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>
33 #include "util/macros.h"
34 #include "util/u_math.h"
36 #include "pan_pretty_print.h"
37 #include "midgard/disassemble.h"
38 #include "bifrost/disassemble.h"
40 #include "pan_encoder.h"
42 int pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
);
44 #define MEMORY_PROP(obj, p) {\
46 char *a = pointer_as_memory_reference(obj->p); \
47 pandecode_prop("%s = %s", #p, a); \
52 #define MEMORY_PROP_DIR(obj, p) {\
54 char *a = pointer_as_memory_reference(obj.p); \
55 pandecode_prop("%s = %s", #p, a); \
60 /* Semantic logging type.
62 * Raw: for raw messages to be printed as is.
63 * Message: for helpful information to be commented out in replays.
64 * Property: for properties of a struct
66 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
69 enum pandecode_log_type
{
75 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
76 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
77 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
79 unsigned pandecode_indent
= 0;
82 pandecode_make_indent(void)
84 for (unsigned i
= 0; i
< pandecode_indent
; ++i
)
89 pandecode_log_typed(enum pandecode_log_type type
, const char *format
, ...)
93 pandecode_make_indent();
95 if (type
== PANDECODE_MESSAGE
)
97 else if (type
== PANDECODE_PROPERTY
)
100 va_start(ap
, format
);
104 if (type
== PANDECODE_PROPERTY
)
109 pandecode_log_cont(const char *format
, ...)
113 va_start(ap
, format
);
118 struct pandecode_flag_info
{
124 pandecode_log_decoded_flags(const struct pandecode_flag_info
*flag_info
,
127 bool decodable_flags_found
= false;
129 for (int i
= 0; flag_info
[i
].name
; i
++) {
130 if ((flags
& flag_info
[i
].flag
) != flag_info
[i
].flag
)
133 if (!decodable_flags_found
) {
134 decodable_flags_found
= true;
136 pandecode_log_cont(" | ");
139 pandecode_log_cont("%s", flag_info
[i
].name
);
141 flags
&= ~flag_info
[i
].flag
;
144 if (decodable_flags_found
) {
146 pandecode_log_cont(" | 0x%" PRIx64
, flags
);
148 pandecode_log_cont("0x%" PRIx64
, flags
);
152 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
153 static const struct pandecode_flag_info gl_enable_flag_info
[] = {
154 FLAG_INFO(OCCLUSION_QUERY
),
155 FLAG_INFO(OCCLUSION_PRECISE
),
156 FLAG_INFO(FRONT_CCW_TOP
),
157 FLAG_INFO(CULL_FACE_FRONT
),
158 FLAG_INFO(CULL_FACE_BACK
),
163 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
164 static const struct pandecode_flag_info clear_flag_info
[] = {
167 FLAG_INFO(SLOW_STENCIL
),
172 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
173 static const struct pandecode_flag_info mask_flag_info
[] = {
182 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
183 static const struct pandecode_flag_info u3_flag_info
[] = {
185 FLAG_INFO(CAN_DISCARD
),
186 FLAG_INFO(HAS_BLEND_SHADER
),
187 FLAG_INFO(DEPTH_TEST
),
191 static const struct pandecode_flag_info u4_flag_info
[] = {
193 FLAG_INFO(NO_DITHER
),
194 FLAG_INFO(DEPTH_RANGE_A
),
195 FLAG_INFO(DEPTH_RANGE_B
),
196 FLAG_INFO(STENCIL_TEST
),
197 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER
),
202 #define FLAG_INFO(flag) { MALI_FRAMEBUFFER_##flag, "MALI_FRAMEBUFFER_" #flag }
203 static const struct pandecode_flag_info fb_fmt_flag_info
[] = {
211 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
212 static const struct pandecode_flag_info mfbd_fmt_flag_info
[] = {
219 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
220 static const struct pandecode_flag_info mfbd_extra_flag_info
[] = {
228 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
229 static const struct pandecode_flag_info shader_midgard1_flag_info
[] = {
231 FLAG_INFO(HELPER_INVOCATIONS
),
232 FLAG_INFO(READS_TILEBUFFER
),
238 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
239 static const struct pandecode_flag_info mfbd_flag_info
[] = {
240 FLAG_INFO(DEPTH_WRITE
),
246 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
247 static const struct pandecode_flag_info sampler_flag_info
[] = {
248 FLAG_INFO(MAG_NEAREST
),
249 FLAG_INFO(MIN_NEAREST
),
250 FLAG_INFO(MIP_LINEAR_1
),
251 FLAG_INFO(MIP_LINEAR_2
),
252 FLAG_INFO(NORM_COORDS
),
257 extern char *replace_fragment
;
258 extern char *replace_vertex
;
261 pandecode_job_type(enum mali_job_type type
)
263 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
267 DEFINE_CASE(SET_VALUE
);
268 DEFINE_CASE(CACHE_FLUSH
);
269 DEFINE_CASE(COMPUTE
);
273 DEFINE_CASE(FRAGMENT
);
275 case JOB_NOT_STARTED
:
276 return "NOT_STARTED";
279 pandecode_log("Warning! Unknown job type %x\n", type
);
287 pandecode_draw_mode(enum mali_draw_mode mode
)
289 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
292 DEFINE_CASE(DRAW_NONE
);
295 DEFINE_CASE(TRIANGLES
);
296 DEFINE_CASE(TRIANGLE_STRIP
);
297 DEFINE_CASE(TRIANGLE_FAN
);
298 DEFINE_CASE(LINE_STRIP
);
299 DEFINE_CASE(LINE_LOOP
);
300 DEFINE_CASE(POLYGON
);
302 DEFINE_CASE(QUAD_STRIP
);
305 return "MALI_TRIANGLES /* XXX: Unknown GL mode, check dump */";
311 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
313 pandecode_func(enum mali_func mode
)
320 DEFINE_CASE(GREATER
);
321 DEFINE_CASE(NOTEQUAL
);
326 return "MALI_FUNC_NEVER /* XXX: Unknown function, check dump */";
331 /* Why is this duplicated? Who knows... */
332 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
334 pandecode_alt_func(enum mali_alt_func mode
)
341 DEFINE_CASE(GREATER
);
342 DEFINE_CASE(NOTEQUAL
);
347 return "MALI_FUNC_NEVER /* XXX: Unknown function, check dump */";
352 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
354 pandecode_stencil_op(enum mali_stencil_op op
)
358 DEFINE_CASE(REPLACE
);
361 DEFINE_CASE(INCR_WRAP
);
362 DEFINE_CASE(DECR_WRAP
);
367 return "MALI_STENCIL_KEEP /* XXX: Unknown stencil op, check dump */";
373 #define DEFINE_CASE(name) case MALI_ATTR_ ## name: return "MALI_ATTR_" #name
374 static char *pandecode_attr_mode(enum mali_attr_mode mode
)
379 DEFINE_CASE(POT_DIVIDE
);
381 DEFINE_CASE(NPOT_DIVIDE
);
383 DEFINE_CASE(INTERNAL
);
385 return "MALI_ATTR_UNUSED /* XXX: Unknown stencil op, check dump */";
391 #define DEFINE_CASE(name) case MALI_CHANNEL_## name: return "MALI_CHANNEL_" #name
393 pandecode_channel(enum mali_channel channel
)
402 DEFINE_CASE(RESERVED_0
);
403 DEFINE_CASE(RESERVED_1
);
406 return "MALI_CHANNEL_ZERO /* XXX: Unknown channel, check dump */";
411 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
413 pandecode_wrap_mode(enum mali_wrap_mode op
)
417 DEFINE_CASE(CLAMP_TO_EDGE
);
418 DEFINE_CASE(CLAMP_TO_BORDER
);
419 DEFINE_CASE(MIRRORED_REPEAT
);
422 return "MALI_WRAP_REPEAT /* XXX: Unknown wrap mode, check dump */";
427 #define DEFINE_CASE(name) case MALI_TEX_## name: return "MALI_TEX_" #name
429 pandecode_texture_type(enum mali_texture_type type
)
438 unreachable("Unknown case");
443 #define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
445 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt
)
449 DEFINE_CASE(UNKNOWN
);
454 unreachable("Invalid case");
459 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
461 pandecode_exception_access(enum mali_exception_access fmt
)
465 DEFINE_CASE(EXECUTE
);
470 unreachable("Invalid case");
475 /* Midgard's tiler descriptor is embedded within the
479 pandecode_midgard_tiler_descriptor(
480 const struct midgard_tiler_descriptor
*t
,
485 pandecode_log(".tiler = {\n");
488 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
489 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
491 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
493 /* We know this name from the kernel, but we never see it nonzero */
495 pandecode_prop("flags = 0x%" PRIx16
" /* XXX: unexpected */", t
->flags
);
497 MEMORY_PROP(t
, polygon_list
);
499 /* The body is offset from the base of the polygon list */
500 assert(t
->polygon_list_body
> t
->polygon_list
);
501 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
503 /* It needs to fit inside the reported size */
504 assert(t
->polygon_list_size
>= body_offset
);
506 /* Check that we fit */
507 struct pandecode_mapped_memory
*plist
=
508 pandecode_find_mapped_gpu_mem_containing(t
->polygon_list
);
510 assert(t
->polygon_list_size
<= plist
->length
);
512 /* Now that we've sanity checked, we'll try to calculate the sizes
513 * ourselves for comparison */
515 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
);
516 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
);
518 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
519 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
520 ref_header
, ref_size
);
521 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
522 pandecode_msg("body offset %d\n", body_offset
);
525 /* The tiler heap has a start and end specified -- it should be
526 * identical to what we have in the BO. The exception is if tiling is
529 MEMORY_PROP(t
, heap_start
);
530 assert(t
->heap_end
>= t
->heap_start
);
532 struct pandecode_mapped_memory
*heap
=
533 pandecode_find_mapped_gpu_mem_containing(t
->heap_start
);
535 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
537 /* Tiling is enabled with a special flag */
538 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
539 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
541 bool tiling_enabled
= hierarchy_mask
;
543 if (tiling_enabled
) {
544 /* When tiling is enabled, the heap should be a tight fit */
545 unsigned heap_offset
= t
->heap_start
- heap
->gpu_va
;
546 if ((heap_offset
+ heap_size
) != heap
->length
) {
547 pandecode_msg("XXX: heap size %d (expected %d)\n",
548 heap_size
, heap
->length
- heap_offset
);
551 /* We should also have no other flags */
553 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
555 /* When tiling is disabled, we should have that flag and no others */
557 if (tiler_flags
!= MALI_TILER_DISABLED
) {
558 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
562 /* We should also have an empty heap */
564 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
568 /* Disabled tiling is used only for clear-only jobs, which are
569 * purely FRAGMENT, so we should never see this for
570 * non-FRAGMENT descriptors. */
573 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
576 /* We've never seen weights used in practice, but we know from the
577 * kernel these fields is there */
579 bool nonzero_weights
= false;
581 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
582 nonzero_weights
|= t
->weights
[w
] != 0x0;
585 if (nonzero_weights
) {
586 pandecode_log(".weights = {");
588 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
589 pandecode_log("%d, ", t
->weights
[w
]);
596 pandecode_log("}\n");
600 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
)
602 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
603 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
605 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
608 pandecode_prop("unknown1 = 0x%" PRIx32
, s
->unknown1
);
609 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->unknown2
);
611 pandecode_log(".format = ");
612 pandecode_log_decoded_flags(fb_fmt_flag_info
, s
->format
);
613 pandecode_log_cont(",\n");
615 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", s
->width
+ 1);
616 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", s
->height
+ 1);
618 MEMORY_PROP(s
, framebuffer
);
619 pandecode_prop("stride = %d", s
->stride
);
621 /* Earlier in the actual commandstream -- right before width -- but we
622 * delay to flow nicer */
624 pandecode_log(".clear_flags = ");
625 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
626 pandecode_log_cont(",\n");
628 if (s
->depth_buffer
| s
->depth_buffer_enable
) {
629 MEMORY_PROP(s
, depth_buffer
);
630 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s
->depth_buffer_enable
));
633 if (s
->stencil_buffer
| s
->stencil_buffer_enable
) {
634 MEMORY_PROP(s
, stencil_buffer
);
635 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s
->stencil_buffer_enable
));
638 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
639 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
640 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
641 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
642 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
645 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
646 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
647 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
648 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
649 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
652 if (s
->clear_stencil
) {
653 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
656 MEMORY_PROP(s
, unknown_address_0
);
657 const struct midgard_tiler_descriptor t
= s
->tiler
;
658 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
);
661 pandecode_log("};\n");
663 pandecode_prop("zero0 = 0x%" PRIx64
, s
->zero0
);
664 pandecode_prop("zero1 = 0x%" PRIx64
, s
->zero1
);
665 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
666 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
668 printf(".zero3 = {");
670 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
671 printf("%X, ", s
->zero3
[i
]);
675 printf(".zero6 = {");
677 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
678 printf("%X, ", s
->zero6
[i
]);
684 pandecode_u32_slide(unsigned name
, const u32
*slide
, unsigned count
)
686 pandecode_log(".unknown%d = {", name
);
688 for (int i
= 0; i
< count
; ++i
)
689 printf("%X, ", slide
[i
]);
691 pandecode_log("},\n");
694 #define SHORT_SLIDE(num) \
695 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
698 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
700 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
701 const struct mali_compute_fbd
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
703 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
713 pandecode_swizzle(unsigned swizzle
)
715 pandecode_prop("swizzle = %s | (%s << 3) | (%s << 6) | (%s << 9)",
716 pandecode_channel((swizzle
>> 0) & 0x7),
717 pandecode_channel((swizzle
>> 3) & 0x7),
718 pandecode_channel((swizzle
>> 6) & 0x7),
719 pandecode_channel((swizzle
>> 9) & 0x7));
723 pandecode_rt_format(struct mali_rt_format format
)
725 pandecode_log(".format = {\n");
728 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
729 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
730 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
732 pandecode_prop("block = %s",
733 pandecode_mfbd_block_format(format
.block
));
735 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
736 MALI_NEGATIVE(format
.nr_channels
));
738 pandecode_log(".flags = ");
739 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
740 pandecode_log_cont(",\n");
742 pandecode_swizzle(format
.swizzle
);
744 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
747 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
750 pandecode_log("},\n");
754 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct bifrost_framebuffer
*fb
)
756 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
759 for (int i
= 0; i
< MALI_NEGATIVE(fb
->rt_count_1
); i
++) {
760 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct bifrost_render_target
);
761 struct pandecode_mapped_memory
*mem
=
762 pandecode_find_mapped_gpu_mem_containing(rt_va
);
763 const struct bifrost_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
765 pandecode_log("{\n");
768 pandecode_rt_format(rt
->format
);
770 if (rt
->format
.block
== MALI_MFBD_BLOCK_AFBC
) {
771 pandecode_log(".afbc = {\n");
774 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
775 pandecode_prop("metadata = %s", a
);
778 pandecode_prop("stride = %d", rt
->afbc
.stride
);
779 pandecode_prop("unk = 0x%" PRIx32
, rt
->afbc
.unk
);
782 pandecode_log("},\n");
784 pandecode_log(".chunknown = {\n");
787 pandecode_prop("unk = 0x%" PRIx64
, rt
->chunknown
.unk
);
789 char *a
= pointer_as_memory_reference(rt
->chunknown
.pointer
);
790 pandecode_prop("pointer = %s", a
);
794 pandecode_log("},\n");
797 MEMORY_PROP(rt
, framebuffer
);
798 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
800 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
801 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
802 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
803 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
804 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
807 if (rt
->zero1
|| rt
->zero2
|| rt
->zero3
) {
808 pandecode_msg("render target zeros tripped\n");
809 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
810 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
811 pandecode_prop("zero3 = 0x%" PRIx32
, rt
->zero3
);
815 pandecode_log("},\n");
819 pandecode_log("};\n");
823 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
)
825 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
826 const struct bifrost_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
828 if (fb
->sample_locations
) {
829 /* The blob stores all possible sample locations in a single buffer
830 * allocated on startup, and just switches the pointer when switching
831 * MSAA state. For now, we just put the data into the cmdstream, but we
832 * should do something like what the blob does with a real driver.
834 * There seem to be 32 slots for sample locations, followed by another
835 * 16. The second 16 is just the center location followed by 15 zeros
836 * in all the cases I've identified (maybe shader vs. depth/color
840 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->sample_locations
);
842 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->sample_locations
);
844 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
847 for (int i
= 0; i
< 32 + 16; i
++) {
848 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
852 pandecode_log("};\n");
855 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
858 pandecode_prop("unk0 = 0x%x", fb
->unk0
);
860 if (fb
->sample_locations
)
861 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
863 /* Assume that unknown1 was emitted in the last job for
865 MEMORY_PROP(fb
, unknown1
);
867 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
868 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
869 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
870 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
872 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
873 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
874 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
875 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
877 pandecode_log(".mfbd_flags = ");
878 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
879 pandecode_log_cont(",\n");
881 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
882 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
884 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
885 MEMORY_PROP(fb
, scratchpad
);
886 const struct midgard_tiler_descriptor t
= fb
->tiler
;
887 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
);
889 if (fb
->zero3
|| fb
->zero4
) {
890 pandecode_msg("framebuffer zeros tripped\n");
891 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
892 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
896 pandecode_log("};\n");
898 gpu_va
+= sizeof(struct bifrost_framebuffer
);
900 if ((fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
) {
901 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
902 const struct bifrost_fb_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
904 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
907 MEMORY_PROP(fbx
, checksum
);
909 if (fbx
->checksum_stride
)
910 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
912 pandecode_log(".flags = ");
913 pandecode_log_decoded_flags(mfbd_extra_flag_info
, fbx
->flags
);
914 pandecode_log_cont(",\n");
916 if (fbx
->flags
& MALI_EXTRA_AFBC_ZS
) {
917 pandecode_log(".ds_afbc = {\n");
920 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
921 pandecode_prop("depth_stencil_afbc_stride = %d",
922 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
923 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
925 if (fbx
->ds_afbc
.zero1
|| fbx
->ds_afbc
.padding
) {
926 pandecode_msg("Depth/stencil AFBC zeros tripped\n");
927 pandecode_prop("zero1 = 0x%" PRIx32
,
929 pandecode_prop("padding = 0x%" PRIx64
,
930 fbx
->ds_afbc
.padding
);
934 pandecode_log("},\n");
936 pandecode_log(".ds_linear = {\n");
939 if (fbx
->ds_linear
.depth
) {
940 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
941 pandecode_prop("depth_stride = %d",
942 fbx
->ds_linear
.depth_stride
);
945 if (fbx
->ds_linear
.stencil
) {
946 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
947 pandecode_prop("stencil_stride = %d",
948 fbx
->ds_linear
.stencil_stride
);
951 if (fbx
->ds_linear
.depth_stride_zero
||
952 fbx
->ds_linear
.stencil_stride_zero
||
953 fbx
->ds_linear
.zero1
|| fbx
->ds_linear
.zero2
) {
954 pandecode_msg("Depth/stencil zeros tripped\n");
955 pandecode_prop("depth_stride_zero = 0x%x",
956 fbx
->ds_linear
.depth_stride_zero
);
957 pandecode_prop("stencil_stride_zero = 0x%x",
958 fbx
->ds_linear
.stencil_stride_zero
);
959 pandecode_prop("zero1 = 0x%" PRIx32
,
960 fbx
->ds_linear
.zero1
);
961 pandecode_prop("zero2 = 0x%" PRIx32
,
962 fbx
->ds_linear
.zero2
);
966 pandecode_log("},\n");
969 if (fbx
->zero3
|| fbx
->zero4
) {
970 pandecode_msg("fb_extra zeros tripped\n");
971 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
972 pandecode_prop("zero4 = 0x%" PRIx64
, fbx
->zero4
);
976 pandecode_log("};\n");
978 gpu_va
+= sizeof(struct bifrost_fb_extra
);
982 pandecode_render_target(gpu_va
, job_no
, fb
);
984 /* Passback the render target count */
985 return MALI_NEGATIVE(fb
->rt_count_1
);
988 /* Just add a comment decoding the shift/odd fields forming the padded vertices
992 pandecode_padded_vertices(unsigned shift
, unsigned k
)
994 unsigned odd
= 2*k
+ 1;
995 unsigned pot
= 1 << shift
;
996 pandecode_msg("padded_num_vertices = %d\n", odd
* pot
);
999 /* Given a magic divisor, recover what we were trying to divide by.
1001 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1002 * 0 <= m < 2^N, for N bits in m.
1004 * Let q represent the number we would like to divide by.
1006 * By definition of a magic divisor for N-bit unsigned integers (a number you
1007 * multiply by to magically get division), m is a number such that:
1009 * (m * x) & (2^N - 1) = floor(x/q).
1010 * for all x on Z where 0 <= x < 2^N
1012 * Ignore the case where any of the above values equals zero; it is irrelevant
1013 * for our purposes (instanced arrays).
1015 * Choose x = q. Then:
1017 * (m * x) & (2^N - 1) = floor(x/q).
1018 * (m * q) & (2^N - 1) = floor(q/q).
1020 * floor(q/q) = floor(1) = 1, therefore:
1022 * (m * q) & (2^N - 1) = 1
1024 * Recall the identity that the bitwise AND of one less than a power-of-two
1025 * equals the modulo with that power of two, i.e. for all x:
1027 * x & (2^N - 1) = x % N
1033 * By definition, a modular multiplicative inverse of a number m is the number
1034 * q such that with respect to a modulos M:
1038 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1043 pandecode_magic_divisor(uint32_t magic
, unsigned shift
, unsigned orig_divisor
, unsigned extra
)
1046 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1047 * shift) the most lame way possible... just repeatedly add.
1048 * Asymptoptically slow but nobody cares in practice, unless you have
1049 * massive numbers of vertices or high divisors. */
1051 unsigned inverse
= 0;
1053 /* Magic implicitly has the highest bit set */
1056 /* Depending on rounding direction */
1061 uint32_t product
= magic
* inverse
;
1073 pandecode_msg("dividing by %d (maybe off by two)\n", inverse
);
1075 /* Recall we're supposed to divide by (gl_level_divisor *
1076 * padded_num_vertices) */
1078 unsigned padded_num_vertices
= inverse
/ orig_divisor
;
1080 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices
);
1085 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1086 mali_ptr addr
, int job_no
, char *suffix
,
1087 int count
, bool varying
)
1089 char *prefix
= varying
? "varyings" : "attributes";
1092 pandecode_msg("no %s\n", prefix
);
1096 union mali_attr
*attr
= pandecode_fetch_gpu_mem(mem
, addr
, sizeof(union mali_attr
) * count
);
1099 snprintf(base
, sizeof(base
), "%s_data_%d%s", prefix
, job_no
, suffix
);
1101 for (int i
= 0; i
< count
; ++i
) {
1102 enum mali_attr_mode mode
= attr
[i
].elements
& 7;
1104 if (mode
== MALI_ATTR_UNUSED
)
1107 mali_ptr raw_elements
= attr
[i
].elements
& ~7;
1109 /* TODO: Do we maybe want to dump the attribute values
1110 * themselves given the specified format? Or is that too hard?
1113 char *a
= pointer_as_memory_reference(raw_elements
);
1114 pandecode_log("mali_ptr %s_%d_p = %s;\n", base
, i
, a
);
1118 pandecode_log("union mali_attr %s_%d[] = {\n", prefix
, job_no
);
1121 for (int i
= 0; i
< count
; ++i
) {
1122 pandecode_log("{\n");
1125 unsigned mode
= attr
[i
].elements
& 7;
1126 pandecode_prop("elements = (%s_%d_p) | %s", base
, i
, pandecode_attr_mode(mode
));
1127 pandecode_prop("shift = %d", attr
[i
].shift
);
1128 pandecode_prop("extra_flags = %d", attr
[i
].extra_flags
);
1129 pandecode_prop("stride = 0x%" PRIx32
, attr
[i
].stride
);
1130 pandecode_prop("size = 0x%" PRIx32
, attr
[i
].size
);
1132 /* Decode further where possible */
1134 if (mode
== MALI_ATTR_MODULO
) {
1135 pandecode_padded_vertices(
1137 attr
[i
].extra_flags
);
1141 pandecode_log("}, \n");
1143 if (mode
== MALI_ATTR_NPOT_DIVIDE
) {
1145 pandecode_log("{\n");
1147 pandecode_prop("unk = 0x%x", attr
[i
].unk
);
1148 pandecode_prop("magic_divisor = 0x%08x", attr
[i
].magic_divisor
);
1149 if (attr
[i
].zero
!= 0)
1150 pandecode_prop("zero = 0x%x /* XXX zero tripped */", attr
[i
].zero
);
1151 pandecode_prop("divisor = %d", attr
[i
].divisor
);
1152 pandecode_magic_divisor(attr
[i
].magic_divisor
, attr
[i
- 1].shift
, attr
[i
].divisor
, attr
[i
- 1].extra_flags
);
1154 pandecode_log("}, \n");
1160 pandecode_log("};\n");
1164 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1166 /* TODO: Decode flags */
1167 mali_ptr shader_ptr
= ptr
& ~15;
1169 char *a
= pointer_as_memory_reference(shader_ptr
);
1170 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1177 all_zero(unsigned *buffer
, unsigned count
)
1179 for (unsigned i
= 0; i
< count
; ++i
) {
1188 pandecode_stencil(const char *name
, const struct mali_stencil_test
*stencil
)
1190 if (all_zero((unsigned *) stencil
, sizeof(stencil
) / sizeof(unsigned)))
1193 const char *func
= pandecode_func(stencil
->func
);
1194 const char *sfail
= pandecode_stencil_op(stencil
->sfail
);
1195 const char *dpfail
= pandecode_stencil_op(stencil
->dpfail
);
1196 const char *dppass
= pandecode_stencil_op(stencil
->dppass
);
1199 pandecode_msg("Stencil zero tripped: %X\n", stencil
->zero
);
1201 pandecode_log(".stencil_%s = {\n", name
);
1203 pandecode_prop("ref = %d", stencil
->ref
);
1204 pandecode_prop("mask = 0x%02X", stencil
->mask
);
1205 pandecode_prop("func = %s", func
);
1206 pandecode_prop("sfail = %s", sfail
);
1207 pandecode_prop("dpfail = %s", dpfail
);
1208 pandecode_prop("dppass = %s", dppass
);
1210 pandecode_log("},\n");
1214 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1217 pandecode_msg("Blend zero tripped: %X\n", blend
->zero1
);
1219 pandecode_log(".equation = {\n");
1222 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1223 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1225 pandecode_log(".color_mask = ");
1226 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1227 pandecode_log_cont(",\n");
1230 pandecode_log("},\n");
1233 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1236 decode_bifrost_constant(u16 constant
)
1238 float lo
= (float) (constant
& 0xFF);
1239 float hi
= (float) (constant
>> 8);
1241 return (hi
/ 255.0) + (lo
/ 65535.0);
1245 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1247 struct bifrost_blend_rt
*b
=
1248 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1250 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1253 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1254 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1255 b
->constant
, decode_bifrost_constant(b
->constant
));
1257 /* TODO figure out blend shader enable bit */
1258 pandecode_blend_equation(&b
->equation
);
1259 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1260 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1261 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1264 pandecode_log("},\n");
1270 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1272 if (all_zero((unsigned *) blend
, sizeof(blend
) / sizeof(unsigned)))
1275 pandecode_log(".blend = {\n");
1279 pandecode_shader_address("shader", blend
->shader
);
1281 pandecode_blend_equation(&blend
->equation
);
1282 pandecode_prop("constant = %f", blend
->constant
);
1286 pandecode_log("},\n");
1288 /* Return blend shader to disassemble if present */
1289 return is_shader
? (blend
->shader
& ~0xF) : 0;
1293 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1295 struct midgard_blend_rt
*b
=
1296 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1298 /* Flags determine presence of blend shader */
1299 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1301 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1304 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1306 union midgard_blend blend
= b
->blend
;
1307 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1310 pandecode_log("};\n");
1316 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1319 char *prefix
= varying
? "varying" : "attribute";
1320 unsigned max_index
= 0;
1321 snprintf(base
, sizeof(base
), "%s_meta", prefix
);
1323 pandecode_log("struct mali_attr_meta %s_%d%s[] = {\n", base
, job_no
, suffix
);
1326 struct mali_attr_meta
*attr_meta
;
1327 mali_ptr p
= varying
? (v
->varying_meta
& ~0xF) : v
->attribute_meta
;
1329 struct pandecode_mapped_memory
*attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
);
1331 for (int i
= 0; i
< count
; ++i
, p
+= sizeof(struct mali_attr_meta
)) {
1332 attr_meta
= pandecode_fetch_gpu_mem(attr_mem
, p
,
1335 pandecode_log("{\n");
1337 pandecode_prop("index = %d", attr_meta
->index
);
1339 if (attr_meta
->index
> max_index
)
1340 max_index
= attr_meta
->index
;
1341 pandecode_swizzle(attr_meta
->swizzle
);
1342 pandecode_prop("format = %s", pandecode_format(attr_meta
->format
));
1344 pandecode_prop("unknown1 = 0x%" PRIx64
, (u64
) attr_meta
->unknown1
);
1345 pandecode_prop("unknown3 = 0x%" PRIx64
, (u64
) attr_meta
->unknown3
);
1346 pandecode_prop("src_offset = %d", attr_meta
->src_offset
);
1348 pandecode_log("},\n");
1353 pandecode_log("};\n");
1355 return count
? (max_index
+ 1) : 0;
1359 pandecode_indices(uintptr_t pindices
, uint32_t index_count
, int job_no
)
1361 struct pandecode_mapped_memory
*imem
= pandecode_find_mapped_gpu_mem_containing(pindices
);
1364 /* Indices are literally just a u32 array :) */
1366 uint32_t *PANDECODE_PTR_VAR(indices
, imem
, pindices
);
1368 pandecode_log("uint32_t indices_%d[] = {\n", job_no
);
1371 for (unsigned i
= 0; i
< (index_count
+ 1); i
+= 3)
1372 pandecode_log("%d, %d, %d,\n",
1378 pandecode_log("};\n");
1382 /* return bits [lo, hi) of word */
1384 bits(u32 word
, u32 lo
, u32 hi
)
1387 return word
; // avoid undefined behavior with the shift
1389 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1393 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool noninstanced
)
1395 pandecode_log_cont("{\n");
1398 /* Decode invocation_count. See the comment before the definition of
1399 * invocation_count for an explanation.
1402 unsigned size_x
= bits(p
->invocation_count
, 0, p
->size_y_shift
) + 1;
1403 unsigned size_y
= bits(p
->invocation_count
, p
->size_y_shift
, p
->size_z_shift
) + 1;
1404 unsigned size_z
= bits(p
->invocation_count
, p
->size_z_shift
, p
->workgroups_x_shift
) + 1;
1406 unsigned groups_x
= bits(p
->invocation_count
, p
->workgroups_x_shift
, p
->workgroups_y_shift
) + 1;
1407 unsigned groups_y
= bits(p
->invocation_count
, p
->workgroups_y_shift
, p
->workgroups_z_shift
) + 1;
1408 unsigned groups_z
= bits(p
->invocation_count
, p
->workgroups_z_shift
, 32) + 1;
1410 /* Even though we have this decoded, we want to ensure that the
1411 * representation is "unique" so we don't lose anything by printing only
1412 * the final result. More specifically, we need to check that we were
1413 * passed something in canonical form, since the definition per the
1414 * hardware is inherently not unique. How? Well, take the resulting
1415 * decode and pack it ourselves! If it is bit exact with what we
1416 * decoded, we're good to go. */
1418 struct mali_vertex_tiler_prefix ref
;
1419 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, noninstanced
);
1422 (p
->invocation_count
== ref
.invocation_count
) &&
1423 (p
->size_y_shift
== ref
.size_y_shift
) &&
1424 (p
->size_z_shift
== ref
.size_z_shift
) &&
1425 (p
->workgroups_x_shift
== ref
.workgroups_x_shift
) &&
1426 (p
->workgroups_y_shift
== ref
.workgroups_y_shift
) &&
1427 (p
->workgroups_z_shift
== ref
.workgroups_z_shift
) &&
1428 (p
->workgroups_x_shift_2
== ref
.workgroups_x_shift_2
);
1431 pandecode_msg("XXX: non-canonical workgroups packing\n");
1432 pandecode_msg("expected: %X, %d, %d, %d, %d, %d\n",
1433 ref
.invocation_count
,
1436 ref
.workgroups_x_shift
,
1437 ref
.workgroups_y_shift
,
1438 ref
.workgroups_z_shift
,
1439 ref
.workgroups_x_shift_2
);
1441 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1442 pandecode_prop("size_y_shift = %d", p
->size_y_shift
);
1443 pandecode_prop("size_z_shift = %d", p
->size_z_shift
);
1444 pandecode_prop("workgroups_x_shift = %d", p
->workgroups_x_shift
);
1445 pandecode_prop("workgroups_y_shift = %d", p
->workgroups_y_shift
);
1446 pandecode_prop("workgroups_z_shift = %d", p
->workgroups_z_shift
);
1447 pandecode_prop("workgroups_x_shift_2 = %d", p
->workgroups_x_shift_2
);
1450 /* Regardless, print the decode */
1451 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1452 size_x
, size_y
, size_z
,
1453 groups_x
, groups_y
, groups_z
);
1456 if (p
->unknown_draw
)
1457 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1459 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1461 if (p
->draw_mode
!= MALI_DRAW_NONE
)
1462 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p
->draw_mode
));
1464 /* Index count only exists for tiler jobs anyway */
1467 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1469 if (p
->offset_bias_correction
)
1470 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1473 pandecode_msg("Zero tripped\n");
1474 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1478 pandecode_log("},\n");
1482 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1484 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1486 struct mali_uniform_buffer_meta
*PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1488 for (int i
= 0; i
< ubufs_count
; i
++) {
1489 mali_ptr ptr
= ubufs
[i
].ptr
<< 2;
1490 struct pandecode_mapped_memory
*umem2
= pandecode_find_mapped_gpu_mem_containing(ptr
);
1491 uint32_t *PANDECODE_PTR_VAR(ubuf
, umem2
, ptr
);
1493 snprintf(name
, sizeof(name
), "ubuf_%d", i
);
1494 /* The blob uses ubuf 0 to upload internal stuff and
1495 * uniforms that won't fit/are accessed indirectly, so
1496 * it puts it in the batchbuffer.
1498 pandecode_log("uint32_t %s_%d[] = {\n", name
, job_no
);
1501 for (int j
= 0; j
<= ubufs
[i
].size
; j
++) {
1502 for (int k
= 0; k
< 4; k
++) {
1504 pandecode_log("0x%"PRIx32
", ", ubuf
[4 * j
+ k
]);
1506 pandecode_log_cont("0x%"PRIx32
", ", ubuf
[4 * j
+ k
]);
1510 pandecode_log_cont("\n");
1514 pandecode_log("};\n");
1517 pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%"PRIx64
"_%d[] = {\n",
1521 for (int i
= 0; i
< ubufs_count
; i
++) {
1522 pandecode_log("{\n");
1524 pandecode_prop("size = MALI_POSITIVE(%d)", ubufs
[i
].size
+ 1);
1525 pandecode_prop("ptr = ubuf_%d_%d_p >> 2", i
, job_no
);
1527 pandecode_log("},\n");
1531 pandecode_log("};\n");
1535 pandecode_scratchpad(uintptr_t pscratchpad
, int job_no
, char *suffix
)
1538 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(pscratchpad
);
1540 struct bifrost_scratchpad
*PANDECODE_PTR_VAR(scratchpad
, mem
, pscratchpad
);
1542 if (scratchpad
->zero
)
1543 pandecode_msg("XXX scratchpad zero tripped");
1545 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64
"_%d%s = {\n", pscratchpad
, job_no
, suffix
);
1548 pandecode_prop("flags = 0x%x", scratchpad
->flags
);
1549 MEMORY_PROP(scratchpad
, gpu_scratchpad
);
1552 pandecode_log("};\n");
1555 static unsigned shader_id
= 0;
1558 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1559 bool is_bifrost
, unsigned nr_regs
)
1561 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1562 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1564 /* Compute maximum possible size */
1565 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1567 /* Print some boilerplate to clearly denote the assembly (which doesn't
1568 * obey indentation rules), and actually do the disassembly! */
1574 snprintf(prefix
, sizeof(prefix
) - 1, "shader%d - %s shader: ",
1576 (type
== JOB_TYPE_TILER
) ? "FRAGMENT" : "VERTEX");
1579 disassemble_bifrost(code
, sz
, false);
1581 disassemble_midgard(code
, sz
, true, nr_regs
, prefix
);
1588 pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix
*p
,
1589 int job_no
, enum mali_job_type job_type
,
1590 char *suffix
, bool is_bifrost
)
1592 mali_ptr shader_meta_ptr
= (u64
) (uintptr_t) (p
->_shader_upper
<< 4);
1593 struct pandecode_mapped_memory
*attr_mem
;
1595 unsigned rt_count
= 1;
1597 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1598 * are the only things actually needed from the FBD, vertex/tiler jobs
1599 * no longer reference the FBD -- instead, this field points to some
1600 * info about the scratchpad.
1603 pandecode_scratchpad(p
->framebuffer
& ~FBD_TYPE
, job_no
, suffix
);
1604 else if (p
->framebuffer
& MALI_MFBD
)
1605 rt_count
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->framebuffer
) & FBD_MASK
, job_no
, false);
1606 else if (job_type
== JOB_TYPE_COMPUTE
)
1607 pandecode_compute_fbd((u64
) (uintptr_t) p
->framebuffer
, job_no
);
1609 pandecode_sfbd((u64
) (uintptr_t) p
->framebuffer
, job_no
, false);
1611 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
1612 int texture_count
= 0, sampler_count
= 0;
1614 if (shader_meta_ptr
) {
1615 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr
);
1616 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, shader_meta_ptr
);
1618 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", shader_meta_ptr
, job_no
, suffix
);
1621 /* Save for dumps */
1622 attribute_count
= s
->attribute_count
;
1623 varying_count
= s
->varying_count
;
1624 texture_count
= s
->texture_count
;
1625 sampler_count
= s
->sampler_count
;
1628 uniform_count
= s
->bifrost2
.uniform_count
;
1629 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
1631 uniform_count
= s
->midgard1
.uniform_count
;
1632 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
1635 mali_ptr shader_ptr
= pandecode_shader_address("shader", s
->shader
);
1637 pandecode_prop("texture_count = %" PRId16
, s
->texture_count
);
1638 pandecode_prop("sampler_count = %" PRId16
, s
->sampler_count
);
1639 pandecode_prop("attribute_count = %" PRId16
, s
->attribute_count
);
1640 pandecode_prop("varying_count = %" PRId16
, s
->varying_count
);
1642 unsigned nr_registers
= 0;
1645 pandecode_log(".bifrost1 = {\n");
1648 pandecode_prop("uniform_buffer_count = %" PRId32
, s
->bifrost1
.uniform_buffer_count
);
1649 pandecode_prop("unk1 = 0x%" PRIx32
, s
->bifrost1
.unk1
);
1652 pandecode_log("},\n");
1654 pandecode_log(".midgard1 = {\n");
1657 pandecode_prop("uniform_count = %" PRId16
, s
->midgard1
.uniform_count
);
1658 pandecode_prop("uniform_buffer_count = %" PRId16
, s
->midgard1
.uniform_buffer_count
);
1659 pandecode_prop("work_count = %" PRId16
, s
->midgard1
.work_count
);
1660 nr_registers
= s
->midgard1
.work_count
;
1662 pandecode_log(".flags = ");
1663 pandecode_log_decoded_flags(shader_midgard1_flag_info
, s
->midgard1
.flags
);
1664 pandecode_log_cont(",\n");
1666 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->midgard1
.unknown2
);
1669 pandecode_log("},\n");
1672 if (s
->depth_units
|| s
->depth_factor
) {
1673 pandecode_prop("depth_factor = %f", s
->depth_factor
);
1674 pandecode_prop("depth_units = %f", s
->depth_units
);
1677 if (s
->alpha_coverage
) {
1678 bool invert_alpha_coverage
= s
->alpha_coverage
& 0xFFF0;
1679 uint16_t inverted_coverage
= invert_alpha_coverage
? ~s
->alpha_coverage
: s
->alpha_coverage
;
1681 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
1682 invert_alpha_coverage
? "~" : "",
1683 MALI_GET_ALPHA_COVERAGE(inverted_coverage
));
1686 if (s
->unknown2_3
|| s
->unknown2_4
) {
1687 pandecode_log(".unknown2_3 = ");
1689 int unknown2_3
= s
->unknown2_3
;
1690 int unknown2_4
= s
->unknown2_4
;
1692 /* We're not quite sure what these flags mean without the depth test, if anything */
1694 if (unknown2_3
& (MALI_DEPTH_TEST
| MALI_DEPTH_FUNC_MASK
)) {
1695 const char *func
= pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3
));
1696 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
1698 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
1701 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
1702 pandecode_log_cont(",\n");
1704 pandecode_log(".unknown2_4 = ");
1705 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
1706 pandecode_log_cont(",\n");
1709 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
1710 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
1711 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
1714 pandecode_stencil("front", &s
->stencil_front
);
1715 pandecode_stencil("back", &s
->stencil_back
);
1718 pandecode_log(".bifrost2 = {\n");
1721 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
1722 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
1723 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
1724 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
1727 pandecode_log("},\n");
1728 } else if (s
->midgard2
.unknown2_7
) {
1729 pandecode_log(".midgard2 = {\n");
1732 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
1734 pandecode_log("},\n");
1738 pandecode_prop("unknown2_8 = 0x%" PRIx32
, s
->unknown2_8
);
1741 /* TODO: Blend shaders routing/disasm */
1743 union midgard_blend blend
= s
->blend
;
1744 pandecode_midgard_blend(&blend
, false);
1748 pandecode_log("};\n");
1750 /* MRT blend fields are used whenever MFBD is used, with
1751 * per-RT descriptors */
1753 if (job_type
== JOB_TYPE_TILER
) {
1754 void* blend_base
= (void *) (s
+ 1);
1756 for (unsigned i
= 0; i
< rt_count
; i
++) {
1757 mali_ptr shader
= 0;
1760 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
1762 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
1765 pandecode_shader_disassemble(shader
, job_no
, job_type
, false, 0);
1769 if (shader_ptr
& ~0xF)
1770 pandecode_shader_disassemble(shader_ptr
, job_no
, job_type
, is_bifrost
, nr_registers
);
1772 pandecode_msg("<no shader>\n");
1775 struct pandecode_mapped_memory
*fmem
= pandecode_find_mapped_gpu_mem_containing(p
->viewport
);
1776 struct mali_viewport
*PANDECODE_PTR_VAR(f
, fmem
, p
->viewport
);
1778 pandecode_log("struct mali_viewport viewport_%"PRIx64
"_%d%s = {\n", p
->viewport
, job_no
, suffix
);
1781 pandecode_prop("clip_minx = %f", f
->clip_minx
);
1782 pandecode_prop("clip_miny = %f", f
->clip_miny
);
1783 pandecode_prop("clip_minz = %f", f
->clip_minz
);
1784 pandecode_prop("clip_maxx = %f", f
->clip_maxx
);
1785 pandecode_prop("clip_maxy = %f", f
->clip_maxy
);
1786 pandecode_prop("clip_maxz = %f", f
->clip_maxz
);
1788 /* Only the higher coordinates are MALI_POSITIVE scaled */
1790 pandecode_prop("viewport0 = { %d, %d }",
1791 f
->viewport0
[0], f
->viewport0
[1]);
1793 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
1794 f
->viewport1
[0] + 1, f
->viewport1
[1] + 1);
1797 pandecode_log("};\n");
1800 if (p
->attribute_meta
) {
1801 unsigned max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
1803 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
1804 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false);
1807 /* Varyings are encoded like attributes but not actually sent; we just
1808 * pass a zero buffer with the right stride/size set, (or whatever)
1809 * since the GPU will write to it itself */
1811 if (p
->varying_meta
) {
1812 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
1816 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
1818 /* Number of descriptors depends on whether there are
1819 * non-internal varyings */
1821 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true);
1824 bool is_compute
= job_type
== JOB_TYPE_COMPUTE
;
1826 if (p
->uniforms
&& !is_compute
) {
1827 int rows
= uniform_count
, width
= 4;
1828 size_t sz
= rows
* width
* sizeof(float);
1830 struct pandecode_mapped_memory
*uniform_mem
= pandecode_find_mapped_gpu_mem_containing(p
->uniforms
);
1831 pandecode_fetch_gpu_mem(uniform_mem
, p
->uniforms
, sz
);
1832 u32
*PANDECODE_PTR_VAR(uniforms
, uniform_mem
, p
->uniforms
);
1834 pandecode_log("u32 uniforms_%d%s[] = {\n", job_no
, suffix
);
1838 for (int row
= 0; row
< rows
; row
++) {
1839 for (int i
= 0; i
< width
; i
++) {
1840 u32 v
= uniforms
[i
];
1842 memcpy(&f
, &v
, sizeof(v
));
1843 pandecode_log_cont("%X /* %f */, ", v
, f
);
1846 pandecode_log_cont("\n");
1852 pandecode_log("};\n");
1853 } else if (p
->uniforms
) {
1854 int rows
= uniform_count
* 2;
1855 size_t sz
= rows
* sizeof(mali_ptr
);
1857 struct pandecode_mapped_memory
*uniform_mem
= pandecode_find_mapped_gpu_mem_containing(p
->uniforms
);
1858 pandecode_fetch_gpu_mem(uniform_mem
, p
->uniforms
, sz
);
1859 mali_ptr
*PANDECODE_PTR_VAR(uniforms
, uniform_mem
, p
->uniforms
);
1861 pandecode_log("mali_ptr uniforms_%d%s[] = {\n", job_no
, suffix
);
1865 for (int row
= 0; row
< rows
; row
++) {
1866 char *a
= pointer_as_memory_reference(uniforms
[row
]);
1867 pandecode_log("%s,\n", a
);
1872 pandecode_log("};\n");
1876 if (p
->uniform_buffers
) {
1877 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
1880 if (p
->texture_trampoline
) {
1881 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(p
->texture_trampoline
);
1884 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
);
1886 pandecode_log("uint64_t texture_trampoline_%"PRIx64
"_%d[] = {\n", p
->texture_trampoline
, job_no
);
1889 for (int tex
= 0; tex
< texture_count
; ++tex
) {
1890 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
1891 char *a
= pointer_as_memory_reference(*u
);
1892 pandecode_log("%s,\n", a
);
1897 pandecode_log("};\n");
1899 /* Now, finally, descend down into the texture descriptor */
1900 for (int tex
= 0; tex
< texture_count
; ++tex
) {
1901 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
1902 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
1905 struct mali_texture_descriptor
*PANDECODE_PTR_VAR(t
, tmem
, *u
);
1907 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64
"_%d_%d = {\n", *u
, job_no
, tex
);
1910 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", t
->width
+ 1);
1911 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", t
->height
+ 1);
1912 pandecode_prop("depth = MALI_POSITIVE(%" PRId16
")", t
->depth
+ 1);
1913 pandecode_prop("array_size = MALI_POSITIVE(%" PRId16
")", t
->array_size
+ 1);
1914 pandecode_prop("unknown3 = %" PRId16
, t
->unknown3
);
1915 pandecode_prop("unknown3A = %" PRId8
, t
->unknown3A
);
1916 pandecode_prop("nr_mipmap_levels = %" PRId8
, t
->nr_mipmap_levels
);
1918 struct mali_texture_format f
= t
->format
;
1920 pandecode_log(".format = {\n");
1923 pandecode_swizzle(f
.swizzle
);
1924 pandecode_prop("format = %s", pandecode_format(f
.format
));
1925 pandecode_prop("type = %s", pandecode_texture_type(f
.type
));
1926 pandecode_prop("srgb = %" PRId32
, f
.srgb
);
1927 pandecode_prop("unknown1 = %" PRId32
, f
.unknown1
);
1928 pandecode_prop("usage2 = 0x%" PRIx32
, f
.usage2
);
1931 pandecode_log("},\n");
1933 pandecode_swizzle(t
->swizzle
);
1935 if (t
->swizzle_zero
) {
1936 /* Shouldn't happen */
1937 pandecode_msg("Swizzle zero tripped but replay will be fine anyway");
1938 pandecode_prop("swizzle_zero = %d", t
->swizzle_zero
);
1941 pandecode_prop("unknown3 = 0x%" PRIx32
, t
->unknown3
);
1943 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
1944 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
1945 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
1947 pandecode_log(".payload = {\n");
1950 /* A bunch of bitmap pointers follow.
1951 * We work out the correct number,
1952 * based on the mipmap/cubemap
1953 * properties, but dump extra
1954 * possibilities to futureproof */
1956 int bitmap_count
= MALI_NEGATIVE(t
->nr_mipmap_levels
);
1957 bool manual_stride
= f
.usage2
& MALI_TEX_MANUAL_STRIDE
;
1959 /* Miptree for each face */
1960 if (f
.type
== MALI_TEX_CUBE
)
1963 /* Array of textures */
1964 bitmap_count
*= MALI_NEGATIVE(t
->array_size
);
1966 /* Stride for each element */
1970 /* Sanity check the size */
1971 int max_count
= sizeof(t
->payload
) / sizeof(t
->payload
[0]);
1972 assert (bitmap_count
<= max_count
);
1974 for (int i
= 0; i
< bitmap_count
; ++i
) {
1975 /* How we dump depends if this is a stride or a pointer */
1977 if ((f
.usage2
& MALI_TEX_MANUAL_STRIDE
) && (i
& 1)) {
1978 /* signed 32-bit snuck in as a 64-bit pointer */
1979 uint64_t stride_set
= t
->payload
[i
];
1980 uint32_t clamped_stride
= stride_set
;
1981 int32_t stride
= clamped_stride
;
1982 assert(stride_set
== clamped_stride
);
1983 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
1985 char *a
= pointer_as_memory_reference(t
->payload
[i
]);
1986 pandecode_log("%s, \n", a
);
1992 pandecode_log("},\n");
1995 pandecode_log("};\n");
2001 if (p
->sampler_descriptor
) {
2002 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->sampler_descriptor
);
2005 struct mali_sampler_descriptor
*s
;
2007 mali_ptr d
= p
->sampler_descriptor
;
2009 for (int i
= 0; i
< sampler_count
; ++i
) {
2010 s
= pandecode_fetch_gpu_mem(smem
, d
+ sizeof(*s
) * i
, sizeof(*s
));
2012 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", d
+ sizeof(*s
) * i
, job_no
, i
);
2015 pandecode_log(".filter_mode = ");
2016 pandecode_log_decoded_flags(sampler_flag_info
, s
->filter_mode
);
2017 pandecode_log_cont(",\n");
2019 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
2020 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
2022 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s
->wrap_s
));
2023 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s
->wrap_t
));
2024 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s
->wrap_r
));
2026 pandecode_prop("compare_func = %s", pandecode_alt_func(s
->compare_func
));
2028 if (s
->zero
|| s
->zero2
) {
2029 pandecode_msg("Zero tripped\n");
2030 pandecode_prop("zero = 0x%X, 0x%X\n", s
->zero
, s
->zero2
);
2033 pandecode_prop("seamless_cube_map = %d", s
->seamless_cube_map
);
2035 pandecode_prop("border_color = { %f, %f, %f, %f }",
2039 s
->border_color
[3]);
2042 pandecode_log("};\n");
2049 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2051 if (!(p
->position_varying
|| p
->occlusion_counter
|| p
->flags
))
2054 pandecode_log(".postfix = {\n");
2057 MEMORY_PROP(p
, position_varying
);
2058 MEMORY_PROP(p
, occlusion_counter
);
2061 pandecode_prop("flags = %d", p
->flags
);
2064 pandecode_log("},\n");
2068 pandecode_vertex_only_bfr(struct bifrost_vertex_only
*v
)
2070 pandecode_log_cont("{\n");
2073 pandecode_prop("unk2 = 0x%x", v
->unk2
);
2075 if (v
->zero0
|| v
->zero1
) {
2076 pandecode_msg("vertex only zero tripped");
2077 pandecode_prop("zero0 = 0x%" PRIx32
, v
->zero0
);
2078 pandecode_prop("zero1 = 0x%" PRIx64
, v
->zero1
);
2082 pandecode_log("}\n");
2086 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2089 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2090 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2092 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no
);
2096 pandecode_msg("tiler heap zero tripped\n");
2097 pandecode_prop("zero = 0x%x", h
->zero
);
2100 for (int i
= 0; i
< 12; i
++) {
2101 if (h
->zeros
[i
] != 0) {
2102 pandecode_msg("tiler heap zero %d tripped, value %x\n",
2107 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2108 MEMORY_PROP(h
, tiler_heap_start
);
2109 MEMORY_PROP(h
, tiler_heap_free
);
2111 /* this might point to the beginning of another buffer, when it's
2112 * really the end of the tiler heap buffer, so we have to be careful
2113 * here. but for zero length, we need the same pointer.
2116 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2117 MEMORY_PROP(h
, tiler_heap_start
);
2119 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2120 pandecode_prop("tiler_heap_end = %s + 1", a
);
2125 pandecode_log("};\n");
2129 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2131 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2132 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2134 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2136 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no
);
2139 if (t
->zero0
|| t
->zero1
) {
2140 pandecode_msg("tiler meta zero tripped");
2141 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2142 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2145 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2146 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2148 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2149 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2151 for (int i
= 0; i
< 12; i
++) {
2152 if (t
->zeros
[i
] != 0) {
2153 pandecode_msg("tiler heap zero %d tripped, value %" PRIx64
"\n",
2159 pandecode_log("};\n");
2163 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2165 pandecode_log(".gl_enables = ");
2167 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2169 pandecode_log_cont(",\n");
2173 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2175 if (u
.pointer
== 0x0)
2178 pandecode_log(".primitive_size = {\n");
2182 pandecode_prop("constant = %f", u
.constant
);
2184 MEMORY_PROP((&u
), pointer
);
2188 pandecode_log("},\n");
2192 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2194 pandecode_log_cont("{\n");
2197 /* TODO: gl_PointSize on Bifrost */
2198 pandecode_primitive_size(t
->primitive_size
, true);
2200 pandecode_gl_enables(t
->gl_enables
, JOB_TYPE_TILER
);
2202 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2203 || t
->zero6
|| t
->zero7
|| t
->zero8
) {
2204 pandecode_msg("tiler only zero tripped");
2205 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2206 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2207 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2208 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2209 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2210 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2211 pandecode_prop("zero7 = 0x%" PRIx32
, t
->zero7
);
2212 pandecode_prop("zero8 = 0x%" PRIx64
, t
->zero8
);
2216 pandecode_log("},\n");
2220 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2221 const struct pandecode_mapped_memory
*mem
,
2222 mali_ptr payload
, int job_no
)
2224 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2226 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true);
2228 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no
);
2231 pandecode_log(".prefix = ");
2232 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2234 pandecode_log(".vertex = ");
2235 pandecode_vertex_only_bfr(&v
->vertex
);
2237 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2240 pandecode_log("};\n");
2246 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2247 const struct pandecode_mapped_memory
*mem
,
2248 mali_ptr payload
, int job_no
)
2250 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2252 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true);
2254 pandecode_indices(t
->prefix
.indices
, t
->prefix
.index_count
, job_no
);
2255 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2257 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no
);
2260 pandecode_log(".prefix = ");
2261 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2263 pandecode_log(".tiler = ");
2264 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2266 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2269 pandecode_log("};\n");
2275 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2276 const struct pandecode_mapped_memory
*mem
,
2277 mali_ptr payload
, int job_no
)
2279 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2281 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false);
2283 pandecode_indices(v
->prefix
.indices
, v
->prefix
.index_count
, job_no
);
2285 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2288 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2289 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2291 bool instanced
= v
->instance_shift
|| v
->instance_odd
;
2292 bool is_graphics
= (h
->job_type
== JOB_TYPE_VERTEX
) || (h
->job_type
== JOB_TYPE_TILER
);
2294 pandecode_log(".prefix = ");
2295 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, !instanced
&& is_graphics
);
2297 pandecode_gl_enables(v
->gl_enables
, h
->job_type
);
2299 if (v
->instance_shift
|| v
->instance_odd
) {
2300 pandecode_prop("instance_shift = 0x%d /* %d */",
2301 v
->instance_shift
, 1 << v
->instance_shift
);
2302 pandecode_prop("instance_odd = 0x%X /* %d */",
2303 v
->instance_odd
, (2 * v
->instance_odd
) + 1);
2305 pandecode_padded_vertices(v
->instance_shift
, v
->instance_odd
);
2308 if (v
->offset_start
)
2309 pandecode_prop("offset_start = %d", v
->offset_start
);
2312 pandecode_msg("Zero tripped\n");
2313 pandecode_prop("zero5 = 0x%" PRIx64
, v
->zero5
);
2316 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2319 pandecode_log("};\n");
2325 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2326 mali_ptr payload
, int job_no
,
2329 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2331 bool fbd_dumped
= false;
2333 if (!is_bifrost
&& (s
->framebuffer
& FBD_TYPE
) == MALI_SFBD
) {
2334 /* Only SFBDs are understood, not MFBDs. We're speculating,
2335 * based on the versioning, kernel code, etc, that the
2336 * difference is between Single FrameBuffer Descriptor and
2337 * Multiple FrmaeBuffer Descriptor; the change apparently lines
2338 * up with multi-framebuffer support being added (T7xx onwards,
2339 * including Gxx). In any event, there's some field shuffling
2340 * that we haven't looked into yet. */
2342 pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true);
2344 } else if ((s
->framebuffer
& FBD_TYPE
) == MALI_MFBD
) {
2345 /* We don't know if Bifrost supports SFBD's at all, since the
2346 * driver never uses them. And the format is different from
2347 * Midgard anyways, due to the tiler heap and scratchpad being
2348 * moved out into separate structures, so it's not clear what a
2349 * Bifrost SFBD would even look like without getting an actual
2350 * trace, which appears impossible.
2353 pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true);
2357 uintptr_t p
= (uintptr_t) s
->framebuffer
& FBD_MASK
;
2358 pandecode_log("struct mali_payload_fragment payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2361 /* See the comments by the macro definitions for mathematical context
2362 * on why this is so weird */
2364 if (MALI_TILE_COORD_FLAGS(s
->max_tile_coord
) || MALI_TILE_COORD_FLAGS(s
->min_tile_coord
))
2365 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
2367 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
2368 MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
,
2369 MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
);
2371 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
2372 (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
,
2373 (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
);
2375 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2376 * we have to fallback on offsets. */
2378 const char *fbd_type
= s
->framebuffer
& MALI_MFBD
? "MALI_MFBD" : "MALI_SFBD";
2381 unsigned extra_flags
= (s
->framebuffer
& ~FBD_MASK
) & ~MALI_MFBD
;
2384 pandecode_prop("framebuffer = framebuffer_%d_p | %s | 0x%X", job_no
,
2385 fbd_type
, extra_flags
);
2387 pandecode_prop("framebuffer = %s | %s | 0x%X", pointer_as_memory_reference(p
),
2388 fbd_type
, extra_flags
);
2391 pandecode_log("};\n");
2396 static int job_descriptor_number
= 0;
2399 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
)
2401 struct mali_job_descriptor_header
*h
;
2403 int start_number
= 0;
2409 struct pandecode_mapped_memory
*mem
=
2410 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2414 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2416 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2417 * high 32-bits of the 64-bit pointer are reused to store
2420 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
2421 h
->job_type
!= JOB_TYPE_FRAGMENT
? 4 : 0;
2422 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
2424 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
,
2427 int job_no
= job_descriptor_number
++;
2430 start_number
= job_no
;
2432 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2435 pandecode_prop("job_type = %s", pandecode_job_type(h
->job_type
));
2437 /* Save for next job fixing */
2438 last_size
= h
->job_descriptor_size
;
2440 if (h
->job_descriptor_size
)
2441 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2443 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2444 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2445 h
->exception_status
,
2446 (h
->exception_status
>> 16) & 0xFFFF,
2447 pandecode_exception_access((h
->exception_status
>> 8) & 0x3),
2448 h
->exception_status
& 0xFF);
2450 if (h
->first_incomplete_task
)
2451 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2453 if (h
->fault_pointer
)
2454 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2457 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2459 pandecode_prop("job_index = %d", h
->job_index
);
2461 if (h
->unknown_flags
)
2462 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2464 if (h
->job_dependency_index_1
)
2465 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2467 if (h
->job_dependency_index_2
)
2468 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2471 pandecode_log("};\n");
2473 /* Do not touch the field yet -- decode the payload first, and
2474 * don't touch that either. This is essential for the uploads
2475 * to occur in sequence and therefore be dynamically allocated
2476 * correctly. Do note the size, however, for that related
2479 switch (h
->job_type
) {
2480 case JOB_TYPE_SET_VALUE
: {
2481 struct mali_payload_set_value
*s
= payload
;
2482 pandecode_log("struct mali_payload_set_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2484 MEMORY_PROP(s
, out
);
2485 pandecode_prop("unknown = 0x%" PRIX64
, s
->unknown
);
2487 pandecode_log("};\n");
2492 case JOB_TYPE_TILER
:
2493 case JOB_TYPE_VERTEX
:
2494 case JOB_TYPE_COMPUTE
:
2496 if (h
->job_type
== JOB_TYPE_TILER
)
2497 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
);
2499 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
);
2501 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
);
2505 case JOB_TYPE_FRAGMENT
:
2506 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
);
2513 /* Handle linkage */
2516 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no
- 1);
2519 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no
);
2521 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no
);
2526 } while ((jc_gpu_va
= h
->job_descriptor_size
? h
->next_job_64
: h
->next_job_32
));
2528 return start_number
;