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 */
541 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
543 MEMORY_PROP(t
, polygon_list
);
545 /* The body is offset from the base of the polygon list */
546 assert(t
->polygon_list_body
> t
->polygon_list
);
547 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
549 /* It needs to fit inside the reported size */
550 assert(t
->polygon_list_size
>= body_offset
);
552 /* Check that we fit */
553 struct pandecode_mapped_memory
*plist
=
554 pandecode_find_mapped_gpu_mem_containing(t
->polygon_list
);
556 assert(t
->polygon_list_size
<= plist
->length
);
558 /* Now that we've sanity checked, we'll try to calculate the sizes
559 * ourselves for comparison */
561 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
);
562 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
);
564 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
565 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
566 ref_header
, ref_size
);
567 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
568 pandecode_msg("body offset %d\n", body_offset
);
571 /* The tiler heap has a start and end specified -- it should be
572 * identical to what we have in the BO. The exception is if tiling is
575 MEMORY_PROP(t
, heap_start
);
576 assert(t
->heap_end
>= t
->heap_start
);
578 struct pandecode_mapped_memory
*heap
=
579 pandecode_find_mapped_gpu_mem_containing(t
->heap_start
);
581 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
583 /* Tiling is enabled with a special flag */
584 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
585 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
587 bool tiling_enabled
= hierarchy_mask
;
589 if (tiling_enabled
) {
590 /* When tiling is enabled, the heap should be a tight fit */
591 unsigned heap_offset
= t
->heap_start
- heap
->gpu_va
;
592 if ((heap_offset
+ heap_size
) != heap
->length
) {
593 pandecode_msg("XXX: heap size %d (expected %d)\n",
594 heap_size
, heap
->length
- heap_offset
);
597 /* We should also have no other flags */
599 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
601 /* When tiling is disabled, we should have that flag and no others */
603 if (tiler_flags
!= MALI_TILER_DISABLED
) {
604 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
608 /* We should also have an empty heap */
610 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
614 /* Disabled tiling is used only for clear-only jobs, which are
615 * purely FRAGMENT, so we should never see this for
616 * non-FRAGMENT descriptors. */
619 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
622 /* We've never seen weights used in practice, but we know from the
623 * kernel these fields is there */
625 bool nonzero_weights
= false;
627 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
628 nonzero_weights
|= t
->weights
[w
] != 0x0;
631 if (nonzero_weights
) {
632 pandecode_log(".weights = {");
634 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
635 pandecode_log("%d, ", t
->weights
[w
]);
642 pandecode_log("}\n");
646 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
)
648 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
649 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
651 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
654 pandecode_prop("unknown1 = 0x%" PRIx32
, s
->unknown1
);
655 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->unknown2
);
657 pandecode_log(".format = ");
658 pandecode_log_decoded_flags(fb_fmt_flag_info
, s
->format
);
659 pandecode_log_cont(",\n");
661 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", s
->width
+ 1);
662 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", s
->height
+ 1);
664 MEMORY_PROP(s
, framebuffer
);
665 pandecode_prop("stride = %d", s
->stride
);
667 /* Earlier in the actual commandstream -- right before width -- but we
668 * delay to flow nicer */
670 pandecode_log(".clear_flags = ");
671 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
672 pandecode_log_cont(",\n");
674 if (s
->depth_buffer
| s
->depth_buffer_enable
) {
675 MEMORY_PROP(s
, depth_buffer
);
676 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s
->depth_buffer_enable
));
679 if (s
->stencil_buffer
| s
->stencil_buffer_enable
) {
680 MEMORY_PROP(s
, stencil_buffer
);
681 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s
->stencil_buffer_enable
));
684 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
685 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
686 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
687 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
688 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
691 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
692 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
693 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
694 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
695 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
698 if (s
->clear_stencil
) {
699 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
702 MEMORY_PROP(s
, unknown_address_0
);
703 const struct midgard_tiler_descriptor t
= s
->tiler
;
704 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
);
707 pandecode_log("};\n");
709 pandecode_prop("zero0 = 0x%" PRIx64
, s
->zero0
);
710 pandecode_prop("zero1 = 0x%" PRIx64
, s
->zero1
);
711 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
712 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
714 printf(".zero3 = {");
716 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
717 printf("%X, ", s
->zero3
[i
]);
721 printf(".zero6 = {");
723 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
724 printf("%X, ", s
->zero6
[i
]);
730 pandecode_u32_slide(unsigned name
, const u32
*slide
, unsigned count
)
732 pandecode_log(".unknown%d = {", name
);
734 for (int i
= 0; i
< count
; ++i
)
735 printf("%X, ", slide
[i
]);
737 pandecode_log("},\n");
740 #define SHORT_SLIDE(num) \
741 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
744 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
746 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
747 const struct mali_compute_fbd
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
749 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
759 pandecode_swizzle(unsigned swizzle
)
761 pandecode_prop("swizzle = %s | (%s << 3) | (%s << 6) | (%s << 9)",
762 pandecode_channel((swizzle
>> 0) & 0x7),
763 pandecode_channel((swizzle
>> 3) & 0x7),
764 pandecode_channel((swizzle
>> 6) & 0x7),
765 pandecode_channel((swizzle
>> 9) & 0x7));
769 pandecode_rt_format(struct mali_rt_format format
)
771 pandecode_log(".format = {\n");
774 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
775 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
776 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
778 pandecode_prop("block = %s",
779 pandecode_mfbd_block_format(format
.block
));
781 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
782 MALI_NEGATIVE(format
.nr_channels
));
784 pandecode_log(".flags = ");
785 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
786 pandecode_log_cont(",\n");
788 pandecode_swizzle(format
.swizzle
);
790 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
791 * which is a faster hardware-based alternative to the wallpaper method
792 * to preserve framebuffer contents across frames. In practice, MFBD
793 * preload is buggy on Midgard, and so this is a chicken bit. If this
794 * bit isn't set, most likely something broke unrelated to preload */
796 if (!format
.no_preload
) {
797 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
798 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
802 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
805 pandecode_log("},\n");
809 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct bifrost_framebuffer
*fb
)
811 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
814 for (int i
= 0; i
< MALI_NEGATIVE(fb
->rt_count_1
); i
++) {
815 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct bifrost_render_target
);
816 struct pandecode_mapped_memory
*mem
=
817 pandecode_find_mapped_gpu_mem_containing(rt_va
);
818 const struct bifrost_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
820 pandecode_log("{\n");
823 pandecode_rt_format(rt
->format
);
825 if (rt
->format
.block
== MALI_MFBD_BLOCK_AFBC
) {
826 pandecode_log(".afbc = {\n");
829 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
830 pandecode_prop("metadata = %s", a
);
833 pandecode_prop("stride = %d", rt
->afbc
.stride
);
834 pandecode_prop("unk = 0x%" PRIx32
, rt
->afbc
.unk
);
837 pandecode_log("},\n");
838 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.unk
) {
839 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
845 MEMORY_PROP(rt
, framebuffer
);
846 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
848 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
849 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
850 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
851 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
852 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
855 if (rt
->zero1
|| rt
->zero2
|| rt
->zero3
) {
856 pandecode_msg("XXX: render target zeros tripped\n");
857 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
858 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
859 pandecode_prop("zero3 = 0x%" PRIx32
, rt
->zero3
);
863 pandecode_log("},\n");
867 pandecode_log("};\n");
871 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
)
873 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
874 const struct bifrost_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
876 if (fb
->sample_locations
) {
877 /* The blob stores all possible sample locations in a single buffer
878 * allocated on startup, and just switches the pointer when switching
879 * MSAA state. For now, we just put the data into the cmdstream, but we
880 * should do something like what the blob does with a real driver.
882 * There seem to be 32 slots for sample locations, followed by another
883 * 16. The second 16 is just the center location followed by 15 zeros
884 * in all the cases I've identified (maybe shader vs. depth/color
888 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->sample_locations
);
890 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->sample_locations
);
892 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
895 for (int i
= 0; i
< 32 + 16; i
++) {
896 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
900 pandecode_log("};\n");
903 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
906 pandecode_prop("unk0 = 0x%x", fb
->unk0
);
908 if (fb
->sample_locations
)
909 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
911 /* Assume that unknown1 was emitted in the last job for
913 MEMORY_PROP(fb
, unknown1
);
915 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
916 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
917 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
918 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
920 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
921 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
922 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
923 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
925 pandecode_log(".mfbd_flags = ");
926 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
927 pandecode_log_cont(",\n");
929 if (fb
->clear_stencil
)
930 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
933 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
935 /* TODO: What is this? Let's not blow up.. */
936 if (fb
->unknown2
!= 0x1F)
937 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
939 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
940 MEMORY_PROP(fb
, scratchpad
);
941 const struct midgard_tiler_descriptor t
= fb
->tiler
;
942 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
);
944 if (fb
->zero3
|| fb
->zero4
) {
945 pandecode_msg("XXX: framebuffer zeros tripped\n");
946 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
947 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
951 pandecode_log("};\n");
953 gpu_va
+= sizeof(struct bifrost_framebuffer
);
955 if ((fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
) {
956 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
957 const struct bifrost_fb_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
959 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
962 MEMORY_PROP(fbx
, checksum
);
964 if (fbx
->checksum_stride
)
965 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
967 pandecode_log(".flags = ");
968 pandecode_log_decoded_flags(mfbd_extra_flag_info
, fbx
->flags
);
969 pandecode_log_cont(",\n");
971 if (fbx
->flags
& MALI_EXTRA_AFBC_ZS
) {
972 pandecode_log(".ds_afbc = {\n");
975 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
976 pandecode_prop("depth_stencil_afbc_stride = %d",
977 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
978 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
980 if (fbx
->ds_afbc
.zero1
|| fbx
->ds_afbc
.padding
) {
981 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
982 pandecode_prop("zero1 = 0x%" PRIx32
,
984 pandecode_prop("padding = 0x%" PRIx64
,
985 fbx
->ds_afbc
.padding
);
989 pandecode_log("},\n");
991 pandecode_log(".ds_linear = {\n");
994 if (fbx
->ds_linear
.depth
) {
995 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
996 pandecode_prop("depth_stride = %d",
997 fbx
->ds_linear
.depth_stride
);
1000 if (fbx
->ds_linear
.stencil
) {
1001 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1002 pandecode_prop("stencil_stride = %d",
1003 fbx
->ds_linear
.stencil_stride
);
1006 if (fbx
->ds_linear
.depth_stride_zero
||
1007 fbx
->ds_linear
.stencil_stride_zero
||
1008 fbx
->ds_linear
.zero1
|| fbx
->ds_linear
.zero2
) {
1009 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1010 pandecode_prop("depth_stride_zero = 0x%x",
1011 fbx
->ds_linear
.depth_stride_zero
);
1012 pandecode_prop("stencil_stride_zero = 0x%x",
1013 fbx
->ds_linear
.stencil_stride_zero
);
1014 pandecode_prop("zero1 = 0x%" PRIx32
,
1015 fbx
->ds_linear
.zero1
);
1016 pandecode_prop("zero2 = 0x%" PRIx32
,
1017 fbx
->ds_linear
.zero2
);
1021 pandecode_log("},\n");
1024 if (fbx
->zero3
|| fbx
->zero4
) {
1025 pandecode_msg("XXX: fb_extra zeros tripped\n");
1026 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1027 pandecode_prop("zero4 = 0x%" PRIx64
, fbx
->zero4
);
1031 pandecode_log("};\n");
1033 gpu_va
+= sizeof(struct bifrost_fb_extra
);
1037 pandecode_render_target(gpu_va
, job_no
, fb
);
1039 /* Passback the render target count */
1040 return MALI_NEGATIVE(fb
->rt_count_1
);
1043 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1047 pandecode_padded_vertices(unsigned shift
, unsigned k
)
1049 unsigned odd
= 2*k
+ 1;
1050 unsigned pot
= 1 << shift
;
1051 pandecode_msg("padded_num_vertices = %d\n", odd
* pot
);
1054 /* Given a magic divisor, recover what we were trying to divide by.
1056 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1057 * 0 <= m < 2^N, for N bits in m.
1059 * Let q represent the number we would like to divide by.
1061 * By definition of a magic divisor for N-bit unsigned integers (a number you
1062 * multiply by to magically get division), m is a number such that:
1064 * (m * x) & (2^N - 1) = floor(x/q).
1065 * for all x on Z where 0 <= x < 2^N
1067 * Ignore the case where any of the above values equals zero; it is irrelevant
1068 * for our purposes (instanced arrays).
1070 * Choose x = q. Then:
1072 * (m * x) & (2^N - 1) = floor(x/q).
1073 * (m * q) & (2^N - 1) = floor(q/q).
1075 * floor(q/q) = floor(1) = 1, therefore:
1077 * (m * q) & (2^N - 1) = 1
1079 * Recall the identity that the bitwise AND of one less than a power-of-two
1080 * equals the modulo with that power of two, i.e. for all x:
1082 * x & (2^N - 1) = x % N
1088 * By definition, a modular multiplicative inverse of a number m is the number
1089 * q such that with respect to a modulos M:
1093 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1098 pandecode_magic_divisor(uint32_t magic
, unsigned shift
, unsigned orig_divisor
, unsigned extra
)
1101 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1102 * shift) the most lame way possible... just repeatedly add.
1103 * Asymptoptically slow but nobody cares in practice, unless you have
1104 * massive numbers of vertices or high divisors. */
1106 unsigned inverse
= 0;
1108 /* Magic implicitly has the highest bit set */
1111 /* Depending on rounding direction */
1116 uint32_t product
= magic
* inverse
;
1128 pandecode_msg("dividing by %d (maybe off by two)\n", inverse
);
1130 /* Recall we're supposed to divide by (gl_level_divisor *
1131 * padded_num_vertices) */
1133 unsigned padded_num_vertices
= inverse
/ orig_divisor
;
1135 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices
);
1140 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1141 mali_ptr addr
, int job_no
, char *suffix
,
1142 int count
, bool varying
)
1144 char *prefix
= varying
? "varyings" : "attributes";
1147 pandecode_msg("no %s\n", prefix
);
1151 union mali_attr
*attr
= pandecode_fetch_gpu_mem(mem
, addr
, sizeof(union mali_attr
) * count
);
1154 snprintf(base
, sizeof(base
), "%s_data_%d%s", prefix
, job_no
, suffix
);
1156 for (int i
= 0; i
< count
; ++i
) {
1157 enum mali_attr_mode mode
= attr
[i
].elements
& 7;
1159 if (mode
== MALI_ATTR_UNUSED
)
1162 mali_ptr raw_elements
= attr
[i
].elements
& ~7;
1164 /* TODO: Do we maybe want to dump the attribute values
1165 * themselves given the specified format? Or is that too hard?
1168 char *a
= pointer_as_memory_reference(raw_elements
);
1169 pandecode_log("mali_ptr %s_%d_p = %s;\n", base
, i
, a
);
1173 pandecode_log("union mali_attr %s_%d[] = {\n", prefix
, job_no
);
1176 for (int i
= 0; i
< count
; ++i
) {
1177 pandecode_log("{\n");
1180 unsigned mode
= attr
[i
].elements
& 7;
1181 pandecode_prop("elements = (%s_%d_p) | %s", base
, i
, pandecode_attr_mode(mode
));
1182 pandecode_prop("shift = %d", attr
[i
].shift
);
1183 pandecode_prop("extra_flags = %d", attr
[i
].extra_flags
);
1184 pandecode_prop("stride = 0x%" PRIx32
, attr
[i
].stride
);
1185 pandecode_prop("size = 0x%" PRIx32
, attr
[i
].size
);
1187 /* Decode further where possible */
1189 if (mode
== MALI_ATTR_MODULO
) {
1190 pandecode_padded_vertices(
1192 attr
[i
].extra_flags
);
1196 pandecode_log("}, \n");
1198 if (mode
== MALI_ATTR_NPOT_DIVIDE
) {
1200 pandecode_log("{\n");
1202 pandecode_prop("unk = 0x%x", attr
[i
].unk
);
1203 pandecode_prop("magic_divisor = 0x%08x", attr
[i
].magic_divisor
);
1204 if (attr
[i
].zero
!= 0)
1205 pandecode_prop("XXX: zero tripped (0x%x)\n", attr
[i
].zero
);
1206 pandecode_prop("divisor = %d", attr
[i
].divisor
);
1207 pandecode_magic_divisor(attr
[i
].magic_divisor
, attr
[i
- 1].shift
, attr
[i
].divisor
, attr
[i
- 1].extra_flags
);
1209 pandecode_log("}, \n");
1215 pandecode_log("};\n");
1219 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1221 /* TODO: Decode flags */
1222 mali_ptr shader_ptr
= ptr
& ~15;
1224 char *a
= pointer_as_memory_reference(shader_ptr
);
1225 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1232 all_zero(unsigned *buffer
, unsigned count
)
1234 for (unsigned i
= 0; i
< count
; ++i
) {
1243 pandecode_stencil(const char *name
, const struct mali_stencil_test
*stencil
)
1245 if (all_zero((unsigned *) stencil
, sizeof(stencil
) / sizeof(unsigned)))
1248 const char *func
= pandecode_func(stencil
->func
);
1249 const char *sfail
= pandecode_stencil_op(stencil
->sfail
);
1250 const char *dpfail
= pandecode_stencil_op(stencil
->dpfail
);
1251 const char *dppass
= pandecode_stencil_op(stencil
->dppass
);
1254 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil
->zero
);
1256 pandecode_log(".stencil_%s = {\n", name
);
1258 pandecode_prop("ref = %d", stencil
->ref
);
1259 pandecode_prop("mask = 0x%02X", stencil
->mask
);
1260 pandecode_prop("func = %s", func
);
1261 pandecode_prop("sfail = %s", sfail
);
1262 pandecode_prop("dpfail = %s", dpfail
);
1263 pandecode_prop("dppass = %s", dppass
);
1265 pandecode_log("},\n");
1269 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1272 pandecode_msg("XXX: blend zero tripped: %X\n", blend
->zero1
);
1274 pandecode_log(".equation = {\n");
1277 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1278 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1280 pandecode_log(".color_mask = ");
1281 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1282 pandecode_log_cont(",\n");
1285 pandecode_log("},\n");
1288 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1291 decode_bifrost_constant(u16 constant
)
1293 float lo
= (float) (constant
& 0xFF);
1294 float hi
= (float) (constant
>> 8);
1296 return (hi
/ 255.0) + (lo
/ 65535.0);
1300 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1302 struct bifrost_blend_rt
*b
=
1303 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1305 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1308 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1309 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1310 b
->constant
, decode_bifrost_constant(b
->constant
));
1312 /* TODO figure out blend shader enable bit */
1313 pandecode_blend_equation(&b
->equation
);
1314 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1315 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1316 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1319 pandecode_log("},\n");
1325 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1327 if (all_zero((unsigned *) blend
, sizeof(blend
) / sizeof(unsigned)))
1330 pandecode_log(".blend = {\n");
1334 pandecode_shader_address("shader", blend
->shader
);
1336 pandecode_blend_equation(&blend
->equation
);
1337 pandecode_prop("constant = %f", blend
->constant
);
1341 pandecode_log("},\n");
1343 /* Return blend shader to disassemble if present */
1344 return is_shader
? (blend
->shader
& ~0xF) : 0;
1348 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1350 struct midgard_blend_rt
*b
=
1351 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1353 /* Flags determine presence of blend shader */
1354 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1356 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1359 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1361 union midgard_blend blend
= b
->blend
;
1362 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1365 pandecode_log("};\n");
1371 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1374 char *prefix
= varying
? "varying" : "attribute";
1375 unsigned max_index
= 0;
1376 snprintf(base
, sizeof(base
), "%s_meta", prefix
);
1378 pandecode_log("struct mali_attr_meta %s_%d%s[] = {\n", base
, job_no
, suffix
);
1381 struct mali_attr_meta
*attr_meta
;
1382 mali_ptr p
= varying
? (v
->varying_meta
& ~0xF) : v
->attribute_meta
;
1384 struct pandecode_mapped_memory
*attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
);
1386 for (int i
= 0; i
< count
; ++i
, p
+= sizeof(struct mali_attr_meta
)) {
1387 attr_meta
= pandecode_fetch_gpu_mem(attr_mem
, p
,
1390 /* If the record is discard, it should be zero for everything else */
1392 if (attr_meta
->format
== MALI_VARYING_DISCARD
) {
1395 attr_meta
->unknown1
|
1396 attr_meta
->unknown3
|
1397 attr_meta
->src_offset
;
1400 pandecode_msg("XXX: expected empty record for varying discard\n");
1402 /* We want to look for a literal 0000 swizzle -- this
1403 * is not encoded with all zeroes, however */
1405 enum mali_channel z
= MALI_CHANNEL_ZERO
;
1406 unsigned zero_swizzle
= z
| (z
<< 3) | (z
<< 6) | (z
<< 9);
1407 bool good_swizzle
= attr_meta
->swizzle
== zero_swizzle
;
1410 pandecode_msg("XXX: expected zero swizzle for discard\n");
1413 pandecode_msg("XXX: cannot discard attribute\n");
1415 /* If we're all good, omit the record */
1416 if (!zero
&& varying
&& good_swizzle
) {
1417 pandecode_log("/* discarded varying */\n");
1422 pandecode_log("{\n");
1424 pandecode_prop("index = %d", attr_meta
->index
);
1426 if (attr_meta
->index
> max_index
)
1427 max_index
= attr_meta
->index
;
1428 pandecode_swizzle(attr_meta
->swizzle
);
1429 pandecode_prop("format = %s", pandecode_format(attr_meta
->format
));
1431 if (attr_meta
->unknown1
!= 0x2) {
1432 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1433 pandecode_prop("unknown1 = 0x%" PRIx64
, (u64
) attr_meta
->unknown1
);
1436 if (attr_meta
->unknown3
) {
1437 pandecode_msg("XXX: unexpected unknown3 set\n");
1438 pandecode_prop("unknown3 = 0x%" PRIx64
, (u64
) attr_meta
->unknown3
);
1441 pandecode_prop("src_offset = %d", attr_meta
->src_offset
);
1443 pandecode_log("},\n");
1448 pandecode_log("};\n");
1450 return count
? (max_index
+ 1) : 0;
1454 pandecode_indices(uintptr_t pindices
, uint32_t index_count
, int job_no
)
1456 struct pandecode_mapped_memory
*imem
= pandecode_find_mapped_gpu_mem_containing(pindices
);
1459 /* Indices are literally just a u32 array :) */
1461 uint32_t *PANDECODE_PTR_VAR(indices
, imem
, pindices
);
1463 pandecode_log("uint32_t indices_%d[] = {\n", job_no
);
1466 for (unsigned i
= 0; i
< (index_count
+ 1); i
+= 3)
1467 pandecode_log("%d, %d, %d,\n",
1473 pandecode_log("};\n");
1477 /* return bits [lo, hi) of word */
1479 bits(u32 word
, u32 lo
, u32 hi
)
1482 return word
; // avoid undefined behavior with the shift
1484 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1488 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool noninstanced
)
1490 pandecode_log_cont("{\n");
1493 /* Decode invocation_count. See the comment before the definition of
1494 * invocation_count for an explanation.
1497 unsigned size_x
= bits(p
->invocation_count
, 0, p
->size_y_shift
) + 1;
1498 unsigned size_y
= bits(p
->invocation_count
, p
->size_y_shift
, p
->size_z_shift
) + 1;
1499 unsigned size_z
= bits(p
->invocation_count
, p
->size_z_shift
, p
->workgroups_x_shift
) + 1;
1501 unsigned groups_x
= bits(p
->invocation_count
, p
->workgroups_x_shift
, p
->workgroups_y_shift
) + 1;
1502 unsigned groups_y
= bits(p
->invocation_count
, p
->workgroups_y_shift
, p
->workgroups_z_shift
) + 1;
1503 unsigned groups_z
= bits(p
->invocation_count
, p
->workgroups_z_shift
, 32) + 1;
1505 /* Even though we have this decoded, we want to ensure that the
1506 * representation is "unique" so we don't lose anything by printing only
1507 * the final result. More specifically, we need to check that we were
1508 * passed something in canonical form, since the definition per the
1509 * hardware is inherently not unique. How? Well, take the resulting
1510 * decode and pack it ourselves! If it is bit exact with what we
1511 * decoded, we're good to go. */
1513 struct mali_vertex_tiler_prefix ref
;
1514 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, noninstanced
);
1517 (p
->invocation_count
== ref
.invocation_count
) &&
1518 (p
->size_y_shift
== ref
.size_y_shift
) &&
1519 (p
->size_z_shift
== ref
.size_z_shift
) &&
1520 (p
->workgroups_x_shift
== ref
.workgroups_x_shift
) &&
1521 (p
->workgroups_y_shift
== ref
.workgroups_y_shift
) &&
1522 (p
->workgroups_z_shift
== ref
.workgroups_z_shift
) &&
1523 (p
->workgroups_x_shift_2
== ref
.workgroups_x_shift_2
);
1526 pandecode_msg("XXX: non-canonical workgroups packing\n");
1527 pandecode_msg("expected: %X, %d, %d, %d, %d, %d\n",
1528 ref
.invocation_count
,
1531 ref
.workgroups_x_shift
,
1532 ref
.workgroups_y_shift
,
1533 ref
.workgroups_z_shift
,
1534 ref
.workgroups_x_shift_2
);
1536 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1537 pandecode_prop("size_y_shift = %d", p
->size_y_shift
);
1538 pandecode_prop("size_z_shift = %d", p
->size_z_shift
);
1539 pandecode_prop("workgroups_x_shift = %d", p
->workgroups_x_shift
);
1540 pandecode_prop("workgroups_y_shift = %d", p
->workgroups_y_shift
);
1541 pandecode_prop("workgroups_z_shift = %d", p
->workgroups_z_shift
);
1542 pandecode_prop("workgroups_x_shift_2 = %d", p
->workgroups_x_shift_2
);
1545 /* Regardless, print the decode */
1546 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1547 size_x
, size_y
, size_z
,
1548 groups_x
, groups_y
, groups_z
);
1551 if (p
->unknown_draw
)
1552 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1554 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1556 if (p
->draw_mode
!= MALI_DRAW_NONE
)
1557 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p
->draw_mode
));
1559 /* Index count only exists for tiler jobs anyway */
1562 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1564 if (p
->offset_bias_correction
)
1565 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1567 /* TODO: Figure out what this is. It's not zero */
1568 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1571 pandecode_log("},\n");
1575 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1577 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1578 struct mali_uniform_buffer_meta
*PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1580 pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%"PRIx64
"_%d[] = {\n",
1584 for (int i
= 0; i
< ubufs_count
; i
++) {
1585 pandecode_log("{\n");
1588 unsigned size
= (ubufs
[i
].size
+ 1) * 16;
1589 mali_ptr addr
= ubufs
[i
].ptr
<< 2;
1591 pandecode_validate_buffer(addr
, size
);
1593 char *ptr
= pointer_as_memory_reference(ubufs
[i
].ptr
<< 2);
1594 pandecode_prop("size = %u", size
);
1595 pandecode_prop("ptr = (%s) >> 2", ptr
);
1597 pandecode_log("},\n");
1602 pandecode_log("};\n");
1606 pandecode_scratchpad(uintptr_t pscratchpad
, int job_no
, char *suffix
)
1609 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(pscratchpad
);
1611 struct bifrost_scratchpad
*PANDECODE_PTR_VAR(scratchpad
, mem
, pscratchpad
);
1613 if (scratchpad
->zero
) {
1614 pandecode_msg("XXX: scratchpad zero tripped");
1615 pandecode_prop("zero = 0x%x\n", scratchpad
->zero
);
1618 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64
"_%d%s = {\n", pscratchpad
, job_no
, suffix
);
1621 pandecode_prop("flags = 0x%x", scratchpad
->flags
);
1622 MEMORY_PROP(scratchpad
, gpu_scratchpad
);
1625 pandecode_log("};\n");
1628 static unsigned shader_id
= 0;
1631 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1632 bool is_bifrost
, unsigned nr_regs
)
1634 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1635 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1637 /* Compute maximum possible size */
1638 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1640 /* Print some boilerplate to clearly denote the assembly (which doesn't
1641 * obey indentation rules), and actually do the disassembly! */
1647 snprintf(prefix
, sizeof(prefix
) - 1, "shader%d - %s shader: ",
1649 (type
== JOB_TYPE_TILER
) ? "FRAGMENT" : "VERTEX");
1652 disassemble_bifrost(code
, sz
, false);
1654 disassemble_midgard(code
, sz
, true, nr_regs
, prefix
);
1661 pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix
*p
,
1662 int job_no
, enum mali_job_type job_type
,
1663 char *suffix
, bool is_bifrost
)
1665 mali_ptr shader_meta_ptr
= (u64
) (uintptr_t) (p
->_shader_upper
<< 4);
1666 struct pandecode_mapped_memory
*attr_mem
;
1668 unsigned rt_count
= 1;
1670 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1671 * are the only things actually needed from the FBD, vertex/tiler jobs
1672 * no longer reference the FBD -- instead, this field points to some
1673 * info about the scratchpad.
1676 pandecode_scratchpad(p
->framebuffer
& ~FBD_TYPE
, job_no
, suffix
);
1677 else if (p
->framebuffer
& MALI_MFBD
)
1678 rt_count
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->framebuffer
) & FBD_MASK
, job_no
, false);
1679 else if (job_type
== JOB_TYPE_COMPUTE
)
1680 pandecode_compute_fbd((u64
) (uintptr_t) p
->framebuffer
, job_no
);
1682 pandecode_sfbd((u64
) (uintptr_t) p
->framebuffer
, job_no
, false);
1684 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
1685 int texture_count
= 0, sampler_count
= 0;
1687 if (shader_meta_ptr
) {
1688 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr
);
1689 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, shader_meta_ptr
);
1691 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", shader_meta_ptr
, job_no
, suffix
);
1694 /* Save for dumps */
1695 attribute_count
= s
->attribute_count
;
1696 varying_count
= s
->varying_count
;
1697 texture_count
= s
->texture_count
;
1698 sampler_count
= s
->sampler_count
;
1701 uniform_count
= s
->bifrost2
.uniform_count
;
1702 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
1704 uniform_count
= s
->midgard1
.uniform_buffer_count
;
1705 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
1708 mali_ptr shader_ptr
= pandecode_shader_address("shader", s
->shader
);
1710 pandecode_prop("texture_count = %" PRId16
, s
->texture_count
);
1711 pandecode_prop("sampler_count = %" PRId16
, s
->sampler_count
);
1712 pandecode_prop("attribute_count = %" PRId16
, s
->attribute_count
);
1713 pandecode_prop("varying_count = %" PRId16
, s
->varying_count
);
1715 unsigned nr_registers
= 0;
1718 pandecode_log(".bifrost1 = {\n");
1721 pandecode_prop("uniform_buffer_count = %" PRId32
, s
->bifrost1
.uniform_buffer_count
);
1722 pandecode_prop("unk1 = 0x%" PRIx32
, s
->bifrost1
.unk1
);
1725 pandecode_log("},\n");
1727 pandecode_log(".midgard1 = {\n");
1730 pandecode_prop("uniform_count = %" PRId16
, s
->midgard1
.uniform_count
);
1731 pandecode_prop("uniform_buffer_count = %" PRId16
, s
->midgard1
.uniform_buffer_count
);
1732 pandecode_prop("work_count = %" PRId16
, s
->midgard1
.work_count
);
1733 nr_registers
= s
->midgard1
.work_count
;
1735 pandecode_log(".flags = ");
1736 pandecode_log_decoded_flags(shader_midgard1_flag_info
, s
->midgard1
.flags
);
1737 pandecode_log_cont(",\n");
1739 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->midgard1
.unknown2
);
1742 pandecode_log("},\n");
1745 if (s
->depth_units
|| s
->depth_factor
) {
1746 pandecode_prop("depth_factor = %f", s
->depth_factor
);
1747 pandecode_prop("depth_units = %f", s
->depth_units
);
1750 if (s
->alpha_coverage
) {
1751 bool invert_alpha_coverage
= s
->alpha_coverage
& 0xFFF0;
1752 uint16_t inverted_coverage
= invert_alpha_coverage
? ~s
->alpha_coverage
: s
->alpha_coverage
;
1754 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
1755 invert_alpha_coverage
? "~" : "",
1756 MALI_GET_ALPHA_COVERAGE(inverted_coverage
));
1759 if (s
->unknown2_3
|| s
->unknown2_4
) {
1760 pandecode_log(".unknown2_3 = ");
1762 int unknown2_3
= s
->unknown2_3
;
1763 int unknown2_4
= s
->unknown2_4
;
1765 /* We're not quite sure what these flags mean without the depth test, if anything */
1767 if (unknown2_3
& (MALI_DEPTH_TEST
| MALI_DEPTH_FUNC_MASK
)) {
1768 const char *func
= pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3
));
1769 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
1771 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
1774 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
1775 pandecode_log_cont(",\n");
1777 pandecode_log(".unknown2_4 = ");
1778 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
1779 pandecode_log_cont(",\n");
1782 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
1783 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
1784 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
1787 pandecode_stencil("front", &s
->stencil_front
);
1788 pandecode_stencil("back", &s
->stencil_back
);
1791 pandecode_log(".bifrost2 = {\n");
1794 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
1795 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
1796 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
1797 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
1800 pandecode_log("},\n");
1801 } else if (s
->midgard2
.unknown2_7
) {
1802 pandecode_log(".midgard2 = {\n");
1805 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
1807 pandecode_log("},\n");
1811 pandecode_prop("unknown2_8 = 0x%" PRIx32
, s
->unknown2_8
);
1814 /* TODO: Blend shaders routing/disasm */
1816 union midgard_blend blend
= s
->blend
;
1817 pandecode_midgard_blend(&blend
, false);
1821 pandecode_log("};\n");
1823 /* MRT blend fields are used whenever MFBD is used, with
1824 * per-RT descriptors */
1826 if (job_type
== JOB_TYPE_TILER
) {
1827 void* blend_base
= (void *) (s
+ 1);
1829 for (unsigned i
= 0; i
< rt_count
; i
++) {
1830 mali_ptr shader
= 0;
1833 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
1835 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
1838 pandecode_shader_disassemble(shader
, job_no
, job_type
, false, 0);
1842 if (shader_ptr
& ~0xF)
1843 pandecode_shader_disassemble(shader_ptr
, job_no
, job_type
, is_bifrost
, nr_registers
);
1845 pandecode_msg("<no shader>\n");
1848 struct pandecode_mapped_memory
*fmem
= pandecode_find_mapped_gpu_mem_containing(p
->viewport
);
1849 struct mali_viewport
*PANDECODE_PTR_VAR(f
, fmem
, p
->viewport
);
1851 pandecode_log("struct mali_viewport viewport_%"PRIx64
"_%d%s = {\n", p
->viewport
, job_no
, suffix
);
1854 pandecode_prop("clip_minx = %f", f
->clip_minx
);
1855 pandecode_prop("clip_miny = %f", f
->clip_miny
);
1856 pandecode_prop("clip_minz = %f", f
->clip_minz
);
1857 pandecode_prop("clip_maxx = %f", f
->clip_maxx
);
1858 pandecode_prop("clip_maxy = %f", f
->clip_maxy
);
1859 pandecode_prop("clip_maxz = %f", f
->clip_maxz
);
1861 /* Only the higher coordinates are MALI_POSITIVE scaled */
1863 pandecode_prop("viewport0 = { %d, %d }",
1864 f
->viewport0
[0], f
->viewport0
[1]);
1866 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
1867 f
->viewport1
[0] + 1, f
->viewport1
[1] + 1);
1870 pandecode_log("};\n");
1873 if (p
->attribute_meta
) {
1874 unsigned max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
1876 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
1877 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false);
1880 /* Varyings are encoded like attributes but not actually sent; we just
1881 * pass a zero buffer with the right stride/size set, (or whatever)
1882 * since the GPU will write to it itself */
1884 if (p
->varying_meta
) {
1885 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
1889 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
1891 /* Number of descriptors depends on whether there are
1892 * non-internal varyings */
1894 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true);
1897 if (p
->uniform_buffers
) {
1898 if (uniform_buffer_count
)
1899 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
1901 pandecode_msg("XXX: UBOs specified but not referenced\n");
1902 } else if (uniform_buffer_count
)
1903 pandecode_msg("XXX: UBOs referenced but not specified\n");
1905 /* We don't want to actually dump uniforms, but we do need to validate
1906 * that the counts we were given are sane */
1910 pandecode_validate_buffer(p
->uniforms
, uniform_count
* 16);
1912 pandecode_msg("XXX: Uniforms specified but not referenced");
1913 } else if (uniform_count
)
1914 pandecode_msg("XXX: UBOs referenced but not specified\n");
1916 if (p
->texture_trampoline
) {
1917 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(p
->texture_trampoline
);
1920 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
);
1922 pandecode_log("uint64_t texture_trampoline_%"PRIx64
"_%d[] = {\n", p
->texture_trampoline
, job_no
);
1925 for (int tex
= 0; tex
< texture_count
; ++tex
) {
1926 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
1927 char *a
= pointer_as_memory_reference(*u
);
1928 pandecode_log("%s,\n", a
);
1933 pandecode_log("};\n");
1935 /* Now, finally, descend down into the texture descriptor */
1936 for (int tex
= 0; tex
< texture_count
; ++tex
) {
1937 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
1938 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
1941 struct mali_texture_descriptor
*PANDECODE_PTR_VAR(t
, tmem
, *u
);
1943 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64
"_%d_%d = {\n", *u
, job_no
, tex
);
1946 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", t
->width
+ 1);
1947 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", t
->height
+ 1);
1948 pandecode_prop("depth = MALI_POSITIVE(%" PRId16
")", t
->depth
+ 1);
1949 pandecode_prop("array_size = MALI_POSITIVE(%" PRId16
")", t
->array_size
+ 1);
1950 pandecode_prop("unknown3 = %" PRId16
, t
->unknown3
);
1951 pandecode_prop("unknown3A = %" PRId8
, t
->unknown3A
);
1952 pandecode_prop("nr_mipmap_levels = %" PRId8
, t
->nr_mipmap_levels
);
1954 struct mali_texture_format f
= t
->format
;
1956 pandecode_log(".format = {\n");
1959 pandecode_swizzle(f
.swizzle
);
1960 pandecode_prop("format = %s", pandecode_format(f
.format
));
1961 pandecode_prop("type = %s", pandecode_texture_type(f
.type
));
1962 pandecode_prop("srgb = %" PRId32
, f
.srgb
);
1963 pandecode_prop("unknown1 = %" PRId32
, f
.unknown1
);
1964 pandecode_prop("usage2 = 0x%" PRIx32
, f
.usage2
);
1967 pandecode_log("},\n");
1969 pandecode_swizzle(t
->swizzle
);
1971 if (t
->swizzle_zero
) {
1972 /* Shouldn't happen */
1973 pandecode_msg("XXX: swizzle zero tripped\n");
1974 pandecode_prop("swizzle_zero = %d", t
->swizzle_zero
);
1977 pandecode_prop("unknown3 = 0x%" PRIx32
, t
->unknown3
);
1979 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
1980 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
1981 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
1983 pandecode_log(".payload = {\n");
1986 /* A bunch of bitmap pointers follow.
1987 * We work out the correct number,
1988 * based on the mipmap/cubemap
1989 * properties, but dump extra
1990 * possibilities to futureproof */
1992 int bitmap_count
= MALI_NEGATIVE(t
->nr_mipmap_levels
);
1993 bool manual_stride
= f
.usage2
& MALI_TEX_MANUAL_STRIDE
;
1995 /* Miptree for each face */
1996 if (f
.type
== MALI_TEX_CUBE
)
1999 /* Array of textures */
2000 bitmap_count
*= MALI_NEGATIVE(t
->array_size
);
2002 /* Stride for each element */
2006 /* Sanity check the size */
2007 int max_count
= sizeof(t
->payload
) / sizeof(t
->payload
[0]);
2008 assert (bitmap_count
<= max_count
);
2010 for (int i
= 0; i
< bitmap_count
; ++i
) {
2011 /* How we dump depends if this is a stride or a pointer */
2013 if ((f
.usage2
& MALI_TEX_MANUAL_STRIDE
) && (i
& 1)) {
2014 /* signed 32-bit snuck in as a 64-bit pointer */
2015 uint64_t stride_set
= t
->payload
[i
];
2016 uint32_t clamped_stride
= stride_set
;
2017 int32_t stride
= clamped_stride
;
2018 assert(stride_set
== clamped_stride
);
2019 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
2021 char *a
= pointer_as_memory_reference(t
->payload
[i
]);
2022 pandecode_log("%s, \n", a
);
2028 pandecode_log("},\n");
2031 pandecode_log("};\n");
2037 if (p
->sampler_descriptor
) {
2038 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->sampler_descriptor
);
2041 struct mali_sampler_descriptor
*s
;
2043 mali_ptr d
= p
->sampler_descriptor
;
2045 for (int i
= 0; i
< sampler_count
; ++i
) {
2046 s
= pandecode_fetch_gpu_mem(smem
, d
+ sizeof(*s
) * i
, sizeof(*s
));
2048 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", d
+ sizeof(*s
) * i
, job_no
, i
);
2051 pandecode_log(".filter_mode = ");
2052 pandecode_log_decoded_flags(sampler_flag_info
, s
->filter_mode
);
2053 pandecode_log_cont(",\n");
2055 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
2056 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
2058 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s
->wrap_s
));
2059 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s
->wrap_t
));
2060 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s
->wrap_r
));
2062 pandecode_prop("compare_func = %s", pandecode_alt_func(s
->compare_func
));
2064 if (s
->zero
|| s
->zero2
) {
2065 pandecode_msg("XXX: sampler zero tripped\n");
2066 pandecode_prop("zero = 0x%X, 0x%X\n", s
->zero
, s
->zero2
);
2069 pandecode_prop("seamless_cube_map = %d", s
->seamless_cube_map
);
2071 pandecode_prop("border_color = { %f, %f, %f, %f }",
2075 s
->border_color
[3]);
2078 pandecode_log("};\n");
2085 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2087 if (!(p
->position_varying
|| p
->occlusion_counter
|| p
->flags
))
2090 pandecode_log(".postfix = {\n");
2093 MEMORY_PROP(p
, position_varying
);
2094 MEMORY_PROP(p
, occlusion_counter
);
2097 pandecode_prop("flags = %d", p
->flags
);
2100 pandecode_log("},\n");
2104 pandecode_vertex_only_bfr(struct bifrost_vertex_only
*v
)
2106 pandecode_log_cont("{\n");
2109 pandecode_prop("unk2 = 0x%x", v
->unk2
);
2111 if (v
->zero0
|| v
->zero1
) {
2112 pandecode_msg("XXX: vertex only zero tripped");
2113 pandecode_prop("zero0 = 0x%" PRIx32
, v
->zero0
);
2114 pandecode_prop("zero1 = 0x%" PRIx64
, v
->zero1
);
2118 pandecode_log("}\n");
2122 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2125 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2126 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2128 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no
);
2132 pandecode_msg("XXX: tiler heap zero tripped\n");
2133 pandecode_prop("zero = 0x%x", h
->zero
);
2136 for (int i
= 0; i
< 12; i
++) {
2137 if (h
->zeros
[i
] != 0) {
2138 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2143 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2144 MEMORY_PROP(h
, tiler_heap_start
);
2145 MEMORY_PROP(h
, tiler_heap_free
);
2147 /* this might point to the beginning of another buffer, when it's
2148 * really the end of the tiler heap buffer, so we have to be careful
2149 * here. but for zero length, we need the same pointer.
2152 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2153 MEMORY_PROP(h
, tiler_heap_start
);
2155 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2156 pandecode_prop("tiler_heap_end = %s + 1", a
);
2161 pandecode_log("};\n");
2165 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2167 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2168 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2170 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2172 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no
);
2175 if (t
->zero0
|| t
->zero1
) {
2176 pandecode_msg("XXX: tiler meta zero tripped\n");
2177 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2178 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2181 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2182 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2184 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2185 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2187 for (int i
= 0; i
< 12; i
++) {
2188 if (t
->zeros
[i
] != 0) {
2189 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2195 pandecode_log("};\n");
2199 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2201 pandecode_log(".gl_enables = ");
2203 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2205 pandecode_log_cont(",\n");
2209 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2211 if (u
.pointer
== 0x0)
2214 pandecode_log(".primitive_size = {\n");
2218 pandecode_prop("constant = %f", u
.constant
);
2220 MEMORY_PROP((&u
), pointer
);
2224 pandecode_log("},\n");
2228 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2230 pandecode_log_cont("{\n");
2233 /* TODO: gl_PointSize on Bifrost */
2234 pandecode_primitive_size(t
->primitive_size
, true);
2236 pandecode_gl_enables(t
->gl_enables
, JOB_TYPE_TILER
);
2238 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2239 || t
->zero6
|| t
->zero7
|| t
->zero8
) {
2240 pandecode_msg("XXX: tiler only zero tripped\n");
2241 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2242 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2243 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2244 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2245 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2246 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2247 pandecode_prop("zero7 = 0x%" PRIx32
, t
->zero7
);
2248 pandecode_prop("zero8 = 0x%" PRIx64
, t
->zero8
);
2252 pandecode_log("},\n");
2256 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2257 const struct pandecode_mapped_memory
*mem
,
2258 mali_ptr payload
, int job_no
)
2260 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2262 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true);
2264 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no
);
2267 pandecode_log(".prefix = ");
2268 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2270 pandecode_log(".vertex = ");
2271 pandecode_vertex_only_bfr(&v
->vertex
);
2273 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2276 pandecode_log("};\n");
2282 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2283 const struct pandecode_mapped_memory
*mem
,
2284 mali_ptr payload
, int job_no
)
2286 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2288 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true);
2290 pandecode_indices(t
->prefix
.indices
, t
->prefix
.index_count
, job_no
);
2291 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2293 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no
);
2296 pandecode_log(".prefix = ");
2297 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2299 pandecode_log(".tiler = ");
2300 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2302 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2305 pandecode_log("};\n");
2311 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2312 const struct pandecode_mapped_memory
*mem
,
2313 mali_ptr payload
, int job_no
)
2315 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2317 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false);
2319 pandecode_indices(v
->prefix
.indices
, v
->prefix
.index_count
, job_no
);
2321 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2324 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2325 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2327 bool instanced
= v
->instance_shift
|| v
->instance_odd
;
2328 bool is_graphics
= (h
->job_type
== JOB_TYPE_VERTEX
) || (h
->job_type
== JOB_TYPE_TILER
);
2330 pandecode_log(".prefix = ");
2331 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, !instanced
&& is_graphics
);
2333 pandecode_gl_enables(v
->gl_enables
, h
->job_type
);
2335 if (v
->instance_shift
|| v
->instance_odd
) {
2336 pandecode_prop("instance_shift = 0x%d /* %d */",
2337 v
->instance_shift
, 1 << v
->instance_shift
);
2338 pandecode_prop("instance_odd = 0x%X /* %d */",
2339 v
->instance_odd
, (2 * v
->instance_odd
) + 1);
2341 pandecode_padded_vertices(v
->instance_shift
, v
->instance_odd
);
2344 if (v
->offset_start
)
2345 pandecode_prop("offset_start = %d", v
->offset_start
);
2348 pandecode_msg("XXX: midgard payload zero tripped\n");
2349 pandecode_prop("zero5 = 0x%" PRIx64
, v
->zero5
);
2352 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2355 pandecode_log("};\n");
2361 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2362 mali_ptr payload
, int job_no
,
2365 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2367 bool fbd_dumped
= false;
2369 if (!is_bifrost
&& (s
->framebuffer
& FBD_TYPE
) == MALI_SFBD
) {
2370 /* Only SFBDs are understood, not MFBDs. We're speculating,
2371 * based on the versioning, kernel code, etc, that the
2372 * difference is between Single FrameBuffer Descriptor and
2373 * Multiple FrmaeBuffer Descriptor; the change apparently lines
2374 * up with multi-framebuffer support being added (T7xx onwards,
2375 * including Gxx). In any event, there's some field shuffling
2376 * that we haven't looked into yet. */
2378 pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true);
2380 } else if ((s
->framebuffer
& FBD_TYPE
) == MALI_MFBD
) {
2381 /* We don't know if Bifrost supports SFBD's at all, since the
2382 * driver never uses them. And the format is different from
2383 * Midgard anyways, due to the tiler heap and scratchpad being
2384 * moved out into separate structures, so it's not clear what a
2385 * Bifrost SFBD would even look like without getting an actual
2386 * trace, which appears impossible.
2389 pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true);
2393 uintptr_t p
= (uintptr_t) s
->framebuffer
& FBD_MASK
;
2394 pandecode_log("struct mali_payload_fragment payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2397 /* See the comments by the macro definitions for mathematical context
2398 * on why this is so weird */
2400 if (MALI_TILE_COORD_FLAGS(s
->max_tile_coord
) || MALI_TILE_COORD_FLAGS(s
->min_tile_coord
))
2401 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
2403 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
2404 MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
,
2405 MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
);
2407 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
2408 (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
,
2409 (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
);
2411 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2412 * we have to fallback on offsets. */
2414 const char *fbd_type
= s
->framebuffer
& MALI_MFBD
? "MALI_MFBD" : "MALI_SFBD";
2417 unsigned extra_flags
= (s
->framebuffer
& ~FBD_MASK
) & ~MALI_MFBD
;
2420 pandecode_prop("framebuffer = framebuffer_%d_p | %s | 0x%X", job_no
,
2421 fbd_type
, extra_flags
);
2423 pandecode_prop("framebuffer = %s | %s | 0x%X", pointer_as_memory_reference(p
),
2424 fbd_type
, extra_flags
);
2427 pandecode_log("};\n");
2432 static int job_descriptor_number
= 0;
2435 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
)
2437 struct mali_job_descriptor_header
*h
;
2439 int start_number
= 0;
2445 struct pandecode_mapped_memory
*mem
=
2446 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2450 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2452 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2453 * high 32-bits of the 64-bit pointer are reused to store
2456 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
2457 h
->job_type
!= JOB_TYPE_FRAGMENT
? 4 : 0;
2458 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
2460 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
,
2463 int job_no
= job_descriptor_number
++;
2466 start_number
= job_no
;
2468 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2471 pandecode_prop("job_type = %s", pandecode_job_type(h
->job_type
));
2473 /* Save for next job fixing */
2474 last_size
= h
->job_descriptor_size
;
2476 if (h
->job_descriptor_size
)
2477 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2479 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2480 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2481 h
->exception_status
,
2482 (h
->exception_status
>> 16) & 0xFFFF,
2483 pandecode_exception_access((h
->exception_status
>> 8) & 0x3),
2484 h
->exception_status
& 0xFF);
2486 if (h
->first_incomplete_task
)
2487 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2489 if (h
->fault_pointer
)
2490 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2493 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2495 pandecode_prop("job_index = %d", h
->job_index
);
2497 if (h
->unknown_flags
)
2498 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2500 if (h
->job_dependency_index_1
)
2501 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2503 if (h
->job_dependency_index_2
)
2504 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2507 pandecode_log("};\n");
2509 /* Do not touch the field yet -- decode the payload first, and
2510 * don't touch that either. This is essential for the uploads
2511 * to occur in sequence and therefore be dynamically allocated
2512 * correctly. Do note the size, however, for that related
2515 switch (h
->job_type
) {
2516 case JOB_TYPE_SET_VALUE
: {
2517 struct mali_payload_set_value
*s
= payload
;
2518 pandecode_log("struct mali_payload_set_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2520 MEMORY_PROP(s
, out
);
2521 pandecode_prop("unknown = 0x%" PRIX64
, s
->unknown
);
2523 pandecode_log("};\n");
2528 case JOB_TYPE_TILER
:
2529 case JOB_TYPE_VERTEX
:
2530 case JOB_TYPE_COMPUTE
:
2532 if (h
->job_type
== JOB_TYPE_TILER
)
2533 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
);
2535 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
);
2537 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
);
2541 case JOB_TYPE_FRAGMENT
:
2542 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
);
2549 /* Handle linkage */
2552 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no
- 1);
2555 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no
);
2557 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no
);
2562 } while ((jc_gpu_va
= h
->job_descriptor_size
? h
->next_job_64
: h
->next_job_32
));
2564 return start_number
;