRAGMENT, so we should never see this for
585 * non-FRAGMENT descriptors. */
588 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
591 /* We've never seen weights used in practice, but we know from the
592 * kernel these fields is there */
594 bool nonzero_weights
= false;
596 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
597 nonzero_weights
|= t
->weights
[w
] != 0x0;
600 if (nonzero_weights
) {
601 pandecode_log(".weights = {");
603 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
604 pandecode_log("%d, ", t
->weights
[w
]);
611 pandecode_log("}\n");
615 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
)
617 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
618 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
620 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
623 pandecode_prop("unknown1 = 0x%" PRIx32
, s
->unknown1
);
624 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->unknown2
);
626 pandecode_log(".format = ");
627 pandecode_log_decoded_flags(fb_fmt_flag_info
, s
->format
);
628 pandecode_log_cont(",\n");
630 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", s
->width
+ 1);
631 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", s
->height
+ 1);
633 MEMORY_PROP(s
, framebuffer
);
634 pandecode_prop("stride = %d", s
->stride
);
636 /* Earlier in the actual commandstream -- right before width -- but we
637 * delay to flow nicer */
639 pandecode_log(".clear_flags = ");
640 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
641 pandecode_log_cont(",\n");
643 if (s
->depth_buffer
| s
->depth_buffer_enable
) {
644 MEMORY_PROP(s
, depth_buffer
);
645 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s
->depth_buffer_enable
));
648 if (s
->stencil_buffer
| s
->stencil_buffer_enable
) {
649 MEMORY_PROP(s
, stencil_buffer
);
650 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s
->stencil_buffer_enable
));
653 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
654 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
655 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
656 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
657 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
660 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
661 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
662 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
663 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
664 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
667 if (s
->clear_stencil
) {
668 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
671 MEMORY_PROP(s
, unknown_address_0
);
672 const struct midgard_tiler_descriptor t
= s
->tiler
;
673 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
);
676 pandecode_log("};\n");
678 pandecode_prop("zero0 = 0x%" PRIx64
, s
->zero0
);
679 pandecode_prop("zero1 = 0x%" PRIx64
, s
->zero1
);
680 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
681 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
683 printf(".zero3 = {");
685 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
686 printf("%X, ", s
->zero3
[i
]);
690 printf(".zero6 = {");
692 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
693 printf("%X, ", s
->zero6
[i
]);
699 pandecode_u32_slide(unsigned name
, const u32
*slide
, unsigned count
)
701 pandecode_log(".unknown%d = {", name
);
703 for (int i
= 0; i
< count
; ++i
)
704 printf("%X, ", slide
[i
]);
706 pandecode_log("},\n");
709 #define SHORT_SLIDE(num) \
710 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
713 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
715 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
716 const struct mali_compute_fbd
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
718 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
727 /* Extracts the number of components associated with a Mali format */
730 pandecode_format_component_count(enum mali_format fmt
)
732 /* Mask out the format class */
733 unsigned top
= fmt
& 0b11100000;
736 case MALI_FORMAT_SNORM
:
737 case MALI_FORMAT_UINT
:
738 case MALI_FORMAT_UNORM
:
739 case MALI_FORMAT_SINT
:
740 return ((fmt
>> 3) & 3) + 1;
747 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
750 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
753 assert(MALI_CHANNEL_RED
== 0);
755 for (unsigned c
= 0; c
< 4; ++c
) {
756 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
758 if (chan
<= MALI_CHANNEL_ALPHA
)
765 /* Validates that a (format, swizzle) pair is valid, in the sense that the
766 * swizzle doesn't access any components that are undefined in the format.
767 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
771 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
773 unsigned nr_comp
= pandecode_format_component_count(fmt
);
774 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
775 unsigned valid_mask
= (1 << nr_comp
) - 1;
776 unsigned invalid_mask
= ~valid_mask
;
778 if (access_mask
& invalid_mask
) {
779 pandecode_msg("XXX: invalid components accessed\n");
783 /* Check for the default non-swizzling swizzle so we can suppress
784 * useless printing for the defaults */
786 unsigned default_swizzles
[4] = {
787 MALI_CHANNEL_RED
| (MALI_CHANNEL_ZERO
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
788 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
789 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ONE
<< 9),
790 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ALPHA
<< 9)
793 return (swizzle
== default_swizzles
[nr_comp
- 1]);
796 /* Maps MALI_RGBA32F to rgba32f, etc */
799 pandecode_format_short(enum mali_format fmt
, bool srgb
)
801 /* We want a type-like format, so cut off the initial MALI_ */
802 char *format
= pandecode_format(fmt
);
803 format
+= strlen("MALI_");
805 unsigned len
= strlen(format
);
806 char *lower_format
= calloc(1, len
+ 1);
808 for (unsigned i
= 0; i
< len
; ++i
)
809 lower_format
[i
] = tolower(format
[i
]);
811 /* Sanity check sRGB flag is applied to RGB, per the name */
812 if (srgb
&& lower_format
[0] != 'r')
813 pandecode_msg("XXX: sRGB applied to non-colour format\n");
815 /* Just prefix with an s, so you get formats like srgba8_unorm */
817 pandecode_log_cont("s");
819 pandecode_log_cont("%s", lower_format
);
824 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
826 /* First, do some validation */
827 bool trivial_swizzle
= pandecode_validate_format_swizzle(
833 /* Next, print the swizzle */
834 pandecode_log_cont(".");
836 static const char components
[] = "rgba01";
838 for (unsigned c
= 0; c
< 4; ++c
) {
839 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
841 if (chan
>= MALI_CHANNEL_RESERVED_0
) {
842 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
845 pandecode_log_cont("%c", components
[chan
]);
850 pandecode_rt_format(struct mali_rt_format format
)
852 pandecode_log(".format = {\n");
855 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
856 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
857 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
859 pandecode_prop("block = %s",
860 pandecode_mfbd_block_format(format
.block
));
862 /* TODO: Map formats so we can check swizzles and print nicely */
863 pandecode_log("swizzle");
864 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
865 pandecode_log_cont(",\n");
867 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
868 MALI_NEGATIVE(format
.nr_channels
));
870 pandecode_log(".flags = ");
871 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
872 pandecode_log_cont(",\n");
874 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
875 * which is a faster hardware-based alternative to the wallpaper method
876 * to preserve framebuffer contents across frames. In practice, MFBD
877 * preload is buggy on Midgard, and so this is a chicken bit. If this
878 * bit isn't set, most likely something broke unrelated to preload */
880 if (!format
.no_preload
) {
881 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
882 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
886 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
889 pandecode_log("},\n");
893 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct bifrost_framebuffer
*fb
)
895 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
898 for (int i
= 0; i
< MALI_NEGATIVE(fb
->rt_count_1
); i
++) {
899 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct bifrost_render_target
);
900 struct pandecode_mapped_memory
*mem
=
901 pandecode_find_mapped_gpu_mem_containing(rt_va
);
902 const struct bifrost_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
904 pandecode_log("{\n");
907 pandecode_rt_format(rt
->format
);
909 if (rt
->format
.block
== MALI_MFBD_BLOCK_AFBC
) {
910 pandecode_log(".afbc = {\n");
913 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
914 pandecode_prop("metadata = %s", a
);
917 pandecode_prop("stride = %d", rt
->afbc
.stride
);
918 pandecode_prop("unk = 0x%" PRIx32
, rt
->afbc
.unk
);
921 pandecode_log("},\n");
922 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.unk
) {
923 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
929 MEMORY_PROP(rt
, framebuffer
);
930 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
932 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
933 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
934 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
935 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
936 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
939 if (rt
->zero1
|| rt
->zero2
|| rt
->zero3
) {
940 pandecode_msg("XXX: render target zeros tripped\n");
941 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
942 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
943 pandecode_prop("zero3 = 0x%" PRIx32
, rt
->zero3
);
947 pandecode_log("},\n");
951 pandecode_log("};\n");
955 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
)
957 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
958 const struct bifrost_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
960 if (fb
->sample_locations
) {
961 /* The blob stores all possible sample locations in a single buffer
962 * allocated on startup, and just switches the pointer when switching
963 * MSAA state. For now, we just put the data into the cmdstream, but we
964 * should do something like what the blob does with a real driver.
966 * There seem to be 32 slots for sample locations, followed by another
967 * 16. The second 16 is just the center location followed by 15 zeros
968 * in all the cases I've identified (maybe shader vs. depth/color
972 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->sample_locations
);
974 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->sample_locations
);
976 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
979 for (int i
= 0; i
< 32 + 16; i
++) {
980 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
984 pandecode_log("};\n");
987 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
990 pandecode_prop("unk0 = 0x%x", fb
->unk0
);
992 if (fb
->sample_locations
)
993 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
995 /* Assume that unknown1 was emitted in the last job for
997 MEMORY_PROP(fb
, unknown1
);
999 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
1000 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
1001 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
1002 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
1004 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
1005 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
1006 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
1007 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
1009 pandecode_log(".mfbd_flags = ");
1010 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
1011 pandecode_log_cont(",\n");
1013 if (fb
->clear_stencil
)
1014 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
1016 if (fb
->clear_depth
)
1017 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
1019 /* TODO: What is this? Let's not blow up.. */
1020 if (fb
->unknown2
!= 0x1F)
1021 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
1023 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
1024 MEMORY_PROP(fb
, scratchpad
);
1025 const struct midgard_tiler_descriptor t
= fb
->tiler
;
1026 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
);
1028 if (fb
->zero3
|| fb
->zero4
) {
1029 pandecode_msg("XXX: framebuffer zeros tripped\n");
1030 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
1031 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
1035 pandecode_log("};\n");
1037 gpu_va
+= sizeof(struct bifrost_framebuffer
);
1039 if ((fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
) {
1040 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1041 const struct bifrost_fb_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
1043 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1046 MEMORY_PROP(fbx
, checksum
);
1048 if (fbx
->checksum_stride
)
1049 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
1051 pandecode_log(".flags = ");
1052 pandecode_log_decoded_flags(mfbd_extra_flag_info
, fbx
->flags
);
1053 pandecode_log_cont(",\n");
1055 if (fbx
->flags
& MALI_EXTRA_AFBC_ZS
) {
1056 pandecode_log(".ds_afbc = {\n");
1059 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
1060 pandecode_prop("depth_stencil_afbc_stride = %d",
1061 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
1062 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
1064 if (fbx
->ds_afbc
.zero1
|| fbx
->ds_afbc
.padding
) {
1065 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1066 pandecode_prop("zero1 = 0x%" PRIx32
,
1067 fbx
->ds_afbc
.zero1
);
1068 pandecode_prop("padding = 0x%" PRIx64
,
1069 fbx
->ds_afbc
.padding
);
1073 pandecode_log("},\n");
1075 pandecode_log(".ds_linear = {\n");
1078 if (fbx
->ds_linear
.depth
) {
1079 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
1080 pandecode_prop("depth_stride = %d",
1081 fbx
->ds_linear
.depth_stride
);
1084 if (fbx
->ds_linear
.stencil
) {
1085 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1086 pandecode_prop("stencil_stride = %d",
1087 fbx
->ds_linear
.stencil_stride
);
1090 if (fbx
->ds_linear
.depth_stride_zero
||
1091 fbx
->ds_linear
.stencil_stride_zero
||
1092 fbx
->ds_linear
.zero1
|| fbx
->ds_linear
.zero2
) {
1093 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1094 pandecode_prop("depth_stride_zero = 0x%x",
1095 fbx
->ds_linear
.depth_stride_zero
);
1096 pandecode_prop("stencil_stride_zero = 0x%x",
1097 fbx
->ds_linear
.stencil_stride_zero
);
1098 pandecode_prop("zero1 = 0x%" PRIx32
,
1099 fbx
->ds_linear
.zero1
);
1100 pandecode_prop("zero2 = 0x%" PRIx32
,
1101 fbx
->ds_linear
.zero2
);
1105 pandecode_log("},\n");
1108 if (fbx
->zero3
|| fbx
->zero4
) {
1109 pandecode_msg("XXX: fb_extra zeros tripped\n");
1110 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1111 pandecode_prop("zero4 = 0x%" PRIx64
, fbx
->zero4
);
1115 pandecode_log("};\n");
1117 gpu_va
+= sizeof(struct bifrost_fb_extra
);
1121 pandecode_render_target(gpu_va
, job_no
, fb
);
1123 /* Passback the render target count */
1124 return MALI_NEGATIVE(fb
->rt_count_1
);
1127 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1131 pandecode_padded_vertices(unsigned shift
, unsigned k
)
1133 unsigned odd
= 2*k
+ 1;
1134 unsigned pot
= 1 << shift
;
1135 pandecode_msg("padded_num_vertices = %d\n", odd
* pot
);
1138 /* Given a magic divisor, recover what we were trying to divide by.
1140 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1141 * 0 <= m < 2^N, for N bits in m.
1143 * Let q represent the number we would like to divide by.
1145 * By definition of a magic divisor for N-bit unsigned integers (a number you
1146 * multiply by to magically get division), m is a number such that:
1148 * (m * x) & (2^N - 1) = floor(x/q).
1149 * for all x on Z where 0 <= x < 2^N
1151 * Ignore the case where any of the above values equals zero; it is irrelevant
1152 * for our purposes (instanced arrays).
1154 * Choose x = q. Then:
1156 * (m * x) & (2^N - 1) = floor(x/q).
1157 * (m * q) & (2^N - 1) = floor(q/q).
1159 * floor(q/q) = floor(1) = 1, therefore:
1161 * (m * q) & (2^N - 1) = 1
1163 * Recall the identity that the bitwise AND of one less than a power-of-two
1164 * equals the modulo with that power of two, i.e. for all x:
1166 * x & (2^N - 1) = x % N
1172 * By definition, a modular multiplicative inverse of a number m is the number
1173 * q such that with respect to a modulos M:
1177 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1182 pandecode_magic_divisor(uint32_t magic
, unsigned shift
, unsigned orig_divisor
, unsigned extra
)
1185 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1186 * shift) the most lame way possible... just repeatedly add.
1187 * Asymptoptically slow but nobody cares in practice, unless you have
1188 * massive numbers of vertices or high divisors. */
1190 unsigned inverse
= 0;
1192 /* Magic implicitly has the highest bit set */
1195 /* Depending on rounding direction */
1200 uint32_t product
= magic
* inverse
;
1212 pandecode_msg("dividing by %d (maybe off by two)\n", inverse
);
1214 /* Recall we're supposed to divide by (gl_level_divisor *
1215 * padded_num_vertices) */
1217 unsigned padded_num_vertices
= inverse
/ orig_divisor
;
1219 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices
);
1224 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1225 mali_ptr addr
, int job_no
, char *suffix
,
1226 int count
, bool varying
)
1228 char *prefix
= varying
? "varying" : "attribute";
1231 pandecode_msg("no %s\n", prefix
);
1235 union mali_attr
*attr
= pandecode_fetch_gpu_mem(mem
, addr
, sizeof(union mali_attr
) * count
);
1237 for (int i
= 0; i
< count
; ++i
) {
1238 enum mali_attr_mode mode
= attr
[i
].elements
& 7;
1240 if (mode
== MALI_ATTR_UNUSED
)
1241 pandecode_msg("XXX: unused attribute record\n");
1243 /* For non-linear records, we need to print the type of record */
1244 if (mode
!= MALI_ATTR_LINEAR
)
1245 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode
));
1247 /* Print the name to link with attr_meta */
1248 pandecode_log_cont("%s_%d", prefix
, i
);
1250 /* Print the stride and size */
1251 pandecode_log_cont("<%u>[%u]", attr
[i
].stride
, attr
[i
].size
);
1253 /* Check: the size must be divisible by the stride */
1254 if (attr
[i
].size
% attr
[i
].stride
)
1255 pandecode_msg("XXX: size not divisible by stride\n");
1257 /* TODO: Sanity check the quotient itself -- it should equal
1258 * vertex count (or something computed from it for instanced)
1259 * which means we can check and elide */
1261 /* Finally, print the pointer */
1262 mali_ptr raw_elements
= attr
[i
].elements
& ~7;
1263 char *a
= pointer_as_memory_reference(raw_elements
);
1264 pandecode_log_cont(" = (%s);\n", a
);
1267 /* Check the pointer */
1268 pandecode_validate_buffer(raw_elements
, attr
[i
].size
);
1270 /* shift/extra_flags exist only for instanced */
1271 if (attr
[i
].shift
| attr
[i
].extra_flags
) {
1272 if (mode
== MALI_ATTR_LINEAR
)
1273 pandecode_msg("XXX: instancing fields set for linear\n");
1275 pandecode_prop("shift = %d", attr
[i
].shift
);
1276 pandecode_prop("extra_flags = %d", attr
[i
].extra_flags
);
1279 /* Decode further where possible */
1281 if (mode
== MALI_ATTR_MODULO
) {
1282 pandecode_padded_vertices(
1284 attr
[i
].extra_flags
);
1287 if (mode
== MALI_ATTR_NPOT_DIVIDE
) {
1289 pandecode_log("{\n");
1291 pandecode_prop("unk = 0x%x", attr
[i
].unk
);
1292 pandecode_prop("magic_divisor = 0x%08x", attr
[i
].magic_divisor
);
1293 if (attr
[i
].zero
!= 0)
1294 pandecode_prop("XXX: zero tripped (0x%x)\n", attr
[i
].zero
);
1295 pandecode_prop("divisor = %d", attr
[i
].divisor
);
1296 pandecode_magic_divisor(attr
[i
].magic_divisor
, attr
[i
- 1].shift
, attr
[i
].divisor
, attr
[i
- 1].extra_flags
);
1298 pandecode_log("}, \n");
1303 pandecode_log("\n");
1307 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1309 /* TODO: Decode flags */
1310 mali_ptr shader_ptr
= ptr
& ~15;
1312 char *a
= pointer_as_memory_reference(shader_ptr
);
1313 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1320 pandecode_stencil(const char *name
, const struct mali_stencil_test
*stencil
)
1322 unsigned any_nonzero
=
1323 stencil
->ref
| stencil
->mask
| stencil
->func
|
1324 stencil
->sfail
| stencil
->dpfail
| stencil
->dppass
;
1326 if (any_nonzero
== 0)
1329 const char *func
= pandecode_func(stencil
->func
);
1330 const char *sfail
= pandecode_stencil_op(stencil
->sfail
);
1331 const char *dpfail
= pandecode_stencil_op(stencil
->dpfail
);
1332 const char *dppass
= pandecode_stencil_op(stencil
->dppass
);
1335 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil
->zero
);
1337 pandecode_log(".stencil_%s = {\n", name
);
1339 pandecode_prop("ref = %d", stencil
->ref
);
1340 pandecode_prop("mask = 0x%02X", stencil
->mask
);
1341 pandecode_prop("func = %s", func
);
1342 pandecode_prop("sfail = %s", sfail
);
1343 pandecode_prop("dpfail = %s", dpfail
);
1344 pandecode_prop("dppass = %s", dppass
);
1346 pandecode_log("},\n");
1350 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1353 pandecode_msg("XXX: blend zero tripped: %X\n", blend
->zero1
);
1355 pandecode_log(".equation = {\n");
1358 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1359 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1361 pandecode_log(".color_mask = ");
1362 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1363 pandecode_log_cont(",\n");
1366 pandecode_log("},\n");
1369 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1372 decode_bifrost_constant(u16 constant
)
1374 float lo
= (float) (constant
& 0xFF);
1375 float hi
= (float) (constant
>> 8);
1377 return (hi
/ 255.0) + (lo
/ 65535.0);
1381 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1383 struct bifrost_blend_rt
*b
=
1384 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1386 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1389 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1390 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1391 b
->constant
, decode_bifrost_constant(b
->constant
));
1393 /* TODO figure out blend shader enable bit */
1394 pandecode_blend_equation(&b
->equation
);
1395 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1396 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1397 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1400 pandecode_log("},\n");
1406 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1408 /* constant/equation is in a union */
1412 pandecode_log(".blend = {\n");
1416 pandecode_shader_address("shader", blend
->shader
);
1418 pandecode_blend_equation(&blend
->equation
);
1419 pandecode_prop("constant = %f", blend
->constant
);
1423 pandecode_log("},\n");
1425 /* Return blend shader to disassemble if present */
1426 return is_shader
? (blend
->shader
& ~0xF) : 0;
1430 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1432 struct midgard_blend_rt
*b
=
1433 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1435 /* Flags determine presence of blend shader */
1436 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1438 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1441 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1443 union midgard_blend blend
= b
->blend
;
1444 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1447 pandecode_log("};\n");
1452 /* Attributes and varyings have descriptor records, which contain information
1453 * about their format and ordering with the attribute/varying buffers. We'll
1454 * want to validate that the combinations specified are self-consistent.
1458 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1461 char *prefix
= varying
? "varying" : "attribute";
1462 unsigned max_index
= 0;
1463 snprintf(base
, sizeof(base
), "%s_meta", prefix
);
1465 struct mali_attr_meta
*attr_meta
;
1466 mali_ptr p
= varying
? (v
->varying_meta
& ~0xF) : v
->attribute_meta
;
1468 struct pandecode_mapped_memory
*attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
);
1470 for (int i
= 0; i
< count
; ++i
, p
+= sizeof(struct mali_attr_meta
)) {
1471 attr_meta
= pandecode_fetch_gpu_mem(attr_mem
, p
,
1474 /* If the record is discard, it should be zero for everything else */
1476 if (attr_meta
->format
== MALI_VARYING_DISCARD
) {
1479 attr_meta
->unknown1
|
1480 attr_meta
->unknown3
|
1481 attr_meta
->src_offset
;
1484 pandecode_msg("XXX: expected empty record for varying discard\n");
1486 /* We want to look for a literal 0000 swizzle -- this
1487 * is not encoded with all zeroes, however */
1489 enum mali_channel z
= MALI_CHANNEL_ZERO
;
1490 unsigned zero_swizzle
= z
| (z
<< 3) | (z
<< 6) | (z
<< 9);
1491 bool good_swizzle
= attr_meta
->swizzle
== zero_swizzle
;
1494 pandecode_msg("XXX: expected zero swizzle for discard\n");
1497 pandecode_msg("XXX: cannot discard attribute\n");
1499 /* If we're all good, omit the record */
1500 if (!zero
&& varying
&& good_swizzle
) {
1501 pandecode_log("/* discarded varying */\n");
1506 if (attr_meta
->index
> max_index
)
1507 max_index
= attr_meta
->index
;
1509 if (attr_meta
->unknown1
!= 0x2) {
1510 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1511 pandecode_prop("unknown1 = 0x%" PRIx64
, (u64
) attr_meta
->unknown1
);
1514 if (attr_meta
->unknown3
) {
1515 pandecode_msg("XXX: unexpected unknown3 set\n");
1516 pandecode_prop("unknown3 = 0x%" PRIx64
, (u64
) attr_meta
->unknown3
);
1519 pandecode_format_short(attr_meta
->format
, false);
1520 pandecode_log_cont(" %s_%u", prefix
, attr_meta
->index
);
1522 if (attr_meta
->src_offset
)
1523 pandecode_log_cont("[%u]", attr_meta
->src_offset
);
1525 pandecode_swizzle(attr_meta
->swizzle
, attr_meta
->format
);
1527 pandecode_log_cont(";\n");
1530 pandecode_log("\n");
1532 return count
? (max_index
+ 1) : 0;
1536 pandecode_indices(uintptr_t pindices
, uint32_t index_count
, int job_no
)
1538 struct pandecode_mapped_memory
*imem
= pandecode_find_mapped_gpu_mem_containing(pindices
);
1541 /* Indices are literally just a u32 array :) */
1543 uint32_t *PANDECODE_PTR_VAR(indices
, imem
, pindices
);
1545 pandecode_log("uint32_t indices_%d[] = {\n", job_no
);
1548 for (unsigned i
= 0; i
< (index_count
+ 1); i
+= 3)
1549 pandecode_log("%d, %d, %d,\n",
1555 pandecode_log("};\n");
1559 /* return bits [lo, hi) of word */
1561 bits(u32 word
, u32 lo
, u32 hi
)
1564 return word
; // avoid undefined behavior with the shift
1566 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1570 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool noninstanced
)
1572 pandecode_log_cont("{\n");
1575 /* Decode invocation_count. See the comment before the definition of
1576 * invocation_count for an explanation.
1579 unsigned size_x
= bits(p
->invocation_count
, 0, p
->size_y_shift
) + 1;
1580 unsigned size_y
= bits(p
->invocation_count
, p
->size_y_shift
, p
->size_z_shift
) + 1;
1581 unsigned size_z
= bits(p
->invocation_count
, p
->size_z_shift
, p
->workgroups_x_shift
) + 1;
1583 unsigned groups_x
= bits(p
->invocation_count
, p
->workgroups_x_shift
, p
->workgroups_y_shift
) + 1;
1584 unsigned groups_y
= bits(p
->invocation_count
, p
->workgroups_y_shift
, p
->workgroups_z_shift
) + 1;
1585 unsigned groups_z
= bits(p
->invocation_count
, p
->workgroups_z_shift
, 32) + 1;
1587 /* Even though we have this decoded, we want to ensure that the
1588 * representation is "unique" so we don't lose anything by printing only
1589 * the final result. More specifically, we need to check that we were
1590 * passed something in canonical form, since the definition per the
1591 * hardware is inherently not unique. How? Well, take the resulting
1592 * decode and pack it ourselves! If it is bit exact with what we
1593 * decoded, we're good to go. */
1595 struct mali_vertex_tiler_prefix ref
;
1596 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, noninstanced
);
1599 (p
->invocation_count
== ref
.invocation_count
) &&
1600 (p
->size_y_shift
== ref
.size_y_shift
) &&
1601 (p
->size_z_shift
== ref
.size_z_shift
) &&
1602 (p
->workgroups_x_shift
== ref
.workgroups_x_shift
) &&
1603 (p
->workgroups_y_shift
== ref
.workgroups_y_shift
) &&
1604 (p
->workgroups_z_shift
== ref
.workgroups_z_shift
) &&
1605 (p
->workgroups_x_shift_2
== ref
.workgroups_x_shift_2
);
1608 pandecode_msg("XXX: non-canonical workgroups packing\n");
1609 pandecode_msg("expected: %X, %d, %d, %d, %d, %d\n",
1610 ref
.invocation_count
,
1613 ref
.workgroups_x_shift
,
1614 ref
.workgroups_y_shift
,
1615 ref
.workgroups_z_shift
,
1616 ref
.workgroups_x_shift_2
);
1618 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1619 pandecode_prop("size_y_shift = %d", p
->size_y_shift
);
1620 pandecode_prop("size_z_shift = %d", p
->size_z_shift
);
1621 pandecode_prop("workgroups_x_shift = %d", p
->workgroups_x_shift
);
1622 pandecode_prop("workgroups_y_shift = %d", p
->workgroups_y_shift
);
1623 pandecode_prop("workgroups_z_shift = %d", p
->workgroups_z_shift
);
1624 pandecode_prop("workgroups_x_shift_2 = %d", p
->workgroups_x_shift_2
);
1627 /* Regardless, print the decode */
1628 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1629 size_x
, size_y
, size_z
,
1630 groups_x
, groups_y
, groups_z
);
1633 if (p
->unknown_draw
)
1634 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1636 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1638 if (p
->draw_mode
!= MALI_DRAW_NONE
)
1639 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p
->draw_mode
));
1641 /* Index count only exists for tiler jobs anyway */
1644 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1646 if (p
->offset_bias_correction
)
1647 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1649 /* TODO: Figure out what this is. It's not zero */
1650 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1653 pandecode_log("},\n");
1657 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1659 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1660 struct mali_uniform_buffer_meta
*PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1662 for (int i
= 0; i
< ubufs_count
; i
++) {
1663 unsigned size
= (ubufs
[i
].size
+ 1) * 16;
1664 mali_ptr addr
= ubufs
[i
].ptr
<< 2;
1666 pandecode_validate_buffer(addr
, size
);
1668 char *ptr
= pointer_as_memory_reference(ubufs
[i
].ptr
<< 2);
1669 pandecode_log("ubuf_%d[%u] = %s;\n", i
, size
, ptr
);
1673 pandecode_log("\n");
1677 pandecode_scratchpad(uintptr_t pscratchpad
, int job_no
, char *suffix
)
1680 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(pscratchpad
);
1682 struct bifrost_scratchpad
*PANDECODE_PTR_VAR(scratchpad
, mem
, pscratchpad
);
1684 if (scratchpad
->zero
) {
1685 pandecode_msg("XXX: scratchpad zero tripped");
1686 pandecode_prop("zero = 0x%x\n", scratchpad
->zero
);
1689 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64
"_%d%s = {\n", pscratchpad
, job_no
, suffix
);
1692 pandecode_prop("flags = 0x%x", scratchpad
->flags
);
1693 MEMORY_PROP(scratchpad
, gpu_scratchpad
);
1696 pandecode_log("};\n");
1699 static unsigned shader_id
= 0;
1701 static struct midgard_disasm_stats
1702 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1703 bool is_bifrost
, unsigned nr_regs
)
1705 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1706 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1708 /* Compute maximum possible size */
1709 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1711 /* Print some boilerplate to clearly denote the assembly (which doesn't
1712 * obey indentation rules), and actually do the disassembly! */
1716 struct midgard_disasm_stats stats
;
1719 disassemble_bifrost(code
, sz
, false);
1721 /* TODO: Extend stats to Bifrost */
1722 stats
.texture_count
= -1;
1723 stats
.sampler_count
= -1;
1724 stats
.attribute_count
= -1;
1725 stats
.varying_count
= -1;
1726 stats
.uniform_count
= -1;
1727 stats
.uniform_buffer_count
= -1;
1728 stats
.work_count
= -1;
1730 stats
.instruction_count
= 0;
1731 stats
.bundle_count
= 0;
1732 stats
.quadword_count
= 0;
1734 stats
= disassemble_midgard(code
, sz
);
1735 stats
.work_count
= nr_regs
;
1738 /* Print shader-db stats */
1740 unsigned nr_threads
=
1741 (stats
.work_count
<= 4) ? 4 :
1742 (stats
.work_count
<= 8) ? 2 :
1745 printf("shader%d - %s shader: "
1746 "%u inst, %u bundles, %u quadwords, "
1747 "%u registers, %u threads, 0 loops\n\n\n",
1749 (type
== JOB_TYPE_TILER
) ? "FRAGMENT" : "VERTEX",
1750 stats
.instruction_count
, stats
.bundle_count
, stats
.quadword_count
,
1751 stats
.work_count
, nr_threads
);
1758 pandecode_texture(mali_ptr u
,
1759 struct pandecode_mapped_memory
*tmem
,
1760 unsigned job_no
, unsigned tex
)
1762 struct mali_texture_descriptor
*PANDECODE_PTR_VAR(t
, tmem
, u
);
1764 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64
"_%d_%d = {\n", u
, job_no
, tex
);
1767 struct mali_texture_format f
= t
->format
;
1769 /* See the definiton of enum mali_texture_type */
1771 bool is_cube
= f
.type
== MALI_TEX_CUBE
;
1772 unsigned dimension
= is_cube
? 2 : f
.type
;
1774 pandecode_make_indent();
1776 /* TODO: Are there others? */
1777 bool is_zs
= f
.format
== MALI_Z32_UNORM
;
1779 /* Recall Z/S switched the meaning of linear/tiled .. */
1780 if (is_zs
&& f
.layout
== MALI_TEXTURE_LINEAR
)
1781 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1783 /* Print the layout. Default is linear; a modifier can denote AFBC or
1784 * u-interleaved/tiled modes */
1786 if (f
.layout
== MALI_TEXTURE_AFBC
)
1787 pandecode_log_cont("afbc");
1788 else if (f
.layout
== MALI_TEXTURE_TILED
)
1789 pandecode_log_cont(is_zs
? "linear" : "tiled");
1790 else if (f
.layout
== MALI_TEXTURE_LINEAR
)
1791 pandecode_log_cont("linear");
1793 pandecode_msg("XXX: invalid texture layout 0x%X\n", f
.layout
);
1795 pandecode_swizzle(t
->swizzle
, f
.format
);
1796 pandecode_log_cont(" ");
1798 /* Distinguish cube/2D with modifier */
1801 pandecode_log_cont("cube ");
1803 pandecode_format_short(f
.format
, f
.srgb
);
1804 pandecode_swizzle(f
.swizzle
, f
.format
);
1806 /* All four width/height/depth/array_size dimensions are present
1807 * regardless of the type of texture, but it is an error to have
1808 * non-zero dimensions for unused dimensions. Verify this. array_size
1809 * can always be set, as can width. */
1811 if (t
->height
&& dimension
< 2)
1812 pandecode_msg("XXX: nonzero height for <2D texture\n");
1814 if (t
->depth
&& dimension
< 3)
1815 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1817 /* Print only the dimensions that are actually there */
1819 pandecode_log_cont(": %d", t
->width
+ 1);
1822 pandecode_log_cont("x%u", t
->height
+ 1);
1825 pandecode_log_cont("x%u", t
->depth
+ 1);
1828 pandecode_log_cont("[%u]", t
->array_size
+ 1);
1831 pandecode_log_cont(" mip %u", t
->levels
);
1833 pandecode_log_cont("\n");
1835 if (f
.unknown1
| f
.zero
) {
1836 pandecode_msg("XXX: texture format zero tripped\n");
1837 pandecode_prop("unknown1 = %" PRId32
, f
.unknown1
);
1838 pandecode_prop("zero = %" PRId32
, f
.zero
);
1842 pandecode_msg("XXX: expected unknown texture bit set\n");
1843 pandecode_prop("unknown2 = %" PRId32
, f
.unknown1
);
1846 if (t
->swizzle_zero
) {
1847 pandecode_msg("XXX: swizzle zero tripped\n");
1848 pandecode_prop("swizzle_zero = %d", t
->swizzle_zero
);
1851 if (t
->unknown3
| t
->unknown3A
| t
->unknown5
| t
->unknown6
| t
->unknown7
) {
1852 pandecode_msg("XXX: texture zero tripped\n");
1853 pandecode_prop("unknown3 = %" PRId16
, t
->unknown3
);
1854 pandecode_prop("unknown3A = %" PRId8
, t
->unknown3A
);
1855 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
1856 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
1857 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
1860 pandecode_log(".payload = {\n");
1863 /* A bunch of bitmap pointers follow.
1864 * We work out the correct number,
1865 * based on the mipmap/cubemap
1866 * properties, but dump extra
1867 * possibilities to futureproof */
1869 int bitmap_count
= MALI_NEGATIVE(t
->levels
);
1871 /* Miptree for each face */
1872 if (f
.type
== MALI_TEX_CUBE
)
1875 /* Array of textures */
1876 bitmap_count
*= MALI_NEGATIVE(t
->array_size
);
1878 /* Stride for each element */
1879 if (f
.manual_stride
)
1882 /* Sanity check the size */
1883 int max_count
= sizeof(t
->payload
) / sizeof(t
->payload
[0]);
1884 assert (bitmap_count
<= max_count
);
1886 for (int i
= 0; i
< bitmap_count
; ++i
) {
1887 /* How we dump depends if this is a stride or a pointer */
1889 if (f
.manual_stride
&& (i
& 1)) {
1890 /* signed 32-bit snuck in as a 64-bit pointer */
1891 uint64_t stride_set
= t
->payload
[i
];
1892 uint32_t clamped_stride
= stride_set
;
1893 int32_t stride
= clamped_stride
;
1894 assert(stride_set
== clamped_stride
);
1895 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
1897 char *a
= pointer_as_memory_reference(t
->payload
[i
]);
1898 pandecode_log("%s, \n", a
);
1904 pandecode_log("},\n");
1907 pandecode_log("};\n");
1911 pandecode_vertex_tiler_postfix_pre(
1912 const struct mali_vertex_tiler_postfix
*p
,
1913 int job_no
, enum mali_job_type job_type
,
1914 char *suffix
, bool is_bifrost
)
1916 mali_ptr shader_meta_ptr
= (u64
) (uintptr_t) (p
->_shader_upper
<< 4);
1917 struct pandecode_mapped_memory
*attr_mem
;
1919 unsigned rt_count
= 1;
1921 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1922 * are the only things actually needed from the FBD, vertex/tiler jobs
1923 * no longer reference the FBD -- instead, this field points to some
1924 * info about the scratchpad.
1927 pandecode_scratchpad(p
->framebuffer
& ~FBD_TYPE
, job_no
, suffix
);
1928 else if (p
->framebuffer
& MALI_MFBD
)
1929 rt_count
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->framebuffer
) & FBD_MASK
, job_no
, false);
1930 else if (job_type
== JOB_TYPE_COMPUTE
)
1931 pandecode_compute_fbd((u64
) (uintptr_t) p
->framebuffer
, job_no
);
1933 pandecode_sfbd((u64
) (uintptr_t) p
->framebuffer
, job_no
, false);
1935 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
1936 int texture_count
= 0, sampler_count
= 0;
1938 if (shader_meta_ptr
) {
1939 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr
);
1940 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, shader_meta_ptr
);
1942 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", shader_meta_ptr
, job_no
, suffix
);
1945 /* Save for dumps */
1946 attribute_count
= s
->attribute_count
;
1947 varying_count
= s
->varying_count
;
1948 texture_count
= s
->texture_count
;
1949 sampler_count
= s
->sampler_count
;
1952 uniform_count
= s
->bifrost2
.uniform_count
;
1953 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
1955 uniform_count
= s
->midgard1
.uniform_buffer_count
;
1956 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
1959 mali_ptr shader_ptr
= pandecode_shader_address("shader", s
->shader
);
1961 pandecode_prop("texture_count = %" PRId16
, s
->texture_count
);
1962 pandecode_prop("sampler_count = %" PRId16
, s
->sampler_count
);
1963 pandecode_prop("attribute_count = %" PRId16
, s
->attribute_count
);
1964 pandecode_prop("varying_count = %" PRId16
, s
->varying_count
);
1966 unsigned nr_registers
= 0;
1969 pandecode_log(".bifrost1 = {\n");
1972 pandecode_prop("uniform_buffer_count = %" PRId32
, s
->bifrost1
.uniform_buffer_count
);
1973 pandecode_prop("unk1 = 0x%" PRIx32
, s
->bifrost1
.unk1
);
1976 pandecode_log("},\n");
1978 pandecode_log(".midgard1 = {\n");
1981 pandecode_prop("uniform_count = %" PRId16
, s
->midgard1
.uniform_count
);
1982 pandecode_prop("uniform_buffer_count = %" PRId16
, s
->midgard1
.uniform_buffer_count
);
1983 pandecode_prop("work_count = %" PRId16
, s
->midgard1
.work_count
);
1984 nr_registers
= s
->midgard1
.work_count
;
1986 pandecode_log(".flags = ");
1987 pandecode_log_decoded_flags(shader_midgard1_flag_info
, s
->midgard1
.flags
);
1988 pandecode_log_cont(",\n");
1990 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->midgard1
.unknown2
);
1993 pandecode_log("},\n");
1996 if (s
->depth_units
|| s
->depth_factor
) {
1997 pandecode_prop("depth_factor = %f", s
->depth_factor
);
1998 pandecode_prop("depth_units = %f", s
->depth_units
);
2001 if (s
->alpha_coverage
) {
2002 bool invert_alpha_coverage
= s
->alpha_coverage
& 0xFFF0;
2003 uint16_t inverted_coverage
= invert_alpha_coverage
? ~s
->alpha_coverage
: s
->alpha_coverage
;
2005 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2006 invert_alpha_coverage
? "~" : "",
2007 MALI_GET_ALPHA_COVERAGE(inverted_coverage
));
2010 if (s
->unknown2_3
|| s
->unknown2_4
) {
2011 pandecode_log(".unknown2_3 = ");
2013 int unknown2_3
= s
->unknown2_3
;
2014 int unknown2_4
= s
->unknown2_4
;
2016 /* We're not quite sure what these flags mean without the depth test, if anything */
2018 if (unknown2_3
& (MALI_DEPTH_TEST
| MALI_DEPTH_FUNC_MASK
)) {
2019 const char *func
= pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3
));
2020 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
2022 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
2025 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
2026 pandecode_log_cont(",\n");
2028 pandecode_log(".unknown2_4 = ");
2029 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
2030 pandecode_log_cont(",\n");
2033 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
2034 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
2035 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
2038 pandecode_stencil("front", &s
->stencil_front
);
2039 pandecode_stencil("back", &s
->stencil_back
);
2042 pandecode_log(".bifrost2 = {\n");
2045 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
2046 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
2047 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
2048 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
2051 pandecode_log("},\n");
2052 } else if (s
->midgard2
.unknown2_7
) {
2053 pandecode_log(".midgard2 = {\n");
2056 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
2058 pandecode_log("},\n");
2062 pandecode_prop("unknown2_8 = 0x%" PRIx32
, s
->unknown2_8
);
2065 /* TODO: Blend shaders routing/disasm */
2067 union midgard_blend blend
= s
->blend
;
2068 pandecode_midgard_blend(&blend
, false);
2072 pandecode_log("};\n");
2074 /* MRT blend fields are used whenever MFBD is used, with
2075 * per-RT descriptors */
2077 if (job_type
== JOB_TYPE_TILER
) {
2078 void* blend_base
= (void *) (s
+ 1);
2080 for (unsigned i
= 0; i
< rt_count
; i
++) {
2081 mali_ptr shader
= 0;
2084 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
2086 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
2089 pandecode_shader_disassemble(shader
, job_no
, job_type
, false, 0);
2093 if (shader_ptr
& ~0xF)
2094 pandecode_shader_disassemble(shader_ptr
, job_no
, job_type
, is_bifrost
, nr_registers
);
2096 pandecode_msg("<no shader>\n");
2099 struct pandecode_mapped_memory
*fmem
= pandecode_find_mapped_gpu_mem_containing(p
->viewport
);
2100 struct mali_viewport
*PANDECODE_PTR_VAR(f
, fmem
, p
->viewport
);
2102 pandecode_log("struct mali_viewport viewport_%"PRIx64
"_%d%s = {\n", p
->viewport
, job_no
, suffix
);
2105 pandecode_prop("clip_minx = %f", f
->clip_minx
);
2106 pandecode_prop("clip_miny = %f", f
->clip_miny
);
2107 pandecode_prop("clip_minz = %f", f
->clip_minz
);
2108 pandecode_prop("clip_maxx = %f", f
->clip_maxx
);
2109 pandecode_prop("clip_maxy = %f", f
->clip_maxy
);
2110 pandecode_prop("clip_maxz = %f", f
->clip_maxz
);
2112 /* Only the higher coordinates are MALI_POSITIVE scaled */
2114 pandecode_prop("viewport0 = { %d, %d }",
2115 f
->viewport0
[0], f
->viewport0
[1]);
2117 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2118 f
->viewport1
[0] + 1, f
->viewport1
[1] + 1);
2121 pandecode_log("};\n");
2124 if (p
->attribute_meta
) {
2125 unsigned max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
2127 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
2128 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false);
2131 /* Varyings are encoded like attributes but not actually sent; we just
2132 * pass a zero buffer with the right stride/size set, (or whatever)
2133 * since the GPU will write to it itself */
2135 if (p
->varying_meta
) {
2136 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
2140 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
2142 /* Number of descriptors depends on whether there are
2143 * non-internal varyings */
2145 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true);
2148 if (p
->uniform_buffers
) {
2149 if (uniform_buffer_count
)
2150 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
2152 pandecode_msg("XXX: UBOs specified but not referenced\n");
2153 } else if (uniform_buffer_count
)
2154 pandecode_msg("XXX: UBOs referenced but not specified\n");
2156 /* We don't want to actually dump uniforms, but we do need to validate
2157 * that the counts we were given are sane */
2161 pandecode_validate_buffer(p
->uniforms
, uniform_count
* 16);
2163 pandecode_msg("XXX: Uniforms specified but not referenced");
2164 } else if (uniform_count
)
2165 pandecode_msg("XXX: UBOs referenced but not specified\n");
2167 if (p
->texture_trampoline
) {
2168 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(p
->texture_trampoline
);
2171 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
);
2173 pandecode_log("uint64_t texture_trampoline_%"PRIx64
"_%d[] = {\n", p
->texture_trampoline
, job_no
);
2176 for (int tex
= 0; tex
< texture_count
; ++tex
) {
2177 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
2178 char *a
= pointer_as_memory_reference(*u
);
2179 pandecode_log("%s,\n", a
);
2184 pandecode_log("};\n");
2186 /* Now, finally, descend down into the texture descriptor */
2187 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
2188 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
2189 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
2191 pandecode_texture(*u
, tmem
, job_no
, tex
);
2196 if (p
->sampler_descriptor
) {
2197 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->sampler_descriptor
);
2200 struct mali_sampler_descriptor
*s
;
2202 mali_ptr d
= p
->sampler_descriptor
;
2204 for (int i
= 0; i
< sampler_count
; ++i
) {
2205 s
= pandecode_fetch_gpu_mem(smem
, d
+ sizeof(*s
) * i
, sizeof(*s
));
2207 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", d
+ sizeof(*s
) * i
, job_no
, i
);
2210 pandecode_log(".filter_mode = ");
2211 pandecode_log_decoded_flags(sampler_flag_info
, s
->filter_mode
);
2212 pandecode_log_cont(",\n");
2214 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
2215 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
2217 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s
->wrap_s
));
2218 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s
->wrap_t
));
2219 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s
->wrap_r
));
2221 pandecode_prop("compare_func = %s", pandecode_alt_func(s
->compare_func
));
2223 if (s
->zero
|| s
->zero2
) {
2224 pandecode_msg("XXX: sampler zero tripped\n");
2225 pandecode_prop("zero = 0x%X, 0x%X\n", s
->zero
, s
->zero2
);
2228 pandecode_prop("seamless_cube_map = %d", s
->seamless_cube_map
);
2230 pandecode_prop("border_color = { %f, %f, %f, %f }",
2234 s
->border_color
[3]);
2237 pandecode_log("};\n");
2244 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2246 if (!(p
->position_varying
|| p
->occlusion_counter
|| p
->flags
))
2249 pandecode_log(".postfix = {\n");
2252 MEMORY_PROP(p
, position_varying
);
2253 MEMORY_PROP(p
, occlusion_counter
);
2256 pandecode_prop("flags = %d", p
->flags
);
2259 pandecode_log("},\n");
2263 pandecode_vertex_only_bfr(struct bifrost_vertex_only
*v
)
2265 pandecode_log_cont("{\n");
2268 pandecode_prop("unk2 = 0x%x", v
->unk2
);
2270 if (v
->zero0
|| v
->zero1
) {
2271 pandecode_msg("XXX: vertex only zero tripped");
2272 pandecode_prop("zero0 = 0x%" PRIx32
, v
->zero0
);
2273 pandecode_prop("zero1 = 0x%" PRIx64
, v
->zero1
);
2277 pandecode_log("}\n");
2281 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2284 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2285 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2287 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no
);
2291 pandecode_msg("XXX: tiler heap zero tripped\n");
2292 pandecode_prop("zero = 0x%x", h
->zero
);
2295 for (int i
= 0; i
< 12; i
++) {
2296 if (h
->zeros
[i
] != 0) {
2297 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2302 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2303 MEMORY_PROP(h
, tiler_heap_start
);
2304 MEMORY_PROP(h
, tiler_heap_free
);
2306 /* this might point to the beginning of another buffer, when it's
2307 * really the end of the tiler heap buffer, so we have to be careful
2308 * here. but for zero length, we need the same pointer.
2311 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2312 MEMORY_PROP(h
, tiler_heap_start
);
2314 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2315 pandecode_prop("tiler_heap_end = %s + 1", a
);
2320 pandecode_log("};\n");
2324 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2326 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2327 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2329 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2331 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no
);
2334 if (t
->zero0
|| t
->zero1
) {
2335 pandecode_msg("XXX: tiler meta zero tripped\n");
2336 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2337 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2340 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2341 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2343 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2344 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2346 for (int i
= 0; i
< 12; i
++) {
2347 if (t
->zeros
[i
] != 0) {
2348 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2354 pandecode_log("};\n");
2358 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2360 pandecode_log(".gl_enables = ");
2362 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2364 pandecode_log_cont(",\n");
2368 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2370 if (u
.pointer
== 0x0)
2373 pandecode_log(".primitive_size = {\n");
2377 pandecode_prop("constant = %f", u
.constant
);
2379 MEMORY_PROP((&u
), pointer
);
2383 pandecode_log("},\n");
2387 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2389 pandecode_log_cont("{\n");
2392 /* TODO: gl_PointSize on Bifrost */
2393 pandecode_primitive_size(t
->primitive_size
, true);
2395 pandecode_gl_enables(t
->gl_enables
, JOB_TYPE_TILER
);
2397 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2398 || t
->zero6
|| t
->zero7
|| t
->zero8
) {
2399 pandecode_msg("XXX: tiler only zero tripped\n");
2400 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2401 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2402 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2403 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2404 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2405 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2406 pandecode_prop("zero7 = 0x%" PRIx32
, t
->zero7
);
2407 pandecode_prop("zero8 = 0x%" PRIx64
, t
->zero8
);
2411 pandecode_log("},\n");
2415 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2416 const struct pandecode_mapped_memory
*mem
,
2417 mali_ptr payload
, int job_no
)
2419 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2421 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true);
2423 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no
);
2426 pandecode_log(".prefix = ");
2427 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2429 pandecode_log(".vertex = ");
2430 pandecode_vertex_only_bfr(&v
->vertex
);
2432 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2435 pandecode_log("};\n");
2441 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2442 const struct pandecode_mapped_memory
*mem
,
2443 mali_ptr payload
, int job_no
)
2445 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2447 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true);
2449 pandecode_indices(t
->prefix
.indices
, t
->prefix
.index_count
, job_no
);
2450 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2452 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no
);
2455 pandecode_log(".prefix = ");
2456 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2458 pandecode_log(".tiler = ");
2459 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2461 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2464 pandecode_log("};\n");
2470 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2471 const struct pandecode_mapped_memory
*mem
,
2472 mali_ptr payload
, int job_no
)
2474 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2476 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false);
2478 pandecode_indices(v
->prefix
.indices
, v
->prefix
.index_count
, job_no
);
2480 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2483 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2484 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2486 bool instanced
= v
->instance_shift
|| v
->instance_odd
;
2487 bool is_graphics
= (h
->job_type
== JOB_TYPE_VERTEX
) || (h
->job_type
== JOB_TYPE_TILER
);
2489 pandecode_log(".prefix = ");
2490 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, !instanced
&& is_graphics
);
2492 pandecode_gl_enables(v
->gl_enables
, h
->job_type
);
2494 if (v
->instance_shift
|| v
->instance_odd
) {
2495 pandecode_prop("instance_shift = 0x%d /* %d */",
2496 v
->instance_shift
, 1 << v
->instance_shift
);
2497 pandecode_prop("instance_odd = 0x%X /* %d */",
2498 v
->instance_odd
, (2 * v
->instance_odd
) + 1);
2500 pandecode_padded_vertices(v
->instance_shift
, v
->instance_odd
);
2503 if (v
->offset_start
)
2504 pandecode_prop("offset_start = %d", v
->offset_start
);
2507 pandecode_msg("XXX: midgard payload zero tripped\n");
2508 pandecode_prop("zero5 = 0x%" PRIx64
, v
->zero5
);
2511 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2514 pandecode_log("};\n");
2520 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2521 mali_ptr payload
, int job_no
,
2524 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2526 bool fbd_dumped
= false;
2528 if (!is_bifrost
&& (s
->framebuffer
& FBD_TYPE
) == MALI_SFBD
) {
2529 /* Only SFBDs are understood, not MFBDs. We're speculating,
2530 * based on the versioning, kernel code, etc, that the
2531 * difference is between Single FrameBuffer Descriptor and
2532 * Multiple FrmaeBuffer Descriptor; the change apparently lines
2533 * up with multi-framebuffer support being added (T7xx onwards,
2534 * including Gxx). In any event, there's some field shuffling
2535 * that we haven't looked into yet. */
2537 pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true);
2539 } else if ((s
->framebuffer
& FBD_TYPE
) == MALI_MFBD
) {
2540 /* We don't know if Bifrost supports SFBD's at all, since the
2541 * driver never uses them. And the format is different from
2542 * Midgard anyways, due to the tiler heap and scratchpad being
2543 * moved out into separate structures, so it's not clear what a
2544 * Bifrost SFBD would even look like without getting an actual
2545 * trace, which appears impossible.
2548 pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true);
2552 uintptr_t p
= (uintptr_t) s
->framebuffer
& FBD_MASK
;
2553 pandecode_log("struct mali_payload_fragment payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2556 /* See the comments by the macro definitions for mathematical context
2557 * on why this is so weird */
2559 if (MALI_TILE_COORD_FLAGS(s
->max_tile_coord
) || MALI_TILE_COORD_FLAGS(s
->min_tile_coord
))
2560 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
2562 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
2563 MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
,
2564 MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
);
2566 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
2567 (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
,
2568 (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
);
2570 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2571 * we have to fallback on offsets. */
2573 const char *fbd_type
= s
->framebuffer
& MALI_MFBD
? "MALI_MFBD" : "MALI_SFBD";
2576 unsigned extra_flags
= (s
->framebuffer
& ~FBD_MASK
) & ~MALI_MFBD
;
2579 pandecode_prop("framebuffer = framebuffer_%d_p | %s | 0x%X", job_no
,
2580 fbd_type
, extra_flags
);
2582 char *a
= pointer_as_memory_reference(p
);
2583 pandecode_prop("framebuffer = %s | %s | 0x%X", a
, fbd_type
, extra_flags
);
2588 pandecode_log("};\n");
2593 static int job_descriptor_number
= 0;
2596 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
)
2598 struct mali_job_descriptor_header
*h
;
2600 int start_number
= 0;
2606 struct pandecode_mapped_memory
*mem
=
2607 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2611 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2613 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2614 * high 32-bits of the 64-bit pointer are reused to store
2617 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
2618 h
->job_type
!= JOB_TYPE_FRAGMENT
? 4 : 0;
2619 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
2621 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
,
2624 int job_no
= job_descriptor_number
++;
2627 start_number
= job_no
;
2629 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2632 pandecode_prop("job_type = %s", pandecode_job_type(h
->job_type
));
2634 /* Save for next job fixing */
2635 last_size
= h
->job_descriptor_size
;
2637 if (h
->job_descriptor_size
)
2638 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2640 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2641 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2642 h
->exception_status
,
2643 (h
->exception_status
>> 16) & 0xFFFF,
2644 pandecode_exception_access((h
->exception_status
>> 8) & 0x3),
2645 h
->exception_status
& 0xFF);
2647 if (h
->first_incomplete_task
)
2648 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2650 if (h
->fault_pointer
)
2651 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2654 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2656 pandecode_prop("job_index = %d", h
->job_index
);
2658 if (h
->unknown_flags
)
2659 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2661 if (h
->job_dependency_index_1
)
2662 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2664 if (h
->job_dependency_index_2
)
2665 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2668 pandecode_log("};\n");
2670 /* Do not touch the field yet -- decode the payload first, and
2671 * don't touch that either. This is essential for the uploads
2672 * to occur in sequence and therefore be dynamically allocated
2673 * correctly. Do note the size, however, for that related
2676 switch (h
->job_type
) {
2677 case JOB_TYPE_SET_VALUE
: {
2678 struct mali_payload_set_value
*s
= payload
;
2679 pandecode_log("struct mali_payload_set_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2681 MEMORY_PROP(s
, out
);
2682 pandecode_prop("unknown = 0x%" PRIX64
, s
->unknown
);
2684 pandecode_log("};\n");
2689 case JOB_TYPE_TILER
:
2690 case JOB_TYPE_VERTEX
:
2691 case JOB_TYPE_COMPUTE
:
2693 if (h
->job_type
== JOB_TYPE_TILER
)
2694 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
);
2696 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
);
2698 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
);
2702 case JOB_TYPE_FRAGMENT
:
2703 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
);
2710 /* Handle linkage */
2713 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no
- 1);
2716 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no
);
2718 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no
);
2723 } while ((jc_gpu_va
= h
->job_descriptor_size
? h
->next_job_64
: h
->next_job_32
));
2725 return start_number
;