2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
26 #include <panfrost-job.h>
34 #include "util/macros.h"
35 #include "util/u_math.h"
37 #include "pan_pretty_print.h"
38 #include "midgard/disassemble.h"
39 #include "bifrost/disassemble.h"
41 #include "pan_encoder.h"
43 static void pandecode_swizzle(unsigned swizzle
, enum mali_format format
);
45 #define MEMORY_PROP(obj, p) {\
47 char *a = pointer_as_memory_reference(obj->p); \
48 pandecode_prop("%s = %s", #p, a); \
53 #define MEMORY_PROP_DIR(obj, p) {\
55 char *a = pointer_as_memory_reference(obj.p); \
56 pandecode_prop("%s = %s", #p, a); \
61 /* 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 %zu at offset %d in buffer of size %zu. "
151 "Overrun by %zu bytes. \n",
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_WRITEMASK
),
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_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
242 static const struct pandecode_flag_info mfbd_fmt_flag_info
[] = {
249 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
250 static const struct pandecode_flag_info mfbd_extra_flag_info
[] = {
258 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
259 static const struct pandecode_flag_info shader_midgard1_flag_info
[] = {
261 FLAG_INFO(READS_TILEBUFFER
),
267 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
268 static const struct pandecode_flag_info mfbd_flag_info
[] = {
269 FLAG_INFO(DEPTH_WRITE
),
275 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
276 static const struct pandecode_flag_info sampler_flag_info
[] = {
277 FLAG_INFO(MAG_NEAREST
),
278 FLAG_INFO(MIN_NEAREST
),
279 FLAG_INFO(MIP_LINEAR_1
),
280 FLAG_INFO(MIP_LINEAR_2
),
281 FLAG_INFO(NORM_COORDS
),
286 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
287 static const struct pandecode_flag_info sfbd_unk1_info
[] = {
294 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
295 static const struct pandecode_flag_info sfbd_unk2_info
[] = {
302 extern char *replace_fragment
;
303 extern char *replace_vertex
;
306 pandecode_job_type(enum mali_job_type type
)
308 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
312 DEFINE_CASE(WRITE_VALUE
);
313 DEFINE_CASE(CACHE_FLUSH
);
314 DEFINE_CASE(COMPUTE
);
318 DEFINE_CASE(FRAGMENT
);
320 case JOB_NOT_STARTED
:
321 return "NOT_STARTED";
324 pandecode_log("Warning! Unknown job type %x\n", type
);
332 pandecode_draw_mode(enum mali_draw_mode mode
)
334 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
337 DEFINE_CASE(DRAW_NONE
);
340 DEFINE_CASE(TRIANGLES
);
341 DEFINE_CASE(TRIANGLE_STRIP
);
342 DEFINE_CASE(TRIANGLE_FAN
);
343 DEFINE_CASE(LINE_STRIP
);
344 DEFINE_CASE(LINE_LOOP
);
345 DEFINE_CASE(POLYGON
);
347 DEFINE_CASE(QUAD_STRIP
);
350 pandecode_msg("XXX: invalid draw mode %X\n", mode
);
357 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
359 pandecode_func(enum mali_func mode
)
366 DEFINE_CASE(GREATER
);
367 DEFINE_CASE(NOTEQUAL
);
372 pandecode_msg("XXX: invalid func %X\n", mode
);
378 /* Why is this duplicated? Who knows... */
379 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
381 pandecode_alt_func(enum mali_alt_func mode
)
388 DEFINE_CASE(GREATER
);
389 DEFINE_CASE(NOTEQUAL
);
394 pandecode_msg("XXX: invalid alt func %X\n", mode
);
400 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
402 pandecode_stencil_op(enum mali_stencil_op op
)
406 DEFINE_CASE(REPLACE
);
409 DEFINE_CASE(INCR_WRAP
);
410 DEFINE_CASE(DECR_WRAP
);
415 pandecode_msg("XXX: invalid stencil op %X\n", op
);
422 static char *pandecode_attr_mode_short(enum mali_attr_mode mode
)
425 /* TODO: Combine to just "instanced" once this can be done
426 * unambiguously in all known cases */
427 case MALI_ATTR_POT_DIVIDE
:
428 return "instanced_pot";
429 case MALI_ATTR_MODULO
:
430 return "instanced_mod";
431 case MALI_ATTR_NPOT_DIVIDE
:
432 return "instanced_npot";
433 case MALI_ATTR_IMAGE
:
435 case MALI_ATTR_INTERNAL
:
438 pandecode_msg("XXX: invalid attribute mode %X\n", mode
);
444 pandecode_special_varying(uint64_t v
)
447 case MALI_VARYING_FRAG_COORD
:
448 return "gl_FragCoord";
449 case MALI_VARYING_FRONT_FACING
:
450 return "gl_FrontFacing";
451 case MALI_VARYING_POINT_COORD
:
452 return "gl_PointCoord";
454 pandecode_msg("XXX: invalid special varying %" PRIx64
"\n", v
);
459 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
461 pandecode_wrap_mode(enum mali_wrap_mode op
)
465 DEFINE_CASE(CLAMP_TO_EDGE
);
466 DEFINE_CASE(CLAMP_TO_BORDER
);
467 DEFINE_CASE(MIRRORED_REPEAT
);
470 pandecode_msg("XXX: invalid wrap mode %X\n", op
);
476 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
478 pandecode_block_format(enum mali_block_format fmt
)
482 DEFINE_CASE(UNKNOWN
);
487 unreachable("Invalid case");
492 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
494 pandecode_exception_access(enum mali_exception_access access
)
498 DEFINE_CASE(EXECUTE
);
503 unreachable("Invalid case");
508 /* Midgard's tiler descriptor is embedded within the
512 pandecode_midgard_tiler_descriptor(
513 const struct midgard_tiler_descriptor
*t
,
519 pandecode_log(".tiler = {\n");
522 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
523 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
525 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
527 /* We know this name from the kernel, but we never see it nonzero */
530 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
532 MEMORY_PROP(t
, polygon_list
);
534 /* The body is offset from the base of the polygon list */
535 assert(t
->polygon_list_body
> t
->polygon_list
);
536 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
538 /* It needs to fit inside the reported size */
539 assert(t
->polygon_list_size
>= body_offset
);
541 /* Check that we fit */
542 struct pandecode_mapped_memory
*plist
=
543 pandecode_find_mapped_gpu_mem_containing(t
->polygon_list
);
545 assert(t
->polygon_list_size
<= plist
->length
);
547 /* Now that we've sanity checked, we'll try to calculate the sizes
548 * ourselves for comparison */
550 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
551 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
553 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
554 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
555 ref_header
, ref_size
);
556 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
557 pandecode_msg("body offset %d\n", body_offset
);
560 /* The tiler heap has a start and end specified -- it should be
561 * identical to what we have in the BO. The exception is if tiling is
564 MEMORY_PROP(t
, heap_start
);
565 assert(t
->heap_end
>= t
->heap_start
);
567 struct pandecode_mapped_memory
*heap
=
568 pandecode_find_mapped_gpu_mem_containing(t
->heap_start
);
570 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
572 /* Tiling is enabled with a special flag */
573 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
574 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
576 bool tiling_enabled
= hierarchy_mask
;
578 if (tiling_enabled
) {
579 /* When tiling is enabled, the heap should be a tight fit */
580 unsigned heap_offset
= t
->heap_start
- heap
->gpu_va
;
581 if ((heap_offset
+ heap_size
) != heap
->length
) {
582 pandecode_msg("XXX: heap size %u (expected %zu)\n",
583 heap_size
, heap
->length
- heap_offset
);
586 /* We should also have no other flags */
588 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
590 /* When tiling is disabled, we should have that flag and no others */
592 if (tiler_flags
!= MALI_TILER_DISABLED
) {
593 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
597 /* We should also have an empty heap */
599 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
603 /* Disabled tiling is used only for clear-only jobs, which are
604 * purely FRAGMENT, so we should never see this for
605 * non-FRAGMENT descriptors. */
608 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
611 /* We've never seen weights used in practice, but we know from the
612 * kernel these fields is there */
614 bool nonzero_weights
= false;
616 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
617 nonzero_weights
|= t
->weights
[w
] != 0x0;
620 if (nonzero_weights
) {
621 pandecode_log(".weights = {");
623 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
624 pandecode_log("%d, ", t
->weights
[w
]);
631 pandecode_log("}\n");
634 /* Information about the framebuffer passed back for
635 * additional analysis */
637 struct pandecode_fbd
{
645 pandecode_sfbd_format(struct mali_sfbd_format format
)
647 pandecode_log(".format = {\n");
650 pandecode_log(".unk1 = ");
651 pandecode_log_decoded_flags(sfbd_unk1_info
, format
.unk1
);
652 pandecode_log_cont(",\n");
654 /* TODO: Map formats so we can check swizzles and print nicely */
655 pandecode_log("swizzle");
656 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
657 pandecode_log_cont(",\n");
659 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
660 MALI_NEGATIVE(format
.nr_channels
));
662 pandecode_log(".unk2 = ");
663 pandecode_log_decoded_flags(sfbd_unk2_info
, format
.unk2
);
664 pandecode_log_cont(",\n");
666 pandecode_prop("block = %s", pandecode_block_format(format
.block
));
668 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
671 pandecode_log("},\n");
674 static struct pandecode_fbd
675 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
, unsigned gpu_id
)
677 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
678 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
680 struct pandecode_fbd info
= {
685 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
688 pandecode_prop("unknown1 = 0x%" PRIx32
, s
->unknown1
);
689 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->unknown2
);
691 pandecode_sfbd_format(s
->format
);
693 info
.width
= s
->width
+ 1;
694 info
.height
= s
->height
+ 1;
696 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", info
.width
);
697 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", info
.height
);
699 MEMORY_PROP(s
, checksum
);
701 if (s
->checksum_stride
)
702 pandecode_prop("checksum_stride = %d", s
->checksum_stride
);
704 MEMORY_PROP(s
, framebuffer
);
705 pandecode_prop("stride = %d", s
->stride
);
707 /* Earlier in the actual commandstream -- right before width -- but we
708 * delay to flow nicer */
710 pandecode_log(".clear_flags = ");
711 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
712 pandecode_log_cont(",\n");
714 if (s
->depth_buffer
) {
715 MEMORY_PROP(s
, depth_buffer
);
716 pandecode_prop("depth_stride = %d", s
->depth_stride
);
719 if (s
->stencil_buffer
) {
720 MEMORY_PROP(s
, stencil_buffer
);
721 pandecode_prop("stencil_stride = %d", s
->stencil_stride
);
724 if (s
->depth_stride_zero
||
725 s
->stencil_stride_zero
||
726 s
->zero7
|| s
->zero8
) {
727 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
728 pandecode_prop("depth_stride_zero = 0x%x",
729 s
->depth_stride_zero
);
730 pandecode_prop("stencil_stride_zero = 0x%x",
731 s
->stencil_stride_zero
);
732 pandecode_prop("zero7 = 0x%" PRIx32
,
734 pandecode_prop("zero8 = 0x%" PRIx32
,
738 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
739 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
740 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
741 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
742 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
745 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
746 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
747 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
748 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
749 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
752 if (s
->clear_stencil
) {
753 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
756 MEMORY_PROP(s
, scratchpad
);
757 const struct midgard_tiler_descriptor t
= s
->tiler
;
759 bool has_hierarchy
= !(gpu_id
== 0x0720 || gpu_id
== 0x0820 || gpu_id
== 0x0830);
760 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
, has_hierarchy
);
763 pandecode_log("};\n");
765 pandecode_prop("zero0 = 0x%" PRIx64
, s
->zero0
);
766 pandecode_prop("zero1 = 0x%" PRIx64
, s
->zero1
);
767 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
768 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
769 pandecode_prop("zero5 = 0x%" PRIx32
, s
->zero5
);
771 printf(".zero3 = {");
773 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
774 printf("%X, ", s
->zero3
[i
]);
778 printf(".zero6 = {");
780 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
781 printf("%X, ", s
->zero6
[i
]);
789 pandecode_u32_slide(unsigned name
, const u32
*slide
, unsigned count
)
791 pandecode_log(".unknown%d = {", name
);
793 for (int i
= 0; i
< count
; ++i
)
794 printf("%X, ", slide
[i
]);
796 pandecode_log("},\n");
799 #define SHORT_SLIDE(num) \
800 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
803 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
805 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
806 const struct mali_compute_fbd
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
808 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
817 /* Extracts the number of components associated with a Mali format */
820 pandecode_format_component_count(enum mali_format fmt
)
822 /* Mask out the format class */
823 unsigned top
= fmt
& 0b11100000;
826 case MALI_FORMAT_SNORM
:
827 case MALI_FORMAT_UINT
:
828 case MALI_FORMAT_UNORM
:
829 case MALI_FORMAT_SINT
:
830 return ((fmt
>> 3) & 3) + 1;
837 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
840 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
843 assert(MALI_CHANNEL_RED
== 0);
845 for (unsigned c
= 0; c
< 4; ++c
) {
846 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
848 if (chan
<= MALI_CHANNEL_ALPHA
)
855 /* Validates that a (format, swizzle) pair is valid, in the sense that the
856 * swizzle doesn't access any components that are undefined in the format.
857 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
861 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
863 unsigned nr_comp
= pandecode_format_component_count(fmt
);
864 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
865 unsigned valid_mask
= (1 << nr_comp
) - 1;
866 unsigned invalid_mask
= ~valid_mask
;
868 if (access_mask
& invalid_mask
) {
869 pandecode_msg("XXX: invalid components accessed\n");
873 /* Check for the default non-swizzling swizzle so we can suppress
874 * useless printing for the defaults */
876 unsigned default_swizzles
[4] = {
877 MALI_CHANNEL_RED
| (MALI_CHANNEL_ZERO
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
878 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
879 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ONE
<< 9),
880 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ALPHA
<< 9)
883 return (swizzle
== default_swizzles
[nr_comp
- 1]);
886 /* Maps MALI_RGBA32F to rgba32f, etc */
889 pandecode_format_short(enum mali_format fmt
, bool srgb
)
891 /* We want a type-like format, so cut off the initial MALI_ */
892 char *format
= pandecode_format(fmt
);
893 format
+= strlen("MALI_");
895 unsigned len
= strlen(format
);
896 char *lower_format
= calloc(1, len
+ 1);
898 for (unsigned i
= 0; i
< len
; ++i
)
899 lower_format
[i
] = tolower(format
[i
]);
901 /* Sanity check sRGB flag is applied to RGB, per the name */
902 if (srgb
&& lower_format
[0] != 'r')
903 pandecode_msg("XXX: sRGB applied to non-colour format\n");
905 /* Just prefix with an s, so you get formats like srgba8_unorm */
907 pandecode_log_cont("s");
909 pandecode_log_cont("%s", lower_format
);
914 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
916 /* First, do some validation */
917 bool trivial_swizzle
= pandecode_validate_format_swizzle(
923 /* Next, print the swizzle */
924 pandecode_log_cont(".");
926 static const char components
[] = "rgba01";
928 for (unsigned c
= 0; c
< 4; ++c
) {
929 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
931 if (chan
>= MALI_CHANNEL_RESERVED_0
) {
932 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
935 pandecode_log_cont("%c", components
[chan
]);
940 pandecode_rt_format(struct mali_rt_format format
)
942 pandecode_log(".format = {\n");
945 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
946 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
947 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
949 pandecode_prop("block = %s", pandecode_block_format(format
.block
));
951 /* TODO: Map formats so we can check swizzles and print nicely */
952 pandecode_log("swizzle");
953 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
954 pandecode_log_cont(",\n");
956 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
957 MALI_NEGATIVE(format
.nr_channels
));
959 pandecode_log(".flags = ");
960 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
961 pandecode_log_cont(",\n");
963 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
964 * which is a faster hardware-based alternative to the wallpaper method
965 * to preserve framebuffer contents across frames. In practice, MFBD
966 * preload is buggy on Midgard, and so this is a chicken bit. If this
967 * bit isn't set, most likely something broke unrelated to preload */
969 if (!format
.no_preload
) {
970 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
971 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
975 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
978 pandecode_log("},\n");
982 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct bifrost_framebuffer
*fb
)
984 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
987 for (int i
= 0; i
< MALI_NEGATIVE(fb
->rt_count_1
); i
++) {
988 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct bifrost_render_target
);
989 struct pandecode_mapped_memory
*mem
=
990 pandecode_find_mapped_gpu_mem_containing(rt_va
);
991 const struct bifrost_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
993 pandecode_log("{\n");
996 pandecode_rt_format(rt
->format
);
998 if (rt
->format
.block
== MALI_BLOCK_AFBC
) {
999 pandecode_log(".afbc = {\n");
1002 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
1003 pandecode_prop("metadata = %s", a
);
1006 pandecode_prop("stride = %d", rt
->afbc
.stride
);
1007 pandecode_prop("unk = 0x%" PRIx32
, rt
->afbc
.unk
);
1010 pandecode_log("},\n");
1011 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.unk
) {
1012 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1018 MEMORY_PROP(rt
, framebuffer
);
1019 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
1021 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
1022 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
1023 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
1024 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
1025 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
1028 if (rt
->zero1
|| rt
->zero2
|| rt
->zero3
) {
1029 pandecode_msg("XXX: render target zeros tripped\n");
1030 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
1031 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
1032 pandecode_prop("zero3 = 0x%" PRIx32
, rt
->zero3
);
1036 pandecode_log("},\n");
1040 pandecode_log("};\n");
1043 static struct pandecode_fbd
1044 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
)
1046 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1047 const struct bifrost_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
1049 struct pandecode_fbd info
;
1051 if (fb
->sample_locations
) {
1052 /* The blob stores all possible sample locations in a single buffer
1053 * allocated on startup, and just switches the pointer when switching
1054 * MSAA state. For now, we just put the data into the cmdstream, but we
1055 * should do something like what the blob does with a real driver.
1057 * There seem to be 32 slots for sample locations, followed by another
1058 * 16. The second 16 is just the center location followed by 15 zeros
1059 * in all the cases I've identified (maybe shader vs. depth/color
1063 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->sample_locations
);
1065 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->sample_locations
);
1067 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
1070 for (int i
= 0; i
< 32 + 16; i
++) {
1071 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
1075 pandecode_log("};\n");
1078 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1081 pandecode_prop("stack_shift = 0x%x", fb
->stack_shift
);
1082 pandecode_prop("unk0 = 0x%x", fb
->unk0
);
1084 if (fb
->sample_locations
)
1085 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
1087 /* Assume that unknown1 was emitted in the last job for
1089 MEMORY_PROP(fb
, unknown1
);
1091 info
.width
= fb
->width1
+ 1;
1092 info
.height
= fb
->height1
+ 1;
1093 info
.rt_count
= fb
->rt_count_1
+ 1;
1095 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
1096 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
1097 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
1098 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
1100 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
1101 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
1102 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
1103 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
1105 pandecode_log(".mfbd_flags = ");
1106 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
1107 pandecode_log_cont(",\n");
1109 if (fb
->clear_stencil
)
1110 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
1112 if (fb
->clear_depth
)
1113 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
1115 /* TODO: What is this? Let's not blow up.. */
1116 if (fb
->unknown2
!= 0x1F)
1117 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
1119 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
1120 MEMORY_PROP(fb
, scratchpad
);
1121 const struct midgard_tiler_descriptor t
= fb
->tiler
;
1122 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
, true);
1124 if (fb
->zero3
|| fb
->zero4
) {
1125 pandecode_msg("XXX: framebuffer zeros tripped\n");
1126 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
1127 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
1131 pandecode_log("};\n");
1133 gpu_va
+= sizeof(struct bifrost_framebuffer
);
1135 info
.has_extra
= (fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
;
1137 if (info
.has_extra
) {
1138 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1139 const struct bifrost_fb_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
1141 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1144 MEMORY_PROP(fbx
, checksum
);
1146 if (fbx
->checksum_stride
)
1147 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
1149 pandecode_log(".flags = ");
1150 pandecode_log_decoded_flags(mfbd_extra_flag_info
, fbx
->flags
);
1151 pandecode_log_cont(",\n");
1153 if (fbx
->flags
& MALI_EXTRA_AFBC_ZS
) {
1154 pandecode_log(".ds_afbc = {\n");
1157 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
1158 pandecode_prop("depth_stencil_afbc_stride = %d",
1159 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
1160 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
1162 if (fbx
->ds_afbc
.zero1
|| fbx
->ds_afbc
.padding
) {
1163 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1164 pandecode_prop("zero1 = 0x%" PRIx32
,
1165 fbx
->ds_afbc
.zero1
);
1166 pandecode_prop("padding = 0x%" PRIx64
,
1167 fbx
->ds_afbc
.padding
);
1171 pandecode_log("},\n");
1173 pandecode_log(".ds_linear = {\n");
1176 if (fbx
->ds_linear
.depth
) {
1177 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
1178 pandecode_prop("depth_stride = %d",
1179 fbx
->ds_linear
.depth_stride
);
1182 if (fbx
->ds_linear
.stencil
) {
1183 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1184 pandecode_prop("stencil_stride = %d",
1185 fbx
->ds_linear
.stencil_stride
);
1188 if (fbx
->ds_linear
.depth_stride_zero
||
1189 fbx
->ds_linear
.stencil_stride_zero
||
1190 fbx
->ds_linear
.zero1
|| fbx
->ds_linear
.zero2
) {
1191 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1192 pandecode_prop("depth_stride_zero = 0x%x",
1193 fbx
->ds_linear
.depth_stride_zero
);
1194 pandecode_prop("stencil_stride_zero = 0x%x",
1195 fbx
->ds_linear
.stencil_stride_zero
);
1196 pandecode_prop("zero1 = 0x%" PRIx32
,
1197 fbx
->ds_linear
.zero1
);
1198 pandecode_prop("zero2 = 0x%" PRIx32
,
1199 fbx
->ds_linear
.zero2
);
1203 pandecode_log("},\n");
1206 if (fbx
->zero3
|| fbx
->zero4
) {
1207 pandecode_msg("XXX: fb_extra zeros tripped\n");
1208 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1209 pandecode_prop("zero4 = 0x%" PRIx64
, fbx
->zero4
);
1213 pandecode_log("};\n");
1215 gpu_va
+= sizeof(struct bifrost_fb_extra
);
1219 pandecode_render_target(gpu_va
, job_no
, fb
);
1224 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1228 pandecode_padded_vertices(unsigned shift
, unsigned k
)
1230 unsigned odd
= 2*k
+ 1;
1231 unsigned pot
= 1 << shift
;
1232 pandecode_msg("padded_num_vertices = %d\n", odd
* pot
);
1235 /* Given a magic divisor, recover what we were trying to divide by.
1237 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1238 * 0 <= m < 2^N, for N bits in m.
1240 * Let q represent the number we would like to divide by.
1242 * By definition of a magic divisor for N-bit unsigned integers (a number you
1243 * multiply by to magically get division), m is a number such that:
1245 * (m * x) & (2^N - 1) = floor(x/q).
1246 * for all x on Z where 0 <= x < 2^N
1248 * Ignore the case where any of the above values equals zero; it is irrelevant
1249 * for our purposes (instanced arrays).
1251 * Choose x = q. Then:
1253 * (m * x) & (2^N - 1) = floor(x/q).
1254 * (m * q) & (2^N - 1) = floor(q/q).
1256 * floor(q/q) = floor(1) = 1, therefore:
1258 * (m * q) & (2^N - 1) = 1
1260 * Recall the identity that the bitwise AND of one less than a power-of-two
1261 * equals the modulo with that power of two, i.e. for all x:
1263 * x & (2^N - 1) = x % N
1269 * By definition, a modular multiplicative inverse of a number m is the number
1270 * q such that with respect to a modulos M:
1274 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1279 pandecode_magic_divisor(uint32_t magic
, unsigned shift
, unsigned orig_divisor
, unsigned extra
)
1282 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1283 * shift) the most lame way possible... just repeatedly add.
1284 * Asymptoptically slow but nobody cares in practice, unless you have
1285 * massive numbers of vertices or high divisors. */
1287 unsigned inverse
= 0;
1289 /* Magic implicitly has the highest bit set */
1292 /* Depending on rounding direction */
1297 uint32_t product
= magic
* inverse
;
1309 pandecode_msg("dividing by %d (maybe off by two)\n", inverse
);
1311 /* Recall we're supposed to divide by (gl_level_divisor *
1312 * padded_num_vertices) */
1314 unsigned padded_num_vertices
= inverse
/ orig_divisor
;
1316 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices
);
1321 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1322 mali_ptr addr
, int job_no
, char *suffix
,
1323 int count
, bool varying
, enum mali_job_type job_type
)
1325 char *prefix
= varying
? "varying" : "attribute";
1329 pandecode_msg("warn: No %s records\n", prefix
);
1333 union mali_attr
*attr
= pandecode_fetch_gpu_mem(mem
, addr
, sizeof(union mali_attr
) * count
);
1335 for (int i
= 0; i
< count
; ++i
) {
1336 /* First, check for special records */
1337 if (attr
[i
].elements
< MALI_VARYING_SPECIAL
) {
1338 /* Special records are always varyings */
1341 pandecode_msg("XXX: Special varying in attribute field\n");
1343 if (job_type
!= JOB_TYPE_TILER
)
1344 pandecode_msg("XXX: Special varying in non-FS\n");
1346 /* We're special, so all fields should be zero */
1347 unsigned zero
= attr
[i
].stride
| attr
[i
].size
;
1348 zero
|= attr
[i
].shift
| attr
[i
].extra_flags
;
1351 pandecode_msg("XXX: Special varying has non-zero fields\n");
1353 /* Print the special varying name */
1354 pandecode_log("varying_%d = %s;", i
, pandecode_special_varying(attr
[i
].elements
));
1359 enum mali_attr_mode mode
= attr
[i
].elements
& 7;
1361 if (mode
== MALI_ATTR_UNUSED
)
1362 pandecode_msg("XXX: unused attribute record\n");
1364 /* For non-linear records, we need to print the type of record */
1365 if (mode
!= MALI_ATTR_LINEAR
)
1366 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode
));
1368 /* Print the name to link with attr_meta */
1369 pandecode_log_cont("%s_%d", prefix
, i
);
1371 /* Print the stride and size */
1372 pandecode_log_cont("<%u>[%u]", attr
[i
].stride
, attr
[i
].size
);
1374 /* TODO: Sanity check the quotient itself. It must be equal to
1375 * (or be greater than, if the driver added padding) the padded
1378 /* Finally, print the pointer */
1379 mali_ptr raw_elements
= attr
[i
].elements
& ~7;
1380 char *a
= pointer_as_memory_reference(raw_elements
);
1381 pandecode_log_cont(" = (%s);\n", a
);
1384 /* Check the pointer */
1385 pandecode_validate_buffer(raw_elements
, attr
[i
].size
);
1387 /* shift/extra_flags exist only for instanced */
1388 if (attr
[i
].shift
| attr
[i
].extra_flags
) {
1389 /* These are set to random values by the blob for
1390 * varyings, most likely a symptom of uninitialized
1391 * memory where the hardware masked the bug. As such we
1392 * put this at a warning, not an error. */
1394 if (mode
== MALI_ATTR_LINEAR
)
1395 pandecode_msg("warn: instancing fields set for linear\n");
1397 pandecode_prop("shift = %d", attr
[i
].shift
);
1398 pandecode_prop("extra_flags = %d", attr
[i
].extra_flags
);
1401 /* Decode further where possible */
1403 if (mode
== MALI_ATTR_MODULO
) {
1404 pandecode_padded_vertices(
1406 attr
[i
].extra_flags
);
1409 if (mode
== MALI_ATTR_NPOT_DIVIDE
) {
1411 pandecode_log("{\n");
1413 pandecode_prop("unk = 0x%x", attr
[i
].unk
);
1414 pandecode_prop("magic_divisor = 0x%08x", attr
[i
].magic_divisor
);
1415 if (attr
[i
].zero
!= 0)
1416 pandecode_prop("XXX: zero tripped (0x%x)\n", attr
[i
].zero
);
1417 pandecode_prop("divisor = %d", attr
[i
].divisor
);
1418 pandecode_magic_divisor(attr
[i
].magic_divisor
, attr
[i
- 1].shift
, attr
[i
].divisor
, attr
[i
- 1].extra_flags
);
1420 pandecode_log("}, \n");
1425 pandecode_log("\n");
1429 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1431 /* TODO: Decode flags */
1432 mali_ptr shader_ptr
= ptr
& ~15;
1434 char *a
= pointer_as_memory_reference(shader_ptr
);
1435 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1442 pandecode_stencil(const char *name
, const struct mali_stencil_test
*stencil
)
1444 unsigned any_nonzero
=
1445 stencil
->ref
| stencil
->mask
| stencil
->func
|
1446 stencil
->sfail
| stencil
->dpfail
| stencil
->dppass
;
1448 if (any_nonzero
== 0)
1451 const char *func
= pandecode_func(stencil
->func
);
1452 const char *sfail
= pandecode_stencil_op(stencil
->sfail
);
1453 const char *dpfail
= pandecode_stencil_op(stencil
->dpfail
);
1454 const char *dppass
= pandecode_stencil_op(stencil
->dppass
);
1457 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil
->zero
);
1459 pandecode_log(".stencil_%s = {\n", name
);
1461 pandecode_prop("ref = %d", stencil
->ref
);
1462 pandecode_prop("mask = 0x%02X", stencil
->mask
);
1463 pandecode_prop("func = %s", func
);
1464 pandecode_prop("sfail = %s", sfail
);
1465 pandecode_prop("dpfail = %s", dpfail
);
1466 pandecode_prop("dppass = %s", dppass
);
1468 pandecode_log("},\n");
1472 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1475 pandecode_msg("XXX: blend zero tripped: %X\n", blend
->zero1
);
1477 pandecode_log(".equation = {\n");
1480 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1481 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1483 pandecode_log(".color_mask = ");
1484 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1485 pandecode_log_cont(",\n");
1488 pandecode_log("},\n");
1491 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1494 decode_bifrost_constant(u16 constant
)
1496 float lo
= (float) (constant
& 0xFF);
1497 float hi
= (float) (constant
>> 8);
1499 return (hi
/ 255.0) + (lo
/ 65535.0);
1503 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1505 struct bifrost_blend_rt
*b
=
1506 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1508 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1511 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1512 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1513 b
->constant
, decode_bifrost_constant(b
->constant
));
1515 /* TODO figure out blend shader enable bit */
1516 pandecode_blend_equation(&b
->equation
);
1517 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1518 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1519 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1522 pandecode_log("},\n");
1528 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1530 /* constant/equation is in a union */
1534 pandecode_log(".blend = {\n");
1538 pandecode_shader_address("shader", blend
->shader
);
1540 pandecode_blend_equation(&blend
->equation
);
1541 pandecode_prop("constant = %f", blend
->constant
);
1545 pandecode_log("},\n");
1547 /* Return blend shader to disassemble if present */
1548 return is_shader
? (blend
->shader
& ~0xF) : 0;
1552 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1554 struct midgard_blend_rt
*b
=
1555 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1557 /* Flags determine presence of blend shader */
1558 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1560 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1563 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1565 union midgard_blend blend
= b
->blend
;
1566 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1569 pandecode_log("};\n");
1574 /* Attributes and varyings have descriptor records, which contain information
1575 * about their format and ordering with the attribute/varying buffers. We'll
1576 * want to validate that the combinations specified are self-consistent.
1580 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1583 char *prefix
= varying
? "varying" : "attribute";
1584 unsigned max_index
= 0;
1585 snprintf(base
, sizeof(base
), "%s_meta", prefix
);
1587 struct mali_attr_meta
*attr_meta
;
1588 mali_ptr p
= varying
? v
->varying_meta
: v
->attribute_meta
;
1590 struct pandecode_mapped_memory
*attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
);
1592 for (int i
= 0; i
< count
; ++i
, p
+= sizeof(struct mali_attr_meta
)) {
1593 attr_meta
= pandecode_fetch_gpu_mem(attr_mem
, p
,
1596 /* If the record is discard, it should be zero for everything else */
1598 if (attr_meta
->format
== MALI_VARYING_DISCARD
) {
1601 attr_meta
->unknown1
|
1602 attr_meta
->unknown3
|
1603 attr_meta
->src_offset
;
1606 pandecode_msg("XXX: expected empty record for varying discard\n");
1608 /* We want to look for a literal 0000 swizzle -- this
1609 * is not encoded with all zeroes, however */
1611 enum mali_channel z
= MALI_CHANNEL_ZERO
;
1612 unsigned zero_swizzle
= z
| (z
<< 3) | (z
<< 6) | (z
<< 9);
1613 bool good_swizzle
= attr_meta
->swizzle
== zero_swizzle
;
1616 pandecode_msg("XXX: expected zero swizzle for discard\n");
1619 pandecode_msg("XXX: cannot discard attribute\n");
1621 /* If we're all good, omit the record */
1622 if (!zero
&& varying
&& good_swizzle
) {
1623 pandecode_log("/* discarded varying */\n");
1628 if (attr_meta
->index
> max_index
)
1629 max_index
= attr_meta
->index
;
1631 if (attr_meta
->unknown1
!= 0x2) {
1632 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1633 pandecode_prop("unknown1 = 0x%" PRIx64
, (u64
) attr_meta
->unknown1
);
1636 if (attr_meta
->unknown3
) {
1637 pandecode_msg("XXX: unexpected unknown3 set\n");
1638 pandecode_prop("unknown3 = 0x%" PRIx64
, (u64
) attr_meta
->unknown3
);
1641 pandecode_format_short(attr_meta
->format
, false);
1642 pandecode_log_cont(" %s_%u", prefix
, attr_meta
->index
);
1644 if (attr_meta
->src_offset
)
1645 pandecode_log_cont("[%u]", attr_meta
->src_offset
);
1647 pandecode_swizzle(attr_meta
->swizzle
, attr_meta
->format
);
1649 pandecode_log_cont(";\n");
1652 pandecode_log("\n");
1654 return count
? (max_index
+ 1) : 0;
1657 /* return bits [lo, hi) of word */
1659 bits(u32 word
, u32 lo
, u32 hi
)
1662 return word
; // avoid undefined behavior with the shift
1664 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1668 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool noninstanced
)
1670 pandecode_log_cont("{\n");
1673 /* Decode invocation_count. See the comment before the definition of
1674 * invocation_count for an explanation.
1677 unsigned size_x
= bits(p
->invocation_count
, 0, p
->size_y_shift
) + 1;
1678 unsigned size_y
= bits(p
->invocation_count
, p
->size_y_shift
, p
->size_z_shift
) + 1;
1679 unsigned size_z
= bits(p
->invocation_count
, p
->size_z_shift
, p
->workgroups_x_shift
) + 1;
1681 unsigned groups_x
= bits(p
->invocation_count
, p
->workgroups_x_shift
, p
->workgroups_y_shift
) + 1;
1682 unsigned groups_y
= bits(p
->invocation_count
, p
->workgroups_y_shift
, p
->workgroups_z_shift
) + 1;
1683 unsigned groups_z
= bits(p
->invocation_count
, p
->workgroups_z_shift
, 32) + 1;
1685 /* Even though we have this decoded, we want to ensure that the
1686 * representation is "unique" so we don't lose anything by printing only
1687 * the final result. More specifically, we need to check that we were
1688 * passed something in canonical form, since the definition per the
1689 * hardware is inherently not unique. How? Well, take the resulting
1690 * decode and pack it ourselves! If it is bit exact with what we
1691 * decoded, we're good to go. */
1693 struct mali_vertex_tiler_prefix ref
;
1694 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, noninstanced
);
1697 (p
->invocation_count
== ref
.invocation_count
) &&
1698 (p
->size_y_shift
== ref
.size_y_shift
) &&
1699 (p
->size_z_shift
== ref
.size_z_shift
) &&
1700 (p
->workgroups_x_shift
== ref
.workgroups_x_shift
) &&
1701 (p
->workgroups_y_shift
== ref
.workgroups_y_shift
) &&
1702 (p
->workgroups_z_shift
== ref
.workgroups_z_shift
) &&
1703 (p
->workgroups_x_shift_2
== ref
.workgroups_x_shift_2
);
1706 pandecode_msg("XXX: non-canonical workgroups packing\n");
1707 pandecode_msg("expected: %X, %d, %d, %d, %d, %d, %d\n",
1708 ref
.invocation_count
,
1711 ref
.workgroups_x_shift
,
1712 ref
.workgroups_y_shift
,
1713 ref
.workgroups_z_shift
,
1714 ref
.workgroups_x_shift_2
);
1716 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1717 pandecode_prop("size_y_shift = %d", p
->size_y_shift
);
1718 pandecode_prop("size_z_shift = %d", p
->size_z_shift
);
1719 pandecode_prop("workgroups_x_shift = %d", p
->workgroups_x_shift
);
1720 pandecode_prop("workgroups_y_shift = %d", p
->workgroups_y_shift
);
1721 pandecode_prop("workgroups_z_shift = %d", p
->workgroups_z_shift
);
1722 pandecode_prop("workgroups_x_shift_2 = %d", p
->workgroups_x_shift_2
);
1725 /* Regardless, print the decode */
1726 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1727 size_x
, size_y
, size_z
,
1728 groups_x
, groups_y
, groups_z
);
1731 if (p
->unknown_draw
)
1732 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1734 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1736 if (p
->draw_mode
!= MALI_DRAW_NONE
)
1737 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p
->draw_mode
));
1739 /* Index count only exists for tiler jobs anyway */
1742 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1745 unsigned index_raw_size
= (p
->unknown_draw
& MALI_DRAW_INDEXED_SIZE
);
1746 index_raw_size
>>= MALI_DRAW_INDEXED_SHIFT
;
1748 /* Validate an index buffer is present if we need one. TODO: verify
1749 * relationship between invocation_count and index_count */
1752 unsigned count
= p
->index_count
;
1755 unsigned size
= (index_raw_size
== 0x3) ? 4 : index_raw_size
;
1757 /* Ensure we got a size, and if so, validate the index buffer
1758 * is large enough to hold a full set of indices of the given
1761 if (!index_raw_size
)
1762 pandecode_msg("XXX: index size missing\n");
1764 pandecode_validate_buffer(p
->indices
, count
* size
);
1765 } else if (index_raw_size
)
1766 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size
);
1768 if (p
->offset_bias_correction
)
1769 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1771 /* TODO: Figure out what this is. It's not zero */
1772 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1775 pandecode_log("},\n");
1779 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1781 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1782 struct mali_uniform_buffer_meta
*PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1784 for (int i
= 0; i
< ubufs_count
; i
++) {
1785 unsigned size
= (ubufs
[i
].size
+ 1) * 16;
1786 mali_ptr addr
= ubufs
[i
].ptr
<< 2;
1788 pandecode_validate_buffer(addr
, size
);
1790 char *ptr
= pointer_as_memory_reference(ubufs
[i
].ptr
<< 2);
1791 pandecode_log("ubuf_%d[%u] = %s;\n", i
, size
, ptr
);
1795 pandecode_log("\n");
1799 pandecode_uniforms(mali_ptr uniforms
, unsigned uniform_count
)
1801 pandecode_validate_buffer(uniforms
, uniform_count
* 16);
1803 char *ptr
= pointer_as_memory_reference(uniforms
);
1804 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count
, ptr
);
1809 pandecode_scratchpad(uintptr_t pscratchpad
, int job_no
, char *suffix
)
1812 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(pscratchpad
);
1814 struct bifrost_scratchpad
*PANDECODE_PTR_VAR(scratchpad
, mem
, pscratchpad
);
1816 if (scratchpad
->zero
) {
1817 pandecode_msg("XXX: scratchpad zero tripped");
1818 pandecode_prop("zero = 0x%x\n", scratchpad
->zero
);
1821 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64
"_%d%s = {\n", pscratchpad
, job_no
, suffix
);
1824 pandecode_prop("flags = 0x%x", scratchpad
->flags
);
1825 MEMORY_PROP(scratchpad
, gpu_scratchpad
);
1828 pandecode_log("};\n");
1831 static unsigned shader_id
= 0;
1833 static struct midgard_disasm_stats
1834 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1835 bool is_bifrost
, unsigned gpu_id
)
1837 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1838 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1840 /* Compute maximum possible size */
1841 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1843 /* Print some boilerplate to clearly denote the assembly (which doesn't
1844 * obey indentation rules), and actually do the disassembly! */
1848 struct midgard_disasm_stats stats
;
1851 disassemble_bifrost(code
, sz
, false);
1853 /* TODO: Extend stats to Bifrost */
1854 stats
.texture_count
= -128;
1855 stats
.sampler_count
= -128;
1856 stats
.attribute_count
= -128;
1857 stats
.varying_count
= -128;
1858 stats
.uniform_count
= -128;
1859 stats
.uniform_buffer_count
= -128;
1860 stats
.work_count
= -128;
1862 stats
.instruction_count
= 0;
1863 stats
.bundle_count
= 0;
1864 stats
.quadword_count
= 0;
1865 stats
.helper_invocations
= false;
1867 stats
= disassemble_midgard(code
, sz
, gpu_id
,
1868 type
== JOB_TYPE_TILER
?
1869 MESA_SHADER_FRAGMENT
: MESA_SHADER_VERTEX
);
1872 /* Print shader-db stats */
1874 unsigned nr_threads
=
1875 (stats
.work_count
<= 4) ? 4 :
1876 (stats
.work_count
<= 8) ? 2 :
1879 printf("shader%d - %s shader: "
1880 "%u inst, %u bundles, %u quadwords, "
1881 "%u registers, %u threads, 0 loops\n\n\n",
1883 (type
== JOB_TYPE_TILER
) ? "FRAGMENT" : "VERTEX",
1884 stats
.instruction_count
, stats
.bundle_count
, stats
.quadword_count
,
1885 stats
.work_count
, nr_threads
);
1892 pandecode_texture(mali_ptr u
,
1893 struct pandecode_mapped_memory
*tmem
,
1894 unsigned job_no
, unsigned tex
)
1896 struct mali_texture_descriptor
*PANDECODE_PTR_VAR(t
, tmem
, u
);
1898 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64
"_%d_%d = {\n", u
, job_no
, tex
);
1901 struct mali_texture_format f
= t
->format
;
1903 /* See the definiton of enum mali_texture_type */
1905 bool is_cube
= f
.type
== MALI_TEX_CUBE
;
1906 unsigned dimension
= is_cube
? 2 : f
.type
;
1908 pandecode_make_indent();
1910 /* TODO: Are there others? */
1911 bool is_zs
= f
.format
== MALI_Z32_UNORM
;
1913 /* Recall Z/S switched the meaning of linear/tiled .. */
1914 if (is_zs
&& f
.layout
== MALI_TEXTURE_LINEAR
)
1915 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1917 /* Print the layout. Default is linear; a modifier can denote AFBC or
1918 * u-interleaved/tiled modes */
1920 if (f
.layout
== MALI_TEXTURE_AFBC
)
1921 pandecode_log_cont("afbc");
1922 else if (f
.layout
== MALI_TEXTURE_TILED
)
1923 pandecode_log_cont(is_zs
? "linear" : "tiled");
1924 else if (f
.layout
== MALI_TEXTURE_LINEAR
)
1925 pandecode_log_cont("linear");
1927 pandecode_msg("XXX: invalid texture layout 0x%X\n", f
.layout
);
1929 pandecode_swizzle(t
->swizzle
, f
.format
);
1930 pandecode_log_cont(" ");
1932 /* Distinguish cube/2D with modifier */
1935 pandecode_log_cont("cube ");
1937 pandecode_format_short(f
.format
, f
.srgb
);
1938 pandecode_swizzle(f
.swizzle
, f
.format
);
1940 /* All four width/height/depth/array_size dimensions are present
1941 * regardless of the type of texture, but it is an error to have
1942 * non-zero dimensions for unused dimensions. Verify this. array_size
1943 * can always be set, as can width. */
1945 if (t
->height
&& dimension
< 2)
1946 pandecode_msg("XXX: nonzero height for <2D texture\n");
1948 if (t
->depth
&& dimension
< 3)
1949 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1951 /* Print only the dimensions that are actually there */
1953 pandecode_log_cont(": %d", t
->width
+ 1);
1956 pandecode_log_cont("x%u", t
->height
+ 1);
1959 pandecode_log_cont("x%u", t
->depth
+ 1);
1962 pandecode_log_cont("[%u]", t
->array_size
+ 1);
1965 pandecode_log_cont(" mip %u", t
->levels
);
1967 pandecode_log_cont("\n");
1969 if (f
.unknown1
| f
.zero
) {
1970 pandecode_msg("XXX: texture format zero tripped\n");
1971 pandecode_prop("unknown1 = %" PRId32
, f
.unknown1
);
1972 pandecode_prop("zero = %" PRId32
, f
.zero
);
1976 pandecode_msg("XXX: expected unknown texture bit set\n");
1977 pandecode_prop("unknown2 = %" PRId32
, f
.unknown1
);
1980 if (t
->swizzle_zero
) {
1981 pandecode_msg("XXX: swizzle zero tripped\n");
1982 pandecode_prop("swizzle_zero = %d", t
->swizzle_zero
);
1985 if (t
->unknown3
| t
->unknown3A
| t
->unknown5
| t
->unknown6
| t
->unknown7
) {
1986 pandecode_msg("XXX: texture zero tripped\n");
1987 pandecode_prop("unknown3 = %" PRId16
, t
->unknown3
);
1988 pandecode_prop("unknown3A = %" PRId8
, t
->unknown3A
);
1989 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
1990 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
1991 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
1994 pandecode_log(".payload = {\n");
1997 /* A bunch of bitmap pointers follow.
1998 * We work out the correct number,
1999 * based on the mipmap/cubemap
2000 * properties, but dump extra
2001 * possibilities to futureproof */
2003 int bitmap_count
= MALI_NEGATIVE(t
->levels
);
2005 /* Miptree for each face */
2006 if (f
.type
== MALI_TEX_CUBE
)
2009 /* Array of textures */
2010 bitmap_count
*= MALI_NEGATIVE(t
->array_size
);
2012 /* Stride for each element */
2013 if (f
.manual_stride
)
2016 /* Sanity check the size */
2017 int max_count
= sizeof(t
->payload
) / sizeof(t
->payload
[0]);
2018 assert (bitmap_count
<= max_count
);
2020 for (int i
= 0; i
< bitmap_count
; ++i
) {
2021 /* How we dump depends if this is a stride or a pointer */
2023 if (f
.manual_stride
&& (i
& 1)) {
2024 /* signed 32-bit snuck in as a 64-bit pointer */
2025 uint64_t stride_set
= t
->payload
[i
];
2026 uint32_t clamped_stride
= stride_set
;
2027 int32_t stride
= clamped_stride
;
2028 assert(stride_set
== clamped_stride
);
2029 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
2031 char *a
= pointer_as_memory_reference(t
->payload
[i
]);
2032 pandecode_log("%s, \n", a
);
2038 pandecode_log("},\n");
2041 pandecode_log("};\n");
2044 /* For shader properties like texture_count, we have a claimed property in the shader_meta, and the actual Truth from static analysis (this may just be an upper limit). We validate accordingly */
2047 pandecode_shader_prop(const char *name
, unsigned claim
, signed truth
, bool fuzzy
)
2056 if ((truth
>= 0) && !fuzzy
) {
2057 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2058 (truth
< claim
) ? "warn" : "XXX",
2059 name
, truth
, claim
);
2060 } else if ((claim
> -truth
) && !fuzzy
) {
2061 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2062 name
, -truth
, claim
);
2063 } else if (fuzzy
&& (claim
< truth
))
2064 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2065 name
, truth
, claim
);
2067 pandecode_log(".%s = %" PRId16
, name
, claim
);
2070 pandecode_log_cont(" /* %u used */", truth
);
2072 pandecode_log_cont(",\n");
2076 pandecode_blend_shader_disassemble(mali_ptr shader
, int job_no
, int job_type
,
2077 bool is_bifrost
, unsigned gpu_id
)
2079 struct midgard_disasm_stats stats
=
2080 pandecode_shader_disassemble(shader
, job_no
, job_type
, is_bifrost
, gpu_id
);
2082 bool has_texture
= (stats
.texture_count
> 0);
2083 bool has_sampler
= (stats
.sampler_count
> 0);
2084 bool has_attribute
= (stats
.attribute_count
> 0);
2085 bool has_varying
= (stats
.varying_count
> 0);
2086 bool has_uniform
= (stats
.uniform_count
> 0);
2087 bool has_ubo
= (stats
.uniform_buffer_count
> 0);
2089 if (has_texture
|| has_sampler
)
2090 pandecode_msg("XXX: blend shader accessing textures\n");
2092 if (has_attribute
|| has_varying
)
2093 pandecode_msg("XXX: blend shader accessing interstage\n");
2095 if (has_uniform
|| has_ubo
)
2096 pandecode_msg("XXX: blend shader accessing uniforms\n");
2100 pandecode_vertex_tiler_postfix_pre(
2101 const struct mali_vertex_tiler_postfix
*p
,
2102 int job_no
, enum mali_job_type job_type
,
2103 char *suffix
, bool is_bifrost
, unsigned gpu_id
)
2105 struct pandecode_mapped_memory
*attr_mem
;
2107 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2108 * are the only things actually needed from the FBD, vertex/tiler jobs
2109 * no longer reference the FBD -- instead, this field points to some
2110 * info about the scratchpad.
2113 struct pandecode_fbd fbd_info
= {
2114 /* Default for Bifrost */
2119 pandecode_scratchpad(p
->framebuffer
& ~FBD_TYPE
, job_no
, suffix
);
2120 else if (p
->framebuffer
& MALI_MFBD
)
2121 fbd_info
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->framebuffer
) & FBD_MASK
, job_no
, false);
2122 else if (job_type
== JOB_TYPE_COMPUTE
)
2123 pandecode_compute_fbd((u64
) (uintptr_t) p
->framebuffer
, job_no
);
2125 fbd_info
= pandecode_sfbd((u64
) (uintptr_t) p
->framebuffer
, job_no
, false, gpu_id
);
2127 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
2128 int texture_count
= 0, sampler_count
= 0;
2131 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->shader
);
2132 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, p
->shader
);
2134 /* Disassemble ahead-of-time to get stats. Initialize with
2135 * stats for the missing-shader case so we get validation
2138 struct midgard_disasm_stats info
= {
2141 .attribute_count
= 0,
2145 .uniform_count
= -128,
2146 .uniform_buffer_count
= 0
2149 if (s
->shader
& ~0xF)
2150 info
= pandecode_shader_disassemble(s
->shader
& ~0xF, job_no
, job_type
, is_bifrost
, gpu_id
);
2152 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", p
->shader
, job_no
, suffix
);
2155 /* Save for dumps */
2156 attribute_count
= s
->attribute_count
;
2157 varying_count
= s
->varying_count
;
2158 texture_count
= s
->texture_count
;
2159 sampler_count
= s
->sampler_count
;
2162 uniform_count
= s
->bifrost2
.uniform_count
;
2163 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
2165 uniform_count
= s
->midgard1
.uniform_count
;
2166 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
2169 pandecode_shader_address("shader", s
->shader
);
2171 pandecode_shader_prop("texture_count", s
->texture_count
, info
.texture_count
, false);
2172 pandecode_shader_prop("sampler_count", s
->sampler_count
, info
.sampler_count
, false);
2173 pandecode_shader_prop("attribute_count", s
->attribute_count
, info
.attribute_count
, false);
2174 pandecode_shader_prop("varying_count", s
->varying_count
, info
.varying_count
, false);
2175 pandecode_shader_prop("uniform_buffer_count",
2176 uniform_buffer_count
,
2177 info
.uniform_buffer_count
, true);
2180 pandecode_shader_prop("uniform_count",
2182 info
.uniform_count
, false);
2184 pandecode_shader_prop("work_count",
2185 s
->midgard1
.work_count
, info
.work_count
, false);
2189 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32
, s
->bifrost1
.unk1
);
2191 bool helpers
= s
->midgard1
.flags
& MALI_HELPER_INVOCATIONS
;
2192 s
->midgard1
.flags
&= ~MALI_HELPER_INVOCATIONS
;
2194 if (helpers
!= info
.helper_invocations
) {
2195 pandecode_msg("XXX: expected helpers %u but got %u\n",
2196 info
.helper_invocations
, helpers
);
2199 pandecode_log(".midgard1.flags = ");
2200 pandecode_log_decoded_flags(shader_midgard1_flag_info
, s
->midgard1
.flags
);
2201 pandecode_log_cont(",\n");
2203 pandecode_prop("midgard1.unknown2 = 0x%" PRIx32
, s
->midgard1
.unknown2
);
2206 if (s
->depth_units
|| s
->depth_factor
) {
2207 pandecode_prop("depth_factor = %f", s
->depth_factor
);
2208 pandecode_prop("depth_units = %f", s
->depth_units
);
2211 if (s
->alpha_coverage
) {
2212 bool invert_alpha_coverage
= s
->alpha_coverage
& 0xFFF0;
2213 uint16_t inverted_coverage
= invert_alpha_coverage
? ~s
->alpha_coverage
: s
->alpha_coverage
;
2215 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2216 invert_alpha_coverage
? "~" : "",
2217 MALI_GET_ALPHA_COVERAGE(inverted_coverage
));
2220 if (s
->unknown2_3
|| s
->unknown2_4
) {
2221 pandecode_log(".unknown2_3 = ");
2223 int unknown2_3
= s
->unknown2_3
;
2224 int unknown2_4
= s
->unknown2_4
;
2226 /* We're not quite sure what these flags mean without the depth test, if anything */
2228 if (unknown2_3
& (MALI_DEPTH_WRITEMASK
| MALI_DEPTH_FUNC_MASK
)) {
2229 const char *func
= pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3
));
2230 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
2232 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
2235 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
2236 pandecode_log_cont(",\n");
2238 pandecode_log(".unknown2_4 = ");
2239 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
2240 pandecode_log_cont(",\n");
2243 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
2244 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
2245 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
2248 pandecode_stencil("front", &s
->stencil_front
);
2249 pandecode_stencil("back", &s
->stencil_back
);
2252 pandecode_log(".bifrost2 = {\n");
2255 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
2256 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
2257 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
2258 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
2261 pandecode_log("},\n");
2262 } else if (s
->midgard2
.unknown2_7
) {
2263 pandecode_log(".midgard2 = {\n");
2266 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
2268 pandecode_log("},\n");
2272 pandecode_prop("unknown2_8 = 0x%" PRIx32
, s
->unknown2_8
);
2275 /* TODO: Blend shaders routing/disasm */
2276 union midgard_blend blend
= s
->blend
;
2277 mali_ptr shader
= pandecode_midgard_blend(&blend
, s
->unknown2_3
& MALI_HAS_BLEND_SHADER
);
2279 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
2283 pandecode_log("};\n");
2285 /* MRT blend fields are used whenever MFBD is used, with
2286 * per-RT descriptors */
2288 if (job_type
== JOB_TYPE_TILER
&& p
->framebuffer
& MALI_MFBD
) {
2289 void* blend_base
= (void *) (s
+ 1);
2291 for (unsigned i
= 0; i
< fbd_info
.rt_count
; i
++) {
2292 mali_ptr shader
= 0;
2295 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
2297 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
2300 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
2305 pandecode_msg("XXX: missing shader descriptor\n");
2308 struct pandecode_mapped_memory
*fmem
= pandecode_find_mapped_gpu_mem_containing(p
->viewport
);
2309 struct mali_viewport
*PANDECODE_PTR_VAR(f
, fmem
, p
->viewport
);
2311 pandecode_log("struct mali_viewport viewport_%"PRIx64
"_%d%s = {\n", p
->viewport
, job_no
, suffix
);
2314 pandecode_prop("clip_minx = %f", f
->clip_minx
);
2315 pandecode_prop("clip_miny = %f", f
->clip_miny
);
2316 pandecode_prop("clip_minz = %f", f
->clip_minz
);
2317 pandecode_prop("clip_maxx = %f", f
->clip_maxx
);
2318 pandecode_prop("clip_maxy = %f", f
->clip_maxy
);
2319 pandecode_prop("clip_maxz = %f", f
->clip_maxz
);
2321 /* Only the higher coordinates are MALI_POSITIVE scaled */
2323 pandecode_prop("viewport0 = { %d, %d }",
2324 f
->viewport0
[0], f
->viewport0
[1]);
2326 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2327 f
->viewport1
[0] + 1, f
->viewport1
[1] + 1);
2330 pandecode_log("};\n");
2333 unsigned max_attr_index
= 0;
2335 if (p
->attribute_meta
)
2336 max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
2338 if (p
->attributes
) {
2339 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
2340 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false, job_type
);
2343 /* Varyings are encoded like attributes but not actually sent; we just
2344 * pass a zero buffer with the right stride/size set, (or whatever)
2345 * since the GPU will write to it itself */
2347 if (p
->varying_meta
) {
2348 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
2352 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
2354 /* Number of descriptors depends on whether there are
2355 * non-internal varyings */
2357 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true, job_type
);
2360 if (p
->uniform_buffers
) {
2361 if (uniform_buffer_count
)
2362 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
2364 pandecode_msg("warn: UBOs specified but not referenced\n");
2365 } else if (uniform_buffer_count
)
2366 pandecode_msg("XXX: UBOs referenced but not specified\n");
2368 /* We don't want to actually dump uniforms, but we do need to validate
2369 * that the counts we were given are sane */
2373 pandecode_uniforms(p
->uniforms
, uniform_count
);
2375 pandecode_msg("warn: Uniforms specified but not referenced\n");
2376 } else if (uniform_count
)
2377 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2379 if (p
->texture_trampoline
) {
2380 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(p
->texture_trampoline
);
2383 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
);
2385 pandecode_log("uint64_t texture_trampoline_%"PRIx64
"_%d[] = {\n", p
->texture_trampoline
, job_no
);
2388 for (int tex
= 0; tex
< texture_count
; ++tex
) {
2389 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
2390 char *a
= pointer_as_memory_reference(*u
);
2391 pandecode_log("%s,\n", a
);
2396 pandecode_log("};\n");
2398 /* Now, finally, descend down into the texture descriptor */
2399 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
2400 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
2401 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
2403 pandecode_texture(*u
, tmem
, job_no
, tex
);
2408 if (p
->sampler_descriptor
) {
2409 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->sampler_descriptor
);
2412 struct mali_sampler_descriptor
*s
;
2414 mali_ptr d
= p
->sampler_descriptor
;
2416 for (int i
= 0; i
< sampler_count
; ++i
) {
2417 s
= pandecode_fetch_gpu_mem(smem
, d
+ sizeof(*s
) * i
, sizeof(*s
));
2419 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", d
+ sizeof(*s
) * i
, job_no
, i
);
2422 pandecode_log(".filter_mode = ");
2423 pandecode_log_decoded_flags(sampler_flag_info
, s
->filter_mode
);
2424 pandecode_log_cont(",\n");
2426 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
2427 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
2430 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s
->lod_bias
));
2432 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s
->wrap_s
));
2433 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s
->wrap_t
));
2434 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s
->wrap_r
));
2436 pandecode_prop("compare_func = %s", pandecode_alt_func(s
->compare_func
));
2438 if (s
->zero
|| s
->zero2
) {
2439 pandecode_msg("XXX: sampler zero tripped\n");
2440 pandecode_prop("zero = 0x%X, 0x%X\n", s
->zero
, s
->zero2
);
2443 pandecode_prop("seamless_cube_map = %d", s
->seamless_cube_map
);
2445 pandecode_prop("border_color = { %f, %f, %f, %f }",
2449 s
->border_color
[3]);
2452 pandecode_log("};\n");
2459 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2461 if (p
->shader
& 0xF)
2462 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p
->shader
& 0xF));
2464 if (!(p
->position_varying
|| p
->occlusion_counter
))
2467 pandecode_log(".postfix = {\n");
2470 MEMORY_PROP(p
, position_varying
);
2471 MEMORY_PROP(p
, occlusion_counter
);
2474 pandecode_log("},\n");
2478 pandecode_vertex_only_bfr(struct bifrost_vertex_only
*v
)
2480 pandecode_log_cont("{\n");
2483 pandecode_prop("unk2 = 0x%x", v
->unk2
);
2485 if (v
->zero0
|| v
->zero1
) {
2486 pandecode_msg("XXX: vertex only zero tripped");
2487 pandecode_prop("zero0 = 0x%" PRIx32
, v
->zero0
);
2488 pandecode_prop("zero1 = 0x%" PRIx64
, v
->zero1
);
2492 pandecode_log("}\n");
2496 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2499 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2500 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2502 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no
);
2506 pandecode_msg("XXX: tiler heap zero tripped\n");
2507 pandecode_prop("zero = 0x%x", h
->zero
);
2510 for (int i
= 0; i
< 12; i
++) {
2511 if (h
->zeros
[i
] != 0) {
2512 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2517 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2518 MEMORY_PROP(h
, tiler_heap_start
);
2519 MEMORY_PROP(h
, tiler_heap_free
);
2521 /* this might point to the beginning of another buffer, when it's
2522 * really the end of the tiler heap buffer, so we have to be careful
2523 * here. but for zero length, we need the same pointer.
2526 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2527 MEMORY_PROP(h
, tiler_heap_start
);
2529 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2530 pandecode_prop("tiler_heap_end = %s + 1", a
);
2535 pandecode_log("};\n");
2539 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2541 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2542 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2544 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2546 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no
);
2549 if (t
->zero0
|| t
->zero1
) {
2550 pandecode_msg("XXX: tiler meta zero tripped\n");
2551 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2552 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2555 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2556 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2558 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2559 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2561 for (int i
= 0; i
< 12; i
++) {
2562 if (t
->zeros
[i
] != 0) {
2563 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2569 pandecode_log("};\n");
2573 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2575 pandecode_log(".gl_enables = ");
2577 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2579 pandecode_log_cont(",\n");
2583 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2585 if (u
.pointer
== 0x0)
2588 pandecode_log(".primitive_size = {\n");
2592 pandecode_prop("constant = %f", u
.constant
);
2594 MEMORY_PROP((&u
), pointer
);
2598 pandecode_log("},\n");
2602 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2604 pandecode_log_cont("{\n");
2607 /* TODO: gl_PointSize on Bifrost */
2608 pandecode_primitive_size(t
->primitive_size
, true);
2610 pandecode_gl_enables(t
->gl_enables
, JOB_TYPE_TILER
);
2612 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2613 || t
->zero6
|| t
->zero7
|| t
->zero8
) {
2614 pandecode_msg("XXX: tiler only zero tripped\n");
2615 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2616 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2617 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2618 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2619 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2620 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2621 pandecode_prop("zero7 = 0x%" PRIx32
, t
->zero7
);
2622 pandecode_prop("zero8 = 0x%" PRIx64
, t
->zero8
);
2626 pandecode_log("},\n");
2630 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2631 const struct pandecode_mapped_memory
*mem
,
2632 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2634 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2636 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2638 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no
);
2641 pandecode_log(".prefix = ");
2642 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2644 pandecode_log(".vertex = ");
2645 pandecode_vertex_only_bfr(&v
->vertex
);
2647 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2650 pandecode_log("};\n");
2656 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2657 const struct pandecode_mapped_memory
*mem
,
2658 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2660 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2662 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2663 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2665 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no
);
2668 pandecode_log(".prefix = ");
2669 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2671 pandecode_log(".tiler = ");
2672 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2674 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2677 pandecode_log("};\n");
2683 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2684 const struct pandecode_mapped_memory
*mem
,
2685 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2687 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2689 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false, gpu_id
);
2691 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2694 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2695 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2697 bool instanced
= v
->instance_shift
|| v
->instance_odd
;
2698 bool is_graphics
= (h
->job_type
== JOB_TYPE_VERTEX
) || (h
->job_type
== JOB_TYPE_TILER
);
2700 pandecode_log(".prefix = ");
2701 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, !instanced
&& is_graphics
);
2703 pandecode_gl_enables(v
->gl_enables
, h
->job_type
);
2705 if (v
->instance_shift
|| v
->instance_odd
) {
2706 pandecode_prop("instance_shift = 0x%d /* %d */",
2707 v
->instance_shift
, 1 << v
->instance_shift
);
2708 pandecode_prop("instance_odd = 0x%X /* %d */",
2709 v
->instance_odd
, (2 * v
->instance_odd
) + 1);
2711 pandecode_padded_vertices(v
->instance_shift
, v
->instance_odd
);
2714 if (v
->offset_start
)
2715 pandecode_prop("offset_start = %d", v
->offset_start
);
2718 pandecode_msg("XXX: midgard payload zero tripped\n");
2719 pandecode_prop("zero5 = 0x%" PRIx64
, v
->zero5
);
2722 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2725 pandecode_log("};\n");
2731 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2732 mali_ptr payload
, int job_no
,
2733 bool is_bifrost
, unsigned gpu_id
)
2735 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2737 bool is_mfbd
= (s
->framebuffer
& FBD_TYPE
) == MALI_MFBD
;
2739 /* Bifrost theoretically may retain support for SFBD on compute jobs,
2740 * but for graphics workloads with a FRAGMENT payload, use MFBD */
2742 if (!is_mfbd
&& is_bifrost
)
2743 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2745 struct pandecode_fbd info
;
2748 info
= pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true);
2750 info
= pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true, gpu_id
);
2752 /* Compute the tag for the tagged pointer. This contains the type of
2753 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2754 * additional structures follow the MFBD header (an extra payload or
2755 * not, as well as a count of render targets) */
2757 unsigned expected_tag
= is_mfbd
? MALI_MFBD
: MALI_SFBD
;
2761 expected_tag
|= MALI_MFBD_TAG_EXTRA
;
2763 expected_tag
|= (MALI_POSITIVE(info
.rt_count
) << 2);
2766 if ((s
->min_tile_coord
| s
->max_tile_coord
) & ~(MALI_X_COORD_MASK
| MALI_Y_COORD_MASK
)) {
2767 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2768 pandecode_prop("min_tile_coord = 0x%X\n", s
->min_tile_coord
);
2769 pandecode_prop("max_tile_coord = 0x%X\n", s
->min_tile_coord
);
2772 /* Extract tile coordinates */
2774 unsigned min_x
= MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2775 unsigned min_y
= MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2777 unsigned max_x
= (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2778 unsigned max_y
= (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2780 /* For the max, we also want the floored (rather than ceiled) version for checking */
2782 unsigned max_x_f
= (MALI_TILE_COORD_X(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2783 unsigned max_y_f
= (MALI_TILE_COORD_Y(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2785 /* Validate the coordinates are well-ordered */
2788 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2789 else if (min_x
> max_x
)
2790 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2793 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2794 else if (min_y
> max_y
)
2795 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2797 /* Validate the coordinates fit inside the framebuffer. We use floor,
2798 * rather than ceil, for the max coordinates, since the tile
2799 * coordinates for something like an 800x600 framebuffer will actually
2800 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2802 if ((min_x
> info
.width
) || (max_x_f
> info
.width
))
2803 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2805 if ((min_y
> info
.height
) || (max_y_f
> info
.height
))
2806 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2808 /* After validation, we print */
2810 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x
, min_y
, max_x
, max_y
);
2812 /* The FBD is a tagged pointer */
2814 unsigned tag
= (s
->framebuffer
& ~FBD_MASK
);
2816 if (tag
!= expected_tag
)
2817 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag
, tag
);
2822 static int job_descriptor_number
= 0;
2825 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
, unsigned gpu_id
)
2827 struct mali_job_descriptor_header
*h
;
2829 int start_number
= 0;
2835 struct pandecode_mapped_memory
*mem
=
2836 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2840 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2842 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2843 * high 32-bits of the 64-bit pointer are reused to store
2846 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
2847 h
->job_type
!= JOB_TYPE_FRAGMENT
? 4 : 0;
2848 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
2850 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
, 256);
2852 int job_no
= job_descriptor_number
++;
2855 start_number
= job_no
;
2857 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2860 pandecode_prop("job_type = %s", pandecode_job_type(h
->job_type
));
2862 /* Save for next job fixing */
2863 last_size
= h
->job_descriptor_size
;
2865 if (h
->job_descriptor_size
)
2866 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2868 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2869 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2870 h
->exception_status
,
2871 (h
->exception_status
>> 16) & 0xFFFF,
2872 pandecode_exception_access((h
->exception_status
>> 8) & 0x3),
2873 h
->exception_status
& 0xFF);
2875 if (h
->first_incomplete_task
)
2876 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2878 if (h
->fault_pointer
)
2879 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2882 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2884 pandecode_prop("job_index = %d", h
->job_index
);
2886 if (h
->unknown_flags
)
2887 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2889 if (h
->job_dependency_index_1
)
2890 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2892 if (h
->job_dependency_index_2
)
2893 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2896 pandecode_log("};\n");
2898 /* Do not touch the field yet -- decode the payload first, and
2899 * don't touch that either. This is essential for the uploads
2900 * to occur in sequence and therefore be dynamically allocated
2901 * correctly. Do note the size, however, for that related
2904 switch (h
->job_type
) {
2905 case JOB_TYPE_WRITE_VALUE
: {
2906 struct mali_payload_write_value
*s
= payload
;
2907 pandecode_log("struct mali_payload_write_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2909 MEMORY_PROP(s
, address
);
2911 if (s
->value_descriptor
!= MALI_WRITE_VALUE_ZERO
) {
2912 pandecode_msg("XXX: unknown value descriptor\n");
2913 pandecode_prop("value_descriptor = 0x%" PRIX32
, s
->value_descriptor
);
2917 pandecode_msg("XXX: set value tripped\n");
2918 pandecode_prop("reserved = 0x%" PRIX32
, s
->reserved
);
2921 pandecode_prop("immediate = 0x%" PRIX64
, s
->immediate
);
2923 pandecode_log("};\n");
2928 case JOB_TYPE_TILER
:
2929 case JOB_TYPE_VERTEX
:
2930 case JOB_TYPE_COMPUTE
:
2932 if (h
->job_type
== JOB_TYPE_TILER
)
2933 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2935 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2937 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2941 case JOB_TYPE_FRAGMENT
:
2942 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
, gpu_id
);
2949 /* Handle linkage */
2952 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no
- 1);
2955 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no
);
2957 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no
);
2962 } while ((jc_gpu_va
= h
->job_descriptor_size
? h
->next_job_64
: h
->next_job_32
));
2964 return start_number
;