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 <midgard_pack.h>
34 #include "util/macros.h"
35 #include "util/u_math.h"
37 #include "midgard/disassemble.h"
38 #include "bifrost/disassemble.h"
40 #include "pan_encoder.h"
42 static void pandecode_swizzle(unsigned swizzle
, enum mali_format format
);
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 #define DUMP_CL(title, T, cl, indent) {\
61 fprintf(pandecode_dump_stream, "%s\n", title); \
62 struct MALI_ ## T temp; \
63 MALI_ ## T ## _unpack((const uint8_t *) cl, &temp); \
64 MALI_ ## T ## _print(pandecode_dump_stream, &temp, indent * 2); \
67 #define DUMP_ADDR(title, T, addr, indent) {\
68 struct pandecode_mapped_memory *mapped_mem = pandecode_find_mapped_gpu_mem_containing(addr); \
69 const uint8_t *cl = pandecode_fetch_gpu_mem(mapped_mem, addr, MALI_ ## T ## _LENGTH); \
70 DUMP_CL(title, T, cl, indent); \
73 FILE *pandecode_dump_stream
;
75 /* Semantic logging type.
77 * Raw: for raw messages to be printed as is.
78 * Message: for helpful information to be commented out in replays.
79 * Property: for properties of a struct
81 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
84 enum pandecode_log_type
{
90 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
91 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
92 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
94 unsigned pandecode_indent
= 0;
97 pandecode_make_indent(void)
99 for (unsigned i
= 0; i
< pandecode_indent
; ++i
)
100 fprintf(pandecode_dump_stream
, " ");
104 pandecode_log_typed(enum pandecode_log_type type
, const char *format
, ...)
108 pandecode_make_indent();
110 if (type
== PANDECODE_MESSAGE
)
111 fprintf(pandecode_dump_stream
, "// ");
112 else if (type
== PANDECODE_PROPERTY
)
113 fprintf(pandecode_dump_stream
, ".");
115 va_start(ap
, format
);
116 vfprintf(pandecode_dump_stream
, format
, ap
);
119 if (type
== PANDECODE_PROPERTY
)
120 fprintf(pandecode_dump_stream
, ",\n");
124 pandecode_log_cont(const char *format
, ...)
128 va_start(ap
, format
);
129 vfprintf(pandecode_dump_stream
, format
, ap
);
133 /* To check for memory safety issues, validates that the given pointer in GPU
134 * memory is valid, containing at least sz bytes. The goal is to eliminate
135 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
136 * overruns) by statically validating pointers.
140 pandecode_validate_buffer(mali_ptr addr
, size_t sz
)
143 pandecode_msg("XXX: null pointer deref");
149 struct pandecode_mapped_memory
*bo
=
150 pandecode_find_mapped_gpu_mem_containing(addr
);
153 pandecode_msg("XXX: invalid memory dereference\n");
159 unsigned offset
= addr
- bo
->gpu_va
;
160 unsigned total
= offset
+ sz
;
162 if (total
> bo
->length
) {
163 pandecode_msg("XXX: buffer overrun. "
164 "Chunk of size %zu at offset %d in buffer of size %zu. "
165 "Overrun by %zu bytes. \n",
166 sz
, offset
, bo
->length
, total
- bo
->length
);
171 struct pandecode_flag_info
{
177 pandecode_log_decoded_flags(const struct pandecode_flag_info
*flag_info
,
180 bool decodable_flags_found
= false;
182 for (int i
= 0; flag_info
[i
].name
; i
++) {
183 if ((flags
& flag_info
[i
].flag
) != flag_info
[i
].flag
)
186 if (!decodable_flags_found
) {
187 decodable_flags_found
= true;
189 pandecode_log_cont(" | ");
192 pandecode_log_cont("%s", flag_info
[i
].name
);
194 flags
&= ~flag_info
[i
].flag
;
197 if (decodable_flags_found
) {
199 pandecode_log_cont(" | 0x%" PRIx64
, flags
);
201 pandecode_log_cont("0x%" PRIx64
, flags
);
205 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
206 static const struct pandecode_flag_info gl_enable_flag_info
[] = {
207 FLAG_INFO(OCCLUSION_QUERY
),
208 FLAG_INFO(OCCLUSION_PRECISE
),
209 FLAG_INFO(FRONT_CCW_TOP
),
210 FLAG_INFO(CULL_FACE_FRONT
),
211 FLAG_INFO(CULL_FACE_BACK
),
216 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
217 static const struct pandecode_flag_info clear_flag_info
[] = {
220 FLAG_INFO(SLOW_STENCIL
),
225 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
226 static const struct pandecode_flag_info mask_flag_info
[] = {
235 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
236 static const struct pandecode_flag_info u3_flag_info
[] = {
238 FLAG_INFO(PER_SAMPLE
),
239 FLAG_INFO(CAN_DISCARD
),
240 FLAG_INFO(HAS_BLEND_SHADER
),
241 FLAG_INFO(DEPTH_WRITEMASK
),
242 FLAG_INFO(DEPTH_CLIP_NEAR
),
243 FLAG_INFO(DEPTH_CLIP_FAR
),
247 static const struct pandecode_flag_info u4_flag_info
[] = {
249 FLAG_INFO(NO_DITHER
),
250 FLAG_INFO(DEPTH_RANGE_A
),
251 FLAG_INFO(DEPTH_RANGE_B
),
252 FLAG_INFO(STENCIL_TEST
),
253 FLAG_INFO(ALPHA_TO_COVERAGE
),
258 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
259 static const struct pandecode_flag_info mfbd_fmt_flag_info
[] = {
265 #define FLAG_INFO(flag) { MALI_AFBC_##flag, "MALI_AFBC_" #flag }
266 static const struct pandecode_flag_info afbc_fmt_flag_info
[] = {
272 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
273 static const struct pandecode_flag_info mfbd_extra_flag_hi_info
[] = {
279 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
280 static const struct pandecode_flag_info mfbd_extra_flag_lo_info
[] = {
286 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
287 static const struct pandecode_flag_info shader_midgard1_flag_lo_info
[] = {
290 FLAG_INFO(READS_TILEBUFFER
),
291 FLAG_INFO(WRITES_GLOBAL
),
296 static const struct pandecode_flag_info shader_midgard1_flag_hi_info
[] = {
298 FLAG_INFO(SUPPRESS_INF_NAN
),
303 #define FLAG_INFO(flag) { MALI_BIFROST_##flag, "MALI_BIFROST_" #flag }
304 static const struct pandecode_flag_info shader_bifrost_info
[] = {
305 FLAG_INFO(FULL_THREAD
),
307 FLAG_INFO(FIRST_ATEST
),
313 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
314 static const struct pandecode_flag_info mfbd_flag_info
[] = {
315 FLAG_INFO(DEPTH_WRITE
),
321 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
322 static const struct pandecode_flag_info sfbd_unk1_info
[] = {
329 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
330 static const struct pandecode_flag_info sfbd_unk2_info
[] = {
337 #define DEFINE_CASE(name) case MALI_## name: return "MALI_" #name
338 static char *pandecode_format(enum mali_format format
)
340 static char unk_format_str
[10];
343 DEFINE_CASE(ETC2_RGB8
);
344 DEFINE_CASE(ETC2_R11_UNORM
);
345 DEFINE_CASE(ETC2_RGBA8
);
346 DEFINE_CASE(ETC2_RG11_UNORM
);
347 DEFINE_CASE(ETC2_R11_SNORM
);
348 DEFINE_CASE(ETC2_RG11_SNORM
);
349 DEFINE_CASE(ETC2_RGB8A1
);
351 DEFINE_CASE(BC1_UNORM
);
352 DEFINE_CASE(BC2_UNORM
);
353 DEFINE_CASE(BC3_UNORM
);
354 DEFINE_CASE(BC4_UNORM
);
355 DEFINE_CASE(BC4_SNORM
);
356 DEFINE_CASE(BC5_UNORM
);
357 DEFINE_CASE(BC5_SNORM
);
358 DEFINE_CASE(BC6H_UF16
);
359 DEFINE_CASE(BC6H_SF16
);
360 DEFINE_CASE(BC7_UNORM
);
361 DEFINE_CASE(ASTC_3D_LDR
);
362 DEFINE_CASE(ASTC_3D_HDR
);
363 DEFINE_CASE(ASTC_2D_LDR
);
364 DEFINE_CASE(ASTC_2D_HDR
);
366 DEFINE_CASE(RGB5_X1_UNORM
);
367 DEFINE_CASE(RGB5_A1_UNORM
);
368 DEFINE_CASE(RGB10_A2_UNORM
);
369 DEFINE_CASE(RGB10_A2_SNORM
);
370 DEFINE_CASE(RGB10_A2UI
);
371 DEFINE_CASE(RGB10_A2I
);
372 DEFINE_CASE(RGB332_UNORM
);
373 DEFINE_CASE(RGB233_UNORM
);
374 DEFINE_CASE(Z24X8_UNORM
);
375 DEFINE_CASE(R32_FIXED
);
376 DEFINE_CASE(RG32_FIXED
);
377 DEFINE_CASE(RGB32_FIXED
);
378 DEFINE_CASE(RGBA32_FIXED
);
379 DEFINE_CASE(R11F_G11F_B10F
);
380 DEFINE_CASE(R9F_G9F_B9F_E5F
);
381 DEFINE_CASE(VARYING_POS
);
382 DEFINE_CASE(VARYING_DISCARD
);
384 DEFINE_CASE(R8_SNORM
);
385 DEFINE_CASE(R16_SNORM
);
386 DEFINE_CASE(R32_SNORM
);
387 DEFINE_CASE(RG8_SNORM
);
388 DEFINE_CASE(RG16_SNORM
);
389 DEFINE_CASE(RG32_SNORM
);
390 DEFINE_CASE(RGB8_SNORM
);
391 DEFINE_CASE(RGB16_SNORM
);
392 DEFINE_CASE(RGB32_SNORM
);
393 DEFINE_CASE(RGBA8_SNORM
);
394 DEFINE_CASE(RGBA16_SNORM
);
395 DEFINE_CASE(RGBA32_SNORM
);
404 DEFINE_CASE(RGB16UI
);
405 DEFINE_CASE(RGB32UI
);
406 DEFINE_CASE(RGBA8UI
);
407 DEFINE_CASE(RGBA16UI
);
408 DEFINE_CASE(RGBA32UI
);
410 DEFINE_CASE(R8_UNORM
);
411 DEFINE_CASE(R16_UNORM
);
412 DEFINE_CASE(R32_UNORM
);
414 DEFINE_CASE(RG8_UNORM
);
415 DEFINE_CASE(RG16_UNORM
);
416 DEFINE_CASE(RG32_UNORM
);
418 DEFINE_CASE(RGB8_UNORM
);
419 DEFINE_CASE(RGB16_UNORM
);
420 DEFINE_CASE(RGB32_UNORM
);
422 DEFINE_CASE(RGBA4_UNORM
);
423 DEFINE_CASE(RGBA8_UNORM
);
424 DEFINE_CASE(RGBA16_UNORM
);
425 DEFINE_CASE(RGBA32_UNORM
);
426 DEFINE_CASE(RGBA32F
);
441 DEFINE_CASE(RGBA16I
);
442 DEFINE_CASE(RGBA32I
);
443 DEFINE_CASE(RGBA16F
);
446 DEFINE_CASE(RGBA8_2
);
447 DEFINE_CASE(RGB10_A2_2
);
449 snprintf(unk_format_str
, sizeof(unk_format_str
), "MALI_0x%02x", format
);
450 return unk_format_str
;
456 static char *pandecode_attr_mode_short(enum mali_attr_mode mode
)
459 /* TODO: Combine to just "instanced" once this can be done
460 * unambiguously in all known cases */
461 case MALI_ATTR_POT_DIVIDE
:
462 return "instanced_pot";
463 case MALI_ATTR_MODULO
:
464 return "instanced_mod";
465 case MALI_ATTR_NPOT_DIVIDE
:
466 return "instanced_npot";
467 case MALI_ATTR_IMAGE
:
470 pandecode_msg("XXX: invalid attribute mode %X\n", mode
);
476 pandecode_special_record(uint64_t v
, bool* attribute
)
479 case MALI_ATTR_VERTEXID
:
481 return "gl_VertexID";
482 case MALI_ATTR_INSTANCEID
:
484 return "gl_InstanceID";
485 case MALI_VARYING_FRAG_COORD
:
486 return "gl_FragCoord";
487 case MALI_VARYING_FRONT_FACING
:
488 return "gl_FrontFacing";
489 case MALI_VARYING_POINT_COORD
:
490 return "gl_PointCoord";
492 pandecode_msg("XXX: invalid special record %" PRIx64
"\n", v
);
498 /* Midgard's tiler descriptor is embedded within the
502 pandecode_midgard_tiler_descriptor(
503 const struct midgard_tiler_descriptor
*t
,
509 pandecode_log(".tiler = {\n");
512 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
513 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
515 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
517 /* We know this name from the kernel, but we never see it nonzero */
520 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
522 MEMORY_PROP(t
, polygon_list
);
524 /* The body is offset from the base of the polygon list */
525 //assert(t->polygon_list_body > t->polygon_list);
526 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
528 /* It needs to fit inside the reported size */
529 //assert(t->polygon_list_size >= body_offset);
531 /* Now that we've sanity checked, we'll try to calculate the sizes
532 * ourselves for comparison */
534 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
535 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
537 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
538 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
539 ref_header
, ref_size
);
540 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
541 pandecode_msg("body offset %d\n", body_offset
);
544 /* The tiler heap has a start and end specified -- it should be
545 * identical to what we have in the BO. The exception is if tiling is
548 MEMORY_PROP(t
, heap_start
);
549 assert(t
->heap_end
>= t
->heap_start
);
551 struct pandecode_mapped_memory
*heap
=
552 pandecode_find_mapped_gpu_mem_containing(t
->heap_start
);
554 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
556 /* Tiling is enabled with a special flag */
557 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
558 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
560 bool tiling_enabled
= hierarchy_mask
;
562 if (tiling_enabled
) {
563 /* When tiling is enabled, the heap should be a tight fit */
564 unsigned heap_offset
= t
->heap_start
- heap
->gpu_va
;
565 if ((heap_offset
+ heap_size
) != heap
->length
) {
566 pandecode_msg("XXX: heap size %u (expected %zu)\n",
567 heap_size
, heap
->length
- heap_offset
);
570 /* We should also have no other flags */
572 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
574 /* When tiling is disabled, we should have that flag and no others */
576 if (tiler_flags
!= MALI_TILER_DISABLED
) {
577 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
581 /* We should also have an empty heap */
583 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
587 /* Disabled tiling is used only for clear-only jobs, which are
588 * purely FRAGMENT, so we should never see this for
589 * non-FRAGMENT descriptors. */
592 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
595 /* We've never seen weights used in practice, but we know from the
596 * kernel these fields is there */
598 bool nonzero_weights
= false;
600 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
601 nonzero_weights
|= t
->weights
[w
] != 0x0;
604 if (nonzero_weights
) {
605 pandecode_log(".weights = { ");
607 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
608 pandecode_log_cont("%d, ", t
->weights
[w
]);
615 pandecode_log("}\n");
618 /* TODO: The Bifrost tiler is not understood at all yet */
621 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer
*fb
)
623 pandecode_log(".tiler = {\n");
626 MEMORY_PROP(fb
, tiler_meta
);
628 for (int i
= 0; i
< 16; i
++) {
629 if (fb
->zeros
[i
] != 0) {
630 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
635 pandecode_log("},\n");
638 pandecode_log("}\n");
642 /* Information about the framebuffer passed back for
643 * additional analysis */
645 struct pandecode_fbd
{
653 pandecode_sfbd_format(struct mali_sfbd_format format
)
655 pandecode_log(".format = {\n");
658 pandecode_log(".unk1 = ");
659 pandecode_log_decoded_flags(sfbd_unk1_info
, format
.unk1
);
660 pandecode_log_cont(",\n");
662 /* TODO: Map formats so we can check swizzles and print nicely */
663 pandecode_log("swizzle");
664 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
665 pandecode_log_cont(",\n");
667 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
668 (format
.nr_channels
+ 1));
670 pandecode_log(".unk2 = ");
671 pandecode_log_decoded_flags(sfbd_unk2_info
, format
.unk2
);
672 pandecode_log_cont(",\n");
674 pandecode_prop("block = %s", mali_block_format_as_str(format
.block
));
676 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
679 pandecode_log("},\n");
683 pandecode_shared_memory(const struct mali_shared_memory
*desc
, bool is_compute
)
685 pandecode_prop("stack_shift = 0x%x", desc
->stack_shift
);
688 pandecode_prop("unk0 = 0x%x", desc
->unk0
);
690 if (desc
->shared_workgroup_count
!= 0x1F) {
691 pandecode_prop("shared_workgroup_count = %d", desc
->shared_workgroup_count
);
693 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
696 if (desc
->shared_unk1
|| desc
->shared_shift
) {
697 pandecode_prop("shared_unk1 = %X", desc
->shared_unk1
);
698 pandecode_prop("shared_shift = %X", desc
->shared_shift
);
701 pandecode_msg("XXX: shared memory configured in noncompute shader");
704 if (desc
->shared_zero
) {
705 pandecode_msg("XXX: shared memory zero tripped\n");
706 pandecode_prop("shared_zero = 0x%" PRIx32
, desc
->shared_zero
);
709 if (desc
->shared_memory
&& !is_compute
)
710 pandecode_msg("XXX: shared memory used in noncompute shader\n");
712 MEMORY_PROP(desc
, scratchpad
);
713 MEMORY_PROP(desc
, shared_memory
);
714 MEMORY_PROP(desc
, unknown1
);
716 if (desc
->scratchpad
) {
717 struct pandecode_mapped_memory
*smem
=
718 pandecode_find_mapped_gpu_mem_containing(desc
->scratchpad
);
720 pandecode_msg("scratchpad size %u\n", smem
->length
);
725 static struct pandecode_fbd
726 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
, unsigned gpu_id
)
728 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
729 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
731 struct pandecode_fbd info
= {
736 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
739 pandecode_log(".shared_memory = {\n");
741 pandecode_shared_memory(&s
->shared_memory
, false);
743 pandecode_log("},\n");
745 pandecode_sfbd_format(s
->format
);
747 info
.width
= s
->width
+ 1;
748 info
.height
= s
->height
+ 1;
750 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", info
.width
);
751 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", info
.height
);
753 MEMORY_PROP(s
, checksum
);
755 if (s
->checksum_stride
)
756 pandecode_prop("checksum_stride = %d", s
->checksum_stride
);
758 MEMORY_PROP(s
, framebuffer
);
759 pandecode_prop("stride = %d", s
->stride
);
761 /* Earlier in the actual commandstream -- right before width -- but we
762 * delay to flow nicer */
764 pandecode_log(".clear_flags = ");
765 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
766 pandecode_log_cont(",\n");
768 if (s
->depth_buffer
) {
769 MEMORY_PROP(s
, depth_buffer
);
770 pandecode_prop("depth_stride = %d", s
->depth_stride
);
773 if (s
->stencil_buffer
) {
774 MEMORY_PROP(s
, stencil_buffer
);
775 pandecode_prop("stencil_stride = %d", s
->stencil_stride
);
778 if (s
->depth_stride_zero
||
779 s
->stencil_stride_zero
||
780 s
->zero7
|| s
->zero8
) {
781 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
782 pandecode_prop("depth_stride_zero = 0x%x",
783 s
->depth_stride_zero
);
784 pandecode_prop("stencil_stride_zero = 0x%x",
785 s
->stencil_stride_zero
);
786 pandecode_prop("zero7 = 0x%" PRIx32
,
788 pandecode_prop("zero8 = 0x%" PRIx32
,
792 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
793 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
794 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
795 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
796 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
799 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
800 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
801 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
802 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
803 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
806 if (s
->clear_stencil
) {
807 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
810 const struct midgard_tiler_descriptor t
= s
->tiler
;
812 bool has_hierarchy
= !(gpu_id
== 0x0720 || gpu_id
== 0x0820 || gpu_id
== 0x0830);
813 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
, has_hierarchy
);
816 pandecode_log("};\n");
818 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
819 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
820 pandecode_prop("zero5 = 0x%" PRIx32
, s
->zero5
);
822 pandecode_log_cont(".zero3 = {");
824 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
825 pandecode_log_cont("%X, ", s
->zero3
[i
]);
827 pandecode_log_cont("},\n");
829 pandecode_log_cont(".zero6 = {");
831 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
832 pandecode_log_cont("%X, ", s
->zero6
[i
]);
834 pandecode_log_cont("},\n");
840 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
842 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
843 const struct mali_shared_memory
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
845 pandecode_log("struct mali_shared_memory shared_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
847 pandecode_shared_memory(s
, true);
849 pandecode_log("},\n");
852 /* Extracts the number of components associated with a Mali format */
855 pandecode_format_component_count(enum mali_format fmt
)
857 /* Mask out the format class */
858 unsigned top
= fmt
& 0b11100000;
861 case MALI_FORMAT_SNORM
:
862 case MALI_FORMAT_UINT
:
863 case MALI_FORMAT_UNORM
:
864 case MALI_FORMAT_SINT
:
865 return ((fmt
>> 3) & 3) + 1;
872 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
875 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
878 assert(MALI_CHANNEL_RED
== 0);
880 for (unsigned c
= 0; c
< 4; ++c
) {
881 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
883 if (chan
<= MALI_CHANNEL_ALPHA
)
890 /* Validates that a (format, swizzle) pair is valid, in the sense that the
891 * swizzle doesn't access any components that are undefined in the format.
892 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
896 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
898 unsigned nr_comp
= pandecode_format_component_count(fmt
);
899 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
900 unsigned valid_mask
= (1 << nr_comp
) - 1;
901 unsigned invalid_mask
= ~valid_mask
;
903 if (access_mask
& invalid_mask
) {
904 pandecode_msg("XXX: invalid components accessed\n");
908 /* Check for the default non-swizzling swizzle so we can suppress
909 * useless printing for the defaults */
911 unsigned default_swizzles
[4] = {
912 MALI_CHANNEL_RED
| (MALI_CHANNEL_ZERO
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
913 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
914 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ONE
<< 9),
915 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ALPHA
<< 9)
918 return (swizzle
== default_swizzles
[nr_comp
- 1]);
921 /* Maps MALI_RGBA32F to rgba32f, etc */
924 pandecode_format_short(enum mali_format fmt
, bool srgb
)
926 /* We want a type-like format, so cut off the initial MALI_ */
927 char *format
= pandecode_format(fmt
);
928 format
+= strlen("MALI_");
930 unsigned len
= strlen(format
);
931 char *lower_format
= calloc(1, len
+ 1);
933 for (unsigned i
= 0; i
< len
; ++i
)
934 lower_format
[i
] = tolower(format
[i
]);
936 /* Sanity check sRGB flag is applied to RGB, per the name */
937 if (srgb
&& lower_format
[0] != 'r')
938 pandecode_msg("XXX: sRGB applied to non-colour format\n");
940 /* Just prefix with an s, so you get formats like srgba8_unorm */
942 pandecode_log_cont("s");
944 pandecode_log_cont("%s", lower_format
);
949 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
951 /* First, do some validation */
952 bool trivial_swizzle
= pandecode_validate_format_swizzle(
958 /* Next, print the swizzle */
959 pandecode_log_cont(".");
961 static const char components
[] = "rgba01";
963 for (unsigned c
= 0; c
< 4; ++c
) {
964 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
966 if (chan
>= MALI_CHANNEL_RESERVED_0
) {
967 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
970 pandecode_log_cont("%c", components
[chan
]);
975 pandecode_rt_format(struct mali_rt_format format
)
977 pandecode_log(".format = {\n");
980 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
981 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
982 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
983 pandecode_prop("unk4 = 0x%" PRIx32
, format
.unk4
);
985 pandecode_prop("block = %s", mali_block_format_as_str(format
.block
));
987 /* TODO: Map formats so we can check swizzles and print nicely */
988 pandecode_log("swizzle");
989 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
990 pandecode_log_cont(",\n");
992 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
993 (format
.nr_channels
+ 1));
995 pandecode_log(".flags = ");
996 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
997 pandecode_log_cont(",\n");
999 pandecode_prop("msaa = %s", mali_msaa_as_str(format
.msaa
));
1001 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
1002 * which is a faster hardware-based alternative to the wallpaper method
1003 * to preserve framebuffer contents across frames. In practice, MFBD
1004 * preload is buggy on Midgard, and so this is a chicken bit. If this
1005 * bit isn't set, most likely something broke unrelated to preload */
1007 if (!format
.no_preload
) {
1008 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
1009 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
1013 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
1016 pandecode_log("},\n");
1020 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct mali_framebuffer
*fb
)
1022 pandecode_log("struct mali_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
1025 for (int i
= 0; i
< (fb
->rt_count_1
+ 1); i
++) {
1026 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct mali_render_target
);
1027 struct pandecode_mapped_memory
*mem
=
1028 pandecode_find_mapped_gpu_mem_containing(rt_va
);
1029 const struct mali_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
1031 pandecode_log("{\n");
1034 pandecode_rt_format(rt
->format
);
1036 if (rt
->format
.block
== MALI_BLOCK_FORMAT_AFBC
) {
1037 pandecode_log(".afbc = {\n");
1040 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
1041 pandecode_prop("metadata = %s", a
);
1044 pandecode_prop("stride = %d", rt
->afbc
.stride
);
1046 pandecode_log(".flags = ");
1047 pandecode_log_decoded_flags(afbc_fmt_flag_info
, rt
->afbc
.flags
);
1048 pandecode_log_cont(",\n");
1051 pandecode_log("},\n");
1052 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.flags
) {
1053 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1059 MEMORY_PROP(rt
, framebuffer
);
1060 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
1062 if (rt
->layer_stride
)
1063 pandecode_prop("layer_stride = %d", rt
->layer_stride
);
1065 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
1066 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
1067 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
1068 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
1069 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
1072 if (rt
->zero1
|| rt
->zero2
) {
1073 pandecode_msg("XXX: render target zeros tripped\n");
1074 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
1075 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
1079 pandecode_log("},\n");
1083 pandecode_log("};\n");
1086 static struct pandecode_fbd
1087 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
, bool is_compute
, bool is_bifrost
)
1089 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1090 const struct mali_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
1092 struct pandecode_fbd info
;
1094 if (is_bifrost
&& fb
->msaa
.sample_locations
) {
1095 /* The blob stores all possible sample locations in a single buffer
1096 * allocated on startup, and just switches the pointer when switching
1097 * MSAA state. For now, we just put the data into the cmdstream, but we
1098 * should do something like what the blob does with a real driver.
1100 * There seem to be 32 slots for sample locations, followed by another
1101 * 16. The second 16 is just the center location followed by 15 zeros
1102 * in all the cases I've identified (maybe shader vs. depth/color
1106 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->msaa
.sample_locations
);
1108 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->msaa
.sample_locations
);
1110 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
1113 for (int i
= 0; i
< 32 + 16; i
++) {
1114 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
1118 pandecode_log("};\n");
1121 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1125 pandecode_log(".msaa = {\n");
1128 if (fb
->msaa
.sample_locations
)
1129 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
1131 pandecode_msg("XXX: sample_locations missing\n");
1133 if (fb
->msaa
.zero1
|| fb
->msaa
.zero2
|| fb
->msaa
.zero4
) {
1134 pandecode_msg("XXX: multisampling zero tripped\n");
1135 pandecode_prop("zero1 = %" PRIx64
, fb
->msaa
.zero1
);
1136 pandecode_prop("zero2 = %" PRIx64
, fb
->msaa
.zero2
);
1137 pandecode_prop("zero4 = %" PRIx64
, fb
->msaa
.zero4
);
1141 pandecode_log("},\n");
1143 pandecode_log(".shared_memory = {\n");
1145 pandecode_shared_memory(&fb
->shared_memory
, is_compute
);
1147 pandecode_log("},\n");
1150 info
.width
= fb
->width1
+ 1;
1151 info
.height
= fb
->height1
+ 1;
1152 info
.rt_count
= fb
->rt_count_1
+ 1;
1154 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
1155 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
1156 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
1157 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
1159 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
1160 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
1161 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
1162 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
1164 pandecode_log(".mfbd_flags = ");
1165 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
1166 pandecode_log_cont(",\n");
1168 if (fb
->clear_stencil
)
1169 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
1171 if (fb
->clear_depth
)
1172 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
1176 pandecode_bifrost_tiler_descriptor(fb
);
1178 const struct midgard_tiler_descriptor t
= fb
->tiler
;
1179 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
, true);
1182 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1184 if (fb
->zero3
|| fb
->zero4
) {
1185 pandecode_msg("XXX: framebuffer zeros tripped\n");
1186 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
1187 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
1191 pandecode_log("};\n");
1193 gpu_va
+= sizeof(struct mali_framebuffer
);
1195 info
.has_extra
= (fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
;
1197 if (info
.has_extra
) {
1198 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1199 const struct mali_framebuffer_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
1201 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1204 MEMORY_PROP(fbx
, checksum
);
1206 if (fbx
->checksum_stride
)
1207 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
1209 pandecode_log(".flags_hi = ");
1210 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info
, fbx
->flags_hi
);
1211 pandecode_log_cont(",\n");
1213 pandecode_log(".flags_lo = ");
1214 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info
, fbx
->flags_lo
);
1215 pandecode_log_cont(",\n");
1217 pandecode_prop("zs_block = %s", mali_block_format_as_str(fbx
->zs_block
));
1218 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx
->zs_samples
+ 1);
1220 if (fbx
->zs_block
== MALI_BLOCK_FORMAT_AFBC
) {
1221 pandecode_log(".ds_afbc = {\n");
1224 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
1225 pandecode_prop("depth_stencil_afbc_stride = %d",
1226 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
1227 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
1229 pandecode_log(".flags = ");
1230 pandecode_log_decoded_flags(afbc_fmt_flag_info
, fbx
->ds_afbc
.flags
);
1231 pandecode_log_cont(",\n");
1233 if (fbx
->ds_afbc
.padding
) {
1234 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1235 pandecode_prop("padding = 0x%" PRIx64
, fbx
->ds_afbc
.padding
);
1239 pandecode_log("},\n");
1241 pandecode_log(".ds_linear = {\n");
1244 if (fbx
->ds_linear
.depth
) {
1245 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
1246 pandecode_prop("depth_stride = %d",
1247 fbx
->ds_linear
.depth_stride
);
1248 pandecode_prop("depth_layer_stride = %d",
1249 fbx
->ds_linear
.depth_layer_stride
);
1250 } else if (fbx
->ds_linear
.depth_stride
|| fbx
->ds_linear
.depth_layer_stride
) {
1251 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx
->ds_linear
.depth_stride
, fbx
->ds_linear
.depth_layer_stride
);
1254 if (fbx
->ds_linear
.stencil
) {
1255 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1256 pandecode_prop("stencil_stride = %d",
1257 fbx
->ds_linear
.stencil_stride
);
1258 pandecode_prop("stencil_layer_stride = %d",
1259 fbx
->ds_linear
.stencil_layer_stride
);
1260 } else if (fbx
->ds_linear
.stencil_stride
|| fbx
->ds_linear
.stencil_layer_stride
) {
1261 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx
->ds_linear
.stencil_stride
, fbx
->ds_linear
.stencil_layer_stride
);
1264 if (fbx
->ds_linear
.depth_stride_zero
||
1265 fbx
->ds_linear
.stencil_stride_zero
) {
1266 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1267 pandecode_prop("depth_stride_zero = 0x%x",
1268 fbx
->ds_linear
.depth_stride_zero
);
1269 pandecode_prop("stencil_stride_zero = 0x%x",
1270 fbx
->ds_linear
.stencil_stride_zero
);
1274 pandecode_log("},\n");
1277 if (fbx
->clear_color_1
| fbx
->clear_color_2
) {
1278 pandecode_prop("clear_color_1 = 0x%" PRIx32
, fbx
->clear_color_1
);
1279 pandecode_prop("clear_color_2 = 0x%" PRIx32
, fbx
->clear_color_2
);
1283 pandecode_msg("XXX: fb_extra zeros tripped\n");
1284 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1288 pandecode_log("};\n");
1290 gpu_va
+= sizeof(struct mali_framebuffer_extra
);
1294 pandecode_render_target(gpu_va
, job_no
, fb
);
1299 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1303 pandecode_padded_vertices(unsigned shift
, unsigned k
)
1305 unsigned odd
= 2*k
+ 1;
1306 unsigned pot
= 1 << shift
;
1307 pandecode_msg("padded_num_vertices = %d\n", odd
* pot
);
1310 /* Given a magic divisor, recover what we were trying to divide by.
1312 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1313 * 0 <= m < 2^N, for N bits in m.
1315 * Let q represent the number we would like to divide by.
1317 * By definition of a magic divisor for N-bit unsigned integers (a number you
1318 * multiply by to magically get division), m is a number such that:
1320 * (m * x) & (2^N - 1) = floor(x/q).
1321 * for all x on Z where 0 <= x < 2^N
1323 * Ignore the case where any of the above values equals zero; it is irrelevant
1324 * for our purposes (instanced arrays).
1326 * Choose x = q. Then:
1328 * (m * x) & (2^N - 1) = floor(x/q).
1329 * (m * q) & (2^N - 1) = floor(q/q).
1331 * floor(q/q) = floor(1) = 1, therefore:
1333 * (m * q) & (2^N - 1) = 1
1335 * Recall the identity that the bitwise AND of one less than a power-of-two
1336 * equals the modulo with that power of two, i.e. for all x:
1338 * x & (2^N - 1) = x % N
1344 * By definition, a modular multiplicative inverse of a number m is the number
1345 * q such that with respect to a modulos M:
1349 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1354 pandecode_magic_divisor(uint32_t magic
, unsigned shift
, unsigned orig_divisor
, unsigned extra
)
1357 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1358 * shift) the most lame way possible... just repeatedly add.
1359 * Asymptoptically slow but nobody cares in practice, unless you have
1360 * massive numbers of vertices or high divisors. */
1362 unsigned inverse
= 0;
1364 /* Magic implicitly has the highest bit set */
1367 /* Depending on rounding direction */
1372 uint32_t product
= magic
* inverse
;
1384 pandecode_msg("dividing by %d (maybe off by two)\n", inverse
);
1386 /* Recall we're supposed to divide by (gl_level_divisor *
1387 * padded_num_vertices) */
1389 unsigned padded_num_vertices
= inverse
/ orig_divisor
;
1391 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices
);
1396 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1397 mali_ptr addr
, int job_no
, char *suffix
,
1398 int count
, bool varying
, enum mali_job_type job_type
)
1400 char *prefix
= varying
? "varying" : "attribute";
1404 pandecode_msg("warn: No %s records\n", prefix
);
1408 union mali_attr
*attr
= pandecode_fetch_gpu_mem(mem
, addr
, sizeof(union mali_attr
) * count
);
1410 for (int i
= 0; i
< count
; ++i
) {
1411 /* First, check for special records */
1412 if (attr
[i
].elements
< MALI_RECORD_SPECIAL
) {
1414 pandecode_msg("XXX: tripped size=%d\n", attr
[i
].size
);
1416 if (attr
[i
].stride
) {
1417 /* gl_InstanceID passes a magic divisor in the
1418 * stride field to divide by the padded vertex
1419 * count. No other records should do so, so
1420 * stride should otherwise be zero. Note that
1421 * stride in the usual attribute sense doesn't
1422 * apply to special records. */
1424 bool has_divisor
= attr
[i
].elements
== MALI_ATTR_INSTANCEID
;
1426 pandecode_log_cont("/* %smagic divisor = %X */ ",
1427 has_divisor
? "" : "XXX: ", attr
[i
].stride
);
1430 if (attr
[i
].shift
|| attr
[i
].extra_flags
) {
1431 /* Attributes use these fields for
1432 * instancing/padding/etc type issues, but
1435 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1436 varying
? "XXX: " : "",
1437 attr
[i
].shift
, attr
[i
].extra_flags
);
1440 /* Print the special record name */
1441 bool attribute
= false;
1442 pandecode_log("%s_%d = %s;\n", prefix
, i
, pandecode_special_record(attr
[i
].elements
, &attribute
));
1445 if (attribute
== varying
)
1446 pandecode_msg("XXX: mismatched special record\n");
1451 enum mali_attr_mode mode
= attr
[i
].elements
& 7;
1453 if (mode
== MALI_ATTR_UNUSED
)
1454 pandecode_msg("XXX: unused attribute record\n");
1456 /* For non-linear records, we need to print the type of record */
1457 if (mode
!= MALI_ATTR_LINEAR
)
1458 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode
));
1460 /* Print the name to link with attr_meta */
1461 pandecode_log_cont("%s_%d", prefix
, i
);
1463 /* Print the stride and size */
1464 pandecode_log_cont("<%u>[%u]", attr
[i
].stride
, attr
[i
].size
);
1466 /* TODO: Sanity check the quotient itself. It must be equal to
1467 * (or be greater than, if the driver added padding) the padded
1470 /* Finally, print the pointer */
1471 mali_ptr raw_elements
= attr
[i
].elements
& ~7;
1472 char *a
= pointer_as_memory_reference(raw_elements
);
1473 pandecode_log_cont(" = (%s);\n", a
);
1476 /* Check the pointer */
1477 pandecode_validate_buffer(raw_elements
, attr
[i
].size
);
1479 /* shift/extra_flags exist only for instanced */
1480 if (attr
[i
].shift
| attr
[i
].extra_flags
) {
1481 /* These are set to random values by the blob for
1482 * varyings, most likely a symptom of uninitialized
1483 * memory where the hardware masked the bug. As such we
1484 * put this at a warning, not an error. */
1486 if (mode
== MALI_ATTR_LINEAR
)
1487 pandecode_msg("warn: instancing fields set for linear\n");
1489 pandecode_prop("shift = %d", attr
[i
].shift
);
1490 pandecode_prop("extra_flags = %d", attr
[i
].extra_flags
);
1493 /* Decode further where possible */
1495 if (mode
== MALI_ATTR_MODULO
) {
1496 pandecode_padded_vertices(
1498 attr
[i
].extra_flags
);
1501 if (mode
== MALI_ATTR_NPOT_DIVIDE
) {
1503 pandecode_log("{\n");
1505 pandecode_prop("unk = 0x%x", attr
[i
].unk
);
1506 pandecode_prop("magic_divisor = 0x%08x", attr
[i
].magic_divisor
);
1507 if (attr
[i
].zero
!= 0)
1508 pandecode_prop("XXX: zero tripped (0x%x)\n", attr
[i
].zero
);
1509 pandecode_prop("divisor = %d", attr
[i
].divisor
);
1510 pandecode_magic_divisor(attr
[i
].magic_divisor
, attr
[i
- 1].shift
, attr
[i
].divisor
, attr
[i
- 1].extra_flags
);
1512 pandecode_log("}, \n");
1517 pandecode_log("\n");
1521 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1523 /* TODO: Decode flags */
1524 mali_ptr shader_ptr
= ptr
& ~15;
1526 char *a
= pointer_as_memory_reference(shader_ptr
);
1527 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1534 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1537 pandecode_msg("XXX: blend zero tripped: %X\n", blend
->zero1
);
1539 pandecode_log(".equation = {\n");
1542 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1543 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1545 pandecode_log(".color_mask = ");
1546 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1547 pandecode_log_cont(",\n");
1550 pandecode_log("},\n");
1553 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1556 decode_bifrost_constant(u16 constant
)
1558 float lo
= (float) (constant
& 0xFF);
1559 float hi
= (float) (constant
>> 8);
1561 return (hi
/ 255.0) + (lo
/ 65535.0);
1565 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1567 struct bifrost_blend_rt
*b
=
1568 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1570 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1573 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1574 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1575 b
->constant
, decode_bifrost_constant(b
->constant
));
1577 /* TODO figure out blend shader enable bit */
1578 pandecode_blend_equation(&b
->equation
);
1580 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1581 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1583 pandecode_log(".format = ");
1584 pandecode_format_short(b
->format
, false);
1585 pandecode_swizzle(b
->swizzle
, b
->format
);
1586 pandecode_log_cont(",\n");
1588 pandecode_prop("swizzle = 0x%" PRIx32
, b
->swizzle
);
1589 pandecode_prop("format = 0x%" PRIx32
, b
->format
);
1592 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1593 pandecode_prop("zero1 = 0x%" PRIx32
, b
->zero1
);
1596 pandecode_log(".shader_type = ");
1597 switch(b
->shader_type
) {
1598 case BIFROST_BLEND_F16
:
1599 pandecode_log_cont("BIFROST_BLEND_F16");
1601 case BIFROST_BLEND_F32
:
1602 pandecode_log_cont("BIFROST_BLEND_F32");
1604 case BIFROST_BLEND_I32
:
1605 pandecode_log_cont("BIFROST_BLEND_I32");
1607 case BIFROST_BLEND_U32
:
1608 pandecode_log_cont("BIFROST_BLEND_U32");
1610 case BIFROST_BLEND_I16
:
1611 pandecode_log_cont("BIFROST_BLEND_I16");
1613 case BIFROST_BLEND_U16
:
1614 pandecode_log_cont("BIFROST_BLEND_U16");
1617 pandecode_log_cont(",\n");
1620 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1621 pandecode_prop("zero2 = 0x%" PRIx32
, b
->zero2
);
1624 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1627 pandecode_log("},\n");
1633 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1635 /* constant/equation is in a union */
1639 pandecode_log(".blend = {\n");
1643 pandecode_shader_address("shader", blend
->shader
);
1645 pandecode_blend_equation(&blend
->equation
);
1646 pandecode_prop("constant = %f", blend
->constant
);
1650 pandecode_log("},\n");
1652 /* Return blend shader to disassemble if present */
1653 return is_shader
? (blend
->shader
& ~0xF) : 0;
1657 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1659 struct midgard_blend_rt
*b
=
1660 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1662 /* Flags determine presence of blend shader */
1663 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1665 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1668 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1670 union midgard_blend blend
= b
->blend
;
1671 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1674 pandecode_log("};\n");
1679 /* Attributes and varyings have descriptor records, which contain information
1680 * about their format and ordering with the attribute/varying buffers. We'll
1681 * want to validate that the combinations specified are self-consistent.
1685 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1688 char *prefix
= varying
? "varying" : "attribute";
1689 unsigned max_index
= 0;
1690 snprintf(base
, sizeof(base
), "%s_meta", prefix
);
1692 struct mali_attr_meta
*attr_meta
;
1693 mali_ptr p
= varying
? v
->varying_meta
: v
->attribute_meta
;
1695 struct pandecode_mapped_memory
*attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
);
1697 for (int i
= 0; i
< count
; ++i
, p
+= sizeof(struct mali_attr_meta
)) {
1698 attr_meta
= pandecode_fetch_gpu_mem(attr_mem
, p
,
1701 /* If the record is discard, it should be zero for everything else */
1703 if (attr_meta
->format
== MALI_VARYING_DISCARD
) {
1706 attr_meta
->unknown1
|
1707 attr_meta
->unknown3
|
1708 attr_meta
->src_offset
;
1711 pandecode_msg("XXX: expected empty record for varying discard\n");
1713 /* We want to look for a literal 0000 swizzle -- this
1714 * is not encoded with all zeroes, however */
1716 enum mali_channel z
= MALI_CHANNEL_ZERO
;
1717 unsigned zero_swizzle
= z
| (z
<< 3) | (z
<< 6) | (z
<< 9);
1718 bool good_swizzle
= attr_meta
->swizzle
== zero_swizzle
;
1721 pandecode_msg("XXX: expected zero swizzle for discard\n");
1724 pandecode_msg("XXX: cannot discard attribute\n");
1726 /* If we're all good, omit the record */
1727 if (!zero
&& varying
&& good_swizzle
) {
1728 pandecode_log("/* discarded varying */\n");
1733 if (attr_meta
->index
> max_index
)
1734 max_index
= attr_meta
->index
;
1736 if (attr_meta
->unknown1
!= 0x2) {
1737 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1738 pandecode_prop("unknown1 = 0x%" PRIx64
, (u64
) attr_meta
->unknown1
);
1741 if (attr_meta
->unknown3
) {
1742 pandecode_msg("XXX: unexpected unknown3 set\n");
1743 pandecode_prop("unknown3 = 0x%" PRIx64
, (u64
) attr_meta
->unknown3
);
1746 pandecode_format_short(attr_meta
->format
, false);
1747 pandecode_log_cont(" %s_%u", prefix
, attr_meta
->index
);
1749 if (attr_meta
->src_offset
)
1750 pandecode_log_cont("[%u]", attr_meta
->src_offset
);
1752 pandecode_swizzle(attr_meta
->swizzle
, attr_meta
->format
);
1754 pandecode_log_cont(";\n");
1757 pandecode_log("\n");
1759 return count
? (max_index
+ 1) : 0;
1762 /* return bits [lo, hi) of word */
1764 bits(u32 word
, u32 lo
, u32 hi
)
1767 return word
; // avoid undefined behavior with the shift
1769 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1773 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool graphics
)
1775 pandecode_log(".prefix = {\n");
1778 /* Decode invocation_count. See the comment before the definition of
1779 * invocation_count for an explanation.
1782 unsigned size_y_shift
= bits(p
->invocation_shifts
, 0, 5);
1783 unsigned size_z_shift
= bits(p
->invocation_shifts
, 5, 10);
1784 unsigned workgroups_x_shift
= bits(p
->invocation_shifts
, 10, 16);
1785 unsigned workgroups_y_shift
= bits(p
->invocation_shifts
, 16, 22);
1786 unsigned workgroups_z_shift
= bits(p
->invocation_shifts
, 22, 28);
1787 unsigned workgroups_x_shift_2
= bits(p
->invocation_shifts
, 28, 32);
1789 unsigned size_x
= bits(p
->invocation_count
, 0, size_y_shift
) + 1;
1790 unsigned size_y
= bits(p
->invocation_count
, size_y_shift
, size_z_shift
) + 1;
1791 unsigned size_z
= bits(p
->invocation_count
, size_z_shift
, workgroups_x_shift
) + 1;
1793 unsigned groups_x
= bits(p
->invocation_count
, workgroups_x_shift
, workgroups_y_shift
) + 1;
1794 unsigned groups_y
= bits(p
->invocation_count
, workgroups_y_shift
, workgroups_z_shift
) + 1;
1795 unsigned groups_z
= bits(p
->invocation_count
, workgroups_z_shift
, 32) + 1;
1797 /* Even though we have this decoded, we want to ensure that the
1798 * representation is "unique" so we don't lose anything by printing only
1799 * the final result. More specifically, we need to check that we were
1800 * passed something in canonical form, since the definition per the
1801 * hardware is inherently not unique. How? Well, take the resulting
1802 * decode and pack it ourselves! If it is bit exact with what we
1803 * decoded, we're good to go. */
1805 struct mali_vertex_tiler_prefix ref
;
1806 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, graphics
);
1809 (p
->invocation_count
== ref
.invocation_count
) &&
1810 (p
->invocation_shifts
== ref
.invocation_shifts
);
1813 pandecode_msg("XXX: non-canonical workgroups packing\n");
1814 pandecode_msg("expected: %X, %X",
1815 ref
.invocation_count
,
1816 ref
.invocation_shifts
);
1818 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1819 pandecode_prop("size_y_shift = %d", size_y_shift
);
1820 pandecode_prop("size_z_shift = %d", size_z_shift
);
1821 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift
);
1822 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift
);
1823 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift
);
1824 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2
);
1827 /* Regardless, print the decode */
1828 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1829 size_x
, size_y
, size_z
,
1830 groups_x
, groups_y
, groups_z
);
1833 if (p
->unknown_draw
)
1834 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1836 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1838 if (p
->draw_mode
!= MALI_DRAW_MODE_NONE
)
1839 pandecode_prop("draw_mode = %s", mali_draw_mode_as_str(p
->draw_mode
));
1841 /* Index count only exists for tiler jobs anyway */
1844 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1847 unsigned index_raw_size
= (p
->unknown_draw
& MALI_DRAW_INDEXED_SIZE
);
1848 index_raw_size
>>= MALI_DRAW_INDEXED_SHIFT
;
1850 /* Validate an index buffer is present if we need one. TODO: verify
1851 * relationship between invocation_count and index_count */
1854 unsigned count
= p
->index_count
;
1857 unsigned size
= (index_raw_size
== 0x3) ? 4 : index_raw_size
;
1859 /* Ensure we got a size, and if so, validate the index buffer
1860 * is large enough to hold a full set of indices of the given
1863 if (!index_raw_size
)
1864 pandecode_msg("XXX: index size missing\n");
1866 pandecode_validate_buffer(p
->indices
, count
* size
);
1867 } else if (index_raw_size
)
1868 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size
);
1870 if (p
->offset_bias_correction
)
1871 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1873 /* TODO: Figure out what this is. It's not zero */
1874 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1877 pandecode_log("},\n");
1881 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1883 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1884 uint64_t *PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1886 for (int i
= 0; i
< ubufs_count
; i
++) {
1887 unsigned size
= (ubufs
[i
] & ((1 << 10) - 1)) * 16;
1888 mali_ptr addr
= (ubufs
[i
] >> 10) << 2;
1890 pandecode_validate_buffer(addr
, size
);
1892 char *ptr
= pointer_as_memory_reference(addr
);
1893 pandecode_log("ubuf_%d[%u] = %s;\n", i
, size
, ptr
);
1897 pandecode_log("\n");
1901 pandecode_uniforms(mali_ptr uniforms
, unsigned uniform_count
)
1903 pandecode_validate_buffer(uniforms
, uniform_count
* 16);
1905 char *ptr
= pointer_as_memory_reference(uniforms
);
1906 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count
, ptr
);
1911 shader_type_for_job(unsigned type
)
1914 case MALI_JOB_TYPE_VERTEX
: return "VERTEX";
1915 case MALI_JOB_TYPE_TILER
: return "FRAGMENT";
1916 case MALI_JOB_TYPE_COMPUTE
: return "COMPUTE";
1922 static unsigned shader_id
= 0;
1924 static struct midgard_disasm_stats
1925 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1926 bool is_bifrost
, unsigned gpu_id
)
1928 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1929 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1931 /* Compute maximum possible size */
1932 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1934 /* Print some boilerplate to clearly denote the assembly (which doesn't
1935 * obey indentation rules), and actually do the disassembly! */
1937 pandecode_log_cont("\n\n");
1939 struct midgard_disasm_stats stats
;
1942 disassemble_bifrost(pandecode_dump_stream
, code
, sz
, true);
1944 /* TODO: Extend stats to Bifrost */
1945 stats
.texture_count
= -128;
1946 stats
.sampler_count
= -128;
1947 stats
.attribute_count
= -128;
1948 stats
.varying_count
= -128;
1949 stats
.uniform_count
= -128;
1950 stats
.uniform_buffer_count
= -128;
1951 stats
.work_count
= -128;
1953 stats
.instruction_count
= 0;
1954 stats
.bundle_count
= 0;
1955 stats
.quadword_count
= 0;
1956 stats
.helper_invocations
= false;
1958 stats
= disassemble_midgard(pandecode_dump_stream
,
1960 type
== MALI_JOB_TYPE_TILER
?
1961 MESA_SHADER_FRAGMENT
: MESA_SHADER_VERTEX
);
1964 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1965 * driver-internal purposes with the blob and interfere */
1967 bool should_shaderdb
= type
!= MALI_JOB_TYPE_COMPUTE
;
1969 if (should_shaderdb
) {
1970 unsigned nr_threads
=
1971 (stats
.work_count
<= 4) ? 4 :
1972 (stats
.work_count
<= 8) ? 2 :
1975 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
1976 "%u inst, %u bundles, %u quadwords, "
1977 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1979 shader_type_for_job(type
),
1980 stats
.instruction_count
, stats
.bundle_count
, stats
.quadword_count
,
1981 stats
.work_count
, nr_threads
);
1989 pandecode_texture_payload(mali_ptr payload
,
1990 enum mali_texture_dimension dim
,
1991 enum mali_texture_layout layout
,
1995 uint16_t array_size
,
1996 struct pandecode_mapped_memory
*tmem
)
1998 pandecode_log(".payload = {\n");
2001 /* A bunch of bitmap pointers follow.
2002 * We work out the correct number,
2003 * based on the mipmap/cubemap
2004 * properties, but dump extra
2005 * possibilities to futureproof */
2007 int bitmap_count
= levels
+ 1;
2009 /* Miptree for each face */
2010 if (dim
== MALI_TEXTURE_DIMENSION_CUBE
)
2013 /* Array of layers */
2014 bitmap_count
*= depth
;
2016 /* Array of textures */
2017 bitmap_count
*= array_size
;
2019 /* Stride for each element */
2023 mali_ptr
*pointers_and_strides
= pandecode_fetch_gpu_mem(tmem
,
2024 payload
, sizeof(mali_ptr
) * bitmap_count
);
2025 for (int i
= 0; i
< bitmap_count
; ++i
) {
2026 /* How we dump depends if this is a stride or a pointer */
2028 if (manual_stride
&& (i
& 1)) {
2029 /* signed 32-bit snuck in as a 64-bit pointer */
2030 uint64_t stride_set
= pointers_and_strides
[i
];
2031 uint32_t clamped_stride
= stride_set
;
2032 int32_t stride
= clamped_stride
;
2033 assert(stride_set
== clamped_stride
);
2034 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
2036 char *a
= pointer_as_memory_reference(pointers_and_strides
[i
]);
2037 pandecode_log("%s, \n", a
);
2043 pandecode_log("},\n");
2047 pandecode_texture(mali_ptr u
,
2048 struct pandecode_mapped_memory
*tmem
,
2049 unsigned job_no
, unsigned tex
)
2051 struct pandecode_mapped_memory
*mapped_mem
= pandecode_find_mapped_gpu_mem_containing(u
);
2052 const uint8_t *cl
= pandecode_fetch_gpu_mem(mapped_mem
, u
, MALI_MIDGARD_TEXTURE_LENGTH
);
2054 struct MALI_MIDGARD_TEXTURE temp
;
2055 MALI_MIDGARD_TEXTURE_unpack(cl
, &temp
);
2056 MALI_MIDGARD_TEXTURE_print(pandecode_dump_stream
, &temp
, 2);
2058 pandecode_texture_payload(u
+ MALI_MIDGARD_TEXTURE_LENGTH
,
2059 temp
.dimension
, temp
.texel_ordering
, temp
.manual_stride
,
2060 temp
.levels
, temp
.depth
, temp
.array_size
, mapped_mem
);
2064 pandecode_bifrost_texture(
2069 struct MALI_BIFROST_TEXTURE temp
;
2070 MALI_BIFROST_TEXTURE_unpack(cl
, &temp
);
2071 MALI_BIFROST_TEXTURE_print(pandecode_dump_stream
, &temp
, 2);
2073 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(temp
.surfaces
);
2074 pandecode_texture_payload(temp
.surfaces
, temp
.dimension
, temp
.texel_ordering
,
2075 true, temp
.levels
, 1, 1, tmem
);
2078 /* 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 */
2081 pandecode_shader_prop(const char *name
, unsigned claim
, signed truth
, bool fuzzy
)
2087 if (fuzzy
&& (truth
< 0))
2088 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name
, claim
, truth
);
2090 if ((truth
>= 0) && !fuzzy
) {
2091 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2092 (truth
< claim
) ? "warn" : "XXX",
2093 name
, truth
, claim
);
2094 } else if ((claim
> -truth
) && !fuzzy
) {
2095 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2096 name
, -truth
, claim
);
2097 } else if (fuzzy
&& (claim
< truth
))
2098 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2099 name
, truth
, claim
);
2101 pandecode_log(".%s = %" PRId16
, name
, claim
);
2104 pandecode_log_cont(" /* %u used */", truth
);
2106 pandecode_log_cont(",\n");
2110 pandecode_blend_shader_disassemble(mali_ptr shader
, int job_no
, int job_type
,
2111 bool is_bifrost
, unsigned gpu_id
)
2113 struct midgard_disasm_stats stats
=
2114 pandecode_shader_disassemble(shader
, job_no
, job_type
, is_bifrost
, gpu_id
);
2116 bool has_texture
= (stats
.texture_count
> 0);
2117 bool has_sampler
= (stats
.sampler_count
> 0);
2118 bool has_attribute
= (stats
.attribute_count
> 0);
2119 bool has_varying
= (stats
.varying_count
> 0);
2120 bool has_uniform
= (stats
.uniform_count
> 0);
2121 bool has_ubo
= (stats
.uniform_buffer_count
> 0);
2123 if (has_texture
|| has_sampler
)
2124 pandecode_msg("XXX: blend shader accessing textures\n");
2126 if (has_attribute
|| has_varying
)
2127 pandecode_msg("XXX: blend shader accessing interstage\n");
2129 if (has_uniform
|| has_ubo
)
2130 pandecode_msg("XXX: blend shader accessing uniforms\n");
2134 pandecode_textures(mali_ptr textures
, unsigned texture_count
, int job_no
, bool is_bifrost
)
2136 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(textures
);
2141 pandecode_log("Textures (%"PRIx64
"):\n", textures
);
2144 const void *cl
= pandecode_fetch_gpu_mem(mmem
,
2145 textures
, MALI_BIFROST_TEXTURE_LENGTH
*
2148 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
2149 pandecode_bifrost_texture(cl
+
2150 MALI_BIFROST_TEXTURE_LENGTH
* tex
,
2154 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
);
2156 for (int tex
= 0; tex
< texture_count
; ++tex
) {
2157 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
2158 char *a
= pointer_as_memory_reference(*u
);
2159 pandecode_log("%s,\n", a
);
2163 /* Now, finally, descend down into the texture descriptor */
2164 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
2165 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
2166 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
2168 pandecode_texture(*u
, tmem
, job_no
, tex
);
2174 pandecode_samplers(mali_ptr samplers
, unsigned sampler_count
, int job_no
, bool is_bifrost
)
2176 for (int i
= 0; i
< sampler_count
; ++i
) {
2178 DUMP_ADDR("Sampler", BIFROST_SAMPLER
, samplers
+ (MALI_BIFROST_SAMPLER_LENGTH
* i
), 1);
2180 DUMP_ADDR("Sampler", MIDGARD_SAMPLER
, samplers
+ (MALI_MIDGARD_SAMPLER_LENGTH
* i
), 1);
2186 pandecode_vertex_tiler_postfix_pre(
2187 const struct mali_vertex_tiler_postfix
*p
,
2188 int job_no
, enum mali_job_type job_type
,
2189 char *suffix
, bool is_bifrost
, unsigned gpu_id
)
2191 struct pandecode_mapped_memory
*attr_mem
;
2193 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2194 * are the only things actually needed from the FBD, vertex/tiler jobs
2195 * no longer reference the FBD -- instead, this field points to some
2196 * info about the scratchpad.
2199 struct pandecode_fbd fbd_info
= {
2200 /* Default for Bifrost */
2205 pandecode_log_cont("\t/* %X %/\n", p
->shared_memory
& 1);
2206 pandecode_compute_fbd(p
->shared_memory
& ~1, job_no
);
2207 } else if (p
->shared_memory
& MALI_MFBD
)
2208 fbd_info
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->shared_memory
) & FBD_MASK
, job_no
, false, job_type
== MALI_JOB_TYPE_COMPUTE
, false);
2209 else if (job_type
== MALI_JOB_TYPE_COMPUTE
)
2210 pandecode_compute_fbd((u64
) (uintptr_t) p
->shared_memory
, job_no
);
2212 fbd_info
= pandecode_sfbd((u64
) (uintptr_t) p
->shared_memory
, job_no
, false, gpu_id
);
2214 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
2215 int texture_count
= 0, sampler_count
= 0;
2218 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->shader
);
2219 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, p
->shader
);
2221 /* Disassemble ahead-of-time to get stats. Initialize with
2222 * stats for the missing-shader case so we get validation
2225 struct midgard_disasm_stats info
= {
2228 .attribute_count
= 0,
2232 .uniform_count
= -128,
2233 .uniform_buffer_count
= 0
2236 if (s
->shader
& ~0xF)
2237 info
= pandecode_shader_disassemble(s
->shader
& ~0xF, job_no
, job_type
, is_bifrost
, gpu_id
);
2239 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", p
->shader
, job_no
, suffix
);
2242 /* Save for dumps */
2243 attribute_count
= s
->attribute_count
;
2244 varying_count
= s
->varying_count
;
2245 texture_count
= s
->texture_count
;
2246 sampler_count
= s
->sampler_count
;
2249 uniform_count
= s
->bifrost2
.uniform_count
;
2250 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
2252 uniform_count
= s
->midgard1
.uniform_count
;
2253 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
2256 pandecode_shader_address("shader", s
->shader
);
2258 pandecode_shader_prop("texture_count", s
->texture_count
, info
.texture_count
, false);
2259 pandecode_shader_prop("sampler_count", s
->sampler_count
, info
.sampler_count
, false);
2260 pandecode_shader_prop("attribute_count", s
->attribute_count
, info
.attribute_count
, false);
2261 pandecode_shader_prop("varying_count", s
->varying_count
, info
.varying_count
, false);
2262 pandecode_shader_prop("uniform_buffer_count",
2263 uniform_buffer_count
,
2264 info
.uniform_buffer_count
, true);
2267 pandecode_shader_prop("uniform_count",
2269 info
.uniform_count
, false);
2271 pandecode_shader_prop("work_count",
2272 s
->midgard1
.work_count
, info
.work_count
, false);
2276 pandecode_log("bifrost1.unk1 = ");
2277 pandecode_log_decoded_flags(shader_bifrost_info
, s
->bifrost1
.unk1
);
2278 pandecode_log_cont(",\n");
2280 bool helpers
= s
->midgard1
.flags_lo
& MALI_HELPER_INVOCATIONS
;
2282 if (helpers
!= info
.helper_invocations
) {
2283 pandecode_msg("XXX: expected helpers %u but got %u\n",
2284 info
.helper_invocations
, helpers
);
2287 pandecode_log(".midgard1.flags_lo = ");
2288 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info
,
2289 s
->midgard1
.flags_lo
& ~MALI_HELPER_INVOCATIONS
);
2290 pandecode_log_cont(",\n");
2292 pandecode_log(".midgard1.flags_hi = ");
2293 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info
, s
->midgard1
.flags_hi
);
2294 pandecode_log_cont(",\n");
2297 if (s
->depth_units
|| s
->depth_factor
) {
2298 pandecode_prop("depth_factor = %f", s
->depth_factor
);
2299 pandecode_prop("depth_units = %f", s
->depth_units
);
2302 if (s
->coverage_mask
)
2303 pandecode_prop("coverage_mask = 0x%X", s
->coverage_mask
);
2306 pandecode_prop(".unknown2_2 = %X", s
->unknown2_2
);
2308 if (s
->unknown2_3
|| s
->unknown2_4
) {
2309 pandecode_log(".unknown2_3 = ");
2311 int unknown2_3
= s
->unknown2_3
;
2312 int unknown2_4
= s
->unknown2_4
;
2314 /* We're not quite sure what these flags mean without the depth test, if anything */
2316 if (unknown2_3
& (MALI_DEPTH_WRITEMASK
| MALI_DEPTH_FUNC_MASK
)) {
2317 const char *func
= mali_func_as_str(MALI_GET_DEPTH_FUNC(unknown2_3
));
2318 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
2320 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
2323 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
2324 pandecode_log_cont(",\n");
2326 pandecode_log(".unknown2_4 = ");
2327 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
2328 pandecode_log_cont(",\n");
2331 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
2332 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
2333 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
2336 DUMP_CL("Stencil front", STENCIL
, &s
->stencil_front
, 1);
2337 DUMP_CL("Stencil back", STENCIL
, &s
->stencil_back
, 1);
2340 pandecode_log(".bifrost2 = {\n");
2343 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
2344 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
2345 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
2346 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
2349 pandecode_log("},\n");
2350 } else if (s
->midgard2
.unknown2_7
) {
2351 pandecode_log(".midgard2 = {\n");
2354 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
2356 pandecode_log("},\n");
2360 pandecode_msg("XXX: shader padding tripped\n");
2361 pandecode_prop("padding = 0x%" PRIx32
, s
->padding
);
2365 /* TODO: Blend shaders routing/disasm */
2366 union midgard_blend blend
= s
->blend
;
2367 mali_ptr shader
= pandecode_midgard_blend(&blend
, s
->unknown2_3
& MALI_HAS_BLEND_SHADER
);
2369 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
2371 pandecode_msg("mdg_blend = %" PRIx64
"\n", s
->blend
.shader
);
2375 pandecode_log("};\n");
2377 /* MRT blend fields are used whenever MFBD is used, with
2378 * per-RT descriptors */
2380 if (job_type
== MALI_JOB_TYPE_TILER
&& (is_bifrost
|| p
->shared_memory
& MALI_MFBD
)) {
2381 void* blend_base
= (void *) (s
+ 1);
2383 for (unsigned i
= 0; i
< fbd_info
.rt_count
; i
++) {
2384 mali_ptr shader
= 0;
2387 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
2389 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
2392 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
2397 pandecode_msg("XXX: missing shader descriptor\n");
2400 DUMP_ADDR("Viewport", VIEWPORT
, p
->viewport
, 1);
2402 unsigned max_attr_index
= 0;
2404 if (p
->attribute_meta
)
2405 max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
2407 if (p
->attributes
) {
2408 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
2409 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false, job_type
);
2412 /* Varyings are encoded like attributes but not actually sent; we just
2413 * pass a zero buffer with the right stride/size set, (or whatever)
2414 * since the GPU will write to it itself */
2416 if (p
->varying_meta
) {
2417 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
2421 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
2423 /* Number of descriptors depends on whether there are
2424 * non-internal varyings */
2426 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true, job_type
);
2429 if (p
->uniform_buffers
) {
2430 if (uniform_buffer_count
)
2431 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
2433 pandecode_msg("warn: UBOs specified but not referenced\n");
2434 } else if (uniform_buffer_count
)
2435 pandecode_msg("XXX: UBOs referenced but not specified\n");
2437 /* We don't want to actually dump uniforms, but we do need to validate
2438 * that the counts we were given are sane */
2442 pandecode_uniforms(p
->uniforms
, uniform_count
);
2444 pandecode_msg("warn: Uniforms specified but not referenced\n");
2445 } else if (uniform_count
)
2446 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2449 pandecode_textures(p
->textures
, texture_count
, job_no
, is_bifrost
);
2451 if (p
->sampler_descriptor
)
2452 pandecode_samplers(p
->sampler_descriptor
, sampler_count
, job_no
, is_bifrost
);
2456 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2458 pandecode_log(".gl_enables = ");
2460 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2462 pandecode_log_cont(",\n");
2466 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2468 if (p
->shader
& 0xF)
2469 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p
->shader
& 0xF));
2471 pandecode_log(".postfix = {\n");
2474 pandecode_gl_enables(p
->gl_enables
, MALI_JOB_TYPE_TILER
);
2475 pandecode_prop("instance_shift = 0x%x", p
->instance_shift
);
2476 pandecode_prop("instance_odd = 0x%x", p
->instance_odd
);
2479 pandecode_msg("XXX: vertex only zero tripped");
2480 pandecode_prop("zero4 = 0x%" PRIx32
, p
->zero4
);
2483 pandecode_prop("offset_start = 0x%x", p
->offset_start
);
2486 pandecode_msg("XXX: vertex only zero tripped");
2487 pandecode_prop("zero5 = 0x%" PRIx64
, p
->zero5
);
2490 MEMORY_PROP(p
, position_varying
);
2491 MEMORY_PROP(p
, occlusion_counter
);
2494 pandecode_log("},\n");
2498 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2500 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2501 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2503 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2507 pandecode_msg("XXX: tiler heap zero tripped\n");
2508 pandecode_prop("zero = 0x%x", h
->zero
);
2511 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2512 MEMORY_PROP(h
, tiler_heap_start
);
2513 MEMORY_PROP(h
, tiler_heap_free
);
2515 /* this might point to the beginning of another buffer, when it's
2516 * really the end of the tiler heap buffer, so we have to be careful
2517 * here. but for zero length, we need the same pointer.
2520 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2521 MEMORY_PROP(h
, tiler_heap_start
);
2523 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2524 pandecode_prop("tiler_heap_end = %s + 1", a
);
2528 for (int i
= 0; i
< 10; i
++) {
2529 if (h
->zeros
[i
] != 0) {
2530 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2535 if (h
->unk1
!= 0x1) {
2536 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2537 pandecode_prop("unk1 = 0x%x", h
->unk1
);
2540 if (h
->unk7e007e
!= 0x7e007e) {
2541 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
2542 pandecode_prop("unk7e007e = 0x%x", h
->unk7e007e
);
2546 pandecode_log("};\n");
2550 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2552 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2553 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2555 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2557 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2560 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32
, t
->tiler_heap_next_start
);
2561 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32
, t
->used_hierarchy_mask
);
2563 if (t
->hierarchy_mask
!= 0xa &&
2564 t
->hierarchy_mask
!= 0x14 &&
2565 t
->hierarchy_mask
!= 0x28 &&
2566 t
->hierarchy_mask
!= 0x50 &&
2567 t
->hierarchy_mask
!= 0xa0)
2568 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
2570 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2572 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2574 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2575 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2578 pandecode_msg("XXX: tiler meta zero tripped\n");
2579 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2582 for (int i
= 0; i
< 12; i
++) {
2583 if (t
->zeros
[i
] != 0) {
2584 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2590 pandecode_log("};\n");
2594 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2596 if (u
.pointer
== 0x0)
2599 pandecode_log(".primitive_size = {\n");
2603 pandecode_prop("constant = %f", u
.constant
);
2605 MEMORY_PROP((&u
), pointer
);
2609 pandecode_log("},\n");
2613 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2615 pandecode_log_cont("{\n");
2618 /* TODO: gl_PointSize on Bifrost */
2619 pandecode_primitive_size(t
->primitive_size
, true);
2621 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2623 pandecode_msg("XXX: tiler only zero tripped\n");
2624 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2625 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2626 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2627 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2628 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2629 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2633 pandecode_log("},\n");
2637 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2638 const struct pandecode_mapped_memory
*mem
,
2639 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2641 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2643 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2645 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2648 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2649 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2652 pandecode_log("};\n");
2658 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2659 const struct pandecode_mapped_memory
*mem
,
2660 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2662 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2664 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2665 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2667 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2670 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2672 pandecode_log(".tiler = ");
2673 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2675 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2678 pandecode_log("};\n");
2684 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2685 const struct pandecode_mapped_memory
*mem
,
2686 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2688 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2689 bool is_graphics
= (h
->job_type
== MALI_JOB_TYPE_VERTEX
) || (h
->job_type
== MALI_JOB_TYPE_TILER
);
2691 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false, gpu_id
);
2693 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2696 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, is_graphics
);
2697 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2699 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2700 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2703 pandecode_log("};\n");
2709 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2710 mali_ptr payload
, int job_no
,
2711 bool is_bifrost
, unsigned gpu_id
)
2713 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2715 bool is_mfbd
= s
->framebuffer
& MALI_MFBD
;
2717 if (!is_mfbd
&& is_bifrost
)
2718 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2720 struct pandecode_fbd info
;
2723 info
= pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true, false, is_bifrost
);
2725 info
= pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true, gpu_id
);
2727 /* Compute the tag for the tagged pointer. This contains the type of
2728 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2729 * additional structures follow the MFBD header (an extra payload or
2730 * not, as well as a count of render targets) */
2732 unsigned expected_tag
= is_mfbd
? MALI_MFBD
: 0;
2736 expected_tag
|= MALI_MFBD_TAG_EXTRA
;
2738 expected_tag
|= (MALI_POSITIVE(info
.rt_count
) << 2);
2741 if ((s
->min_tile_coord
| s
->max_tile_coord
) & ~(MALI_X_COORD_MASK
| MALI_Y_COORD_MASK
)) {
2742 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2743 pandecode_prop("min_tile_coord = 0x%X\n", s
->min_tile_coord
);
2744 pandecode_prop("max_tile_coord = 0x%X\n", s
->max_tile_coord
);
2747 /* Extract tile coordinates */
2749 unsigned min_x
= MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2750 unsigned min_y
= MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2752 unsigned max_x
= (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2753 unsigned max_y
= (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2755 /* For the max, we also want the floored (rather than ceiled) version for checking */
2757 unsigned max_x_f
= (MALI_TILE_COORD_X(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2758 unsigned max_y_f
= (MALI_TILE_COORD_Y(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2760 /* Validate the coordinates are well-ordered */
2763 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2764 else if (min_x
> max_x
)
2765 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2768 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2769 else if (min_y
> max_y
)
2770 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2772 /* Validate the coordinates fit inside the framebuffer. We use floor,
2773 * rather than ceil, for the max coordinates, since the tile
2774 * coordinates for something like an 800x600 framebuffer will actually
2775 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2777 if ((min_x
> info
.width
) || (max_x_f
> info
.width
))
2778 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2780 if ((min_y
> info
.height
) || (max_y_f
> info
.height
))
2781 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2783 /* After validation, we print */
2785 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x
, min_y
, max_x
, max_y
);
2787 /* The FBD is a tagged pointer */
2789 unsigned tag
= (s
->framebuffer
& ~FBD_MASK
);
2791 if (tag
!= expected_tag
)
2792 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag
, tag
);
2797 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
2798 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
2799 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
2800 * might wish to combine this with the bifrost parameter) because some details
2801 * are model-specific even within a particular architecture. Minimal traces
2802 * *only* examine the job descriptors, skipping printing entirely if there is
2803 * no faults, and only descends into the payload if there are faults. This is
2804 * useful for looking for faults without the overhead of invasive traces. */
2807 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
, unsigned gpu_id
, bool minimal
)
2809 pandecode_dump_file_open();
2811 struct mali_job_descriptor_header
*h
;
2812 unsigned job_descriptor_number
= 0;
2815 struct pandecode_mapped_memory
*mem
=
2816 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2820 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2822 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2823 * high 32-bits of the 64-bit pointer are reused to store
2826 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
2827 h
->job_type
!= MALI_JOB_TYPE_FRAGMENT
? 4 : 0;
2828 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
2830 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
, 256);
2832 int job_no
= job_descriptor_number
++;
2834 /* If the job is good to go, skip it in minimal mode */
2835 if (minimal
&& (h
->exception_status
== 0x0 || h
->exception_status
== 0x1))
2838 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2841 pandecode_prop("job_type = %s", mali_job_type_as_str(h
->job_type
));
2843 if (h
->job_descriptor_size
)
2844 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2846 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2847 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2848 h
->exception_status
,
2849 (h
->exception_status
>> 16) & 0xFFFF,
2850 mali_exception_access_as_str((h
->exception_status
>> 8) & 0x3),
2851 h
->exception_status
& 0xFF);
2853 if (h
->first_incomplete_task
)
2854 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2856 if (h
->fault_pointer
)
2857 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2860 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2862 pandecode_prop("job_index = %d", h
->job_index
);
2864 if (h
->unknown_flags
)
2865 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2867 if (h
->job_dependency_index_1
)
2868 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2870 if (h
->job_dependency_index_2
)
2871 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2874 pandecode_log("};\n");
2876 switch (h
->job_type
) {
2877 case MALI_JOB_TYPE_WRITE_VALUE
: {
2878 struct mali_payload_write_value
*s
= payload
;
2879 pandecode_log("struct mali_payload_write_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2881 MEMORY_PROP(s
, address
);
2883 if (s
->value_descriptor
!= MALI_WRITE_VALUE_ZERO
) {
2884 pandecode_msg("XXX: unknown value descriptor\n");
2885 pandecode_prop("value_descriptor = 0x%" PRIX32
, s
->value_descriptor
);
2889 pandecode_msg("XXX: set value tripped\n");
2890 pandecode_prop("reserved = 0x%" PRIX32
, s
->reserved
);
2893 pandecode_prop("immediate = 0x%" PRIX64
, s
->immediate
);
2895 pandecode_log("};\n");
2900 case MALI_JOB_TYPE_TILER
:
2901 case MALI_JOB_TYPE_VERTEX
:
2902 case MALI_JOB_TYPE_COMPUTE
:
2904 if (h
->job_type
== MALI_JOB_TYPE_TILER
)
2905 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2907 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2909 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2913 case MALI_JOB_TYPE_FRAGMENT
:
2914 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
, gpu_id
);
2920 } while ((jc_gpu_va
= h
->next_job
));
2922 pandecode_map_read_write();