2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
26 #include <panfrost-job.h>
34 #include "util/macros.h"
35 #include "util/u_math.h"
37 #include "pan_pretty_print.h"
38 #include "midgard/disassemble.h"
39 #include "bifrost/disassemble.h"
41 #include "pan_encoder.h"
43 int pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
);
45 #define MEMORY_PROP(obj, p) {\
47 char *a = pointer_as_memory_reference(obj->p); \
48 pandecode_prop("%s = %s", #p, a); \
53 #define MEMORY_PROP_DIR(obj, p) {\
55 char *a = pointer_as_memory_reference(obj.p); \
56 pandecode_prop("%s = %s", #p, a); \
61 /* Semantic logging type.
63 * Raw: for raw messages to be printed as is.
64 * Message: for helpful information to be commented out in replays.
65 * Property: for properties of a struct
67 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
70 enum pandecode_log_type
{
76 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
77 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
78 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
80 unsigned pandecode_indent
= 0;
83 pandecode_make_indent(void)
85 for (unsigned i
= 0; i
< pandecode_indent
; ++i
)
90 pandecode_log_typed(enum pandecode_log_type type
, const char *format
, ...)
94 pandecode_make_indent();
96 if (type
== PANDECODE_MESSAGE
)
98 else if (type
== PANDECODE_PROPERTY
)
101 va_start(ap
, format
);
105 if (type
== PANDECODE_PROPERTY
)
110 pandecode_log_cont(const char *format
, ...)
114 va_start(ap
, format
);
119 /* To check for memory safety issues, validates that the given pointer in GPU
120 * memory is valid, containing at least sz bytes. The goal is to eliminate
121 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
122 * overruns) by statically validating pointers.
126 pandecode_validate_buffer(mali_ptr addr
, size_t sz
)
129 pandecode_msg("XXX: null pointer deref");
135 struct pandecode_mapped_memory
*bo
=
136 pandecode_find_mapped_gpu_mem_containing(addr
);
139 pandecode_msg("XXX: invalid memory dereference\n");
145 unsigned offset
= addr
- bo
->gpu_va
;
146 unsigned total
= offset
+ sz
;
148 if (total
> bo
->length
) {
149 pandecode_msg("XXX: buffer overrun. "
150 "Chunk of size %zu at offset %d in buffer of size %zu. "
151 "Overrun by %zu bytes. \n",
152 sz
, offset
, bo
->length
, total
- bo
->length
);
157 struct pandecode_flag_info
{
163 pandecode_log_decoded_flags(const struct pandecode_flag_info
*flag_info
,
166 bool decodable_flags_found
= false;
168 for (int i
= 0; flag_info
[i
].name
; i
++) {
169 if ((flags
& flag_info
[i
].flag
) != flag_info
[i
].flag
)
172 if (!decodable_flags_found
) {
173 decodable_flags_found
= true;
175 pandecode_log_cont(" | ");
178 pandecode_log_cont("%s", flag_info
[i
].name
);
180 flags
&= ~flag_info
[i
].flag
;
183 if (decodable_flags_found
) {
185 pandecode_log_cont(" | 0x%" PRIx64
, flags
);
187 pandecode_log_cont("0x%" PRIx64
, flags
);
191 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
192 static const struct pandecode_flag_info gl_enable_flag_info
[] = {
193 FLAG_INFO(OCCLUSION_QUERY
),
194 FLAG_INFO(OCCLUSION_PRECISE
),
195 FLAG_INFO(FRONT_CCW_TOP
),
196 FLAG_INFO(CULL_FACE_FRONT
),
197 FLAG_INFO(CULL_FACE_BACK
),
202 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
203 static const struct pandecode_flag_info clear_flag_info
[] = {
206 FLAG_INFO(SLOW_STENCIL
),
211 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
212 static const struct pandecode_flag_info mask_flag_info
[] = {
221 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
222 static const struct pandecode_flag_info u3_flag_info
[] = {
224 FLAG_INFO(CAN_DISCARD
),
225 FLAG_INFO(HAS_BLEND_SHADER
),
226 FLAG_INFO(DEPTH_WRITEMASK
),
230 static const struct pandecode_flag_info u4_flag_info
[] = {
232 FLAG_INFO(NO_DITHER
),
233 FLAG_INFO(DEPTH_RANGE_A
),
234 FLAG_INFO(DEPTH_RANGE_B
),
235 FLAG_INFO(STENCIL_TEST
),
236 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER
),
241 #define FLAG_INFO(flag) { MALI_FRAMEBUFFER_##flag, "MALI_FRAMEBUFFER_" #flag }
242 static const struct pandecode_flag_info fb_fmt_flag_info
[] = {
250 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
251 static const struct pandecode_flag_info mfbd_fmt_flag_info
[] = {
258 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
259 static const struct pandecode_flag_info mfbd_extra_flag_info
[] = {
267 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
268 static const struct pandecode_flag_info shader_midgard1_flag_info
[] = {
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 static char *pandecode_attr_mode_short(enum mali_attr_mode mode
)
418 /* TODO: Combine to just "instanced" once this can be done
419 * unambiguously in all known cases */
420 case MALI_ATTR_POT_DIVIDE
:
421 return "instanced_pot";
422 case MALI_ATTR_MODULO
:
423 return "instanced_mod";
424 case MALI_ATTR_NPOT_DIVIDE
:
425 return "instanced_npot";
426 case MALI_ATTR_IMAGE
:
428 case MALI_ATTR_INTERNAL
:
431 pandecode_msg("XXX: invalid attribute mode %X\n", mode
);
437 pandecode_special_varying(uint64_t v
)
440 case MALI_VARYING_FRAG_COORD
:
441 return "gl_FragCoord";
442 case MALI_VARYING_FRONT_FACING
:
443 return "gl_FrontFacing";
444 case MALI_VARYING_POINT_COORD
:
445 return "gl_PointCoord";
447 pandecode_msg("XXX: invalid special varying %" PRIx64
"\n", v
);
452 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
454 pandecode_wrap_mode(enum mali_wrap_mode op
)
458 DEFINE_CASE(CLAMP_TO_EDGE
);
459 DEFINE_CASE(CLAMP_TO_BORDER
);
460 DEFINE_CASE(MIRRORED_REPEAT
);
463 pandecode_msg("XXX: invalid wrap mode %X\n", op
);
469 #define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
471 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt
)
475 DEFINE_CASE(UNKNOWN
);
480 unreachable("Invalid case");
485 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
487 pandecode_exception_access(enum mali_exception_access access
)
491 DEFINE_CASE(EXECUTE
);
496 unreachable("Invalid case");
501 /* Midgard's tiler descriptor is embedded within the
505 pandecode_midgard_tiler_descriptor(
506 const struct midgard_tiler_descriptor
*t
,
511 pandecode_log(".tiler = {\n");
514 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
515 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
517 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
519 /* We know this name from the kernel, but we never see it nonzero */
522 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
524 MEMORY_PROP(t
, polygon_list
);
526 /* The body is offset from the base of the polygon list */
527 assert(t
->polygon_list_body
> t
->polygon_list
);
528 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
530 /* It needs to fit inside the reported size */
531 assert(t
->polygon_list_size
>= body_offset
);
533 /* Check that we fit */
534 struct pandecode_mapped_memory
*plist
=
535 pandecode_find_mapped_gpu_mem_containing(t
->polygon_list
);
537 assert(t
->polygon_list_size
<= plist
->length
);
539 /* Now that we've sanity checked, we'll try to calculate the sizes
540 * ourselves for comparison */
542 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
);
543 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
);
545 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
546 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
547 ref_header
, ref_size
);
548 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
549 pandecode_msg("body offset %d\n", body_offset
);
552 /* The tiler heap has a start and end specified -- it should be
553 * identical to what we have in the BO. The exception is if tiling is
556 MEMORY_PROP(t
, heap_start
);
557 assert(t
->heap_end
>= t
->heap_start
);
559 struct pandecode_mapped_memory
*heap
=
560 pandecode_find_mapped_gpu_mem_containing(t
->heap_start
);
562 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
564 /* Tiling is enabled with a special flag */
565 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
566 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
568 bool tiling_enabled
= hierarchy_mask
;
570 if (tiling_enabled
) {
571 /* When tiling is enabled, the heap should be a tight fit */
572 unsigned heap_offset
= t
->heap_start
- heap
->gpu_va
;
573 if ((heap_offset
+ heap_size
) != heap
->length
) {
574 pandecode_msg("XXX: heap size %u (expected %zu)\n",
575 heap_size
, heap
->length
- heap_offset
);
578 /* We should also have no other flags */
580 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
582 /* When tiling is disabled, we should have that flag and no others */
584 if (tiler_flags
!= MALI_TILER_DISABLED
) {
585 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
589 /* We should also have an empty heap */
591 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
595 /* Disabled tiling is used only for clear-only jobs, which are
596 * purely FRAGMENT, so we should never see this for
597 * non-FRAGMENT descriptors. */
600 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
603 /* We've never seen weights used in practice, but we know from the
604 * kernel these fields is there */
606 bool nonzero_weights
= false;
608 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
609 nonzero_weights
|= t
->weights
[w
] != 0x0;
612 if (nonzero_weights
) {
613 pandecode_log(".weights = {");
615 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
616 pandecode_log("%d, ", t
->weights
[w
]);
623 pandecode_log("}\n");
626 /* Information about the framebuffer passed back for
627 * additional analysis */
629 struct pandecode_fbd
{
636 static struct pandecode_fbd
637 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
)
639 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
640 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
642 struct pandecode_fbd info
= {
647 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
650 pandecode_prop("unknown1 = 0x%" PRIx32
, s
->unknown1
);
651 pandecode_prop("unknown2 = 0x%" PRIx32
, s
->unknown2
);
653 pandecode_log(".format = ");
654 pandecode_log_decoded_flags(fb_fmt_flag_info
, s
->format
);
655 pandecode_log_cont(",\n");
657 info
.width
= s
->width
+ 1;
658 info
.height
= s
->height
+ 1;
660 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", info
.width
);
661 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", info
.height
);
663 MEMORY_PROP(s
, checksum
);
665 if (s
->checksum_stride
)
666 pandecode_prop("checksum_stride = %d", s
->checksum_stride
);
668 MEMORY_PROP(s
, framebuffer
);
669 pandecode_prop("stride = %d", s
->stride
);
671 /* Earlier in the actual commandstream -- right before width -- but we
672 * delay to flow nicer */
674 pandecode_log(".clear_flags = ");
675 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
676 pandecode_log_cont(",\n");
678 if (s
->depth_buffer
| s
->depth_buffer_enable
) {
679 MEMORY_PROP(s
, depth_buffer
);
680 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s
->depth_buffer_enable
));
683 if (s
->stencil_buffer
| s
->stencil_buffer_enable
) {
684 MEMORY_PROP(s
, stencil_buffer
);
685 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s
->stencil_buffer_enable
));
688 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
689 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
690 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
691 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
692 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
695 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
696 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
697 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
698 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
699 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
702 if (s
->clear_stencil
) {
703 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
706 MEMORY_PROP(s
, unknown_address_0
);
707 const struct midgard_tiler_descriptor t
= s
->tiler
;
708 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
);
711 pandecode_log("};\n");
713 pandecode_prop("zero0 = 0x%" PRIx64
, s
->zero0
);
714 pandecode_prop("zero1 = 0x%" PRIx64
, s
->zero1
);
715 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
716 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
717 pandecode_prop("zero5 = 0x%" PRIx32
, s
->zero4
);
719 printf(".zero3 = {");
721 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
722 printf("%X, ", s
->zero3
[i
]);
726 printf(".zero6 = {");
728 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
729 printf("%X, ", s
->zero6
[i
]);
737 pandecode_u32_slide(unsigned name
, const u32
*slide
, unsigned count
)
739 pandecode_log(".unknown%d = {", name
);
741 for (int i
= 0; i
< count
; ++i
)
742 printf("%X, ", slide
[i
]);
744 pandecode_log("},\n");
747 #define SHORT_SLIDE(num) \
748 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
751 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
753 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
754 const struct mali_compute_fbd
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
756 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
765 /* Extracts the number of components associated with a Mali format */
768 pandecode_format_component_count(enum mali_format fmt
)
770 /* Mask out the format class */
771 unsigned top
= fmt
& 0b11100000;
774 case MALI_FORMAT_SNORM
:
775 case MALI_FORMAT_UINT
:
776 case MALI_FORMAT_UNORM
:
777 case MALI_FORMAT_SINT
:
778 return ((fmt
>> 3) & 3) + 1;
785 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
788 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
791 assert(MALI_CHANNEL_RED
== 0);
793 for (unsigned c
= 0; c
< 4; ++c
) {
794 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
796 if (chan
<= MALI_CHANNEL_ALPHA
)
803 /* Validates that a (format, swizzle) pair is valid, in the sense that the
804 * swizzle doesn't access any components that are undefined in the format.
805 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
809 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
811 unsigned nr_comp
= pandecode_format_component_count(fmt
);
812 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
813 unsigned valid_mask
= (1 << nr_comp
) - 1;
814 unsigned invalid_mask
= ~valid_mask
;
816 if (access_mask
& invalid_mask
) {
817 pandecode_msg("XXX: invalid components accessed\n");
821 /* Check for the default non-swizzling swizzle so we can suppress
822 * useless printing for the defaults */
824 unsigned default_swizzles
[4] = {
825 MALI_CHANNEL_RED
| (MALI_CHANNEL_ZERO
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
826 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
827 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ONE
<< 9),
828 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ALPHA
<< 9)
831 return (swizzle
== default_swizzles
[nr_comp
- 1]);
834 /* Maps MALI_RGBA32F to rgba32f, etc */
837 pandecode_format_short(enum mali_format fmt
, bool srgb
)
839 /* We want a type-like format, so cut off the initial MALI_ */
840 char *format
= pandecode_format(fmt
);
841 format
+= strlen("MALI_");
843 unsigned len
= strlen(format
);
844 char *lower_format
= calloc(1, len
+ 1);
846 for (unsigned i
= 0; i
< len
; ++i
)
847 lower_format
[i
] = tolower(format
[i
]);
849 /* Sanity check sRGB flag is applied to RGB, per the name */
850 if (srgb
&& lower_format
[0] != 'r')
851 pandecode_msg("XXX: sRGB applied to non-colour format\n");
853 /* Just prefix with an s, so you get formats like srgba8_unorm */
855 pandecode_log_cont("s");
857 pandecode_log_cont("%s", lower_format
);
862 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
864 /* First, do some validation */
865 bool trivial_swizzle
= pandecode_validate_format_swizzle(
871 /* Next, print the swizzle */
872 pandecode_log_cont(".");
874 static const char components
[] = "rgba01";
876 for (unsigned c
= 0; c
< 4; ++c
) {
877 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
879 if (chan
>= MALI_CHANNEL_RESERVED_0
) {
880 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
883 pandecode_log_cont("%c", components
[chan
]);
888 pandecode_rt_format(struct mali_rt_format format
)
890 pandecode_log(".format = {\n");
893 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
894 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
895 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
897 pandecode_prop("block = %s",
898 pandecode_mfbd_block_format(format
.block
));
900 /* TODO: Map formats so we can check swizzles and print nicely */
901 pandecode_log("swizzle");
902 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
903 pandecode_log_cont(",\n");
905 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
906 MALI_NEGATIVE(format
.nr_channels
));
908 pandecode_log(".flags = ");
909 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
910 pandecode_log_cont(",\n");
912 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
913 * which is a faster hardware-based alternative to the wallpaper method
914 * to preserve framebuffer contents across frames. In practice, MFBD
915 * preload is buggy on Midgard, and so this is a chicken bit. If this
916 * bit isn't set, most likely something broke unrelated to preload */
918 if (!format
.no_preload
) {
919 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
920 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
924 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
927 pandecode_log("},\n");
931 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct bifrost_framebuffer
*fb
)
933 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
936 for (int i
= 0; i
< MALI_NEGATIVE(fb
->rt_count_1
); i
++) {
937 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct bifrost_render_target
);
938 struct pandecode_mapped_memory
*mem
=
939 pandecode_find_mapped_gpu_mem_containing(rt_va
);
940 const struct bifrost_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
942 pandecode_log("{\n");
945 pandecode_rt_format(rt
->format
);
947 if (rt
->format
.block
== MALI_MFBD_BLOCK_AFBC
) {
948 pandecode_log(".afbc = {\n");
951 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
952 pandecode_prop("metadata = %s", a
);
955 pandecode_prop("stride = %d", rt
->afbc
.stride
);
956 pandecode_prop("unk = 0x%" PRIx32
, rt
->afbc
.unk
);
959 pandecode_log("},\n");
960 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.unk
) {
961 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
967 MEMORY_PROP(rt
, framebuffer
);
968 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
970 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
971 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
972 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
973 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
974 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
977 if (rt
->zero1
|| rt
->zero2
|| rt
->zero3
) {
978 pandecode_msg("XXX: render target zeros tripped\n");
979 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
980 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
981 pandecode_prop("zero3 = 0x%" PRIx32
, rt
->zero3
);
985 pandecode_log("},\n");
989 pandecode_log("};\n");
992 static struct pandecode_fbd
993 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
)
995 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
996 const struct bifrost_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
998 struct pandecode_fbd info
;
1000 if (fb
->sample_locations
) {
1001 /* The blob stores all possible sample locations in a single buffer
1002 * allocated on startup, and just switches the pointer when switching
1003 * MSAA state. For now, we just put the data into the cmdstream, but we
1004 * should do something like what the blob does with a real driver.
1006 * There seem to be 32 slots for sample locations, followed by another
1007 * 16. The second 16 is just the center location followed by 15 zeros
1008 * in all the cases I've identified (maybe shader vs. depth/color
1012 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->sample_locations
);
1014 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->sample_locations
);
1016 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
1019 for (int i
= 0; i
< 32 + 16; i
++) {
1020 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
1024 pandecode_log("};\n");
1027 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1030 pandecode_prop("unk0 = 0x%x", fb
->unk0
);
1032 if (fb
->sample_locations
)
1033 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
1035 /* Assume that unknown1 was emitted in the last job for
1037 MEMORY_PROP(fb
, unknown1
);
1039 info
.width
= fb
->width1
+ 1;
1040 info
.height
= fb
->height1
+ 1;
1041 info
.rt_count
= fb
->rt_count_1
+ 1;
1043 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
1044 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
1045 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
1046 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
1048 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
1049 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
1050 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
1051 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
1053 pandecode_log(".mfbd_flags = ");
1054 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
1055 pandecode_log_cont(",\n");
1057 if (fb
->clear_stencil
)
1058 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
1060 if (fb
->clear_depth
)
1061 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
1063 /* TODO: What is this? Let's not blow up.. */
1064 if (fb
->unknown2
!= 0x1F)
1065 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
1067 pandecode_prop("unknown2 = 0x%x", fb
->unknown2
);
1068 MEMORY_PROP(fb
, scratchpad
);
1069 const struct midgard_tiler_descriptor t
= fb
->tiler
;
1070 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
);
1072 if (fb
->zero3
|| fb
->zero4
) {
1073 pandecode_msg("XXX: framebuffer zeros tripped\n");
1074 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
1075 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
1079 pandecode_log("};\n");
1081 gpu_va
+= sizeof(struct bifrost_framebuffer
);
1083 info
.has_extra
= (fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
;
1085 if (info
.has_extra
) {
1086 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1087 const struct bifrost_fb_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
1089 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1092 MEMORY_PROP(fbx
, checksum
);
1094 if (fbx
->checksum_stride
)
1095 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
1097 pandecode_log(".flags = ");
1098 pandecode_log_decoded_flags(mfbd_extra_flag_info
, fbx
->flags
);
1099 pandecode_log_cont(",\n");
1101 if (fbx
->flags
& MALI_EXTRA_AFBC_ZS
) {
1102 pandecode_log(".ds_afbc = {\n");
1105 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
1106 pandecode_prop("depth_stencil_afbc_stride = %d",
1107 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
1108 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
1110 if (fbx
->ds_afbc
.zero1
|| fbx
->ds_afbc
.padding
) {
1111 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1112 pandecode_prop("zero1 = 0x%" PRIx32
,
1113 fbx
->ds_afbc
.zero1
);
1114 pandecode_prop("padding = 0x%" PRIx64
,
1115 fbx
->ds_afbc
.padding
);
1119 pandecode_log("},\n");
1121 pandecode_log(".ds_linear = {\n");
1124 if (fbx
->ds_linear
.depth
) {
1125 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
1126 pandecode_prop("depth_stride = %d",
1127 fbx
->ds_linear
.depth_stride
);
1130 if (fbx
->ds_linear
.stencil
) {
1131 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1132 pandecode_prop("stencil_stride = %d",
1133 fbx
->ds_linear
.stencil_stride
);
1136 if (fbx
->ds_linear
.depth_stride_zero
||
1137 fbx
->ds_linear
.stencil_stride_zero
||
1138 fbx
->ds_linear
.zero1
|| fbx
->ds_linear
.zero2
) {
1139 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1140 pandecode_prop("depth_stride_zero = 0x%x",
1141 fbx
->ds_linear
.depth_stride_zero
);
1142 pandecode_prop("stencil_stride_zero = 0x%x",
1143 fbx
->ds_linear
.stencil_stride_zero
);
1144 pandecode_prop("zero1 = 0x%" PRIx32
,
1145 fbx
->ds_linear
.zero1
);
1146 pandecode_prop("zero2 = 0x%" PRIx32
,
1147 fbx
->ds_linear
.zero2
);
1151 pandecode_log("},\n");
1154 if (fbx
->zero3
|| fbx
->zero4
) {
1155 pandecode_msg("XXX: fb_extra zeros tripped\n");
1156 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1157 pandecode_prop("zero4 = 0x%" PRIx64
, fbx
->zero4
);
1161 pandecode_log("};\n");
1163 gpu_va
+= sizeof(struct bifrost_fb_extra
);
1167 pandecode_render_target(gpu_va
, job_no
, fb
);
1172 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1176 pandecode_padded_vertices(unsigned shift
, unsigned k
)
1178 unsigned odd
= 2*k
+ 1;
1179 unsigned pot
= 1 << shift
;
1180 pandecode_msg("padded_num_vertices = %d\n", odd
* pot
);
1183 /* Given a magic divisor, recover what we were trying to divide by.
1185 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1186 * 0 <= m < 2^N, for N bits in m.
1188 * Let q represent the number we would like to divide by.
1190 * By definition of a magic divisor for N-bit unsigned integers (a number you
1191 * multiply by to magically get division), m is a number such that:
1193 * (m * x) & (2^N - 1) = floor(x/q).
1194 * for all x on Z where 0 <= x < 2^N
1196 * Ignore the case where any of the above values equals zero; it is irrelevant
1197 * for our purposes (instanced arrays).
1199 * Choose x = q. Then:
1201 * (m * x) & (2^N - 1) = floor(x/q).
1202 * (m * q) & (2^N - 1) = floor(q/q).
1204 * floor(q/q) = floor(1) = 1, therefore:
1206 * (m * q) & (2^N - 1) = 1
1208 * Recall the identity that the bitwise AND of one less than a power-of-two
1209 * equals the modulo with that power of two, i.e. for all x:
1211 * x & (2^N - 1) = x % N
1217 * By definition, a modular multiplicative inverse of a number m is the number
1218 * q such that with respect to a modulos M:
1222 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1227 pandecode_magic_divisor(uint32_t magic
, unsigned shift
, unsigned orig_divisor
, unsigned extra
)
1230 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1231 * shift) the most lame way possible... just repeatedly add.
1232 * Asymptoptically slow but nobody cares in practice, unless you have
1233 * massive numbers of vertices or high divisors. */
1235 unsigned inverse
= 0;
1237 /* Magic implicitly has the highest bit set */
1240 /* Depending on rounding direction */
1245 uint32_t product
= magic
* inverse
;
1257 pandecode_msg("dividing by %d (maybe off by two)\n", inverse
);
1259 /* Recall we're supposed to divide by (gl_level_divisor *
1260 * padded_num_vertices) */
1262 unsigned padded_num_vertices
= inverse
/ orig_divisor
;
1264 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices
);
1269 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1270 mali_ptr addr
, int job_no
, char *suffix
,
1271 int count
, bool varying
, enum mali_job_type job_type
)
1273 char *prefix
= varying
? "varying" : "attribute";
1277 pandecode_msg("warn: No %s records\n", prefix
);
1281 union mali_attr
*attr
= pandecode_fetch_gpu_mem(mem
, addr
, sizeof(union mali_attr
) * count
);
1283 for (int i
= 0; i
< count
; ++i
) {
1284 /* First, check for special records */
1285 if (attr
[i
].elements
< MALI_VARYING_SPECIAL
) {
1286 /* Special records are always varyings */
1289 pandecode_msg("XXX: Special varying in attribute field\n");
1291 if (job_type
!= JOB_TYPE_TILER
)
1292 pandecode_msg("XXX: Special varying in non-FS\n");
1294 /* We're special, so all fields should be zero */
1295 unsigned zero
= attr
[i
].stride
| attr
[i
].size
;
1296 zero
|= attr
[i
].shift
| attr
[i
].extra_flags
;
1299 pandecode_msg("XXX: Special varying has non-zero fields\n");
1301 /* Print the special varying name */
1302 pandecode_log("varying_%d = %s;", i
, pandecode_special_varying(attr
[i
].elements
));
1307 enum mali_attr_mode mode
= attr
[i
].elements
& 7;
1309 if (mode
== MALI_ATTR_UNUSED
)
1310 pandecode_msg("XXX: unused attribute record\n");
1312 /* For non-linear records, we need to print the type of record */
1313 if (mode
!= MALI_ATTR_LINEAR
)
1314 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode
));
1316 /* Print the name to link with attr_meta */
1317 pandecode_log_cont("%s_%d", prefix
, i
);
1319 /* Print the stride and size */
1320 pandecode_log_cont("<%u>[%u]", attr
[i
].stride
, attr
[i
].size
);
1322 /* TODO: Sanity check the quotient itself. It must be equal to
1323 * (or be greater than, if the driver added padding) the padded
1326 /* Finally, print the pointer */
1327 mali_ptr raw_elements
= attr
[i
].elements
& ~7;
1328 char *a
= pointer_as_memory_reference(raw_elements
);
1329 pandecode_log_cont(" = (%s);\n", a
);
1332 /* Check the pointer */
1333 pandecode_validate_buffer(raw_elements
, attr
[i
].size
);
1335 /* shift/extra_flags exist only for instanced */
1336 if (attr
[i
].shift
| attr
[i
].extra_flags
) {
1337 /* These are set to random values by the blob for
1338 * varyings, most likely a symptom of uninitialized
1339 * memory where the hardware masked the bug. As such we
1340 * put this at a warning, not an error. */
1342 if (mode
== MALI_ATTR_LINEAR
)
1343 pandecode_msg("warn: instancing fields set for linear\n");
1345 pandecode_prop("shift = %d", attr
[i
].shift
);
1346 pandecode_prop("extra_flags = %d", attr
[i
].extra_flags
);
1349 /* Decode further where possible */
1351 if (mode
== MALI_ATTR_MODULO
) {
1352 pandecode_padded_vertices(
1354 attr
[i
].extra_flags
);
1357 if (mode
== MALI_ATTR_NPOT_DIVIDE
) {
1359 pandecode_log("{\n");
1361 pandecode_prop("unk = 0x%x", attr
[i
].unk
);
1362 pandecode_prop("magic_divisor = 0x%08x", attr
[i
].magic_divisor
);
1363 if (attr
[i
].zero
!= 0)
1364 pandecode_prop("XXX: zero tripped (0x%x)\n", attr
[i
].zero
);
1365 pandecode_prop("divisor = %d", attr
[i
].divisor
);
1366 pandecode_magic_divisor(attr
[i
].magic_divisor
, attr
[i
- 1].shift
, attr
[i
].divisor
, attr
[i
- 1].extra_flags
);
1368 pandecode_log("}, \n");
1373 pandecode_log("\n");
1377 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1379 /* TODO: Decode flags */
1380 mali_ptr shader_ptr
= ptr
& ~15;
1382 char *a
= pointer_as_memory_reference(shader_ptr
);
1383 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1390 pandecode_stencil(const char *name
, const struct mali_stencil_test
*stencil
)
1392 unsigned any_nonzero
=
1393 stencil
->ref
| stencil
->mask
| stencil
->func
|
1394 stencil
->sfail
| stencil
->dpfail
| stencil
->dppass
;
1396 if (any_nonzero
== 0)
1399 const char *func
= pandecode_func(stencil
->func
);
1400 const char *sfail
= pandecode_stencil_op(stencil
->sfail
);
1401 const char *dpfail
= pandecode_stencil_op(stencil
->dpfail
);
1402 const char *dppass
= pandecode_stencil_op(stencil
->dppass
);
1405 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil
->zero
);
1407 pandecode_log(".stencil_%s = {\n", name
);
1409 pandecode_prop("ref = %d", stencil
->ref
);
1410 pandecode_prop("mask = 0x%02X", stencil
->mask
);
1411 pandecode_prop("func = %s", func
);
1412 pandecode_prop("sfail = %s", sfail
);
1413 pandecode_prop("dpfail = %s", dpfail
);
1414 pandecode_prop("dppass = %s", dppass
);
1416 pandecode_log("},\n");
1420 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1423 pandecode_msg("XXX: blend zero tripped: %X\n", blend
->zero1
);
1425 pandecode_log(".equation = {\n");
1428 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1429 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1431 pandecode_log(".color_mask = ");
1432 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1433 pandecode_log_cont(",\n");
1436 pandecode_log("},\n");
1439 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1442 decode_bifrost_constant(u16 constant
)
1444 float lo
= (float) (constant
& 0xFF);
1445 float hi
= (float) (constant
>> 8);
1447 return (hi
/ 255.0) + (lo
/ 65535.0);
1451 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1453 struct bifrost_blend_rt
*b
=
1454 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1456 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1459 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1460 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1461 b
->constant
, decode_bifrost_constant(b
->constant
));
1463 /* TODO figure out blend shader enable bit */
1464 pandecode_blend_equation(&b
->equation
);
1465 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1466 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1467 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1470 pandecode_log("},\n");
1476 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1478 /* constant/equation is in a union */
1482 pandecode_log(".blend = {\n");
1486 pandecode_shader_address("shader", blend
->shader
);
1488 pandecode_blend_equation(&blend
->equation
);
1489 pandecode_prop("constant = %f", blend
->constant
);
1493 pandecode_log("},\n");
1495 /* Return blend shader to disassemble if present */
1496 return is_shader
? (blend
->shader
& ~0xF) : 0;
1500 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1502 struct midgard_blend_rt
*b
=
1503 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1505 /* Flags determine presence of blend shader */
1506 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1508 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1511 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1513 union midgard_blend blend
= b
->blend
;
1514 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1517 pandecode_log("};\n");
1522 /* Attributes and varyings have descriptor records, which contain information
1523 * about their format and ordering with the attribute/varying buffers. We'll
1524 * want to validate that the combinations specified are self-consistent.
1528 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1531 char *prefix
= varying
? "varying" : "attribute";
1532 unsigned max_index
= 0;
1533 snprintf(base
, sizeof(base
), "%s_meta", prefix
);
1535 struct mali_attr_meta
*attr_meta
;
1536 mali_ptr p
= varying
? v
->varying_meta
: v
->attribute_meta
;
1538 struct pandecode_mapped_memory
*attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
);
1540 for (int i
= 0; i
< count
; ++i
, p
+= sizeof(struct mali_attr_meta
)) {
1541 attr_meta
= pandecode_fetch_gpu_mem(attr_mem
, p
,
1544 /* If the record is discard, it should be zero for everything else */
1546 if (attr_meta
->format
== MALI_VARYING_DISCARD
) {
1549 attr_meta
->unknown1
|
1550 attr_meta
->unknown3
|
1551 attr_meta
->src_offset
;
1554 pandecode_msg("XXX: expected empty record for varying discard\n");
1556 /* We want to look for a literal 0000 swizzle -- this
1557 * is not encoded with all zeroes, however */
1559 enum mali_channel z
= MALI_CHANNEL_ZERO
;
1560 unsigned zero_swizzle
= z
| (z
<< 3) | (z
<< 6) | (z
<< 9);
1561 bool good_swizzle
= attr_meta
->swizzle
== zero_swizzle
;
1564 pandecode_msg("XXX: expected zero swizzle for discard\n");
1567 pandecode_msg("XXX: cannot discard attribute\n");
1569 /* If we're all good, omit the record */
1570 if (!zero
&& varying
&& good_swizzle
) {
1571 pandecode_log("/* discarded varying */\n");
1576 if (attr_meta
->index
> max_index
)
1577 max_index
= attr_meta
->index
;
1579 if (attr_meta
->unknown1
!= 0x2) {
1580 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1581 pandecode_prop("unknown1 = 0x%" PRIx64
, (u64
) attr_meta
->unknown1
);
1584 if (attr_meta
->unknown3
) {
1585 pandecode_msg("XXX: unexpected unknown3 set\n");
1586 pandecode_prop("unknown3 = 0x%" PRIx64
, (u64
) attr_meta
->unknown3
);
1589 pandecode_format_short(attr_meta
->format
, false);
1590 pandecode_log_cont(" %s_%u", prefix
, attr_meta
->index
);
1592 if (attr_meta
->src_offset
)
1593 pandecode_log_cont("[%u]", attr_meta
->src_offset
);
1595 pandecode_swizzle(attr_meta
->swizzle
, attr_meta
->format
);
1597 pandecode_log_cont(";\n");
1600 pandecode_log("\n");
1602 return count
? (max_index
+ 1) : 0;
1605 /* return bits [lo, hi) of word */
1607 bits(u32 word
, u32 lo
, u32 hi
)
1610 return word
; // avoid undefined behavior with the shift
1612 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1616 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool noninstanced
)
1618 pandecode_log_cont("{\n");
1621 /* Decode invocation_count. See the comment before the definition of
1622 * invocation_count for an explanation.
1625 unsigned size_x
= bits(p
->invocation_count
, 0, p
->size_y_shift
) + 1;
1626 unsigned size_y
= bits(p
->invocation_count
, p
->size_y_shift
, p
->size_z_shift
) + 1;
1627 unsigned size_z
= bits(p
->invocation_count
, p
->size_z_shift
, p
->workgroups_x_shift
) + 1;
1629 unsigned groups_x
= bits(p
->invocation_count
, p
->workgroups_x_shift
, p
->workgroups_y_shift
) + 1;
1630 unsigned groups_y
= bits(p
->invocation_count
, p
->workgroups_y_shift
, p
->workgroups_z_shift
) + 1;
1631 unsigned groups_z
= bits(p
->invocation_count
, p
->workgroups_z_shift
, 32) + 1;
1633 /* Even though we have this decoded, we want to ensure that the
1634 * representation is "unique" so we don't lose anything by printing only
1635 * the final result. More specifically, we need to check that we were
1636 * passed something in canonical form, since the definition per the
1637 * hardware is inherently not unique. How? Well, take the resulting
1638 * decode and pack it ourselves! If it is bit exact with what we
1639 * decoded, we're good to go. */
1641 struct mali_vertex_tiler_prefix ref
;
1642 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, noninstanced
);
1645 (p
->invocation_count
== ref
.invocation_count
) &&
1646 (p
->size_y_shift
== ref
.size_y_shift
) &&
1647 (p
->size_z_shift
== ref
.size_z_shift
) &&
1648 (p
->workgroups_x_shift
== ref
.workgroups_x_shift
) &&
1649 (p
->workgroups_y_shift
== ref
.workgroups_y_shift
) &&
1650 (p
->workgroups_z_shift
== ref
.workgroups_z_shift
) &&
1651 (p
->workgroups_x_shift_2
== ref
.workgroups_x_shift_2
);
1654 pandecode_msg("XXX: non-canonical workgroups packing\n");
1655 pandecode_msg("expected: %X, %d, %d, %d, %d, %d, %d\n",
1656 ref
.invocation_count
,
1659 ref
.workgroups_x_shift
,
1660 ref
.workgroups_y_shift
,
1661 ref
.workgroups_z_shift
,
1662 ref
.workgroups_x_shift_2
);
1664 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1665 pandecode_prop("size_y_shift = %d", p
->size_y_shift
);
1666 pandecode_prop("size_z_shift = %d", p
->size_z_shift
);
1667 pandecode_prop("workgroups_x_shift = %d", p
->workgroups_x_shift
);
1668 pandecode_prop("workgroups_y_shift = %d", p
->workgroups_y_shift
);
1669 pandecode_prop("workgroups_z_shift = %d", p
->workgroups_z_shift
);
1670 pandecode_prop("workgroups_x_shift_2 = %d", p
->workgroups_x_shift_2
);
1673 /* Regardless, print the decode */
1674 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1675 size_x
, size_y
, size_z
,
1676 groups_x
, groups_y
, groups_z
);
1679 if (p
->unknown_draw
)
1680 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1682 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1684 if (p
->draw_mode
!= MALI_DRAW_NONE
)
1685 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p
->draw_mode
));
1687 /* Index count only exists for tiler jobs anyway */
1690 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1693 unsigned index_raw_size
= (p
->unknown_draw
& MALI_DRAW_INDEXED_SIZE
);
1694 index_raw_size
>>= MALI_DRAW_INDEXED_SHIFT
;
1696 /* Validate an index buffer is present if we need one. TODO: verify
1697 * relationship between invocation_count and index_count */
1700 unsigned count
= p
->index_count
;
1703 unsigned size
= (index_raw_size
== 0x3) ? 4 : index_raw_size
;
1705 /* Ensure we got a size, and if so, validate the index buffer
1706 * is large enough to hold a full set of indices of the given
1709 if (!index_raw_size
)
1710 pandecode_msg("XXX: index size missing\n");
1712 pandecode_validate_buffer(p
->indices
, count
* size
);
1713 } else if (index_raw_size
)
1714 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size
);
1716 if (p
->offset_bias_correction
)
1717 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1719 /* TODO: Figure out what this is. It's not zero */
1720 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1723 pandecode_log("},\n");
1727 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1729 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1730 struct mali_uniform_buffer_meta
*PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1732 for (int i
= 0; i
< ubufs_count
; i
++) {
1733 unsigned size
= (ubufs
[i
].size
+ 1) * 16;
1734 mali_ptr addr
= ubufs
[i
].ptr
<< 2;
1736 pandecode_validate_buffer(addr
, size
);
1738 char *ptr
= pointer_as_memory_reference(ubufs
[i
].ptr
<< 2);
1739 pandecode_log("ubuf_%d[%u] = %s;\n", i
, size
, ptr
);
1743 pandecode_log("\n");
1747 pandecode_uniforms(mali_ptr uniforms
, unsigned uniform_count
)
1749 pandecode_validate_buffer(uniforms
, uniform_count
* 16);
1751 char *ptr
= pointer_as_memory_reference(uniforms
);
1752 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count
, ptr
);
1757 pandecode_scratchpad(uintptr_t pscratchpad
, int job_no
, char *suffix
)
1760 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(pscratchpad
);
1762 struct bifrost_scratchpad
*PANDECODE_PTR_VAR(scratchpad
, mem
, pscratchpad
);
1764 if (scratchpad
->zero
) {
1765 pandecode_msg("XXX: scratchpad zero tripped");
1766 pandecode_prop("zero = 0x%x\n", scratchpad
->zero
);
1769 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64
"_%d%s = {\n", pscratchpad
, job_no
, suffix
);
1772 pandecode_prop("flags = 0x%x", scratchpad
->flags
);
1773 MEMORY_PROP(scratchpad
, gpu_scratchpad
);
1776 pandecode_log("};\n");
1779 static unsigned shader_id
= 0;
1781 static struct midgard_disasm_stats
1782 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1785 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1786 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1788 /* Compute maximum possible size */
1789 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1791 /* Print some boilerplate to clearly denote the assembly (which doesn't
1792 * obey indentation rules), and actually do the disassembly! */
1796 struct midgard_disasm_stats stats
;
1799 disassemble_bifrost(code
, sz
, false);
1801 /* TODO: Extend stats to Bifrost */
1802 stats
.texture_count
= -128;
1803 stats
.sampler_count
= -128;
1804 stats
.attribute_count
= -128;
1805 stats
.varying_count
= -128;
1806 stats
.uniform_count
= -128;
1807 stats
.uniform_buffer_count
= -128;
1808 stats
.work_count
= -128;
1810 stats
.instruction_count
= 0;
1811 stats
.bundle_count
= 0;
1812 stats
.quadword_count
= 0;
1813 stats
.helper_invocations
= false;
1815 stats
= disassemble_midgard(code
, sz
);
1818 /* Print shader-db stats */
1820 unsigned nr_threads
=
1821 (stats
.work_count
<= 4) ? 4 :
1822 (stats
.work_count
<= 8) ? 2 :
1825 printf("shader%d - %s shader: "
1826 "%u inst, %u bundles, %u quadwords, "
1827 "%u registers, %u threads, 0 loops\n\n\n",
1829 (type
== JOB_TYPE_TILER
) ? "FRAGMENT" : "VERTEX",
1830 stats
.instruction_count
, stats
.bundle_count
, stats
.quadword_count
,
1831 stats
.work_count
, nr_threads
);
1838 pandecode_texture(mali_ptr u
,
1839 struct pandecode_mapped_memory
*tmem
,
1840 unsigned job_no
, unsigned tex
)
1842 struct mali_texture_descriptor
*PANDECODE_PTR_VAR(t
, tmem
, u
);
1844 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64
"_%d_%d = {\n", u
, job_no
, tex
);
1847 struct mali_texture_format f
= t
->format
;
1849 /* See the definiton of enum mali_texture_type */
1851 bool is_cube
= f
.type
== MALI_TEX_CUBE
;
1852 unsigned dimension
= is_cube
? 2 : f
.type
;
1854 pandecode_make_indent();
1856 /* TODO: Are there others? */
1857 bool is_zs
= f
.format
== MALI_Z32_UNORM
;
1859 /* Recall Z/S switched the meaning of linear/tiled .. */
1860 if (is_zs
&& f
.layout
== MALI_TEXTURE_LINEAR
)
1861 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1863 /* Print the layout. Default is linear; a modifier can denote AFBC or
1864 * u-interleaved/tiled modes */
1866 if (f
.layout
== MALI_TEXTURE_AFBC
)
1867 pandecode_log_cont("afbc");
1868 else if (f
.layout
== MALI_TEXTURE_TILED
)
1869 pandecode_log_cont(is_zs
? "linear" : "tiled");
1870 else if (f
.layout
== MALI_TEXTURE_LINEAR
)
1871 pandecode_log_cont("linear");
1873 pandecode_msg("XXX: invalid texture layout 0x%X\n", f
.layout
);
1875 pandecode_swizzle(t
->swizzle
, f
.format
);
1876 pandecode_log_cont(" ");
1878 /* Distinguish cube/2D with modifier */
1881 pandecode_log_cont("cube ");
1883 pandecode_format_short(f
.format
, f
.srgb
);
1884 pandecode_swizzle(f
.swizzle
, f
.format
);
1886 /* All four width/height/depth/array_size dimensions are present
1887 * regardless of the type of texture, but it is an error to have
1888 * non-zero dimensions for unused dimensions. Verify this. array_size
1889 * can always be set, as can width. */
1891 if (t
->height
&& dimension
< 2)
1892 pandecode_msg("XXX: nonzero height for <2D texture\n");
1894 if (t
->depth
&& dimension
< 3)
1895 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1897 /* Print only the dimensions that are actually there */
1899 pandecode_log_cont(": %d", t
->width
+ 1);
1902 pandecode_log_cont("x%u", t
->height
+ 1);
1905 pandecode_log_cont("x%u", t
->depth
+ 1);
1908 pandecode_log_cont("[%u]", t
->array_size
+ 1);
1911 pandecode_log_cont(" mip %u", t
->levels
);
1913 pandecode_log_cont("\n");
1915 if (f
.unknown1
| f
.zero
) {
1916 pandecode_msg("XXX: texture format zero tripped\n");
1917 pandecode_prop("unknown1 = %" PRId32
, f
.unknown1
);
1918 pandecode_prop("zero = %" PRId32
, f
.zero
);
1922 pandecode_msg("XXX: expected unknown texture bit set\n");
1923 pandecode_prop("unknown2 = %" PRId32
, f
.unknown1
);
1926 if (t
->swizzle_zero
) {
1927 pandecode_msg("XXX: swizzle zero tripped\n");
1928 pandecode_prop("swizzle_zero = %d", t
->swizzle_zero
);
1931 if (t
->unknown3
| t
->unknown3A
| t
->unknown5
| t
->unknown6
| t
->unknown7
) {
1932 pandecode_msg("XXX: texture zero tripped\n");
1933 pandecode_prop("unknown3 = %" PRId16
, t
->unknown3
);
1934 pandecode_prop("unknown3A = %" PRId8
, t
->unknown3A
);
1935 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
1936 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
1937 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
1940 pandecode_log(".payload = {\n");
1943 /* A bunch of bitmap pointers follow.
1944 * We work out the correct number,
1945 * based on the mipmap/cubemap
1946 * properties, but dump extra
1947 * possibilities to futureproof */
1949 int bitmap_count
= MALI_NEGATIVE(t
->levels
);
1951 /* Miptree for each face */
1952 if (f
.type
== MALI_TEX_CUBE
)
1955 /* Array of textures */
1956 bitmap_count
*= MALI_NEGATIVE(t
->array_size
);
1958 /* Stride for each element */
1959 if (f
.manual_stride
)
1962 /* Sanity check the size */
1963 int max_count
= sizeof(t
->payload
) / sizeof(t
->payload
[0]);
1964 assert (bitmap_count
<= max_count
);
1966 for (int i
= 0; i
< bitmap_count
; ++i
) {
1967 /* How we dump depends if this is a stride or a pointer */
1969 if (f
.manual_stride
&& (i
& 1)) {
1970 /* signed 32-bit snuck in as a 64-bit pointer */
1971 uint64_t stride_set
= t
->payload
[i
];
1972 uint32_t clamped_stride
= stride_set
;
1973 int32_t stride
= clamped_stride
;
1974 assert(stride_set
== clamped_stride
);
1975 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
1977 char *a
= pointer_as_memory_reference(t
->payload
[i
]);
1978 pandecode_log("%s, \n", a
);
1984 pandecode_log("},\n");
1987 pandecode_log("};\n");
1990 /* For shader properties like texture_count, we have a claimed property in the shader_meta, and the actual Truth from static analysis (this may just be an upper limit). We validate accordingly */
1993 pandecode_shader_prop(const char *name
, unsigned claim
, signed truth
, bool fuzzy
)
2002 if ((truth
>= 0) && !fuzzy
) {
2003 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2004 (truth
< claim
) ? "warn" : "XXX",
2005 name
, truth
, claim
);
2006 } else if ((claim
> -truth
) && !fuzzy
) {
2007 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2008 name
, -truth
, claim
);
2009 } else if (fuzzy
&& (claim
< truth
))
2010 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2011 name
, truth
, claim
);
2013 pandecode_log(".%s = %" PRId16
, name
, claim
);
2016 pandecode_log_cont(" /* %u used */", truth
);
2018 pandecode_log_cont(",\n");
2022 pandecode_vertex_tiler_postfix_pre(
2023 const struct mali_vertex_tiler_postfix
*p
,
2024 int job_no
, enum mali_job_type job_type
,
2025 char *suffix
, bool is_bifrost
)
2027 struct pandecode_mapped_memory
*attr_mem
;
2029 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2030 * are the only things actually needed from the FBD, vertex/tiler jobs
2031 * no longer reference the FBD -- instead, this field points to some
2032 * info about the scratchpad.
2035 struct pandecode_fbd fbd_info
= {
2036 /* Default for Bifrost */
2041 pandecode_scratchpad(p
->framebuffer
& ~FBD_TYPE
, job_no
, suffix
);
2042 else if (p
->framebuffer
& MALI_MFBD
)
2043 fbd_info
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->framebuffer
) & FBD_MASK
, job_no
, false);
2044 else if (job_type
== JOB_TYPE_COMPUTE
)
2045 pandecode_compute_fbd((u64
) (uintptr_t) p
->framebuffer
, job_no
);
2047 fbd_info
= pandecode_sfbd((u64
) (uintptr_t) p
->framebuffer
, job_no
, false);
2049 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
2050 int texture_count
= 0, sampler_count
= 0;
2053 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->shader
);
2054 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, p
->shader
);
2056 /* Disassemble ahead-of-time to get stats. Initialize with
2057 * stats for the missing-shader case so we get validation
2060 struct midgard_disasm_stats info
= {
2063 .attribute_count
= 0,
2067 .uniform_count
= -128,
2068 .uniform_buffer_count
= 0
2071 if (s
->shader
& ~0xF)
2072 info
= pandecode_shader_disassemble(s
->shader
& ~0xF, job_no
, job_type
, is_bifrost
);
2074 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", p
->shader
, job_no
, suffix
);
2077 /* Save for dumps */
2078 attribute_count
= s
->attribute_count
;
2079 varying_count
= s
->varying_count
;
2080 texture_count
= s
->texture_count
;
2081 sampler_count
= s
->sampler_count
;
2084 uniform_count
= s
->bifrost2
.uniform_count
;
2085 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
2087 uniform_count
= s
->midgard1
.uniform_count
;
2088 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
2091 pandecode_shader_address("shader", s
->shader
);
2093 pandecode_shader_prop("texture_count", s
->texture_count
, info
.texture_count
, false);
2094 pandecode_shader_prop("sampler_count", s
->sampler_count
, info
.sampler_count
, false);
2095 pandecode_shader_prop("attribute_count", s
->attribute_count
, info
.attribute_count
, false);
2096 pandecode_shader_prop("varying_count", s
->varying_count
, info
.varying_count
, false);
2097 pandecode_shader_prop("uniform_buffer_count",
2098 uniform_buffer_count
,
2099 info
.uniform_buffer_count
, true);
2102 pandecode_shader_prop("uniform_count",
2104 info
.uniform_count
, false);
2106 pandecode_shader_prop("work_count",
2107 s
->midgard1
.work_count
, info
.work_count
, false);
2111 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32
, s
->bifrost1
.unk1
);
2113 bool helpers
= s
->midgard1
.flags
& MALI_HELPER_INVOCATIONS
;
2114 s
->midgard1
.flags
&= ~MALI_HELPER_INVOCATIONS
;
2116 if (helpers
!= info
.helper_invocations
) {
2117 pandecode_msg("XXX: expected helpers %u but got %u\n",
2118 info
.helper_invocations
, helpers
);
2121 pandecode_log(".midgard1.flags = ");
2122 pandecode_log_decoded_flags(shader_midgard1_flag_info
, s
->midgard1
.flags
);
2123 pandecode_log_cont(",\n");
2125 pandecode_prop("midgard1.unknown2 = 0x%" PRIx32
, s
->midgard1
.unknown2
);
2128 if (s
->depth_units
|| s
->depth_factor
) {
2129 pandecode_prop("depth_factor = %f", s
->depth_factor
);
2130 pandecode_prop("depth_units = %f", s
->depth_units
);
2133 if (s
->alpha_coverage
) {
2134 bool invert_alpha_coverage
= s
->alpha_coverage
& 0xFFF0;
2135 uint16_t inverted_coverage
= invert_alpha_coverage
? ~s
->alpha_coverage
: s
->alpha_coverage
;
2137 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2138 invert_alpha_coverage
? "~" : "",
2139 MALI_GET_ALPHA_COVERAGE(inverted_coverage
));
2142 if (s
->unknown2_3
|| s
->unknown2_4
) {
2143 pandecode_log(".unknown2_3 = ");
2145 int unknown2_3
= s
->unknown2_3
;
2146 int unknown2_4
= s
->unknown2_4
;
2148 /* We're not quite sure what these flags mean without the depth test, if anything */
2150 if (unknown2_3
& (MALI_DEPTH_WRITEMASK
| MALI_DEPTH_FUNC_MASK
)) {
2151 const char *func
= pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3
));
2152 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
2154 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
2157 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
2158 pandecode_log_cont(",\n");
2160 pandecode_log(".unknown2_4 = ");
2161 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
2162 pandecode_log_cont(",\n");
2165 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
2166 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
2167 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
2170 pandecode_stencil("front", &s
->stencil_front
);
2171 pandecode_stencil("back", &s
->stencil_back
);
2174 pandecode_log(".bifrost2 = {\n");
2177 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
2178 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
2179 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
2180 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
2183 pandecode_log("},\n");
2184 } else if (s
->midgard2
.unknown2_7
) {
2185 pandecode_log(".midgard2 = {\n");
2188 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
2190 pandecode_log("},\n");
2194 pandecode_prop("unknown2_8 = 0x%" PRIx32
, s
->unknown2_8
);
2197 /* TODO: Blend shaders routing/disasm */
2199 union midgard_blend blend
= s
->blend
;
2200 pandecode_midgard_blend(&blend
, false);
2204 pandecode_log("};\n");
2206 /* MRT blend fields are used whenever MFBD is used, with
2207 * per-RT descriptors */
2209 if (job_type
== JOB_TYPE_TILER
) {
2210 void* blend_base
= (void *) (s
+ 1);
2212 for (unsigned i
= 0; i
< fbd_info
.rt_count
; i
++) {
2213 mali_ptr shader
= 0;
2216 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
2218 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
2220 if (shader
& ~0xF) {
2221 struct midgard_disasm_stats stats
=
2222 pandecode_shader_disassemble(shader
, job_no
, job_type
, false);
2224 bool has_texture
= (stats
.texture_count
> 0);
2225 bool has_sampler
= (stats
.sampler_count
> 0);
2226 bool has_attribute
= (stats
.attribute_count
> 0);
2227 bool has_varying
= (stats
.varying_count
> 0);
2228 bool has_uniform
= (stats
.uniform_count
> 0);
2229 bool has_ubo
= (stats
.uniform_buffer_count
> 0);
2231 if (has_texture
|| has_sampler
)
2232 pandecode_msg("XXX: blend shader accessing textures\n");
2234 if (has_attribute
|| has_varying
)
2235 pandecode_msg("XXX: blend shader accessing interstage\n");
2237 if (has_uniform
|| has_ubo
)
2238 pandecode_msg("XXX: blend shader accessing uniforms\n");
2244 pandecode_msg("XXX: missing shader descriptor\n");
2247 struct pandecode_mapped_memory
*fmem
= pandecode_find_mapped_gpu_mem_containing(p
->viewport
);
2248 struct mali_viewport
*PANDECODE_PTR_VAR(f
, fmem
, p
->viewport
);
2250 pandecode_log("struct mali_viewport viewport_%"PRIx64
"_%d%s = {\n", p
->viewport
, job_no
, suffix
);
2253 pandecode_prop("clip_minx = %f", f
->clip_minx
);
2254 pandecode_prop("clip_miny = %f", f
->clip_miny
);
2255 pandecode_prop("clip_minz = %f", f
->clip_minz
);
2256 pandecode_prop("clip_maxx = %f", f
->clip_maxx
);
2257 pandecode_prop("clip_maxy = %f", f
->clip_maxy
);
2258 pandecode_prop("clip_maxz = %f", f
->clip_maxz
);
2260 /* Only the higher coordinates are MALI_POSITIVE scaled */
2262 pandecode_prop("viewport0 = { %d, %d }",
2263 f
->viewport0
[0], f
->viewport0
[1]);
2265 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2266 f
->viewport1
[0] + 1, f
->viewport1
[1] + 1);
2269 pandecode_log("};\n");
2272 unsigned max_attr_index
= 0;
2274 if (p
->attribute_meta
)
2275 max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
2277 if (p
->attributes
) {
2278 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
2279 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false, job_type
);
2282 /* Varyings are encoded like attributes but not actually sent; we just
2283 * pass a zero buffer with the right stride/size set, (or whatever)
2284 * since the GPU will write to it itself */
2286 if (p
->varying_meta
) {
2287 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
2291 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
2293 /* Number of descriptors depends on whether there are
2294 * non-internal varyings */
2296 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true, job_type
);
2299 if (p
->uniform_buffers
) {
2300 if (uniform_buffer_count
)
2301 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
2303 pandecode_msg("warn: UBOs specified but not referenced\n");
2304 } else if (uniform_buffer_count
)
2305 pandecode_msg("XXX: UBOs referenced but not specified\n");
2307 /* We don't want to actually dump uniforms, but we do need to validate
2308 * that the counts we were given are sane */
2312 pandecode_uniforms(p
->uniforms
, uniform_count
);
2314 pandecode_msg("warn: Uniforms specified but not referenced\n");
2315 } else if (uniform_count
)
2316 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2318 if (p
->texture_trampoline
) {
2319 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(p
->texture_trampoline
);
2322 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
);
2324 pandecode_log("uint64_t texture_trampoline_%"PRIx64
"_%d[] = {\n", p
->texture_trampoline
, job_no
);
2327 for (int tex
= 0; tex
< texture_count
; ++tex
) {
2328 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
2329 char *a
= pointer_as_memory_reference(*u
);
2330 pandecode_log("%s,\n", a
);
2335 pandecode_log("};\n");
2337 /* Now, finally, descend down into the texture descriptor */
2338 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
2339 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, p
->texture_trampoline
+ tex
* sizeof(mali_ptr
));
2340 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
2342 pandecode_texture(*u
, tmem
, job_no
, tex
);
2347 if (p
->sampler_descriptor
) {
2348 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->sampler_descriptor
);
2351 struct mali_sampler_descriptor
*s
;
2353 mali_ptr d
= p
->sampler_descriptor
;
2355 for (int i
= 0; i
< sampler_count
; ++i
) {
2356 s
= pandecode_fetch_gpu_mem(smem
, d
+ sizeof(*s
) * i
, sizeof(*s
));
2358 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", d
+ sizeof(*s
) * i
, job_no
, i
);
2361 pandecode_log(".filter_mode = ");
2362 pandecode_log_decoded_flags(sampler_flag_info
, s
->filter_mode
);
2363 pandecode_log_cont(",\n");
2365 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
2366 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
2368 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s
->wrap_s
));
2369 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s
->wrap_t
));
2370 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s
->wrap_r
));
2372 pandecode_prop("compare_func = %s", pandecode_alt_func(s
->compare_func
));
2374 if (s
->zero
|| s
->zero2
) {
2375 pandecode_msg("XXX: sampler zero tripped\n");
2376 pandecode_prop("zero = 0x%X, 0x%X\n", s
->zero
, s
->zero2
);
2379 pandecode_prop("seamless_cube_map = %d", s
->seamless_cube_map
);
2381 pandecode_prop("border_color = { %f, %f, %f, %f }",
2385 s
->border_color
[3]);
2388 pandecode_log("};\n");
2395 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2397 if (p
->shader
& 0xF)
2398 pandecode_msg("warn: shader tagged %X\n", p
->shader
& 0xF);
2400 if (!(p
->position_varying
|| p
->occlusion_counter
))
2403 pandecode_log(".postfix = {\n");
2406 MEMORY_PROP(p
, position_varying
);
2407 MEMORY_PROP(p
, occlusion_counter
);
2410 pandecode_log("},\n");
2414 pandecode_vertex_only_bfr(struct bifrost_vertex_only
*v
)
2416 pandecode_log_cont("{\n");
2419 pandecode_prop("unk2 = 0x%x", v
->unk2
);
2421 if (v
->zero0
|| v
->zero1
) {
2422 pandecode_msg("XXX: vertex only zero tripped");
2423 pandecode_prop("zero0 = 0x%" PRIx32
, v
->zero0
);
2424 pandecode_prop("zero1 = 0x%" PRIx64
, v
->zero1
);
2428 pandecode_log("}\n");
2432 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2435 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2436 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2438 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no
);
2442 pandecode_msg("XXX: tiler heap zero tripped\n");
2443 pandecode_prop("zero = 0x%x", h
->zero
);
2446 for (int i
= 0; i
< 12; i
++) {
2447 if (h
->zeros
[i
] != 0) {
2448 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2453 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2454 MEMORY_PROP(h
, tiler_heap_start
);
2455 MEMORY_PROP(h
, tiler_heap_free
);
2457 /* this might point to the beginning of another buffer, when it's
2458 * really the end of the tiler heap buffer, so we have to be careful
2459 * here. but for zero length, we need the same pointer.
2462 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2463 MEMORY_PROP(h
, tiler_heap_start
);
2465 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2466 pandecode_prop("tiler_heap_end = %s + 1", a
);
2471 pandecode_log("};\n");
2475 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2477 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2478 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2480 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2482 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no
);
2485 if (t
->zero0
|| t
->zero1
) {
2486 pandecode_msg("XXX: tiler meta zero tripped\n");
2487 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2488 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2491 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2492 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2494 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2495 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2497 for (int i
= 0; i
< 12; i
++) {
2498 if (t
->zeros
[i
] != 0) {
2499 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2505 pandecode_log("};\n");
2509 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2511 pandecode_log(".gl_enables = ");
2513 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2515 pandecode_log_cont(",\n");
2519 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2521 if (u
.pointer
== 0x0)
2524 pandecode_log(".primitive_size = {\n");
2528 pandecode_prop("constant = %f", u
.constant
);
2530 MEMORY_PROP((&u
), pointer
);
2534 pandecode_log("},\n");
2538 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2540 pandecode_log_cont("{\n");
2543 /* TODO: gl_PointSize on Bifrost */
2544 pandecode_primitive_size(t
->primitive_size
, true);
2546 pandecode_gl_enables(t
->gl_enables
, JOB_TYPE_TILER
);
2548 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2549 || t
->zero6
|| t
->zero7
|| t
->zero8
) {
2550 pandecode_msg("XXX: tiler only zero tripped\n");
2551 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2552 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2553 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2554 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2555 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2556 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2557 pandecode_prop("zero7 = 0x%" PRIx32
, t
->zero7
);
2558 pandecode_prop("zero8 = 0x%" PRIx64
, t
->zero8
);
2562 pandecode_log("},\n");
2566 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2567 const struct pandecode_mapped_memory
*mem
,
2568 mali_ptr payload
, int job_no
)
2570 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2572 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true);
2574 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no
);
2577 pandecode_log(".prefix = ");
2578 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2580 pandecode_log(".vertex = ");
2581 pandecode_vertex_only_bfr(&v
->vertex
);
2583 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2586 pandecode_log("};\n");
2592 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2593 const struct pandecode_mapped_memory
*mem
,
2594 mali_ptr payload
, int job_no
)
2596 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2598 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true);
2599 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2601 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no
);
2604 pandecode_log(".prefix = ");
2605 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2607 pandecode_log(".tiler = ");
2608 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2610 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2613 pandecode_log("};\n");
2619 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2620 const struct pandecode_mapped_memory
*mem
,
2621 mali_ptr payload
, int job_no
)
2623 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2625 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false);
2627 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2630 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2631 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2633 bool instanced
= v
->instance_shift
|| v
->instance_odd
;
2634 bool is_graphics
= (h
->job_type
== JOB_TYPE_VERTEX
) || (h
->job_type
== JOB_TYPE_TILER
);
2636 pandecode_log(".prefix = ");
2637 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, !instanced
&& is_graphics
);
2639 pandecode_gl_enables(v
->gl_enables
, h
->job_type
);
2641 if (v
->instance_shift
|| v
->instance_odd
) {
2642 pandecode_prop("instance_shift = 0x%d /* %d */",
2643 v
->instance_shift
, 1 << v
->instance_shift
);
2644 pandecode_prop("instance_odd = 0x%X /* %d */",
2645 v
->instance_odd
, (2 * v
->instance_odd
) + 1);
2647 pandecode_padded_vertices(v
->instance_shift
, v
->instance_odd
);
2650 if (v
->offset_start
)
2651 pandecode_prop("offset_start = %d", v
->offset_start
);
2654 pandecode_msg("XXX: midgard payload zero tripped\n");
2655 pandecode_prop("zero5 = 0x%" PRIx64
, v
->zero5
);
2658 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2661 pandecode_log("};\n");
2667 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2668 mali_ptr payload
, int job_no
,
2671 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2673 bool is_mfbd
= (s
->framebuffer
& FBD_TYPE
) == MALI_MFBD
;
2675 /* Bifrost theoretically may retain support for SFBD on compute jobs,
2676 * but for graphics workloads with a FRAGMENT payload, use MFBD */
2678 if (!is_mfbd
&& is_bifrost
)
2679 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2681 struct pandecode_fbd info
;
2684 info
= pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true);
2686 info
= pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true);
2688 /* Compute the tag for the tagged pointer. This contains the type of
2689 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2690 * additional structures follow the MFBD header (an extra payload or
2691 * not, as well as a count of render targets) */
2693 unsigned expected_tag
= is_mfbd
? MALI_MFBD
: MALI_SFBD
;
2697 expected_tag
|= MALI_MFBD_TAG_EXTRA
;
2699 expected_tag
|= (MALI_POSITIVE(info
.rt_count
) << 2);
2702 if ((s
->min_tile_coord
| s
->max_tile_coord
) & ~(MALI_X_COORD_MASK
| MALI_Y_COORD_MASK
)) {
2703 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2704 pandecode_prop("min_tile_coord = 0x%X\n", s
->min_tile_coord
);
2705 pandecode_prop("max_tile_coord = 0x%X\n", s
->min_tile_coord
);
2708 /* Extract tile coordinates */
2710 unsigned min_x
= MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2711 unsigned min_y
= MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2713 unsigned max_x
= (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2714 unsigned max_y
= (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2716 /* For the max, we also want the floored (rather than ceiled) version for checking */
2718 unsigned max_x_f
= (MALI_TILE_COORD_X(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2719 unsigned max_y_f
= (MALI_TILE_COORD_Y(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2721 /* Validate the coordinates are well-ordered */
2724 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2725 else if (min_x
> max_x
)
2726 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2729 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2730 else if (min_y
> max_y
)
2731 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2733 /* Validate the coordinates fit inside the framebuffer. We use floor,
2734 * rather than ceil, for the max coordinates, since the tile
2735 * coordinates for something like an 800x600 framebuffer will actually
2736 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2738 if ((min_x
> info
.width
) || (max_x_f
> info
.width
))
2739 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2741 if ((min_y
> info
.height
) || (max_y_f
> info
.height
))
2742 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2744 /* After validation, we print */
2746 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x
, min_y
, max_x
, max_y
);
2748 /* The FBD is a tagged pointer */
2750 unsigned tag
= (s
->framebuffer
& ~FBD_MASK
);
2752 if (tag
!= expected_tag
)
2753 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag
, tag
);
2758 static int job_descriptor_number
= 0;
2761 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
)
2763 struct mali_job_descriptor_header
*h
;
2765 int start_number
= 0;
2771 struct pandecode_mapped_memory
*mem
=
2772 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2776 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2778 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2779 * high 32-bits of the 64-bit pointer are reused to store
2782 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
2783 h
->job_type
!= JOB_TYPE_FRAGMENT
? 4 : 0;
2784 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
2786 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
, 256);
2788 int job_no
= job_descriptor_number
++;
2791 start_number
= job_no
;
2793 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2796 pandecode_prop("job_type = %s", pandecode_job_type(h
->job_type
));
2798 /* Save for next job fixing */
2799 last_size
= h
->job_descriptor_size
;
2801 if (h
->job_descriptor_size
)
2802 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2804 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2805 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2806 h
->exception_status
,
2807 (h
->exception_status
>> 16) & 0xFFFF,
2808 pandecode_exception_access((h
->exception_status
>> 8) & 0x3),
2809 h
->exception_status
& 0xFF);
2811 if (h
->first_incomplete_task
)
2812 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2814 if (h
->fault_pointer
)
2815 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2818 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2820 pandecode_prop("job_index = %d", h
->job_index
);
2822 if (h
->unknown_flags
)
2823 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2825 if (h
->job_dependency_index_1
)
2826 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2828 if (h
->job_dependency_index_2
)
2829 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2832 pandecode_log("};\n");
2834 /* Do not touch the field yet -- decode the payload first, and
2835 * don't touch that either. This is essential for the uploads
2836 * to occur in sequence and therefore be dynamically allocated
2837 * correctly. Do note the size, however, for that related
2840 switch (h
->job_type
) {
2841 case JOB_TYPE_SET_VALUE
: {
2842 struct mali_payload_set_value
*s
= payload
;
2843 pandecode_log("struct mali_payload_set_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2845 MEMORY_PROP(s
, out
);
2846 pandecode_prop("unknown = 0x%" PRIX64
, s
->unknown
);
2848 pandecode_log("};\n");
2853 case JOB_TYPE_TILER
:
2854 case JOB_TYPE_VERTEX
:
2855 case JOB_TYPE_COMPUTE
:
2857 if (h
->job_type
== JOB_TYPE_TILER
)
2858 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
);
2860 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
);
2862 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
);
2866 case JOB_TYPE_FRAGMENT
:
2867 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
);
2874 /* Handle linkage */
2877 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no
- 1);
2880 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no
);
2882 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no
);
2887 } while ((jc_gpu_va
= h
->job_descriptor_size
? h
->next_job_64
: h
->next_job_32
));
2889 return start_number
;