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(const struct midgard_tiler_descriptor
*t
)
481 pandecode_log(".tiler = {\n");
484 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
485 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
486 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
488 MEMORY_PROP(t
, polygon_list
);
489 MEMORY_PROP(t
, polygon_list_body
);
491 MEMORY_PROP(t
, heap_start
);
493 if (t
->heap_start
== t
->heap_end
) {
494 /* Print identically to show symmetry for empty tiler heaps */
495 MEMORY_PROP(t
, heap_end
);
497 /* Points to the end of a buffer */
498 char *a
= pointer_as_memory_reference(t
->heap_end
- 1);
499 pandecode_prop("heap_end = %s + 1", a
);
503 bool nonzero_weights
= false;
505 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
506 nonzero_weights
|= t
->weights
[w
] != 0x0;
509 if (nonzero_weights
) {
510 pandecode_log(".weights = {");
512 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
513 pandecode_log("%d, ", t
->weights
[w
]);
520 pandecode_log("}\n");
524 pandecode_sfbd(uint64_t gpu_va
, int job_no
)
526 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
527 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
529 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
532 pandecode_prop("unknown1 = 0x%" PRIx32
, s
->unknown1
);
533 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->unknown2
);
535 pandecode_log(".format = ");
536 pandecode_log_decoded_flags(fb_fmt_flag_info
, s
->format
);
537 pandecode_log_cont(",\n");
539 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", s
->width
+ 1);
540 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", s
->height
+ 1);
542 MEMORY_PROP(s
, framebuffer
);
543 pandecode_prop("stride = %d", s
->stride
);
545 /* Earlier in the actual commandstream -- right before width -- but we
546 * delay to flow nicer */
548 pandecode_log(".clear_flags = ");
549 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
550 pandecode_log_cont(",\n");
552 if (s
->depth_buffer
| s
->depth_buffer_enable
) {
553 MEMORY_PROP(s
, depth_buffer
);
554 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s
->depth_buffer_enable
));
557 if (s
->stencil_buffer
| s
->stencil_buffer_enable
) {
558 MEMORY_PROP(s
, stencil_buffer
);
559 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s
->stencil_buffer_enable
));
562 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
563 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
564 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
565 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
566 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
569 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
570 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
571 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
572 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
573 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
576 if (s
->clear_stencil
) {
577 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
580 MEMORY_PROP(s
, unknown_address_0
);
581 const struct midgard_tiler_descriptor t
= s
->tiler
;
582 pandecode_midgard_tiler_descriptor(&t
);
585 pandecode_log("};\n");
587 pandecode_prop("zero0 = 0x%" PRIx64
, s
->zero0
);
588 pandecode_prop("zero1 = 0x%" PRIx64
, s
->zero1
);
589 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
590 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
592 printf(".zero3 = {");
594 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
595 printf("%X, ", s
->zero3
[i
]);
599 printf(".zero6 = {");
601 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
602 printf("%X, ", s
->zero6
[i
]);
608 pandecode_u32_slide(unsigned name
, const u32
*slide
, unsigned count
)
610 pandecode_log(".unknown%d = {", name
);
612 for (int i
= 0; i
< count
; ++i
)
613 printf("%X, ", slide
[i
]);
615 pandecode_log("},\n");
618 #define SHORT_SLIDE(num) \
619 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
622 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
624 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
625 const struct mali_compute_fbd
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
627 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
637 pandecode_swizzle(unsigned swizzle
)
639 pandecode_prop("swizzle = %s | (%s << 3) | (%s << 6) | (%s << 9)",
640 pandecode_channel((swizzle
>> 0) & 0x7),
641 pandecode_channel((swizzle
>> 3) & 0x7),
642 pandecode_channel((swizzle
>> 6) & 0x7),
643 pandecode_channel((swizzle
>> 9) & 0x7));
647 pandecode_rt_format(struct mali_rt_format format
)
649 pandecode_log(".format = {\n");
652 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
653 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
654 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
656 pandecode_prop("block = %s",
657 pandecode_mfbd_block_format(format
.block
));
659 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
660 MALI_NEGATIVE(format
.nr_channels
));
662 pandecode_log(".flags = ");
663 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
664 pandecode_log_cont(",\n");
666 pandecode_swizzle(format
.swizzle
);
668 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
671 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
674 pandecode_log("},\n");
678 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct bifrost_framebuffer
*fb
)
680 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
683 for (int i
= 0; i
< MALI_NEGATIVE(fb
->rt_count_1
); i
++) {
684 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct bifrost_render_target
);
685 struct pandecode_mapped_memory
*mem
=
686 pandecode_find_mapped_gpu_mem_containing(rt_va
);
687 const struct bifrost_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
689 pandecode_log("{\n");
692 pandecode_rt_format(rt
->format
);
694 if (rt
->format
.block
== MALI_MFBD_BLOCK_AFBC
) {
695 pandecode_log(".afbc = {\n");
698 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
699 pandecode_prop("metadata = %s", a
);
702 pandecode_prop("stride = %d", rt
->afbc
.stride
);
703 pandecode_prop("unk = 0x%" PRIx32
, rt
->afbc
.unk
);
706 pandecode_log("},\n");
708 pandecode_log(".chunknown = {\n");
711 pandecode_prop("unk = 0x%" PRIx64
, rt
->chunknown
.unk
);
713 char *a
= pointer_as_memory_reference(rt
->chunknown
.pointer
);
714 pandecode_prop("pointer = %s", a
);
718 pandecode_log("},\n");
721 MEMORY_PROP(rt
, framebuffer
);
722 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
724 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
725 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
726 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
727 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
728 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
731 if (rt
->zero1
|| rt
->zero2
|| rt
->zero3
) {
732 pandecode_msg("render target zeros tripped\n");
733 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
734 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
735 pandecode_prop("zero3 = 0x%" PRIx32
, rt
->zero3
);
739 pandecode_log("},\n");
743 pandecode_log("};\n");
747 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool with_render_targets
)
749 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
750 const struct bifrost_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
752 if (fb
->sample_locations
) {
753 /* The blob stores all possible sample locations in a single buffer
754 * allocated on startup, and just switches the pointer when switching
755 * MSAA state. For now, we just put the data into the cmdstream, but we
756 * should do something like what the blob does with a real driver.
758 * There seem to be 32 slots for sample locations, followed by another
759 * 16. The second 16 is just the center location followed by 15 zeros
760 * in all the cases I've identified (maybe shader vs. depth/color
764 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->sample_locations
);
766 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->sample_locations
);
768 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
771 for (int i
= 0; i
< 32 + 16; i
++) {
772 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
776 pandecode_log("};\n");
779 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
782 pandecode_prop("unk0 = 0x%x", fb
->unk0
);
784 if (fb
->sample_locations
)
785 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
787 /* Assume that unknown1 was emitted in the last job for
789 MEMORY_PROP(fb
, unknown1
);
791 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
792 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
793 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
794 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
796 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
797 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
798 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
799 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
801 pandecode_log(".mfbd_flags = ");
802 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
803 pandecode_log_cont(",\n");
805 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
806 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
808 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
809 MEMORY_PROP(fb
, scratchpad
);
810 const struct midgard_tiler_descriptor t
= fb
->tiler
;
811 pandecode_midgard_tiler_descriptor(&t
);
813 if (fb
->zero3
|| fb
->zero4
) {
814 pandecode_msg("framebuffer zeros tripped\n");
815 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
816 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
820 pandecode_log("};\n");
822 gpu_va
+= sizeof(struct bifrost_framebuffer
);
824 if ((fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && with_render_targets
) {
825 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
826 const struct bifrost_fb_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
828 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
831 MEMORY_PROP(fbx
, checksum
);
833 if (fbx
->checksum_stride
)
834 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
836 pandecode_log(".flags = ");
837 pandecode_log_decoded_flags(mfbd_extra_flag_info
, fbx
->flags
);
838 pandecode_log_cont(",\n");
840 if (fbx
->flags
& MALI_EXTRA_AFBC_ZS
) {
841 pandecode_log(".ds_afbc = {\n");
844 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
845 pandecode_prop("depth_stencil_afbc_stride = %d",
846 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
847 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
849 if (fbx
->ds_afbc
.zero1
|| fbx
->ds_afbc
.padding
) {
850 pandecode_msg("Depth/stencil AFBC zeros tripped\n");
851 pandecode_prop("zero1 = 0x%" PRIx32
,
853 pandecode_prop("padding = 0x%" PRIx64
,
854 fbx
->ds_afbc
.padding
);
858 pandecode_log("},\n");
860 pandecode_log(".ds_linear = {\n");
863 if (fbx
->ds_linear
.depth
) {
864 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
865 pandecode_prop("depth_stride = %d",
866 fbx
->ds_linear
.depth_stride
);
869 if (fbx
->ds_linear
.stencil
) {
870 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
871 pandecode_prop("stencil_stride = %d",
872 fbx
->ds_linear
.stencil_stride
);
875 if (fbx
->ds_linear
.depth_stride_zero
||
876 fbx
->ds_linear
.stencil_stride_zero
||
877 fbx
->ds_linear
.zero1
|| fbx
->ds_linear
.zero2
) {
878 pandecode_msg("Depth/stencil zeros tripped\n");
879 pandecode_prop("depth_stride_zero = 0x%x",
880 fbx
->ds_linear
.depth_stride_zero
);
881 pandecode_prop("stencil_stride_zero = 0x%x",
882 fbx
->ds_linear
.stencil_stride_zero
);
883 pandecode_prop("zero1 = 0x%" PRIx32
,
884 fbx
->ds_linear
.zero1
);
885 pandecode_prop("zero2 = 0x%" PRIx32
,
886 fbx
->ds_linear
.zero2
);
890 pandecode_log("},\n");
893 if (fbx
->zero3
|| fbx
->zero4
) {
894 pandecode_msg("fb_extra zeros tripped\n");
895 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
896 pandecode_prop("zero4 = 0x%" PRIx64
, fbx
->zero4
);
900 pandecode_log("};\n");
902 gpu_va
+= sizeof(struct bifrost_fb_extra
);
905 if (with_render_targets
)
906 pandecode_render_target(gpu_va
, job_no
, fb
);
908 /* Passback the render target count */
909 return MALI_NEGATIVE(fb
->rt_count_1
);
912 /* Just add a comment decoding the shift/odd fields forming the padded vertices
916 pandecode_padded_vertices(unsigned shift
, unsigned k
)
918 unsigned odd
= 2*k
+ 1;
919 unsigned pot
= 1 << shift
;
920 pandecode_msg("padded_num_vertices = %d\n", odd
* pot
);
923 /* Given a magic divisor, recover what we were trying to divide by.
925 * Let m represent the magic divisor. By definition, m is an element on Z, whre
926 * 0 <= m < 2^N, for N bits in m.
928 * Let q represent the number we would like to divide by.
930 * By definition of a magic divisor for N-bit unsigned integers (a number you
931 * multiply by to magically get division), m is a number such that:
933 * (m * x) & (2^N - 1) = floor(x/q).
934 * for all x on Z where 0 <= x < 2^N
936 * Ignore the case where any of the above values equals zero; it is irrelevant
937 * for our purposes (instanced arrays).
939 * Choose x = q. Then:
941 * (m * x) & (2^N - 1) = floor(x/q).
942 * (m * q) & (2^N - 1) = floor(q/q).
944 * floor(q/q) = floor(1) = 1, therefore:
946 * (m * q) & (2^N - 1) = 1
948 * Recall the identity that the bitwise AND of one less than a power-of-two
949 * equals the modulo with that power of two, i.e. for all x:
951 * x & (2^N - 1) = x % N
957 * By definition, a modular multiplicative inverse of a number m is the number
958 * q such that with respect to a modulos M:
962 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
967 pandecode_magic_divisor(uint32_t magic
, unsigned shift
, unsigned orig_divisor
, unsigned extra
)
970 /* Compute the modular inverse of `magic` with respect to 2^(32 -
971 * shift) the most lame way possible... just repeatedly add.
972 * Asymptoptically slow but nobody cares in practice, unless you have
973 * massive numbers of vertices or high divisors. */
975 unsigned inverse
= 0;
977 /* Magic implicitly has the highest bit set */
980 /* Depending on rounding direction */
985 uint32_t product
= magic
* inverse
;
997 pandecode_msg("dividing by %d (maybe off by two)\n", inverse
);
999 /* Recall we're supposed to divide by (gl_level_divisor *
1000 * padded_num_vertices) */
1002 unsigned padded_num_vertices
= inverse
/ orig_divisor
;
1004 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices
);
1009 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1010 mali_ptr addr
, int job_no
, char *suffix
,
1011 int count
, bool varying
)
1013 char *prefix
= varying
? "varyings" : "attributes";
1016 pandecode_msg("no %s\n", prefix
);
1020 union mali_attr
*attr
= pandecode_fetch_gpu_mem(mem
, addr
, sizeof(union mali_attr
) * count
);
1023 snprintf(base
, sizeof(base
), "%s_data_%d%s", prefix
, job_no
, suffix
);
1025 for (int i
= 0; i
< count
; ++i
) {
1026 enum mali_attr_mode mode
= attr
[i
].elements
& 7;
1028 if (mode
== MALI_ATTR_UNUSED
)
1031 mali_ptr raw_elements
= attr
[i
].elements
& ~7;
1033 /* TODO: Do we maybe want to dump the attribute values
1034 * themselves given the specified format? Or is that too hard?
1037 char *a
= pointer_as_memory_reference(raw_elements
);
1038 pandecode_log("mali_ptr %s_%d_p = %s;\n", base
, i
, a
);
1042 pandecode_log("union mali_attr %s_%d[] = {\n", prefix
, job_no
);
1045 for (int i
= 0; i
< count
; ++i
) {
1046 pandecode_log("{\n");
1049 unsigned mode
= attr
[i
].elements
& 7;
1050 pandecode_prop("elements = (%s_%d_p) | %s", base
, i
, pandecode_attr_mode(mode
));
1051 pandecode_prop("shift = %d", attr
[i
].shift
);
1052 pandecode_prop("extra_flags = %d", attr
[i
].extra_flags
);
1053 pandecode_prop("stride = 0x%" PRIx32
, attr
[i
].stride
);
1054 pandecode_prop("size = 0x%" PRIx32
, attr
[i
].size
);
1056 /* Decode further where possible */
1058 if (mode
== MALI_ATTR_MODULO
) {
1059 pandecode_padded_vertices(
1061 attr
[i
].extra_flags
);
1065 pandecode_log("}, \n");
1067 if (mode
== MALI_ATTR_NPOT_DIVIDE
) {
1069 pandecode_log("{\n");
1071 pandecode_prop("unk = 0x%x", attr
[i
].unk
);
1072 pandecode_prop("magic_divisor = 0x%08x", attr
[i
].magic_divisor
);
1073 if (attr
[i
].zero
!= 0)
1074 pandecode_prop("zero = 0x%x /* XXX zero tripped */", attr
[i
].zero
);
1075 pandecode_prop("divisor = %d", attr
[i
].divisor
);
1076 pandecode_magic_divisor(attr
[i
].magic_divisor
, attr
[i
- 1].shift
, attr
[i
].divisor
, attr
[i
- 1].extra_flags
);
1078 pandecode_log("}, \n");
1084 pandecode_log("};\n");
1088 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1090 /* TODO: Decode flags */
1091 mali_ptr shader_ptr
= ptr
& ~15;
1093 char *a
= pointer_as_memory_reference(shader_ptr
);
1094 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1101 all_zero(unsigned *buffer
, unsigned count
)
1103 for (unsigned i
= 0; i
< count
; ++i
) {
1112 pandecode_stencil(const char *name
, const struct mali_stencil_test
*stencil
)
1114 if (all_zero((unsigned *) stencil
, sizeof(stencil
) / sizeof(unsigned)))
1117 const char *func
= pandecode_func(stencil
->func
);
1118 const char *sfail
= pandecode_stencil_op(stencil
->sfail
);
1119 const char *dpfail
= pandecode_stencil_op(stencil
->dpfail
);
1120 const char *dppass
= pandecode_stencil_op(stencil
->dppass
);
1123 pandecode_msg("Stencil zero tripped: %X\n", stencil
->zero
);
1125 pandecode_log(".stencil_%s = {\n", name
);
1127 pandecode_prop("ref = %d", stencil
->ref
);
1128 pandecode_prop("mask = 0x%02X", stencil
->mask
);
1129 pandecode_prop("func = %s", func
);
1130 pandecode_prop("sfail = %s", sfail
);
1131 pandecode_prop("dpfail = %s", dpfail
);
1132 pandecode_prop("dppass = %s", dppass
);
1134 pandecode_log("},\n");
1138 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1141 pandecode_msg("Blend zero tripped: %X\n", blend
->zero1
);
1143 pandecode_log(".equation = {\n");
1146 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1147 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1149 pandecode_log(".color_mask = ");
1150 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1151 pandecode_log_cont(",\n");
1154 pandecode_log("},\n");
1157 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1160 decode_bifrost_constant(u16 constant
)
1162 float lo
= (float) (constant
& 0xFF);
1163 float hi
= (float) (constant
>> 8);
1165 return (hi
/ 255.0) + (lo
/ 65535.0);
1169 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1171 struct bifrost_blend_rt
*b
=
1172 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1174 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1177 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1178 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1179 b
->constant
, decode_bifrost_constant(b
->constant
));
1181 /* TODO figure out blend shader enable bit */
1182 pandecode_blend_equation(&b
->equation
);
1183 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1184 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1185 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1188 pandecode_log("},\n");
1194 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1196 if (all_zero((unsigned *) blend
, sizeof(blend
) / sizeof(unsigned)))
1199 pandecode_log(".blend = {\n");
1203 pandecode_shader_address("shader", blend
->shader
);
1205 pandecode_blend_equation(&blend
->equation
);
1206 pandecode_prop("constant = %f", blend
->constant
);
1210 pandecode_log("},\n");
1212 /* Return blend shader to disassemble if present */
1213 return is_shader
? (blend
->shader
& ~0xF) : 0;
1217 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1219 struct midgard_blend_rt
*b
=
1220 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1222 /* Flags determine presence of blend shader */
1223 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1225 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1228 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1230 union midgard_blend blend
= b
->blend
;
1231 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1234 pandecode_log("};\n");
1240 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1243 char *prefix
= varying
? "varying" : "attribute";
1244 unsigned max_index
= 0;
1245 snprintf(base
, sizeof(base
), "%s_meta", prefix
);
1247 pandecode_log("struct mali_attr_meta %s_%d%s[] = {\n", base
, job_no
, suffix
);
1250 struct mali_attr_meta
*attr_meta
;
1251 mali_ptr p
= varying
? (v
->varying_meta
& ~0xF) : v
->attribute_meta
;
1253 struct pandecode_mapped_memory
*attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
);
1255 for (int i
= 0; i
< count
; ++i
, p
+= sizeof(struct mali_attr_meta
)) {
1256 attr_meta
= pandecode_fetch_gpu_mem(attr_mem
, p
,
1259 pandecode_log("{\n");
1261 pandecode_prop("index = %d", attr_meta
->index
);
1263 if (attr_meta
->index
> max_index
)
1264 max_index
= attr_meta
->index
;
1265 pandecode_swizzle(attr_meta
->swizzle
);
1266 pandecode_prop("format = %s", pandecode_format(attr_meta
->format
));
1268 pandecode_prop("unknown1 = 0x%" PRIx64
, (u64
) attr_meta
->unknown1
);
1269 pandecode_prop("unknown3 = 0x%" PRIx64
, (u64
) attr_meta
->unknown3
);
1270 pandecode_prop("src_offset = %d", attr_meta
->src_offset
);
1272 pandecode_log("},\n");
1277 pandecode_log("};\n");
1279 return count
? (max_index
+ 1) : 0;
1283 pandecode_indices(uintptr_t pindices
, uint32_t index_count
, int job_no
)
1285 struct pandecode_mapped_memory
*imem
= pandecode_find_mapped_gpu_mem_containing(pindices
);
1288 /* Indices are literally just a u32 array :) */
1290 uint32_t *PANDECODE_PTR_VAR(indices
, imem
, pindices
);
1292 pandecode_log("uint32_t indices_%d[] = {\n", job_no
);
1295 for (unsigned i
= 0; i
< (index_count
+ 1); i
+= 3)
1296 pandecode_log("%d, %d, %d,\n",
1302 pandecode_log("};\n");
1306 /* return bits [lo, hi) of word */
1308 bits(u32 word
, u32 lo
, u32 hi
)
1311 return word
; // avoid undefined behavior with the shift
1313 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1317 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool noninstanced
)
1319 pandecode_log_cont("{\n");
1322 /* Decode invocation_count. See the comment before the definition of
1323 * invocation_count for an explanation.
1326 unsigned size_x
= bits(p
->invocation_count
, 0, p
->size_y_shift
) + 1;
1327 unsigned size_y
= bits(p
->invocation_count
, p
->size_y_shift
, p
->size_z_shift
) + 1;
1328 unsigned size_z
= bits(p
->invocation_count
, p
->size_z_shift
, p
->workgroups_x_shift
) + 1;
1330 unsigned groups_x
= bits(p
->invocation_count
, p
->workgroups_x_shift
, p
->workgroups_y_shift
) + 1;
1331 unsigned groups_y
= bits(p
->invocation_count
, p
->workgroups_y_shift
, p
->workgroups_z_shift
) + 1;
1332 unsigned groups_z
= bits(p
->invocation_count
, p
->workgroups_z_shift
, 32) + 1;
1334 /* Even though we have this decoded, we want to ensure that the
1335 * representation is "unique" so we don't lose anything by printing only
1336 * the final result. More specifically, we need to check that we were
1337 * passed something in canonical form, since the definition per the
1338 * hardware is inherently not unique. How? Well, take the resulting
1339 * decode and pack it ourselves! If it is bit exact with what we
1340 * decoded, we're good to go. */
1342 struct mali_vertex_tiler_prefix ref
;
1343 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, noninstanced
);
1346 (p
->invocation_count
== ref
.invocation_count
) &&
1347 (p
->size_y_shift
== ref
.size_y_shift
) &&
1348 (p
->size_z_shift
== ref
.size_z_shift
) &&
1349 (p
->workgroups_x_shift
== ref
.workgroups_x_shift
) &&
1350 (p
->workgroups_y_shift
== ref
.workgroups_y_shift
) &&
1351 (p
->workgroups_z_shift
== ref
.workgroups_z_shift
) &&
1352 (p
->workgroups_x_shift_2
== ref
.workgroups_x_shift_2
);
1355 pandecode_msg("XXX: non-canonical workgroups packing\n");
1356 pandecode_msg("expected: %X, %d, %d, %d, %d, %d\n",
1357 ref
.invocation_count
,
1360 ref
.workgroups_x_shift
,
1361 ref
.workgroups_y_shift
,
1362 ref
.workgroups_z_shift
,
1363 ref
.workgroups_x_shift_2
);
1365 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1366 pandecode_prop("size_y_shift = %d", p
->size_y_shift
);
1367 pandecode_prop("size_z_shift = %d", p
->size_z_shift
);
1368 pandecode_prop("workgroups_x_shift = %d", p
->workgroups_x_shift
);
1369 pandecode_prop("workgroups_y_shift = %d", p
->workgroups_y_shift
);
1370 pandecode_prop("workgroups_z_shift = %d", p
->workgroups_z_shift
);
1371 pandecode_prop("workgroups_x_shift_2 = %d", p
->workgroups_x_shift_2
);
1374 /* Regardless, print the decode */
1375 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1376 size_x
, size_y
, size_z
,
1377 groups_x
, groups_y
, groups_z
);
1380 if (p
->unknown_draw
)
1381 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1383 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1385 if (p
->draw_mode
!= MALI_DRAW_NONE
)
1386 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p
->draw_mode
));
1388 /* Index count only exists for tiler jobs anyway */
1391 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1393 if (p
->offset_bias_correction
)
1394 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1397 pandecode_msg("Zero tripped\n");
1398 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1402 pandecode_log("},\n");
1406 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1408 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1410 struct mali_uniform_buffer_meta
*PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1412 for (int i
= 0; i
< ubufs_count
; i
++) {
1413 mali_ptr ptr
= ubufs
[i
].ptr
<< 2;
1414 struct pandecode_mapped_memory
*umem2
= pandecode_find_mapped_gpu_mem_containing(ptr
);
1415 uint32_t *PANDECODE_PTR_VAR(ubuf
, umem2
, ptr
);
1417 snprintf(name
, sizeof(name
), "ubuf_%d", i
);
1418 /* The blob uses ubuf 0 to upload internal stuff and
1419 * uniforms that won't fit/are accessed indirectly, so
1420 * it puts it in the batchbuffer.
1422 pandecode_log("uint32_t %s_%d[] = {\n", name
, job_no
);
1425 for (int j
= 0; j
<= ubufs
[i
].size
; j
++) {
1426 for (int k
= 0; k
< 4; k
++) {
1428 pandecode_log("0x%"PRIx32
", ", ubuf
[4 * j
+ k
]);
1430 pandecode_log_cont("0x%"PRIx32
", ", ubuf
[4 * j
+ k
]);
1434 pandecode_log_cont("\n");
1438 pandecode_log("};\n");
1441 pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%"PRIx64
"_%d[] = {\n",
1445 for (int i
= 0; i
< ubufs_count
; i
++) {
1446 pandecode_log("{\n");
1448 pandecode_prop("size = MALI_POSITIVE(%d)", ubufs
[i
].size
+ 1);
1449 pandecode_prop("ptr = ubuf_%d_%d_p >> 2", i
, job_no
);
1451 pandecode_log("},\n");
1455 pandecode_log("};\n");
1459 pandecode_scratchpad(uintptr_t pscratchpad
, int job_no
, char *suffix
)
1462 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(pscratchpad
);
1464 struct bifrost_scratchpad
*PANDECODE_PTR_VAR(scratchpad
, mem
, pscratchpad
);
1466 if (scratchpad
->zero
)
1467 pandecode_msg("XXX scratchpad zero tripped");
1469 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64
"_%d%s = {\n", pscratchpad
, job_no
, suffix
);
1472 pandecode_prop("flags = 0x%x", scratchpad
->flags
);
1473 MEMORY_PROP(scratchpad
, gpu_scratchpad
);
1476 pandecode_log("};\n");
1479 static unsigned shader_id
= 0;
1482 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1483 bool is_bifrost
, unsigned nr_regs
)
1485 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1486 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1488 /* Compute maximum possible size */
1489 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1491 /* Print some boilerplate to clearly denote the assembly (which doesn't
1492 * obey indentation rules), and actually do the disassembly! */
1498 snprintf(prefix
, sizeof(prefix
) - 1, "shader%d - %s shader: ",
1500 (type
== JOB_TYPE_TILER
) ? "FRAGMENT" : "VERTEX");
1503 disassemble_bifrost(code
, sz
, false);
1505 disassemble_midgard(code
, sz
, true, nr_regs
, prefix
);
1512 pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix
*p
,
1513 int job_no
, enum mali_job_type job_type
,
1514 char *suffix
, bool is_bifrost
)
1516 mali_ptr shader_meta_ptr
= (u64
) (uintptr_t) (p
->_shader_upper
<< 4);
1517 struct pandecode_mapped_memory
*attr_mem
;
1519 unsigned rt_count
= 1;
1521 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1522 * are the only things actually needed from the FBD, vertex/tiler jobs
1523 * no longer reference the FBD -- instead, this field points to some
1524 * info about the scratchpad.
1527 pandecode_scratchpad(p
->framebuffer
& ~FBD_TYPE
, job_no
, suffix
);
1528 else if (p
->framebuffer
& MALI_MFBD
)
1529 rt_count
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->framebuffer
) & FBD_MASK
, job_no
, false);
1530 else if (job_type
== JOB_TYPE_COMPUTE
)
1531 pandecode_compute_fbd((u64
) (uintptr_t) p
->framebuffer
, job_no
);
1533 pandecode_sfbd((u64
) (uintptr_t) p
->framebuffer
, job_no
);
1535 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
1536 int texture_count
= 0, sampler_count
= 0;
1538 if (shader_meta_ptr
) {
1539 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr
);
1540 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, shader_meta_ptr
);
1542 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", shader_meta_ptr
, job_no
, suffix
);
1545 /* Save for dumps */
1546 attribute_count
= s
->attribute_count
;
1547 varying_count
= s
->varying_count
;
1548 texture_count
= s
->texture_count
;
1549 sampler_count
= s
->sampler_count
;
1552 uniform_count
= s
->bifrost2
.uniform_count
;
1553 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
1555 uniform_count
= s
->midgard1
.uniform_count
;
1556 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
1559 mali_ptr shader_ptr
= pandecode_shader_address("shader", s
->shader
);
1561 pandecode_prop("texture_count = %" PRId16
, s
->texture_count
);
1562 pandecode_prop("sampler_count = %" PRId16
, s
->sampler_count
);
1563 pandecode_prop("attribute_count = %" PRId16
, s
->attribute_count
);
1564 pandecode_prop("varying_count = %" PRId16
, s
->varying_count
);
1566 unsigned nr_registers
= 0;
1569 pandecode_log(".bifrost1 = {\n");
1572 pandecode_prop("uniform_buffer_count = %" PRId32
, s
->bifrost1
.uniform_buffer_count
);
1573 pandecode_prop("unk1 = 0x%" PRIx32
, s
->bifrost1
.unk1
);
1576 pandecode_log("},\n");
1578 pandecode_log(".midgard1 = {\n");
1581 pandecode_prop("uniform_count = %" PRId16
, s
->midgard1
.uniform_count
);
1582 pandecode_prop("uniform_buffer_count = %" PRId16
, s
->midgard1
.uniform_buffer_count
);
1583 pandecode_prop("work_count = %" PRId16
, s
->midgard1
.work_count
);
1584 nr_registers
= s
->midgard1
.work_count
;
1586 pandecode_log(".flags = ");
1587 pandecode_log_decoded_flags(shader_midgard1_flag_info
, s
->midgard1
.flags
);
1588 pandecode_log_cont(",\n");
1590 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->midgard1
.unknown2
);
1593 pandecode_log("},\n");
1596 if (s
->depth_units
|| s
->depth_factor
) {
1597 pandecode_prop("depth_factor = %f", s
->depth_factor
);
1598 pandecode_prop("depth_units = %f", s
->depth_units
);
1601 if (s
->alpha_coverage
) {
1602 bool invert_alpha_coverage
= s
->alpha_coverage
& 0xFFF0;
1603 uint16_t inverted_coverage
= invert_alpha_coverage
? ~s
->alpha_coverage
: s
->alpha_coverage
;
1605 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
1606 invert_alpha_coverage
? "~" : "",
1607 MALI_GET_ALPHA_COVERAGE(inverted_coverage
));
1610 if (s
->unknown2_3
|| s
->unknown2_4
) {
1611 pandecode_log(".unknown2_3 = ");
1613 int unknown2_3
= s
->unknown2_3
;
1614 int unknown2_4
= s
->unknown2_4
;
1616 /* We're not quite sure what these flags mean without the depth test, if anything */
1618 if (unknown2_3
& (MALI_DEPTH_TEST
| MALI_DEPTH_FUNC_MASK
)) {
1619 const char *func
= pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3
));
1620 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
1622 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
1625 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
1626 pandecode_log_cont(",\n");
1628 pandecode_log(".unknown2_4 = ");
1629 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
1630 pandecode_log_cont(",\n");
1633 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
1634 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
1635 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
1638 pandecode_stencil("front", &s
->stencil_front
);
1639 pandecode_stencil("back", &s
->stencil_back
);
1642 pandecode_log(".bifrost2 = {\n");
1645 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
1646 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
1647 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
1648 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
1651 pandecode_log("},\n");
1652 } else if (s
->midgard2
.unknown2_7
) {
1653 pandecode_log(".midgard2 = {\n");
1656 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
1658 pandecode_log("},\n");
1662 pandecode_prop("unknown2_8 = 0x%" PRIx32
, s
->unknown2_8
);
1665 /* TODO: Blend shaders routing/disasm */
1667 union midgard_blend blend
= s
->blend
;
1668 pandecode_midgard_blend(&blend
, false);
1672 pandecode_log("};\n");
1674 /* MRT blend fields are used whenever MFBD is used, with
1675 * per-RT descriptors */
1677 if (job_type
== JOB_TYPE_TILER
) {
1678 void* blend_base
= (void *) (s
+ 1);
1680 for (unsigned i
= 0; i
< rt_count
; i
++) {
1681 mali_ptr shader
= 0;
1684 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
1686 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
1689 pandecode_shader_disassemble(shader
, job_no
, job_type
, false, 0);
1693 if (shader_ptr
& ~0xF)
1694 pandecode_shader_disassemble(shader_ptr
, job_no
, job_type
, is_bifrost
, nr_registers
);
1696 pandecode_msg("<no shader>\n");
1699 struct pandecode_mapped_memory
*fmem
= pandecode_find_mapped_gpu_mem_containing(p
->viewport
);
1700 struct mali_viewport
*PANDECODE_PTR_VAR(f
, fmem
, p
->viewport
);
1702 pandecode_log("struct mali_viewport viewport_%"PRIx64
"_%d%s = {\n", p
->viewport
, job_no
, suffix
);
1705 pandecode_prop("clip_minx = %f", f
->clip_minx
);
1706 pandecode_prop("clip_miny = %f", f
->clip_miny
);
1707 pandecode_prop("clip_minz = %f", f
->clip_minz
);
1708 pandecode_prop("clip_maxx = %f", f
->clip_maxx
);
1709 pandecode_prop("clip_maxy = %f", f
->clip_maxy
);
1710 pandecode_prop("clip_maxz = %f", f
->clip_maxz
);
1712 /* Only the higher coordinates are MALI_POSITIVE scaled */
1714 pandecode_prop("viewport0 = { %d, %d }",
1715 f
->viewport0
[0], f
->viewport0
[1]);
1717 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
1718 f
->viewport1
[0] + 1, f
->viewport1
[1] + 1);
1721 pandecode_log("};\n");
1724 if (p
->attribute_meta
) {
1725 unsigned max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
1727 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
1728 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
+ 1, false);
1731 /* Varyings are encoded like attributes but not actually sent; we just
1732 * pass a zero buffer with the right stride/size set, (or whatever)
1733 * since the GPU will write to it itself */
1735 if (p
->varying_meta
) {
1736 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
1740 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
1742 /* Number of descriptors depends on whether there are
1743 * non-internal varyings */
1745 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true);
1748 bool is_compute
= job_type
== JOB_TYPE_COMPUTE
;
1750 if (p
->uniforms
&& !is_compute
) {
1751 int rows
= uniform_count
, width
= 4;
1752 size_t sz
= rows
* width
* sizeof(float);
1754 struct pandecode_mapped_memory
*uniform_mem
= pandecode_find_mapped_gpu_mem_containing(p
->uniforms
);
1755 pandecode_fetch_gpu_mem(uniform_mem
, p
->uniforms
, sz
);
1756 u32
*PANDECODE_PTR_VAR(uniforms
, uniform_mem
, p
->uniforms
);
1758 pandecode_log("u32 uniforms_%d%s[] = {\n", job_no
, suffix
);
1762 for (int row
= 0; row
< rows
; row
++) {
1763 for (int i
= 0; i
< width
; i
++) {
1764 u32 v
= uniforms
[i
];
1766 memcpy(&f
, &v
, sizeof(v
));
1767 pandecode_log_cont("%X /* %f */, ", v
, f
);
1770 pandecode_log_cont("\n");
1776 pandecode_log("};\n");
1777 } else if (p
->uniforms
) {
1778 int rows
= uniform_count
* 2;
1779 size_t sz
= rows
* sizeof(mali_ptr
);
1781 struct pandecode_mapped_memory
*uniform_mem
= pandecode_find_mapped_gpu_mem_containing(p
->uniforms
);
1782 pandecode_fetch_gpu_mem(uniform_mem
, p
->uniforms
, sz
);
1783 mali_ptr
*PANDECODE_PTR_VAR(uniforms
, uniform_mem
, p
->uniforms
);
1785 pandecode_log("mali_ptr uniforms_%d%s[] = {\n", job_no
, suffix
);
1789 for (int row
= 0; row
< rows
; row
++) {
1790 char *a
= pointer_as_memory_reference(uniforms
[row
]);
1791 pandecode_log("%s,\n", a
);
1796 pandecode_log("};\n");
1800 if (p
->uniform_buffers
) {
1801 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
1804 if (p
->texture_trampoline
) {
1805 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(p
->texture_trampoline
);
1808 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
);
1810 pandecode_log("uint64_t texture_trampoline_%"PRIx64
"_%d[] = {\n", p
->texture_trampoline
, job_no
);
1813 for (int tex
= 0; tex
< texture_count
; ++tex
) {
1814 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
1815 char *a
= pointer_as_memory_reference(*u
);
1816 pandecode_log("%s,\n", a
);
1821 pandecode_log("};\n");
1823 /* Now, finally, descend down into the texture descriptor */
1824 for (int tex
= 0; tex
< texture_count
; ++tex
) {
1825 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
1826 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
1829 struct mali_texture_descriptor
*PANDECODE_PTR_VAR(t
, tmem
, *u
);
1831 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64
"_%d_%d = {\n", *u
, job_no
, tex
);
1834 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", t
->width
+ 1);
1835 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", t
->height
+ 1);
1836 pandecode_prop("depth = MALI_POSITIVE(%" PRId16
")", t
->depth
+ 1);
1837 pandecode_prop("array_size = MALI_POSITIVE(%" PRId16
")", t
->array_size
+ 1);
1838 pandecode_prop("unknown3 = %" PRId16
, t
->unknown3
);
1839 pandecode_prop("unknown3A = %" PRId8
, t
->unknown3A
);
1840 pandecode_prop("nr_mipmap_levels = %" PRId8
, t
->nr_mipmap_levels
);
1842 struct mali_texture_format f
= t
->format
;
1844 pandecode_log(".format = {\n");
1847 pandecode_swizzle(f
.swizzle
);
1848 pandecode_prop("format = %s", pandecode_format(f
.format
));
1849 pandecode_prop("type = %s", pandecode_texture_type(f
.type
));
1850 pandecode_prop("srgb = %" PRId32
, f
.srgb
);
1851 pandecode_prop("unknown1 = %" PRId32
, f
.unknown1
);
1852 pandecode_prop("usage2 = 0x%" PRIx32
, f
.usage2
);
1855 pandecode_log("},\n");
1857 pandecode_swizzle(t
->swizzle
);
1859 if (t
->swizzle_zero
) {
1860 /* Shouldn't happen */
1861 pandecode_msg("Swizzle zero tripped but replay will be fine anyway");
1862 pandecode_prop("swizzle_zero = %d", t
->swizzle_zero
);
1865 pandecode_prop("unknown3 = 0x%" PRIx32
, t
->unknown3
);
1867 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
1868 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
1869 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
1871 pandecode_log(".payload = {\n");
1874 /* A bunch of bitmap pointers follow.
1875 * We work out the correct number,
1876 * based on the mipmap/cubemap
1877 * properties, but dump extra
1878 * possibilities to futureproof */
1880 int bitmap_count
= MALI_NEGATIVE(t
->nr_mipmap_levels
);
1881 bool manual_stride
= f
.usage2
& MALI_TEX_MANUAL_STRIDE
;
1883 /* Miptree for each face */
1884 if (f
.type
== MALI_TEX_CUBE
)
1887 /* Array of textures */
1888 bitmap_count
*= MALI_NEGATIVE(t
->array_size
);
1890 /* Stride for each element */
1894 /* Sanity check the size */
1895 int max_count
= sizeof(t
->payload
) / sizeof(t
->payload
[0]);
1896 assert (bitmap_count
<= max_count
);
1898 /* Dump more to be safe, but not _that_ much more */
1899 int safe_count
= MIN2(bitmap_count
* 2, max_count
);
1901 for (int i
= 0; i
< safe_count
; ++i
) {
1902 char *prefix
= (i
>= bitmap_count
) ? "// " : "";
1904 /* How we dump depends if this is a stride or a pointer */
1906 if ((f
.usage2
& MALI_TEX_MANUAL_STRIDE
) && (i
& 1)) {
1907 /* signed 32-bit snuck in as a 64-bit pointer */
1908 uint64_t stride_set
= t
->payload
[i
];
1909 uint32_t clamped_stride
= stride_set
;
1910 int32_t stride
= clamped_stride
;
1911 assert(stride_set
== clamped_stride
);
1912 pandecode_log("%s(mali_ptr) %d /* stride */, \n", prefix
, stride
);
1914 char *a
= pointer_as_memory_reference(t
->payload
[i
]);
1915 pandecode_log("%s%s, \n", prefix
, a
);
1921 pandecode_log("},\n");
1924 pandecode_log("};\n");
1930 if (p
->sampler_descriptor
) {
1931 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->sampler_descriptor
);
1934 struct mali_sampler_descriptor
*s
;
1936 mali_ptr d
= p
->sampler_descriptor
;
1938 for (int i
= 0; i
< sampler_count
; ++i
) {
1939 s
= pandecode_fetch_gpu_mem(smem
, d
+ sizeof(*s
) * i
, sizeof(*s
));
1941 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", d
+ sizeof(*s
) * i
, job_no
, i
);
1944 pandecode_log(".filter_mode = ");
1945 pandecode_log_decoded_flags(sampler_flag_info
, s
->filter_mode
);
1946 pandecode_log_cont(",\n");
1948 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
1949 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
1951 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s
->wrap_s
));
1952 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s
->wrap_t
));
1953 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s
->wrap_r
));
1955 pandecode_prop("compare_func = %s", pandecode_alt_func(s
->compare_func
));
1957 if (s
->zero
|| s
->zero2
) {
1958 pandecode_msg("Zero tripped\n");
1959 pandecode_prop("zero = 0x%X, 0x%X\n", s
->zero
, s
->zero2
);
1962 pandecode_prop("seamless_cube_map = %d", s
->seamless_cube_map
);
1964 pandecode_prop("border_color = { %f, %f, %f, %f }",
1968 s
->border_color
[3]);
1971 pandecode_log("};\n");
1978 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
1980 if (!(p
->position_varying
|| p
->occlusion_counter
|| p
->flags
))
1983 pandecode_log(".postfix = {\n");
1986 MEMORY_PROP(p
, position_varying
);
1987 MEMORY_PROP(p
, occlusion_counter
);
1990 pandecode_prop("flags = %d", p
->flags
);
1993 pandecode_log("},\n");
1997 pandecode_vertex_only_bfr(struct bifrost_vertex_only
*v
)
1999 pandecode_log_cont("{\n");
2002 pandecode_prop("unk2 = 0x%x", v
->unk2
);
2004 if (v
->zero0
|| v
->zero1
) {
2005 pandecode_msg("vertex only zero tripped");
2006 pandecode_prop("zero0 = 0x%" PRIx32
, v
->zero0
);
2007 pandecode_prop("zero1 = 0x%" PRIx64
, v
->zero1
);
2011 pandecode_log("}\n");
2015 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2018 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2019 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2021 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no
);
2025 pandecode_msg("tiler heap zero tripped\n");
2026 pandecode_prop("zero = 0x%x", h
->zero
);
2029 for (int i
= 0; i
< 12; i
++) {
2030 if (h
->zeros
[i
] != 0) {
2031 pandecode_msg("tiler heap zero %d tripped, value %x\n",
2036 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2037 MEMORY_PROP(h
, tiler_heap_start
);
2038 MEMORY_PROP(h
, tiler_heap_free
);
2040 /* this might point to the beginning of another buffer, when it's
2041 * really the end of the tiler heap buffer, so we have to be careful
2042 * here. but for zero length, we need the same pointer.
2045 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2046 MEMORY_PROP(h
, tiler_heap_start
);
2048 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2049 pandecode_prop("tiler_heap_end = %s + 1", a
);
2054 pandecode_log("};\n");
2058 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2060 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2061 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2063 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2065 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no
);
2068 if (t
->zero0
|| t
->zero1
) {
2069 pandecode_msg("tiler meta zero tripped");
2070 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2071 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2074 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2075 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2077 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2078 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2080 for (int i
= 0; i
< 12; i
++) {
2081 if (t
->zeros
[i
] != 0) {
2082 pandecode_msg("tiler heap zero %d tripped, value %" PRIx64
"\n",
2088 pandecode_log("};\n");
2092 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2094 pandecode_log(".gl_enables = ");
2096 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2098 pandecode_log_cont(",\n");
2102 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2104 if (u
.pointer
== 0x0)
2107 pandecode_log(".primitive_size = {\n");
2111 pandecode_prop("constant = %f", u
.constant
);
2113 MEMORY_PROP((&u
), pointer
);
2117 pandecode_log("},\n");
2121 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2123 pandecode_log_cont("{\n");
2126 /* TODO: gl_PointSize on Bifrost */
2127 pandecode_primitive_size(t
->primitive_size
, true);
2129 pandecode_gl_enables(t
->gl_enables
, JOB_TYPE_TILER
);
2131 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2132 || t
->zero6
|| t
->zero7
|| t
->zero8
) {
2133 pandecode_msg("tiler only zero tripped");
2134 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2135 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2136 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2137 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2138 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2139 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2140 pandecode_prop("zero7 = 0x%" PRIx32
, t
->zero7
);
2141 pandecode_prop("zero8 = 0x%" PRIx64
, t
->zero8
);
2145 pandecode_log("},\n");
2149 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2150 const struct pandecode_mapped_memory
*mem
,
2151 mali_ptr payload
, int job_no
)
2153 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2155 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true);
2157 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no
);
2160 pandecode_log(".prefix = ");
2161 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2163 pandecode_log(".vertex = ");
2164 pandecode_vertex_only_bfr(&v
->vertex
);
2166 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2169 pandecode_log("};\n");
2175 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2176 const struct pandecode_mapped_memory
*mem
,
2177 mali_ptr payload
, int job_no
)
2179 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2181 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true);
2183 pandecode_indices(t
->prefix
.indices
, t
->prefix
.index_count
, job_no
);
2184 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2186 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no
);
2189 pandecode_log(".prefix = ");
2190 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2192 pandecode_log(".tiler = ");
2193 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2195 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2198 pandecode_log("};\n");
2204 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2205 const struct pandecode_mapped_memory
*mem
,
2206 mali_ptr payload
, int job_no
)
2208 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2210 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false);
2212 pandecode_indices(v
->prefix
.indices
, v
->prefix
.index_count
, job_no
);
2214 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2217 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2218 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2220 bool instanced
= v
->instance_shift
|| v
->instance_odd
;
2221 bool is_graphics
= (h
->job_type
== JOB_TYPE_VERTEX
) || (h
->job_type
== JOB_TYPE_TILER
);
2223 pandecode_log(".prefix = ");
2224 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, !instanced
&& is_graphics
);
2226 pandecode_gl_enables(v
->gl_enables
, h
->job_type
);
2228 if (v
->instance_shift
|| v
->instance_odd
) {
2229 pandecode_prop("instance_shift = 0x%d /* %d */",
2230 v
->instance_shift
, 1 << v
->instance_shift
);
2231 pandecode_prop("instance_odd = 0x%X /* %d */",
2232 v
->instance_odd
, (2 * v
->instance_odd
) + 1);
2234 pandecode_padded_vertices(v
->instance_shift
, v
->instance_odd
);
2237 if (v
->offset_start
)
2238 pandecode_prop("offset_start = %d", v
->offset_start
);
2241 pandecode_msg("Zero tripped\n");
2242 pandecode_prop("zero5 = 0x%" PRIx64
, v
->zero5
);
2245 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2248 pandecode_log("};\n");
2254 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2255 mali_ptr payload
, int job_no
,
2258 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2260 bool fbd_dumped
= false;
2262 if (!is_bifrost
&& (s
->framebuffer
& FBD_TYPE
) == MALI_SFBD
) {
2263 /* Only SFBDs are understood, not MFBDs. We're speculating,
2264 * based on the versioning, kernel code, etc, that the
2265 * difference is between Single FrameBuffer Descriptor and
2266 * Multiple FrmaeBuffer Descriptor; the change apparently lines
2267 * up with multi-framebuffer support being added (T7xx onwards,
2268 * including Gxx). In any event, there's some field shuffling
2269 * that we haven't looked into yet. */
2271 pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
);
2273 } else if ((s
->framebuffer
& FBD_TYPE
) == MALI_MFBD
) {
2274 /* We don't know if Bifrost supports SFBD's at all, since the
2275 * driver never uses them. And the format is different from
2276 * Midgard anyways, due to the tiler heap and scratchpad being
2277 * moved out into separate structures, so it's not clear what a
2278 * Bifrost SFBD would even look like without getting an actual
2279 * trace, which appears impossible.
2282 pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true);
2286 uintptr_t p
= (uintptr_t) s
->framebuffer
& FBD_MASK
;
2287 pandecode_log("struct mali_payload_fragment payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2290 /* See the comments by the macro definitions for mathematical context
2291 * on why this is so weird */
2293 if (MALI_TILE_COORD_FLAGS(s
->max_tile_coord
) || MALI_TILE_COORD_FLAGS(s
->min_tile_coord
))
2294 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
2296 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
2297 MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
,
2298 MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
);
2300 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
2301 (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
,
2302 (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
);
2304 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2305 * we have to fallback on offsets. */
2307 const char *fbd_type
= s
->framebuffer
& MALI_MFBD
? "MALI_MFBD" : "MALI_SFBD";
2310 unsigned extra_flags
= (s
->framebuffer
& ~FBD_MASK
) & ~MALI_MFBD
;
2313 pandecode_prop("framebuffer = framebuffer_%d_p | %s | 0x%X", job_no
,
2314 fbd_type
, extra_flags
);
2316 pandecode_prop("framebuffer = %s | %s | 0x%X", pointer_as_memory_reference(p
),
2317 fbd_type
, extra_flags
);
2320 pandecode_log("};\n");
2325 static int job_descriptor_number
= 0;
2328 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
)
2330 struct mali_job_descriptor_header
*h
;
2332 int start_number
= 0;
2338 struct pandecode_mapped_memory
*mem
=
2339 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2343 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2345 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2346 * high 32-bits of the 64-bit pointer are reused to store
2349 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
2350 h
->job_type
!= JOB_TYPE_FRAGMENT
? 4 : 0;
2351 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
2353 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
,
2356 int job_no
= job_descriptor_number
++;
2359 start_number
= job_no
;
2361 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2364 pandecode_prop("job_type = %s", pandecode_job_type(h
->job_type
));
2366 /* Save for next job fixing */
2367 last_size
= h
->job_descriptor_size
;
2369 if (h
->job_descriptor_size
)
2370 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2372 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2373 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2374 h
->exception_status
,
2375 (h
->exception_status
>> 16) & 0xFFFF,
2376 pandecode_exception_access((h
->exception_status
>> 8) & 0x3),
2377 h
->exception_status
& 0xFF);
2379 if (h
->first_incomplete_task
)
2380 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2382 if (h
->fault_pointer
)
2383 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2386 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2388 pandecode_prop("job_index = %d", h
->job_index
);
2390 if (h
->unknown_flags
)
2391 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2393 if (h
->job_dependency_index_1
)
2394 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2396 if (h
->job_dependency_index_2
)
2397 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2400 pandecode_log("};\n");
2402 /* Do not touch the field yet -- decode the payload first, and
2403 * don't touch that either. This is essential for the uploads
2404 * to occur in sequence and therefore be dynamically allocated
2405 * correctly. Do note the size, however, for that related
2408 switch (h
->job_type
) {
2409 case JOB_TYPE_SET_VALUE
: {
2410 struct mali_payload_set_value
*s
= payload
;
2411 pandecode_log("struct mali_payload_set_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2413 MEMORY_PROP(s
, out
);
2414 pandecode_prop("unknown = 0x%" PRIX64
, s
->unknown
);
2416 pandecode_log("};\n");
2421 case JOB_TYPE_TILER
:
2422 case JOB_TYPE_VERTEX
:
2423 case JOB_TYPE_COMPUTE
:
2425 if (h
->job_type
== JOB_TYPE_TILER
)
2426 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
);
2428 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
);
2430 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
);
2434 case JOB_TYPE_FRAGMENT
:
2435 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
);
2442 /* Handle linkage */
2445 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no
- 1);
2448 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no
);
2450 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no
);
2455 } while ((jc_gpu_va
= h
->job_descriptor_size
? h
->next_job_64
: h
->next_job_32
));
2457 return start_number
;