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 /* To check for memory safety issues, validates that the given pointer in GPU
119 * memory is valid, containing at least sz bytes. The goal is to eliminate
120 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
121 * overruns) by statically validating pointers.
125 pandecode_validate_buffer(mali_ptr addr
, size_t sz
)
128 pandecode_msg("XXX: null pointer deref");
134 struct pandecode_mapped_memory
*bo
=
135 pandecode_find_mapped_gpu_mem_containing(addr
);
138 pandecode_msg("XXX: invalid memory dereference\n");
144 unsigned offset
= addr
- bo
->gpu_va
;
145 unsigned total
= offset
+ sz
;
147 if (total
> bo
->length
) {
148 pandecode_msg("XXX: buffer overrun."
149 "Chunk of size %d at offset %d in buffer of size %d. "
150 "Overrun by %d bytes.",
151 sz
, offset
, bo
->length
, total
- bo
->length
);
156 struct pandecode_flag_info
{
162 pandecode_log_decoded_flags(const struct pandecode_flag_info
*flag_info
,
165 bool decodable_flags_found
= false;
167 for (int i
= 0; flag_info
[i
].name
; i
++) {
168 if ((flags
& flag_info
[i
].flag
) != flag_info
[i
].flag
)
171 if (!decodable_flags_found
) {
172 decodable_flags_found
= true;
174 pandecode_log_cont(" | ");
177 pandecode_log_cont("%s", flag_info
[i
].name
);
179 flags
&= ~flag_info
[i
].flag
;
182 if (decodable_flags_found
) {
184 pandecode_log_cont(" | 0x%" PRIx64
, flags
);
186 pandecode_log_cont("0x%" PRIx64
, flags
);
190 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
191 static const struct pandecode_flag_info gl_enable_flag_info
[] = {
192 FLAG_INFO(OCCLUSION_QUERY
),
193 FLAG_INFO(OCCLUSION_PRECISE
),
194 FLAG_INFO(FRONT_CCW_TOP
),
195 FLAG_INFO(CULL_FACE_FRONT
),
196 FLAG_INFO(CULL_FACE_BACK
),
201 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
202 static const struct pandecode_flag_info clear_flag_info
[] = {
205 FLAG_INFO(SLOW_STENCIL
),
210 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
211 static const struct pandecode_flag_info mask_flag_info
[] = {
220 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
221 static const struct pandecode_flag_info u3_flag_info
[] = {
223 FLAG_INFO(CAN_DISCARD
),
224 FLAG_INFO(HAS_BLEND_SHADER
),
225 FLAG_INFO(DEPTH_TEST
),
229 static const struct pandecode_flag_info u4_flag_info
[] = {
231 FLAG_INFO(NO_DITHER
),
232 FLAG_INFO(DEPTH_RANGE_A
),
233 FLAG_INFO(DEPTH_RANGE_B
),
234 FLAG_INFO(STENCIL_TEST
),
235 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER
),
240 #define FLAG_INFO(flag) { MALI_FRAMEBUFFER_##flag, "MALI_FRAMEBUFFER_" #flag }
241 static const struct pandecode_flag_info fb_fmt_flag_info
[] = {
249 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
250 static const struct pandecode_flag_info mfbd_fmt_flag_info
[] = {
257 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
258 static const struct pandecode_flag_info mfbd_extra_flag_info
[] = {
266 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
267 static const struct pandecode_flag_info shader_midgard1_flag_info
[] = {
269 FLAG_INFO(HELPER_INVOCATIONS
),
270 FLAG_INFO(READS_TILEBUFFER
),
276 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
277 static const struct pandecode_flag_info mfbd_flag_info
[] = {
278 FLAG_INFO(DEPTH_WRITE
),
284 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
285 static const struct pandecode_flag_info sampler_flag_info
[] = {
286 FLAG_INFO(MAG_NEAREST
),
287 FLAG_INFO(MIN_NEAREST
),
288 FLAG_INFO(MIP_LINEAR_1
),
289 FLAG_INFO(MIP_LINEAR_2
),
290 FLAG_INFO(NORM_COORDS
),
295 extern char *replace_fragment
;
296 extern char *replace_vertex
;
299 pandecode_job_type(enum mali_job_type type
)
301 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
305 DEFINE_CASE(SET_VALUE
);
306 DEFINE_CASE(CACHE_FLUSH
);
307 DEFINE_CASE(COMPUTE
);
311 DEFINE_CASE(FRAGMENT
);
313 case JOB_NOT_STARTED
:
314 return "NOT_STARTED";
317 pandecode_log("Warning! Unknown job type %x\n", type
);
325 pandecode_draw_mode(enum mali_draw_mode mode
)
327 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
330 DEFINE_CASE(DRAW_NONE
);
333 DEFINE_CASE(TRIANGLES
);
334 DEFINE_CASE(TRIANGLE_STRIP
);
335 DEFINE_CASE(TRIANGLE_FAN
);
336 DEFINE_CASE(LINE_STRIP
);
337 DEFINE_CASE(LINE_LOOP
);
338 DEFINE_CASE(POLYGON
);
340 DEFINE_CASE(QUAD_STRIP
);
343 pandecode_msg("XXX: invalid draw mode %X\n", mode
);
350 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
352 pandecode_func(enum mali_func mode
)
359 DEFINE_CASE(GREATER
);
360 DEFINE_CASE(NOTEQUAL
);
365 pandecode_msg("XXX: invalid func %X\n", mode
);
371 /* Why is this duplicated? Who knows... */
372 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
374 pandecode_alt_func(enum mali_alt_func mode
)
381 DEFINE_CASE(GREATER
);
382 DEFINE_CASE(NOTEQUAL
);
387 pandecode_msg("XXX: invalid alt func %X\n", mode
);
393 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
395 pandecode_stencil_op(enum mali_stencil_op op
)
399 DEFINE_CASE(REPLACE
);
402 DEFINE_CASE(INCR_WRAP
);
403 DEFINE_CASE(DECR_WRAP
);
408 pandecode_msg("XXX: invalid stencil op %X\n", op
);
415 #define DEFINE_CASE(name) case MALI_ATTR_ ## name: return "MALI_ATTR_" #name
416 static char *pandecode_attr_mode(enum mali_attr_mode mode
)
421 DEFINE_CASE(POT_DIVIDE
);
423 DEFINE_CASE(NPOT_DIVIDE
);
425 DEFINE_CASE(INTERNAL
);
427 pandecode_msg("XXX: invalid attribute mode %X\n", mode
);
434 #define DEFINE_CASE(name) case MALI_CHANNEL_## name: return "MALI_CHANNEL_" #name
436 pandecode_channel(enum mali_channel channel
)
445 DEFINE_CASE(RESERVED_0
);
446 DEFINE_CASE(RESERVED_1
);
449 pandecode_msg("XXX: invalid channel %X\n", channel
);
455 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
457 pandecode_wrap_mode(enum mali_wrap_mode op
)
461 DEFINE_CASE(CLAMP_TO_EDGE
);
462 DEFINE_CASE(CLAMP_TO_BORDER
);
463 DEFINE_CASE(MIRRORED_REPEAT
);
466 pandecode_msg("XXX: invalid wrap mode %X\n", op
);
472 #define DEFINE_CASE(name) case MALI_TEX_## name: return "MALI_TEX_" #name
474 pandecode_texture_type(enum mali_texture_type type
)
483 unreachable("Unknown case");
488 #define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
490 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt
)
494 DEFINE_CASE(UNKNOWN
);
499 unreachable("Invalid case");
504 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
506 pandecode_exception_access(enum mali_exception_access access
)
510 DEFINE_CASE(EXECUTE
);
515 unreachable("Invalid case");
520 /* Midgard's tiler descriptor is embedded within the
524 pandecode_midgard_tiler_descriptor(
525 const struct midgard_tiler_descriptor
*t
,
530 pandecode_log(".tiler = {\n");
533 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
534 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
536 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
538 /* We know this name from the kernel, but we never see it nonzero */
540 pandecode_prop("flags = 0x%" PRIx16
" /* XXX: unexpected */", t
->flags
);
542 MEMORY_PROP(t
, polygon_list
);
544 /* The body is offset from the base of the polygon list */
545 assert(t
->polygon_list_body
> t
->polygon_list
);
546 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
548 /* It needs to fit inside the reported size */
549 assert(t
->polygon_list_size
>= body_offset
);
551 /* Check that we fit */
552 struct pandecode_mapped_memory
*plist
=
553 pandecode_find_mapped_gpu_mem_containing(t
->polygon_list
);
555 assert(t
->polygon_list_size
<= plist
->length
);
557 /* Now that we've sanity checked, we'll try to calculate the sizes
558 * ourselves for comparison */
560 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
);
561 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
);
563 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
564 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
565 ref_header
, ref_size
);
566 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
567 pandecode_msg("body offset %d\n", body_offset
);
570 /* The tiler heap has a start and end specified -- it should be
571 * identical to what we have in the BO. The exception is if tiling is
574 MEMORY_PROP(t
, heap_start
);
575 assert(t
->heap_end
>= t
->heap_start
);
577 struct pandecode_mapped_memory
*heap
=
578 pandecode_find_mapped_gpu_mem_containing(t
->heap_start
);
580 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
582 /* Tiling is enabled with a special flag */
583 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
584 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
586 bool tiling_enabled
= hierarchy_mask
;
588 if (tiling_enabled
) {
589 /* When tiling is enabled, the heap should be a tight fit */
590 unsigned heap_offset
= t
->heap_start
- heap
->gpu_va
;
591 if ((heap_offset
+ heap_size
) != heap
->length
) {
592 pandecode_msg("XXX: heap size %d (expected %d)\n",
593 heap_size
, heap
->length
- heap_offset
);
596 /* We should also have no other flags */
598 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
600 /* When tiling is disabled, we should have that flag and no others */
602 if (tiler_flags
!= MALI_TILER_DISABLED
) {
603 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
607 /* We should also have an empty heap */
609 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
613 /* Disabled tiling is used only for clear-only jobs, which are
614 * purely FRAGMENT, so we should never see this for
615 * non-FRAGMENT descriptors. */
618 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
621 /* We've never seen weights used in practice, but we know from the
622 * kernel these fields is there */
624 bool nonzero_weights
= false;
626 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
627 nonzero_weights
|= t
->weights
[w
] != 0x0;
630 if (nonzero_weights
) {
631 pandecode_log(".weights = {");
633 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
634 pandecode_log("%d, ", t
->weights
[w
]);
641 pandecode_log("}\n");
645 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
)
647 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
648 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
650 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
653 pandecode_prop("unknown1 = 0x%" PRIx32
, s
->unknown1
);
654 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->unknown2
);
656 pandecode_log(".format = ");
657 pandecode_log_decoded_flags(fb_fmt_flag_info
, s
->format
);
658 pandecode_log_cont(",\n");
660 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", s
->width
+ 1);
661 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", s
->height
+ 1);
663 MEMORY_PROP(s
, framebuffer
);
664 pandecode_prop("stride = %d", s
->stride
);
666 /* Earlier in the actual commandstream -- right before width -- but we
667 * delay to flow nicer */
669 pandecode_log(".clear_flags = ");
670 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
671 pandecode_log_cont(",\n");
673 if (s
->depth_buffer
| s
->depth_buffer_enable
) {
674 MEMORY_PROP(s
, depth_buffer
);
675 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s
->depth_buffer_enable
));
678 if (s
->stencil_buffer
| s
->stencil_buffer_enable
) {
679 MEMORY_PROP(s
, stencil_buffer
);
680 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s
->stencil_buffer_enable
));
683 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
684 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
685 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
686 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
687 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
690 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
691 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
692 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
693 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
694 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
697 if (s
->clear_stencil
) {
698 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
701 MEMORY_PROP(s
, unknown_address_0
);
702 const struct midgard_tiler_descriptor t
= s
->tiler
;
703 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
);
706 pandecode_log("};\n");
708 pandecode_prop("zero0 = 0x%" PRIx64
, s
->zero0
);
709 pandecode_prop("zero1 = 0x%" PRIx64
, s
->zero1
);
710 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
711 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
713 printf(".zero3 = {");
715 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
716 printf("%X, ", s
->zero3
[i
]);
720 printf(".zero6 = {");
722 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
723 printf("%X, ", s
->zero6
[i
]);
729 pandecode_u32_slide(unsigned name
, const u32
*slide
, unsigned count
)
731 pandecode_log(".unknown%d = {", name
);
733 for (int i
= 0; i
< count
; ++i
)
734 printf("%X, ", slide
[i
]);
736 pandecode_log("},\n");
739 #define SHORT_SLIDE(num) \
740 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
743 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
745 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
746 const struct mali_compute_fbd
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
748 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
758 pandecode_swizzle(unsigned swizzle
)
760 pandecode_prop("swizzle = %s | (%s << 3) | (%s << 6) | (%s << 9)",
761 pandecode_channel((swizzle
>> 0) & 0x7),
762 pandecode_channel((swizzle
>> 3) & 0x7),
763 pandecode_channel((swizzle
>> 6) & 0x7),
764 pandecode_channel((swizzle
>> 9) & 0x7));
768 pandecode_rt_format(struct mali_rt_format format
)
770 pandecode_log(".format = {\n");
773 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
774 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
775 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
777 pandecode_prop("block = %s",
778 pandecode_mfbd_block_format(format
.block
));
780 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
781 MALI_NEGATIVE(format
.nr_channels
));
783 pandecode_log(".flags = ");
784 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
785 pandecode_log_cont(",\n");
787 pandecode_swizzle(format
.swizzle
);
789 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
790 * which is a faster hardware-based alternative to the wallpaper method
791 * to preserve framebuffer contents across frames. In practice, MFBD
792 * preload is buggy on Midgard, and so this is a chicken bit. If this
793 * bit isn't set, most likely something broke unrelated to preload */
795 if (!format
.no_preload
) {
796 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
797 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
801 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
804 pandecode_log("},\n");
808 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct bifrost_framebuffer
*fb
)
810 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
813 for (int i
= 0; i
< MALI_NEGATIVE(fb
->rt_count_1
); i
++) {
814 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct bifrost_render_target
);
815 struct pandecode_mapped_memory
*mem
=
816 pandecode_find_mapped_gpu_mem_containing(rt_va
);
817 const struct bifrost_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
819 pandecode_log("{\n");
822 pandecode_rt_format(rt
->format
);
824 if (rt
->format
.block
== MALI_MFBD_BLOCK_AFBC
) {
825 pandecode_log(".afbc = {\n");
828 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
829 pandecode_prop("metadata = %s", a
);
832 pandecode_prop("stride = %d", rt
->afbc
.stride
);
833 pandecode_prop("unk = 0x%" PRIx32
, rt
->afbc
.unk
);
836 pandecode_log("},\n");
837 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.unk
) {
838 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
844 MEMORY_PROP(rt
, framebuffer
);
845 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
847 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
848 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
849 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
850 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
851 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
854 if (rt
->zero1
|| rt
->zero2
|| rt
->zero3
) {
855 pandecode_msg("XXX: render target zeros tripped\n");
856 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
857 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
858 pandecode_prop("zero3 = 0x%" PRIx32
, rt
->zero3
);
862 pandecode_log("},\n");
866 pandecode_log("};\n");
870 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
)
872 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
873 const struct bifrost_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
875 if (fb
->sample_locations
) {
876 /* The blob stores all possible sample locations in a single buffer
877 * allocated on startup, and just switches the pointer when switching
878 * MSAA state. For now, we just put the data into the cmdstream, but we
879 * should do something like what the blob does with a real driver.
881 * There seem to be 32 slots for sample locations, followed by another
882 * 16. The second 16 is just the center location followed by 15 zeros
883 * in all the cases I've identified (maybe shader vs. depth/color
887 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->sample_locations
);
889 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->sample_locations
);
891 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
894 for (int i
= 0; i
< 32 + 16; i
++) {
895 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
899 pandecode_log("};\n");
902 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
905 pandecode_prop("unk0 = 0x%x", fb
->unk0
);
907 if (fb
->sample_locations
)
908 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
910 /* Assume that unknown1 was emitted in the last job for
912 MEMORY_PROP(fb
, unknown1
);
914 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
915 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
916 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
917 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
919 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
920 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
921 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
922 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
924 pandecode_log(".mfbd_flags = ");
925 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
926 pandecode_log_cont(",\n");
928 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
929 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
931 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
932 MEMORY_PROP(fb
, scratchpad
);
933 const struct midgard_tiler_descriptor t
= fb
->tiler
;
934 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
);
936 if (fb
->zero3
|| fb
->zero4
) {
937 pandecode_msg("XXX: framebuffer zeros tripped\n");
938 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
939 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
943 pandecode_log("};\n");
945 gpu_va
+= sizeof(struct bifrost_framebuffer
);
947 if ((fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
) {
948 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
949 const struct bifrost_fb_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
951 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
954 MEMORY_PROP(fbx
, checksum
);
956 if (fbx
->checksum_stride
)
957 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
959 pandecode_log(".flags = ");
960 pandecode_log_decoded_flags(mfbd_extra_flag_info
, fbx
->flags
);
961 pandecode_log_cont(",\n");
963 if (fbx
->flags
& MALI_EXTRA_AFBC_ZS
) {
964 pandecode_log(".ds_afbc = {\n");
967 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
968 pandecode_prop("depth_stencil_afbc_stride = %d",
969 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
970 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
972 if (fbx
->ds_afbc
.zero1
|| fbx
->ds_afbc
.padding
) {
973 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
974 pandecode_prop("zero1 = 0x%" PRIx32
,
976 pandecode_prop("padding = 0x%" PRIx64
,
977 fbx
->ds_afbc
.padding
);
981 pandecode_log("},\n");
983 pandecode_log(".ds_linear = {\n");
986 if (fbx
->ds_linear
.depth
) {
987 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
988 pandecode_prop("depth_stride = %d",
989 fbx
->ds_linear
.depth_stride
);
992 if (fbx
->ds_linear
.stencil
) {
993 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
994 pandecode_prop("stencil_stride = %d",
995 fbx
->ds_linear
.stencil_stride
);
998 if (fbx
->ds_linear
.depth_stride_zero
||
999 fbx
->ds_linear
.stencil_stride_zero
||
1000 fbx
->ds_linear
.zero1
|| fbx
->ds_linear
.zero2
) {
1001 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1002 pandecode_prop("depth_stride_zero = 0x%x",
1003 fbx
->ds_linear
.depth_stride_zero
);
1004 pandecode_prop("stencil_stride_zero = 0x%x",
1005 fbx
->ds_linear
.stencil_stride_zero
);
1006 pandecode_prop("zero1 = 0x%" PRIx32
,
1007 fbx
->ds_linear
.zero1
);
1008 pandecode_prop("zero2 = 0x%" PRIx32
,
1009 fbx
->ds_linear
.zero2
);
1013 pandecode_log("},\n");
1016 if (fbx
->zero3
|| fbx
->zero4
) {
1017 pandecode_msg("XXX: fb_extra zeros tripped\n");
1018 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1019 pandecode_prop("zero4 = 0x%" PRIx64
, fbx
->zero4
);
1023 pandecode_log("};\n");
1025 gpu_va
+= sizeof(struct bifrost_fb_extra
);
1029 pandecode_render_target(gpu_va
, job_no
, fb
);
1031 /* Passback the render target count */
1032 return MALI_NEGATIVE(fb
->rt_count_1
);
1035 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1039 pandecode_padded_vertices(unsigned shift
, unsigned k
)
1041 unsigned odd
= 2*k
+ 1;
1042 unsigned pot
= 1 << shift
;
1043 pandecode_msg("padded_num_vertices = %d\n", odd
* pot
);
1046 /* Given a magic divisor, recover what we were trying to divide by.
1048 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1049 * 0 <= m < 2^N, for N bits in m.
1051 * Let q represent the number we would like to divide by.
1053 * By definition of a magic divisor for N-bit unsigned integers (a number you
1054 * multiply by to magically get division), m is a number such that:
1056 * (m * x) & (2^N - 1) = floor(x/q).
1057 * for all x on Z where 0 <= x < 2^N
1059 * Ignore the case where any of the above values equals zero; it is irrelevant
1060 * for our purposes (instanced arrays).
1062 * Choose x = q. Then:
1064 * (m * x) & (2^N - 1) = floor(x/q).
1065 * (m * q) & (2^N - 1) = floor(q/q).
1067 * floor(q/q) = floor(1) = 1, therefore:
1069 * (m * q) & (2^N - 1) = 1
1071 * Recall the identity that the bitwise AND of one less than a power-of-two
1072 * equals the modulo with that power of two, i.e. for all x:
1074 * x & (2^N - 1) = x % N
1080 * By definition, a modular multiplicative inverse of a number m is the number
1081 * q such that with respect to a modulos M:
1085 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1090 pandecode_magic_divisor(uint32_t magic
, unsigned shift
, unsigned orig_divisor
, unsigned extra
)
1093 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1094 * shift) the most lame way possible... just repeatedly add.
1095 * Asymptoptically slow but nobody cares in practice, unless you have
1096 * massive numbers of vertices or high divisors. */
1098 unsigned inverse
= 0;
1100 /* Magic implicitly has the highest bit set */
1103 /* Depending on rounding direction */
1108 uint32_t product
= magic
* inverse
;
1120 pandecode_msg("dividing by %d (maybe off by two)\n", inverse
);
1122 /* Recall we're supposed to divide by (gl_level_divisor *
1123 * padded_num_vertices) */
1125 unsigned padded_num_vertices
= inverse
/ orig_divisor
;
1127 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices
);
1132 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1133 mali_ptr addr
, int job_no
, char *suffix
,
1134 int count
, bool varying
)
1136 char *prefix
= varying
? "varyings" : "attributes";
1139 pandecode_msg("no %s\n", prefix
);
1143 union mali_attr
*attr
= pandecode_fetch_gpu_mem(mem
, addr
, sizeof(union mali_attr
) * count
);
1146 snprintf(base
, sizeof(base
), "%s_data_%d%s", prefix
, job_no
, suffix
);
1148 for (int i
= 0; i
< count
; ++i
) {
1149 enum mali_attr_mode mode
= attr
[i
].elements
& 7;
1151 if (mode
== MALI_ATTR_UNUSED
)
1154 mali_ptr raw_elements
= attr
[i
].elements
& ~7;
1156 /* TODO: Do we maybe want to dump the attribute values
1157 * themselves given the specified format? Or is that too hard?
1160 char *a
= pointer_as_memory_reference(raw_elements
);
1161 pandecode_log("mali_ptr %s_%d_p = %s;\n", base
, i
, a
);
1165 pandecode_log("union mali_attr %s_%d[] = {\n", prefix
, job_no
);
1168 for (int i
= 0; i
< count
; ++i
) {
1169 pandecode_log("{\n");
1172 unsigned mode
= attr
[i
].elements
& 7;
1173 pandecode_prop("elements = (%s_%d_p) | %s", base
, i
, pandecode_attr_mode(mode
));
1174 pandecode_prop("shift = %d", attr
[i
].shift
);
1175 pandecode_prop("extra_flags = %d", attr
[i
].extra_flags
);
1176 pandecode_prop("stride = 0x%" PRIx32
, attr
[i
].stride
);
1177 pandecode_prop("size = 0x%" PRIx32
, attr
[i
].size
);
1179 /* Decode further where possible */
1181 if (mode
== MALI_ATTR_MODULO
) {
1182 pandecode_padded_vertices(
1184 attr
[i
].extra_flags
);
1188 pandecode_log("}, \n");
1190 if (mode
== MALI_ATTR_NPOT_DIVIDE
) {
1192 pandecode_log("{\n");
1194 pandecode_prop("unk = 0x%x", attr
[i
].unk
);
1195 pandecode_prop("magic_divisor = 0x%08x", attr
[i
].magic_divisor
);
1196 if (attr
[i
].zero
!= 0)
1197 pandecode_prop("XXX: zero tripped (0x%x)\n", attr
[i
].zero
);
1198 pandecode_prop("divisor = %d", attr
[i
].divisor
);
1199 pandecode_magic_divisor(attr
[i
].magic_divisor
, attr
[i
- 1].shift
, attr
[i
].divisor
, attr
[i
- 1].extra_flags
);
1201 pandecode_log("}, \n");
1207 pandecode_log("};\n");
1211 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1213 /* TODO: Decode flags */
1214 mali_ptr shader_ptr
= ptr
& ~15;
1216 char *a
= pointer_as_memory_reference(shader_ptr
);
1217 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1224 all_zero(unsigned *buffer
, unsigned count
)
1226 for (unsigned i
= 0; i
< count
; ++i
) {
1235 pandecode_stencil(const char *name
, const struct mali_stencil_test
*stencil
)
1237 if (all_zero((unsigned *) stencil
, sizeof(stencil
) / sizeof(unsigned)))
1240 const char *func
= pandecode_func(stencil
->func
);
1241 const char *sfail
= pandecode_stencil_op(stencil
->sfail
);
1242 const char *dpfail
= pandecode_stencil_op(stencil
->dpfail
);
1243 const char *dppass
= pandecode_stencil_op(stencil
->dppass
);
1246 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil
->zero
);
1248 pandecode_log(".stencil_%s = {\n", name
);
1250 pandecode_prop("ref = %d", stencil
->ref
);
1251 pandecode_prop("mask = 0x%02X", stencil
->mask
);
1252 pandecode_prop("func = %s", func
);
1253 pandecode_prop("sfail = %s", sfail
);
1254 pandecode_prop("dpfail = %s", dpfail
);
1255 pandecode_prop("dppass = %s", dppass
);
1257 pandecode_log("},\n");
1261 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1264 pandecode_msg("XXX: blend zero tripped: %X\n", blend
->zero1
);
1266 pandecode_log(".equation = {\n");
1269 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1270 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1272 pandecode_log(".color_mask = ");
1273 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1274 pandecode_log_cont(",\n");
1277 pandecode_log("},\n");
1280 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1283 decode_bifrost_constant(u16 constant
)
1285 float lo
= (float) (constant
& 0xFF);
1286 float hi
= (float) (constant
>> 8);
1288 return (hi
/ 255.0) + (lo
/ 65535.0);
1292 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1294 struct bifrost_blend_rt
*b
=
1295 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1297 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1300 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1301 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1302 b
->constant
, decode_bifrost_constant(b
->constant
));
1304 /* TODO figure out blend shader enable bit */
1305 pandecode_blend_equation(&b
->equation
);
1306 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1307 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1308 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1311 pandecode_log("},\n");
1317 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1319 if (all_zero((unsigned *) blend
, sizeof(blend
) / sizeof(unsigned)))
1322 pandecode_log(".blend = {\n");
1326 pandecode_shader_address("shader", blend
->shader
);
1328 pandecode_blend_equation(&blend
->equation
);
1329 pandecode_prop("constant = %f", blend
->constant
);
1333 pandecode_log("},\n");
1335 /* Return blend shader to disassemble if present */
1336 return is_shader
? (blend
->shader
& ~0xF) : 0;
1340 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1342 struct midgard_blend_rt
*b
=
1343 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1345 /* Flags determine presence of blend shader */
1346 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1348 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1351 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1353 union midgard_blend blend
= b
->blend
;
1354 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1357 pandecode_log("};\n");
1363 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1366 char *prefix
= varying
? "varying" : "attribute";
1367 unsigned max_index
= 0;
1368 snprintf(base
, sizeof(base
), "%s_meta", prefix
);
1370 pandecode_log("struct mali_attr_meta %s_%d%s[] = {\n", base
, job_no
, suffix
);
1373 struct mali_attr_meta
*attr_meta
;
1374 mali_ptr p
= varying
? (v
->varying_meta
& ~0xF) : v
->attribute_meta
;
1376 struct pandecode_mapped_memory
*attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
);
1378 for (int i
= 0; i
< count
; ++i
, p
+= sizeof(struct mali_attr_meta
)) {
1379 attr_meta
= pandecode_fetch_gpu_mem(attr_mem
, p
,
1382 pandecode_log("{\n");
1384 pandecode_prop("index = %d", attr_meta
->index
);
1386 if (attr_meta
->index
> max_index
)
1387 max_index
= attr_meta
->index
;
1388 pandecode_swizzle(attr_meta
->swizzle
);
1389 pandecode_prop("format = %s", pandecode_format(attr_meta
->format
));
1391 pandecode_prop("unknown1 = 0x%" PRIx64
, (u64
) attr_meta
->unknown1
);
1392 pandecode_prop("unknown3 = 0x%" PRIx64
, (u64
) attr_meta
->unknown3
);
1393 pandecode_prop("src_offset = %d", attr_meta
->src_offset
);
1395 pandecode_log("},\n");
1400 pandecode_log("};\n");
1402 return count
? (max_index
+ 1) : 0;
1406 pandecode_indices(uintptr_t pindices
, uint32_t index_count
, int job_no
)
1408 struct pandecode_mapped_memory
*imem
= pandecode_find_mapped_gpu_mem_containing(pindices
);
1411 /* Indices are literally just a u32 array :) */
1413 uint32_t *PANDECODE_PTR_VAR(indices
, imem
, pindices
);
1415 pandecode_log("uint32_t indices_%d[] = {\n", job_no
);
1418 for (unsigned i
= 0; i
< (index_count
+ 1); i
+= 3)
1419 pandecode_log("%d, %d, %d,\n",
1425 pandecode_log("};\n");
1429 /* return bits [lo, hi) of word */
1431 bits(u32 word
, u32 lo
, u32 hi
)
1434 return word
; // avoid undefined behavior with the shift
1436 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1440 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool noninstanced
)
1442 pandecode_log_cont("{\n");
1445 /* Decode invocation_count. See the comment before the definition of
1446 * invocation_count for an explanation.
1449 unsigned size_x
= bits(p
->invocation_count
, 0, p
->size_y_shift
) + 1;
1450 unsigned size_y
= bits(p
->invocation_count
, p
->size_y_shift
, p
->size_z_shift
) + 1;
1451 unsigned size_z
= bits(p
->invocation_count
, p
->size_z_shift
, p
->workgroups_x_shift
) + 1;
1453 unsigned groups_x
= bits(p
->invocation_count
, p
->workgroups_x_shift
, p
->workgroups_y_shift
) + 1;
1454 unsigned groups_y
= bits(p
->invocation_count
, p
->workgroups_y_shift
, p
->workgroups_z_shift
) + 1;
1455 unsigned groups_z
= bits(p
->invocation_count
, p
->workgroups_z_shift
, 32) + 1;
1457 /* Even though we have this decoded, we want to ensure that the
1458 * representation is "unique" so we don't lose anything by printing only
1459 * the final result. More specifically, we need to check that we were
1460 * passed something in canonical form, since the definition per the
1461 * hardware is inherently not unique. How? Well, take the resulting
1462 * decode and pack it ourselves! If it is bit exact with what we
1463 * decoded, we're good to go. */
1465 struct mali_vertex_tiler_prefix ref
;
1466 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, noninstanced
);
1469 (p
->invocation_count
== ref
.invocation_count
) &&
1470 (p
->size_y_shift
== ref
.size_y_shift
) &&
1471 (p
->size_z_shift
== ref
.size_z_shift
) &&
1472 (p
->workgroups_x_shift
== ref
.workgroups_x_shift
) &&
1473 (p
->workgroups_y_shift
== ref
.workgroups_y_shift
) &&
1474 (p
->workgroups_z_shift
== ref
.workgroups_z_shift
) &&
1475 (p
->workgroups_x_shift_2
== ref
.workgroups_x_shift_2
);
1478 pandecode_msg("XXX: non-canonical workgroups packing\n");
1479 pandecode_msg("expected: %X, %d, %d, %d, %d, %d\n",
1480 ref
.invocation_count
,
1483 ref
.workgroups_x_shift
,
1484 ref
.workgroups_y_shift
,
1485 ref
.workgroups_z_shift
,
1486 ref
.workgroups_x_shift_2
);
1488 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1489 pandecode_prop("size_y_shift = %d", p
->size_y_shift
);
1490 pandecode_prop("size_z_shift = %d", p
->size_z_shift
);
1491 pandecode_prop("workgroups_x_shift = %d", p
->workgroups_x_shift
);
1492 pandecode_prop("workgroups_y_shift = %d", p
->workgroups_y_shift
);
1493 pandecode_prop("workgroups_z_shift = %d", p
->workgroups_z_shift
);
1494 pandecode_prop("workgroups_x_shift_2 = %d", p
->workgroups_x_shift_2
);
1497 /* Regardless, print the decode */
1498 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1499 size_x
, size_y
, size_z
,
1500 groups_x
, groups_y
, groups_z
);
1503 if (p
->unknown_draw
)
1504 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1506 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1508 if (p
->draw_mode
!= MALI_DRAW_NONE
)
1509 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p
->draw_mode
));
1511 /* Index count only exists for tiler jobs anyway */
1514 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1516 if (p
->offset_bias_correction
)
1517 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1520 pandecode_msg("XXX: payload zero tripped\n");
1521 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1525 pandecode_log("},\n");
1529 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1531 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1532 struct mali_uniform_buffer_meta
*PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1534 pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%"PRIx64
"_%d[] = {\n",
1538 for (int i
= 0; i
< ubufs_count
; i
++) {
1539 pandecode_log("{\n");
1542 unsigned size
= (ubufs
[i
].size
+ 1) * 16;
1543 mali_ptr addr
= ubufs
[i
].ptr
<< 2;
1545 pandecode_validate_buffer(addr
, size
);
1547 char *ptr
= pointer_as_memory_reference(ubufs
[i
].ptr
<< 2);
1548 pandecode_prop("size = %u", size
);
1549 pandecode_prop("ptr = (%s) >> 2", ptr
);
1551 pandecode_log("},\n");
1556 pandecode_log("};\n");
1560 pandecode_scratchpad(uintptr_t pscratchpad
, int job_no
, char *suffix
)
1563 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(pscratchpad
);
1565 struct bifrost_scratchpad
*PANDECODE_PTR_VAR(scratchpad
, mem
, pscratchpad
);
1567 if (scratchpad
->zero
) {
1568 pandecode_msg("XXX: scratchpad zero tripped");
1569 pandecode_prop("zero = 0x%x\n", scratchpad
->zero
);
1572 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64
"_%d%s = {\n", pscratchpad
, job_no
, suffix
);
1575 pandecode_prop("flags = 0x%x", scratchpad
->flags
);
1576 MEMORY_PROP(scratchpad
, gpu_scratchpad
);
1579 pandecode_log("};\n");
1582 static unsigned shader_id
= 0;
1585 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1586 bool is_bifrost
, unsigned nr_regs
)
1588 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1589 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1591 /* Compute maximum possible size */
1592 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1594 /* Print some boilerplate to clearly denote the assembly (which doesn't
1595 * obey indentation rules), and actually do the disassembly! */
1601 snprintf(prefix
, sizeof(prefix
) - 1, "shader%d - %s shader: ",
1603 (type
== JOB_TYPE_TILER
) ? "FRAGMENT" : "VERTEX");
1606 disassemble_bifrost(code
, sz
, false);
1608 disassemble_midgard(code
, sz
, true, nr_regs
, prefix
);
1615 pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix
*p
,
1616 int job_no
, enum mali_job_type job_type
,
1617 char *suffix
, bool is_bifrost
)
1619 mali_ptr shader_meta_ptr
= (u64
) (uintptr_t) (p
->_shader_upper
<< 4);
1620 struct pandecode_mapped_memory
*attr_mem
;
1622 unsigned rt_count
= 1;
1624 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1625 * are the only things actually needed from the FBD, vertex/tiler jobs
1626 * no longer reference the FBD -- instead, this field points to some
1627 * info about the scratchpad.
1630 pandecode_scratchpad(p
->framebuffer
& ~FBD_TYPE
, job_no
, suffix
);
1631 else if (p
->framebuffer
& MALI_MFBD
)
1632 rt_count
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->framebuffer
) & FBD_MASK
, job_no
, false);
1633 else if (job_type
== JOB_TYPE_COMPUTE
)
1634 pandecode_compute_fbd((u64
) (uintptr_t) p
->framebuffer
, job_no
);
1636 pandecode_sfbd((u64
) (uintptr_t) p
->framebuffer
, job_no
, false);
1638 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
1639 int texture_count
= 0, sampler_count
= 0;
1641 if (shader_meta_ptr
) {
1642 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr
);
1643 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, shader_meta_ptr
);
1645 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", shader_meta_ptr
, job_no
, suffix
);
1648 /* Save for dumps */
1649 attribute_count
= s
->attribute_count
;
1650 varying_count
= s
->varying_count
;
1651 texture_count
= s
->texture_count
;
1652 sampler_count
= s
->sampler_count
;
1655 uniform_count
= s
->bifrost2
.uniform_count
;
1656 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
1658 uniform_count
= s
->midgard1
.uniform_buffer_count
;
1659 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
1662 mali_ptr shader_ptr
= pandecode_shader_address("shader", s
->shader
);
1664 pandecode_prop("texture_count = %" PRId16
, s
->texture_count
);
1665 pandecode_prop("sampler_count = %" PRId16
, s
->sampler_count
);
1666 pandecode_prop("attribute_count = %" PRId16
, s
->attribute_count
);
1667 pandecode_prop("varying_count = %" PRId16
, s
->varying_count
);
1669 unsigned nr_registers
= 0;
1672 pandecode_log(".bifrost1 = {\n");
1675 pandecode_prop("uniform_buffer_count = %" PRId32
, s
->bifrost1
.uniform_buffer_count
);
1676 pandecode_prop("unk1 = 0x%" PRIx32
, s
->bifrost1
.unk1
);
1679 pandecode_log("},\n");
1681 pandecode_log(".midgard1 = {\n");
1684 pandecode_prop("uniform_count = %" PRId16
, s
->midgard1
.uniform_count
);
1685 pandecode_prop("uniform_buffer_count = %" PRId16
, s
->midgard1
.uniform_buffer_count
);
1686 pandecode_prop("work_count = %" PRId16
, s
->midgard1
.work_count
);
1687 nr_registers
= s
->midgard1
.work_count
;
1689 pandecode_log(".flags = ");
1690 pandecode_log_decoded_flags(shader_midgard1_flag_info
, s
->midgard1
.flags
);
1691 pandecode_log_cont(",\n");
1693 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->midgard1
.unknown2
);
1696 pandecode_log("},\n");
1699 if (s
->depth_units
|| s
->depth_factor
) {
1700 pandecode_prop("depth_factor = %f", s
->depth_factor
);
1701 pandecode_prop("depth_units = %f", s
->depth_units
);
1704 if (s
->alpha_coverage
) {
1705 bool invert_alpha_coverage
= s
->alpha_coverage
& 0xFFF0;
1706 uint16_t inverted_coverage
= invert_alpha_coverage
? ~s
->alpha_coverage
: s
->alpha_coverage
;
1708 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
1709 invert_alpha_coverage
? "~" : "",
1710 MALI_GET_ALPHA_COVERAGE(inverted_coverage
));
1713 if (s
->unknown2_3
|| s
->unknown2_4
) {
1714 pandecode_log(".unknown2_3 = ");
1716 int unknown2_3
= s
->unknown2_3
;
1717 int unknown2_4
= s
->unknown2_4
;
1719 /* We're not quite sure what these flags mean without the depth test, if anything */
1721 if (unknown2_3
& (MALI_DEPTH_TEST
| MALI_DEPTH_FUNC_MASK
)) {
1722 const char *func
= pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3
));
1723 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
1725 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
1728 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
1729 pandecode_log_cont(",\n");
1731 pandecode_log(".unknown2_4 = ");
1732 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
1733 pandecode_log_cont(",\n");
1736 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
1737 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
1738 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
1741 pandecode_stencil("front", &s
->stencil_front
);
1742 pandecode_stencil("back", &s
->stencil_back
);
1745 pandecode_log(".bifrost2 = {\n");
1748 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
1749 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
1750 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
1751 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
1754 pandecode_log("},\n");
1755 } else if (s
->midgard2
.unknown2_7
) {
1756 pandecode_log(".midgard2 = {\n");
1759 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
1761 pandecode_log("},\n");
1765 pandecode_prop("unknown2_8 = 0x%" PRIx32
, s
->unknown2_8
);
1768 /* TODO: Blend shaders routing/disasm */
1770 union midgard_blend blend
= s
->blend
;
1771 pandecode_midgard_blend(&blend
, false);
1775 pandecode_log("};\n");
1777 /* MRT blend fields are used whenever MFBD is used, with
1778 * per-RT descriptors */
1780 if (job_type
== JOB_TYPE_TILER
) {
1781 void* blend_base
= (void *) (s
+ 1);
1783 for (unsigned i
= 0; i
< rt_count
; i
++) {
1784 mali_ptr shader
= 0;
1787 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
1789 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
1792 pandecode_shader_disassemble(shader
, job_no
, job_type
, false, 0);
1796 if (shader_ptr
& ~0xF)
1797 pandecode_shader_disassemble(shader_ptr
, job_no
, job_type
, is_bifrost
, nr_registers
);
1799 pandecode_msg("<no shader>\n");
1802 struct pandecode_mapped_memory
*fmem
= pandecode_find_mapped_gpu_mem_containing(p
->viewport
);
1803 struct mali_viewport
*PANDECODE_PTR_VAR(f
, fmem
, p
->viewport
);
1805 pandecode_log("struct mali_viewport viewport_%"PRIx64
"_%d%s = {\n", p
->viewport
, job_no
, suffix
);
1808 pandecode_prop("clip_minx = %f", f
->clip_minx
);
1809 pandecode_prop("clip_miny = %f", f
->clip_miny
);
1810 pandecode_prop("clip_minz = %f", f
->clip_minz
);
1811 pandecode_prop("clip_maxx = %f", f
->clip_maxx
);
1812 pandecode_prop("clip_maxy = %f", f
->clip_maxy
);
1813 pandecode_prop("clip_maxz = %f", f
->clip_maxz
);
1815 /* Only the higher coordinates are MALI_POSITIVE scaled */
1817 pandecode_prop("viewport0 = { %d, %d }",
1818 f
->viewport0
[0], f
->viewport0
[1]);
1820 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
1821 f
->viewport1
[0] + 1, f
->viewport1
[1] + 1);
1824 pandecode_log("};\n");
1827 if (p
->attribute_meta
) {
1828 unsigned max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
1830 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
1831 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false);
1834 /* Varyings are encoded like attributes but not actually sent; we just
1835 * pass a zero buffer with the right stride/size set, (or whatever)
1836 * since the GPU will write to it itself */
1838 if (p
->varying_meta
) {
1839 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
1843 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
1845 /* Number of descriptors depends on whether there are
1846 * non-internal varyings */
1848 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true);
1851 if (p
->uniform_buffers
) {
1852 if (uniform_buffer_count
)
1853 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
1855 pandecode_msg("XXX: UBOs specified but not referenced\n");
1856 } else if (uniform_buffer_count
)
1857 pandecode_msg("XXX: UBOs referenced but not specified\n");
1859 /* We don't want to actually dump uniforms, but we do need to validate
1860 * that the counts we were given are sane */
1864 pandecode_validate_buffer(p
->uniforms
, uniform_count
* 16);
1866 pandecode_msg("XXX: Uniforms specified but not referenced");
1867 } else if (uniform_count
)
1868 pandecode_msg("XXX: UBOs referenced but not specified\n");
1870 if (p
->texture_trampoline
) {
1871 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(p
->texture_trampoline
);
1874 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
);
1876 pandecode_log("uint64_t texture_trampoline_%"PRIx64
"_%d[] = {\n", p
->texture_trampoline
, job_no
);
1879 for (int tex
= 0; tex
< texture_count
; ++tex
) {
1880 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
1881 char *a
= pointer_as_memory_reference(*u
);
1882 pandecode_log("%s,\n", a
);
1887 pandecode_log("};\n");
1889 /* Now, finally, descend down into the texture descriptor */
1890 for (int tex
= 0; tex
< texture_count
; ++tex
) {
1891 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
1892 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
1895 struct mali_texture_descriptor
*PANDECODE_PTR_VAR(t
, tmem
, *u
);
1897 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64
"_%d_%d = {\n", *u
, job_no
, tex
);
1900 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", t
->width
+ 1);
1901 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", t
->height
+ 1);
1902 pandecode_prop("depth = MALI_POSITIVE(%" PRId16
")", t
->depth
+ 1);
1903 pandecode_prop("array_size = MALI_POSITIVE(%" PRId16
")", t
->array_size
+ 1);
1904 pandecode_prop("unknown3 = %" PRId16
, t
->unknown3
);
1905 pandecode_prop("unknown3A = %" PRId8
, t
->unknown3A
);
1906 pandecode_prop("nr_mipmap_levels = %" PRId8
, t
->nr_mipmap_levels
);
1908 struct mali_texture_format f
= t
->format
;
1910 pandecode_log(".format = {\n");
1913 pandecode_swizzle(f
.swizzle
);
1914 pandecode_prop("format = %s", pandecode_format(f
.format
));
1915 pandecode_prop("type = %s", pandecode_texture_type(f
.type
));
1916 pandecode_prop("srgb = %" PRId32
, f
.srgb
);
1917 pandecode_prop("unknown1 = %" PRId32
, f
.unknown1
);
1918 pandecode_prop("usage2 = 0x%" PRIx32
, f
.usage2
);
1921 pandecode_log("},\n");
1923 pandecode_swizzle(t
->swizzle
);
1925 if (t
->swizzle_zero
) {
1926 /* Shouldn't happen */
1927 pandecode_msg("XXX: swizzle zero tripped\n");
1928 pandecode_prop("swizzle_zero = %d", t
->swizzle_zero
);
1931 pandecode_prop("unknown3 = 0x%" PRIx32
, t
->unknown3
);
1933 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
1934 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
1935 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
1937 pandecode_log(".payload = {\n");
1940 /* A bunch of bitmap pointers follow.
1941 * We work out the correct number,
1942 * based on the mipmap/cubemap
1943 * properties, but dump extra
1944 * possibilities to futureproof */
1946 int bitmap_count
= MALI_NEGATIVE(t
->nr_mipmap_levels
);
1947 bool manual_stride
= f
.usage2
& MALI_TEX_MANUAL_STRIDE
;
1949 /* Miptree for each face */
1950 if (f
.type
== MALI_TEX_CUBE
)
1953 /* Array of textures */
1954 bitmap_count
*= MALI_NEGATIVE(t
->array_size
);
1956 /* Stride for each element */
1960 /* Sanity check the size */
1961 int max_count
= sizeof(t
->payload
) / sizeof(t
->payload
[0]);
1962 assert (bitmap_count
<= max_count
);
1964 for (int i
= 0; i
< bitmap_count
; ++i
) {
1965 /* How we dump depends if this is a stride or a pointer */
1967 if ((f
.usage2
& MALI_TEX_MANUAL_STRIDE
) && (i
& 1)) {
1968 /* signed 32-bit snuck in as a 64-bit pointer */
1969 uint64_t stride_set
= t
->payload
[i
];
1970 uint32_t clamped_stride
= stride_set
;
1971 int32_t stride
= clamped_stride
;
1972 assert(stride_set
== clamped_stride
);
1973 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
1975 char *a
= pointer_as_memory_reference(t
->payload
[i
]);
1976 pandecode_log("%s, \n", a
);
1982 pandecode_log("},\n");
1985 pandecode_log("};\n");
1991 if (p
->sampler_descriptor
) {
1992 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->sampler_descriptor
);
1995 struct mali_sampler_descriptor
*s
;
1997 mali_ptr d
= p
->sampler_descriptor
;
1999 for (int i
= 0; i
< sampler_count
; ++i
) {
2000 s
= pandecode_fetch_gpu_mem(smem
, d
+ sizeof(*s
) * i
, sizeof(*s
));
2002 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", d
+ sizeof(*s
) * i
, job_no
, i
);
2005 pandecode_log(".filter_mode = ");
2006 pandecode_log_decoded_flags(sampler_flag_info
, s
->filter_mode
);
2007 pandecode_log_cont(",\n");
2009 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
2010 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
2012 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s
->wrap_s
));
2013 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s
->wrap_t
));
2014 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s
->wrap_r
));
2016 pandecode_prop("compare_func = %s", pandecode_alt_func(s
->compare_func
));
2018 if (s
->zero
|| s
->zero2
) {
2019 pandecode_msg("XXX: sampler zero tripped\n");
2020 pandecode_prop("zero = 0x%X, 0x%X\n", s
->zero
, s
->zero2
);
2023 pandecode_prop("seamless_cube_map = %d", s
->seamless_cube_map
);
2025 pandecode_prop("border_color = { %f, %f, %f, %f }",
2029 s
->border_color
[3]);
2032 pandecode_log("};\n");
2039 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2041 if (!(p
->position_varying
|| p
->occlusion_counter
|| p
->flags
))
2044 pandecode_log(".postfix = {\n");
2047 MEMORY_PROP(p
, position_varying
);
2048 MEMORY_PROP(p
, occlusion_counter
);
2051 pandecode_prop("flags = %d", p
->flags
);
2054 pandecode_log("},\n");
2058 pandecode_vertex_only_bfr(struct bifrost_vertex_only
*v
)
2060 pandecode_log_cont("{\n");
2063 pandecode_prop("unk2 = 0x%x", v
->unk2
);
2065 if (v
->zero0
|| v
->zero1
) {
2066 pandecode_msg("XXX: vertex only zero tripped");
2067 pandecode_prop("zero0 = 0x%" PRIx32
, v
->zero0
);
2068 pandecode_prop("zero1 = 0x%" PRIx64
, v
->zero1
);
2072 pandecode_log("}\n");
2076 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2079 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2080 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2082 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no
);
2086 pandecode_msg("XXX: tiler heap zero tripped\n");
2087 pandecode_prop("zero = 0x%x", h
->zero
);
2090 for (int i
= 0; i
< 12; i
++) {
2091 if (h
->zeros
[i
] != 0) {
2092 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2097 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2098 MEMORY_PROP(h
, tiler_heap_start
);
2099 MEMORY_PROP(h
, tiler_heap_free
);
2101 /* this might point to the beginning of another buffer, when it's
2102 * really the end of the tiler heap buffer, so we have to be careful
2103 * here. but for zero length, we need the same pointer.
2106 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2107 MEMORY_PROP(h
, tiler_heap_start
);
2109 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2110 pandecode_prop("tiler_heap_end = %s + 1", a
);
2115 pandecode_log("};\n");
2119 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2121 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2122 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2124 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2126 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no
);
2129 if (t
->zero0
|| t
->zero1
) {
2130 pandecode_msg("XXX: tiler meta zero tripped\n");
2131 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2132 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2135 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2136 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2138 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2139 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2141 for (int i
= 0; i
< 12; i
++) {
2142 if (t
->zeros
[i
] != 0) {
2143 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2149 pandecode_log("};\n");
2153 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2155 pandecode_log(".gl_enables = ");
2157 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2159 pandecode_log_cont(",\n");
2163 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2165 if (u
.pointer
== 0x0)
2168 pandecode_log(".primitive_size = {\n");
2172 pandecode_prop("constant = %f", u
.constant
);
2174 MEMORY_PROP((&u
), pointer
);
2178 pandecode_log("},\n");
2182 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2184 pandecode_log_cont("{\n");
2187 /* TODO: gl_PointSize on Bifrost */
2188 pandecode_primitive_size(t
->primitive_size
, true);
2190 pandecode_gl_enables(t
->gl_enables
, JOB_TYPE_TILER
);
2192 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2193 || t
->zero6
|| t
->zero7
|| t
->zero8
) {
2194 pandecode_msg("XXX: tiler only zero tripped\n");
2195 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2196 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2197 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2198 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2199 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2200 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2201 pandecode_prop("zero7 = 0x%" PRIx32
, t
->zero7
);
2202 pandecode_prop("zero8 = 0x%" PRIx64
, t
->zero8
);
2206 pandecode_log("},\n");
2210 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2211 const struct pandecode_mapped_memory
*mem
,
2212 mali_ptr payload
, int job_no
)
2214 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2216 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true);
2218 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no
);
2221 pandecode_log(".prefix = ");
2222 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2224 pandecode_log(".vertex = ");
2225 pandecode_vertex_only_bfr(&v
->vertex
);
2227 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2230 pandecode_log("};\n");
2236 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2237 const struct pandecode_mapped_memory
*mem
,
2238 mali_ptr payload
, int job_no
)
2240 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2242 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true);
2244 pandecode_indices(t
->prefix
.indices
, t
->prefix
.index_count
, job_no
);
2245 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2247 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no
);
2250 pandecode_log(".prefix = ");
2251 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2253 pandecode_log(".tiler = ");
2254 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2256 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2259 pandecode_log("};\n");
2265 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2266 const struct pandecode_mapped_memory
*mem
,
2267 mali_ptr payload
, int job_no
)
2269 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2271 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false);
2273 pandecode_indices(v
->prefix
.indices
, v
->prefix
.index_count
, job_no
);
2275 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2278 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2279 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2281 bool instanced
= v
->instance_shift
|| v
->instance_odd
;
2282 bool is_graphics
= (h
->job_type
== JOB_TYPE_VERTEX
) || (h
->job_type
== JOB_TYPE_TILER
);
2284 pandecode_log(".prefix = ");
2285 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, !instanced
&& is_graphics
);
2287 pandecode_gl_enables(v
->gl_enables
, h
->job_type
);
2289 if (v
->instance_shift
|| v
->instance_odd
) {
2290 pandecode_prop("instance_shift = 0x%d /* %d */",
2291 v
->instance_shift
, 1 << v
->instance_shift
);
2292 pandecode_prop("instance_odd = 0x%X /* %d */",
2293 v
->instance_odd
, (2 * v
->instance_odd
) + 1);
2295 pandecode_padded_vertices(v
->instance_shift
, v
->instance_odd
);
2298 if (v
->offset_start
)
2299 pandecode_prop("offset_start = %d", v
->offset_start
);
2302 pandecode_msg("XXX: midgard payload zero tripped\n");
2303 pandecode_prop("zero5 = 0x%" PRIx64
, v
->zero5
);
2306 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2309 pandecode_log("};\n");
2315 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2316 mali_ptr payload
, int job_no
,
2319 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2321 bool fbd_dumped
= false;
2323 if (!is_bifrost
&& (s
->framebuffer
& FBD_TYPE
) == MALI_SFBD
) {
2324 /* Only SFBDs are understood, not MFBDs. We're speculating,
2325 * based on the versioning, kernel code, etc, that the
2326 * difference is between Single FrameBuffer Descriptor and
2327 * Multiple FrmaeBuffer Descriptor; the change apparently lines
2328 * up with multi-framebuffer support being added (T7xx onwards,
2329 * including Gxx). In any event, there's some field shuffling
2330 * that we haven't looked into yet. */
2332 pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true);
2334 } else if ((s
->framebuffer
& FBD_TYPE
) == MALI_MFBD
) {
2335 /* We don't know if Bifrost supports SFBD's at all, since the
2336 * driver never uses them. And the format is different from
2337 * Midgard anyways, due to the tiler heap and scratchpad being
2338 * moved out into separate structures, so it's not clear what a
2339 * Bifrost SFBD would even look like without getting an actual
2340 * trace, which appears impossible.
2343 pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true);
2347 uintptr_t p
= (uintptr_t) s
->framebuffer
& FBD_MASK
;
2348 pandecode_log("struct mali_payload_fragment payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2351 /* See the comments by the macro definitions for mathematical context
2352 * on why this is so weird */
2354 if (MALI_TILE_COORD_FLAGS(s
->max_tile_coord
) || MALI_TILE_COORD_FLAGS(s
->min_tile_coord
))
2355 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
2357 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
2358 MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
,
2359 MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
);
2361 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
2362 (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
,
2363 (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
);
2365 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2366 * we have to fallback on offsets. */
2368 const char *fbd_type
= s
->framebuffer
& MALI_MFBD
? "MALI_MFBD" : "MALI_SFBD";
2371 unsigned extra_flags
= (s
->framebuffer
& ~FBD_MASK
) & ~MALI_MFBD
;
2374 pandecode_prop("framebuffer = framebuffer_%d_p | %s | 0x%X", job_no
,
2375 fbd_type
, extra_flags
);
2377 pandecode_prop("framebuffer = %s | %s | 0x%X", pointer_as_memory_reference(p
),
2378 fbd_type
, extra_flags
);
2381 pandecode_log("};\n");
2386 static int job_descriptor_number
= 0;
2389 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
)
2391 struct mali_job_descriptor_header
*h
;
2393 int start_number
= 0;
2399 struct pandecode_mapped_memory
*mem
=
2400 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2404 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2406 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2407 * high 32-bits of the 64-bit pointer are reused to store
2410 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
2411 h
->job_type
!= JOB_TYPE_FRAGMENT
? 4 : 0;
2412 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
2414 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
,
2417 int job_no
= job_descriptor_number
++;
2420 start_number
= job_no
;
2422 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2425 pandecode_prop("job_type = %s", pandecode_job_type(h
->job_type
));
2427 /* Save for next job fixing */
2428 last_size
= h
->job_descriptor_size
;
2430 if (h
->job_descriptor_size
)
2431 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2433 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2434 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2435 h
->exception_status
,
2436 (h
->exception_status
>> 16) & 0xFFFF,
2437 pandecode_exception_access((h
->exception_status
>> 8) & 0x3),
2438 h
->exception_status
& 0xFF);
2440 if (h
->first_incomplete_task
)
2441 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2443 if (h
->fault_pointer
)
2444 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2447 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2449 pandecode_prop("job_index = %d", h
->job_index
);
2451 if (h
->unknown_flags
)
2452 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2454 if (h
->job_dependency_index_1
)
2455 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2457 if (h
->job_dependency_index_2
)
2458 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2461 pandecode_log("};\n");
2463 /* Do not touch the field yet -- decode the payload first, and
2464 * don't touch that either. This is essential for the uploads
2465 * to occur in sequence and therefore be dynamically allocated
2466 * correctly. Do note the size, however, for that related
2469 switch (h
->job_type
) {
2470 case JOB_TYPE_SET_VALUE
: {
2471 struct mali_payload_set_value
*s
= payload
;
2472 pandecode_log("struct mali_payload_set_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2474 MEMORY_PROP(s
, out
);
2475 pandecode_prop("unknown = 0x%" PRIX64
, s
->unknown
);
2477 pandecode_log("};\n");
2482 case JOB_TYPE_TILER
:
2483 case JOB_TYPE_VERTEX
:
2484 case JOB_TYPE_COMPUTE
:
2486 if (h
->job_type
== JOB_TYPE_TILER
)
2487 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
);
2489 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
);
2491 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
);
2495 case JOB_TYPE_FRAGMENT
:
2496 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
);
2503 /* Handle linkage */
2506 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no
- 1);
2509 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no
);
2511 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no
);
2516 } while ((jc_gpu_va
= h
->job_descriptor_size
? h
->next_job_64
: h
->next_job_32
));
2518 return start_number
;