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 #define DEFINE_CASE(name) case MALI_MSAA_ ## name: return "MALI_MSAA_" #name
458 pandecode_msaa_mode(enum mali_msaa_mode mode
)
462 DEFINE_CASE(AVERAGE
);
463 DEFINE_CASE(MULTIPLE
);
464 DEFINE_CASE(LAYERED
);
466 unreachable("Impossible");
472 static char *pandecode_attr_mode_short(enum mali_attr_mode mode
)
475 /* TODO: Combine to just "instanced" once this can be done
476 * unambiguously in all known cases */
477 case MALI_ATTR_POT_DIVIDE
:
478 return "instanced_pot";
479 case MALI_ATTR_MODULO
:
480 return "instanced_mod";
481 case MALI_ATTR_NPOT_DIVIDE
:
482 return "instanced_npot";
483 case MALI_ATTR_IMAGE
:
486 pandecode_msg("XXX: invalid attribute mode %X\n", mode
);
492 pandecode_special_record(uint64_t v
, bool* attribute
)
495 case MALI_ATTR_VERTEXID
:
497 return "gl_VertexID";
498 case MALI_ATTR_INSTANCEID
:
500 return "gl_InstanceID";
501 case MALI_VARYING_FRAG_COORD
:
502 return "gl_FragCoord";
503 case MALI_VARYING_FRONT_FACING
:
504 return "gl_FrontFacing";
505 case MALI_VARYING_POINT_COORD
:
506 return "gl_PointCoord";
508 pandecode_msg("XXX: invalid special record %" PRIx64
"\n", v
);
513 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
515 pandecode_block_format(enum mali_block_format fmt
)
519 DEFINE_CASE(UNKNOWN
);
524 unreachable("Invalid case");
529 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
531 pandecode_exception_access(unsigned access
)
535 DEFINE_CASE(EXECUTE
);
540 unreachable("Invalid case");
545 /* Midgard's tiler descriptor is embedded within the
549 pandecode_midgard_tiler_descriptor(
550 const struct midgard_tiler_descriptor
*t
,
556 pandecode_log(".tiler = {\n");
559 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
560 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
562 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
564 /* We know this name from the kernel, but we never see it nonzero */
567 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
569 MEMORY_PROP(t
, polygon_list
);
571 /* The body is offset from the base of the polygon list */
572 //assert(t->polygon_list_body > t->polygon_list);
573 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
575 /* It needs to fit inside the reported size */
576 //assert(t->polygon_list_size >= body_offset);
578 /* Now that we've sanity checked, we'll try to calculate the sizes
579 * ourselves for comparison */
581 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
582 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
584 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
585 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
586 ref_header
, ref_size
);
587 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
588 pandecode_msg("body offset %d\n", body_offset
);
591 /* The tiler heap has a start and end specified -- it should be
592 * identical to what we have in the BO. The exception is if tiling is
595 MEMORY_PROP(t
, heap_start
);
596 assert(t
->heap_end
>= t
->heap_start
);
598 struct pandecode_mapped_memory
*heap
=
599 pandecode_find_mapped_gpu_mem_containing(t
->heap_start
);
601 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
603 /* Tiling is enabled with a special flag */
604 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
605 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
607 bool tiling_enabled
= hierarchy_mask
;
609 if (tiling_enabled
) {
610 /* When tiling is enabled, the heap should be a tight fit */
611 unsigned heap_offset
= t
->heap_start
- heap
->gpu_va
;
612 if ((heap_offset
+ heap_size
) != heap
->length
) {
613 pandecode_msg("XXX: heap size %u (expected %zu)\n",
614 heap_size
, heap
->length
- heap_offset
);
617 /* We should also have no other flags */
619 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
621 /* When tiling is disabled, we should have that flag and no others */
623 if (tiler_flags
!= MALI_TILER_DISABLED
) {
624 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
628 /* We should also have an empty heap */
630 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
634 /* Disabled tiling is used only for clear-only jobs, which are
635 * purely FRAGMENT, so we should never see this for
636 * non-FRAGMENT descriptors. */
639 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
642 /* We've never seen weights used in practice, but we know from the
643 * kernel these fields is there */
645 bool nonzero_weights
= false;
647 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
648 nonzero_weights
|= t
->weights
[w
] != 0x0;
651 if (nonzero_weights
) {
652 pandecode_log(".weights = { ");
654 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
655 pandecode_log_cont("%d, ", t
->weights
[w
]);
662 pandecode_log("}\n");
665 /* TODO: The Bifrost tiler is not understood at all yet */
668 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer
*fb
)
670 pandecode_log(".tiler = {\n");
673 MEMORY_PROP(fb
, tiler_meta
);
675 for (int i
= 0; i
< 16; i
++) {
676 if (fb
->zeros
[i
] != 0) {
677 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
682 pandecode_log("},\n");
685 pandecode_log("}\n");
689 /* Information about the framebuffer passed back for
690 * additional analysis */
692 struct pandecode_fbd
{
700 pandecode_sfbd_format(struct mali_sfbd_format format
)
702 pandecode_log(".format = {\n");
705 pandecode_log(".unk1 = ");
706 pandecode_log_decoded_flags(sfbd_unk1_info
, format
.unk1
);
707 pandecode_log_cont(",\n");
709 /* TODO: Map formats so we can check swizzles and print nicely */
710 pandecode_log("swizzle");
711 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
712 pandecode_log_cont(",\n");
714 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
715 (format
.nr_channels
+ 1));
717 pandecode_log(".unk2 = ");
718 pandecode_log_decoded_flags(sfbd_unk2_info
, format
.unk2
);
719 pandecode_log_cont(",\n");
721 pandecode_prop("block = %s", pandecode_block_format(format
.block
));
723 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
726 pandecode_log("},\n");
730 pandecode_shared_memory(const struct mali_shared_memory
*desc
, bool is_compute
)
732 pandecode_prop("stack_shift = 0x%x", desc
->stack_shift
);
735 pandecode_prop("unk0 = 0x%x", desc
->unk0
);
737 if (desc
->shared_workgroup_count
!= 0x1F) {
738 pandecode_prop("shared_workgroup_count = %d", desc
->shared_workgroup_count
);
740 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
743 if (desc
->shared_unk1
|| desc
->shared_shift
) {
744 pandecode_prop("shared_unk1 = %X", desc
->shared_unk1
);
745 pandecode_prop("shared_shift = %X", desc
->shared_shift
);
748 pandecode_msg("XXX: shared memory configured in noncompute shader");
751 if (desc
->shared_zero
) {
752 pandecode_msg("XXX: shared memory zero tripped\n");
753 pandecode_prop("shared_zero = 0x%" PRIx32
, desc
->shared_zero
);
756 if (desc
->shared_memory
&& !is_compute
)
757 pandecode_msg("XXX: shared memory used in noncompute shader\n");
759 MEMORY_PROP(desc
, scratchpad
);
760 MEMORY_PROP(desc
, shared_memory
);
761 MEMORY_PROP(desc
, unknown1
);
763 if (desc
->scratchpad
) {
764 struct pandecode_mapped_memory
*smem
=
765 pandecode_find_mapped_gpu_mem_containing(desc
->scratchpad
);
767 pandecode_msg("scratchpad size %u\n", smem
->length
);
772 static struct pandecode_fbd
773 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
, unsigned gpu_id
)
775 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
776 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
778 struct pandecode_fbd info
= {
783 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
786 pandecode_log(".shared_memory = {\n");
788 pandecode_shared_memory(&s
->shared_memory
, false);
790 pandecode_log("},\n");
792 pandecode_sfbd_format(s
->format
);
794 info
.width
= s
->width
+ 1;
795 info
.height
= s
->height
+ 1;
797 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", info
.width
);
798 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", info
.height
);
800 MEMORY_PROP(s
, checksum
);
802 if (s
->checksum_stride
)
803 pandecode_prop("checksum_stride = %d", s
->checksum_stride
);
805 MEMORY_PROP(s
, framebuffer
);
806 pandecode_prop("stride = %d", s
->stride
);
808 /* Earlier in the actual commandstream -- right before width -- but we
809 * delay to flow nicer */
811 pandecode_log(".clear_flags = ");
812 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
813 pandecode_log_cont(",\n");
815 if (s
->depth_buffer
) {
816 MEMORY_PROP(s
, depth_buffer
);
817 pandecode_prop("depth_stride = %d", s
->depth_stride
);
820 if (s
->stencil_buffer
) {
821 MEMORY_PROP(s
, stencil_buffer
);
822 pandecode_prop("stencil_stride = %d", s
->stencil_stride
);
825 if (s
->depth_stride_zero
||
826 s
->stencil_stride_zero
||
827 s
->zero7
|| s
->zero8
) {
828 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
829 pandecode_prop("depth_stride_zero = 0x%x",
830 s
->depth_stride_zero
);
831 pandecode_prop("stencil_stride_zero = 0x%x",
832 s
->stencil_stride_zero
);
833 pandecode_prop("zero7 = 0x%" PRIx32
,
835 pandecode_prop("zero8 = 0x%" PRIx32
,
839 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
840 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
841 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
842 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
843 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
846 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
847 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
848 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
849 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
850 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
853 if (s
->clear_stencil
) {
854 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
857 const struct midgard_tiler_descriptor t
= s
->tiler
;
859 bool has_hierarchy
= !(gpu_id
== 0x0720 || gpu_id
== 0x0820 || gpu_id
== 0x0830);
860 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
, has_hierarchy
);
863 pandecode_log("};\n");
865 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
866 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
867 pandecode_prop("zero5 = 0x%" PRIx32
, s
->zero5
);
869 pandecode_log_cont(".zero3 = {");
871 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
872 pandecode_log_cont("%X, ", s
->zero3
[i
]);
874 pandecode_log_cont("},\n");
876 pandecode_log_cont(".zero6 = {");
878 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
879 pandecode_log_cont("%X, ", s
->zero6
[i
]);
881 pandecode_log_cont("},\n");
887 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
889 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
890 const struct mali_shared_memory
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
892 pandecode_log("struct mali_shared_memory shared_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
894 pandecode_shared_memory(s
, true);
896 pandecode_log("},\n");
899 /* Extracts the number of components associated with a Mali format */
902 pandecode_format_component_count(enum mali_format fmt
)
904 /* Mask out the format class */
905 unsigned top
= fmt
& 0b11100000;
908 case MALI_FORMAT_SNORM
:
909 case MALI_FORMAT_UINT
:
910 case MALI_FORMAT_UNORM
:
911 case MALI_FORMAT_SINT
:
912 return ((fmt
>> 3) & 3) + 1;
919 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
922 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
925 assert(MALI_CHANNEL_RED
== 0);
927 for (unsigned c
= 0; c
< 4; ++c
) {
928 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
930 if (chan
<= MALI_CHANNEL_ALPHA
)
937 /* Validates that a (format, swizzle) pair is valid, in the sense that the
938 * swizzle doesn't access any components that are undefined in the format.
939 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
943 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
945 unsigned nr_comp
= pandecode_format_component_count(fmt
);
946 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
947 unsigned valid_mask
= (1 << nr_comp
) - 1;
948 unsigned invalid_mask
= ~valid_mask
;
950 if (access_mask
& invalid_mask
) {
951 pandecode_msg("XXX: invalid components accessed\n");
955 /* Check for the default non-swizzling swizzle so we can suppress
956 * useless printing for the defaults */
958 unsigned default_swizzles
[4] = {
959 MALI_CHANNEL_RED
| (MALI_CHANNEL_ZERO
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
960 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_ZERO
<< 6) | (MALI_CHANNEL_ONE
<< 9),
961 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ONE
<< 9),
962 MALI_CHANNEL_RED
| (MALI_CHANNEL_GREEN
<< 3) | (MALI_CHANNEL_BLUE
<< 6) | (MALI_CHANNEL_ALPHA
<< 9)
965 return (swizzle
== default_swizzles
[nr_comp
- 1]);
968 /* Maps MALI_RGBA32F to rgba32f, etc */
971 pandecode_format_short(enum mali_format fmt
, bool srgb
)
973 /* We want a type-like format, so cut off the initial MALI_ */
974 char *format
= pandecode_format(fmt
);
975 format
+= strlen("MALI_");
977 unsigned len
= strlen(format
);
978 char *lower_format
= calloc(1, len
+ 1);
980 for (unsigned i
= 0; i
< len
; ++i
)
981 lower_format
[i
] = tolower(format
[i
]);
983 /* Sanity check sRGB flag is applied to RGB, per the name */
984 if (srgb
&& lower_format
[0] != 'r')
985 pandecode_msg("XXX: sRGB applied to non-colour format\n");
987 /* Just prefix with an s, so you get formats like srgba8_unorm */
989 pandecode_log_cont("s");
991 pandecode_log_cont("%s", lower_format
);
996 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
998 /* First, do some validation */
999 bool trivial_swizzle
= pandecode_validate_format_swizzle(
1002 if (trivial_swizzle
)
1005 /* Next, print the swizzle */
1006 pandecode_log_cont(".");
1008 static const char components
[] = "rgba01";
1010 for (unsigned c
= 0; c
< 4; ++c
) {
1011 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
1013 if (chan
>= MALI_CHANNEL_RESERVED_0
) {
1014 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
1017 pandecode_log_cont("%c", components
[chan
]);
1022 pandecode_rt_format(struct mali_rt_format format
)
1024 pandecode_log(".format = {\n");
1027 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
1028 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
1029 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
1030 pandecode_prop("unk4 = 0x%" PRIx32
, format
.unk4
);
1032 pandecode_prop("block = %s", pandecode_block_format(format
.block
));
1034 /* TODO: Map formats so we can check swizzles and print nicely */
1035 pandecode_log("swizzle");
1036 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
1037 pandecode_log_cont(",\n");
1039 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
1040 (format
.nr_channels
+ 1));
1042 pandecode_log(".flags = ");
1043 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
1044 pandecode_log_cont(",\n");
1046 pandecode_prop("msaa = %s", pandecode_msaa_mode(format
.msaa
));
1048 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
1049 * which is a faster hardware-based alternative to the wallpaper method
1050 * to preserve framebuffer contents across frames. In practice, MFBD
1051 * preload is buggy on Midgard, and so this is a chicken bit. If this
1052 * bit isn't set, most likely something broke unrelated to preload */
1054 if (!format
.no_preload
) {
1055 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
1056 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
1060 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
1063 pandecode_log("},\n");
1067 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct mali_framebuffer
*fb
)
1069 pandecode_log("struct mali_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
1072 for (int i
= 0; i
< (fb
->rt_count_1
+ 1); i
++) {
1073 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct mali_render_target
);
1074 struct pandecode_mapped_memory
*mem
=
1075 pandecode_find_mapped_gpu_mem_containing(rt_va
);
1076 const struct mali_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
1078 pandecode_log("{\n");
1081 pandecode_rt_format(rt
->format
);
1083 if (rt
->format
.block
== MALI_BLOCK_AFBC
) {
1084 pandecode_log(".afbc = {\n");
1087 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
1088 pandecode_prop("metadata = %s", a
);
1091 pandecode_prop("stride = %d", rt
->afbc
.stride
);
1093 pandecode_log(".flags = ");
1094 pandecode_log_decoded_flags(afbc_fmt_flag_info
, rt
->afbc
.flags
);
1095 pandecode_log_cont(",\n");
1098 pandecode_log("},\n");
1099 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.flags
) {
1100 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1106 MEMORY_PROP(rt
, framebuffer
);
1107 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
1109 if (rt
->layer_stride
)
1110 pandecode_prop("layer_stride = %d", rt
->layer_stride
);
1112 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
1113 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
1114 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
1115 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
1116 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
1119 if (rt
->zero1
|| rt
->zero2
) {
1120 pandecode_msg("XXX: render target zeros tripped\n");
1121 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
1122 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
1126 pandecode_log("},\n");
1130 pandecode_log("};\n");
1133 static struct pandecode_fbd
1134 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
, bool is_compute
, bool is_bifrost
)
1136 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1137 const struct mali_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
1139 struct pandecode_fbd info
;
1141 if (is_bifrost
&& fb
->msaa
.sample_locations
) {
1142 /* The blob stores all possible sample locations in a single buffer
1143 * allocated on startup, and just switches the pointer when switching
1144 * MSAA state. For now, we just put the data into the cmdstream, but we
1145 * should do something like what the blob does with a real driver.
1147 * There seem to be 32 slots for sample locations, followed by another
1148 * 16. The second 16 is just the center location followed by 15 zeros
1149 * in all the cases I've identified (maybe shader vs. depth/color
1153 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->msaa
.sample_locations
);
1155 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->msaa
.sample_locations
);
1157 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
1160 for (int i
= 0; i
< 32 + 16; i
++) {
1161 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
1165 pandecode_log("};\n");
1168 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1172 pandecode_log(".msaa = {\n");
1175 if (fb
->msaa
.sample_locations
)
1176 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
1178 pandecode_msg("XXX: sample_locations missing\n");
1180 if (fb
->msaa
.zero1
|| fb
->msaa
.zero2
|| fb
->msaa
.zero4
) {
1181 pandecode_msg("XXX: multisampling zero tripped\n");
1182 pandecode_prop("zero1 = %" PRIx64
, fb
->msaa
.zero1
);
1183 pandecode_prop("zero2 = %" PRIx64
, fb
->msaa
.zero2
);
1184 pandecode_prop("zero4 = %" PRIx64
, fb
->msaa
.zero4
);
1188 pandecode_log("},\n");
1190 pandecode_log(".shared_memory = {\n");
1192 pandecode_shared_memory(&fb
->shared_memory
, is_compute
);
1194 pandecode_log("},\n");
1197 info
.width
= fb
->width1
+ 1;
1198 info
.height
= fb
->height1
+ 1;
1199 info
.rt_count
= fb
->rt_count_1
+ 1;
1201 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
1202 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
1203 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
1204 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
1206 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
1207 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
1208 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
1209 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
1211 pandecode_log(".mfbd_flags = ");
1212 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
1213 pandecode_log_cont(",\n");
1215 if (fb
->clear_stencil
)
1216 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
1218 if (fb
->clear_depth
)
1219 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
1223 pandecode_bifrost_tiler_descriptor(fb
);
1225 const struct midgard_tiler_descriptor t
= fb
->tiler
;
1226 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
, true);
1229 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1231 if (fb
->zero3
|| fb
->zero4
) {
1232 pandecode_msg("XXX: framebuffer zeros tripped\n");
1233 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
1234 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
1238 pandecode_log("};\n");
1240 gpu_va
+= sizeof(struct mali_framebuffer
);
1242 info
.has_extra
= (fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
;
1244 if (info
.has_extra
) {
1245 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1246 const struct mali_framebuffer_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
1248 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1251 MEMORY_PROP(fbx
, checksum
);
1253 if (fbx
->checksum_stride
)
1254 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
1256 pandecode_log(".flags_hi = ");
1257 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info
, fbx
->flags_hi
);
1258 pandecode_log_cont(",\n");
1260 pandecode_log(".flags_lo = ");
1261 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info
, fbx
->flags_lo
);
1262 pandecode_log_cont(",\n");
1264 pandecode_prop("zs_block = %s", pandecode_block_format(fbx
->zs_block
));
1265 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx
->zs_samples
+ 1);
1267 if (fbx
->zs_block
== MALI_BLOCK_AFBC
) {
1268 pandecode_log(".ds_afbc = {\n");
1271 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
1272 pandecode_prop("depth_stencil_afbc_stride = %d",
1273 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
1274 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
1276 pandecode_log(".flags = ");
1277 pandecode_log_decoded_flags(afbc_fmt_flag_info
, fbx
->ds_afbc
.flags
);
1278 pandecode_log_cont(",\n");
1280 if (fbx
->ds_afbc
.padding
) {
1281 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1282 pandecode_prop("padding = 0x%" PRIx64
, fbx
->ds_afbc
.padding
);
1286 pandecode_log("},\n");
1288 pandecode_log(".ds_linear = {\n");
1291 if (fbx
->ds_linear
.depth
) {
1292 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
1293 pandecode_prop("depth_stride = %d",
1294 fbx
->ds_linear
.depth_stride
);
1295 pandecode_prop("depth_layer_stride = %d",
1296 fbx
->ds_linear
.depth_layer_stride
);
1297 } else if (fbx
->ds_linear
.depth_stride
|| fbx
->ds_linear
.depth_layer_stride
) {
1298 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx
->ds_linear
.depth_stride
, fbx
->ds_linear
.depth_layer_stride
);
1301 if (fbx
->ds_linear
.stencil
) {
1302 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1303 pandecode_prop("stencil_stride = %d",
1304 fbx
->ds_linear
.stencil_stride
);
1305 pandecode_prop("stencil_layer_stride = %d",
1306 fbx
->ds_linear
.stencil_layer_stride
);
1307 } else if (fbx
->ds_linear
.stencil_stride
|| fbx
->ds_linear
.stencil_layer_stride
) {
1308 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx
->ds_linear
.stencil_stride
, fbx
->ds_linear
.stencil_layer_stride
);
1311 if (fbx
->ds_linear
.depth_stride_zero
||
1312 fbx
->ds_linear
.stencil_stride_zero
) {
1313 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1314 pandecode_prop("depth_stride_zero = 0x%x",
1315 fbx
->ds_linear
.depth_stride_zero
);
1316 pandecode_prop("stencil_stride_zero = 0x%x",
1317 fbx
->ds_linear
.stencil_stride_zero
);
1321 pandecode_log("},\n");
1324 if (fbx
->clear_color_1
| fbx
->clear_color_2
) {
1325 pandecode_prop("clear_color_1 = 0x%" PRIx32
, fbx
->clear_color_1
);
1326 pandecode_prop("clear_color_2 = 0x%" PRIx32
, fbx
->clear_color_2
);
1330 pandecode_msg("XXX: fb_extra zeros tripped\n");
1331 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1335 pandecode_log("};\n");
1337 gpu_va
+= sizeof(struct mali_framebuffer_extra
);
1341 pandecode_render_target(gpu_va
, job_no
, fb
);
1346 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1350 pandecode_padded_vertices(unsigned shift
, unsigned k
)
1352 unsigned odd
= 2*k
+ 1;
1353 unsigned pot
= 1 << shift
;
1354 pandecode_msg("padded_num_vertices = %d\n", odd
* pot
);
1357 /* Given a magic divisor, recover what we were trying to divide by.
1359 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1360 * 0 <= m < 2^N, for N bits in m.
1362 * Let q represent the number we would like to divide by.
1364 * By definition of a magic divisor for N-bit unsigned integers (a number you
1365 * multiply by to magically get division), m is a number such that:
1367 * (m * x) & (2^N - 1) = floor(x/q).
1368 * for all x on Z where 0 <= x < 2^N
1370 * Ignore the case where any of the above values equals zero; it is irrelevant
1371 * for our purposes (instanced arrays).
1373 * Choose x = q. Then:
1375 * (m * x) & (2^N - 1) = floor(x/q).
1376 * (m * q) & (2^N - 1) = floor(q/q).
1378 * floor(q/q) = floor(1) = 1, therefore:
1380 * (m * q) & (2^N - 1) = 1
1382 * Recall the identity that the bitwise AND of one less than a power-of-two
1383 * equals the modulo with that power of two, i.e. for all x:
1385 * x & (2^N - 1) = x % N
1391 * By definition, a modular multiplicative inverse of a number m is the number
1392 * q such that with respect to a modulos M:
1396 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1401 pandecode_magic_divisor(uint32_t magic
, unsigned shift
, unsigned orig_divisor
, unsigned extra
)
1404 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1405 * shift) the most lame way possible... just repeatedly add.
1406 * Asymptoptically slow but nobody cares in practice, unless you have
1407 * massive numbers of vertices or high divisors. */
1409 unsigned inverse
= 0;
1411 /* Magic implicitly has the highest bit set */
1414 /* Depending on rounding direction */
1419 uint32_t product
= magic
* inverse
;
1431 pandecode_msg("dividing by %d (maybe off by two)\n", inverse
);
1433 /* Recall we're supposed to divide by (gl_level_divisor *
1434 * padded_num_vertices) */
1436 unsigned padded_num_vertices
= inverse
/ orig_divisor
;
1438 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices
);
1443 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1444 mali_ptr addr
, int job_no
, char *suffix
,
1445 int count
, bool varying
, enum mali_job_type job_type
)
1447 char *prefix
= varying
? "varying" : "attribute";
1451 pandecode_msg("warn: No %s records\n", prefix
);
1455 union mali_attr
*attr
= pandecode_fetch_gpu_mem(mem
, addr
, sizeof(union mali_attr
) * count
);
1457 for (int i
= 0; i
< count
; ++i
) {
1458 /* First, check for special records */
1459 if (attr
[i
].elements
< MALI_RECORD_SPECIAL
) {
1461 pandecode_msg("XXX: tripped size=%d\n", attr
[i
].size
);
1463 if (attr
[i
].stride
) {
1464 /* gl_InstanceID passes a magic divisor in the
1465 * stride field to divide by the padded vertex
1466 * count. No other records should do so, so
1467 * stride should otherwise be zero. Note that
1468 * stride in the usual attribute sense doesn't
1469 * apply to special records. */
1471 bool has_divisor
= attr
[i
].elements
== MALI_ATTR_INSTANCEID
;
1473 pandecode_log_cont("/* %smagic divisor = %X */ ",
1474 has_divisor
? "" : "XXX: ", attr
[i
].stride
);
1477 if (attr
[i
].shift
|| attr
[i
].extra_flags
) {
1478 /* Attributes use these fields for
1479 * instancing/padding/etc type issues, but
1482 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1483 varying
? "XXX: " : "",
1484 attr
[i
].shift
, attr
[i
].extra_flags
);
1487 /* Print the special record name */
1488 bool attribute
= false;
1489 pandecode_log("%s_%d = %s;\n", prefix
, i
, pandecode_special_record(attr
[i
].elements
, &attribute
));
1492 if (attribute
== varying
)
1493 pandecode_msg("XXX: mismatched special record\n");
1498 enum mali_attr_mode mode
= attr
[i
].elements
& 7;
1500 if (mode
== MALI_ATTR_UNUSED
)
1501 pandecode_msg("XXX: unused attribute record\n");
1503 /* For non-linear records, we need to print the type of record */
1504 if (mode
!= MALI_ATTR_LINEAR
)
1505 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode
));
1507 /* Print the name to link with attr_meta */
1508 pandecode_log_cont("%s_%d", prefix
, i
);
1510 /* Print the stride and size */
1511 pandecode_log_cont("<%u>[%u]", attr
[i
].stride
, attr
[i
].size
);
1513 /* TODO: Sanity check the quotient itself. It must be equal to
1514 * (or be greater than, if the driver added padding) the padded
1517 /* Finally, print the pointer */
1518 mali_ptr raw_elements
= attr
[i
].elements
& ~7;
1519 char *a
= pointer_as_memory_reference(raw_elements
);
1520 pandecode_log_cont(" = (%s);\n", a
);
1523 /* Check the pointer */
1524 pandecode_validate_buffer(raw_elements
, attr
[i
].size
);
1526 /* shift/extra_flags exist only for instanced */
1527 if (attr
[i
].shift
| attr
[i
].extra_flags
) {
1528 /* These are set to random values by the blob for
1529 * varyings, most likely a symptom of uninitialized
1530 * memory where the hardware masked the bug. As such we
1531 * put this at a warning, not an error. */
1533 if (mode
== MALI_ATTR_LINEAR
)
1534 pandecode_msg("warn: instancing fields set for linear\n");
1536 pandecode_prop("shift = %d", attr
[i
].shift
);
1537 pandecode_prop("extra_flags = %d", attr
[i
].extra_flags
);
1540 /* Decode further where possible */
1542 if (mode
== MALI_ATTR_MODULO
) {
1543 pandecode_padded_vertices(
1545 attr
[i
].extra_flags
);
1548 if (mode
== MALI_ATTR_NPOT_DIVIDE
) {
1550 pandecode_log("{\n");
1552 pandecode_prop("unk = 0x%x", attr
[i
].unk
);
1553 pandecode_prop("magic_divisor = 0x%08x", attr
[i
].magic_divisor
);
1554 if (attr
[i
].zero
!= 0)
1555 pandecode_prop("XXX: zero tripped (0x%x)\n", attr
[i
].zero
);
1556 pandecode_prop("divisor = %d", attr
[i
].divisor
);
1557 pandecode_magic_divisor(attr
[i
].magic_divisor
, attr
[i
- 1].shift
, attr
[i
].divisor
, attr
[i
- 1].extra_flags
);
1559 pandecode_log("}, \n");
1564 pandecode_log("\n");
1568 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1570 /* TODO: Decode flags */
1571 mali_ptr shader_ptr
= ptr
& ~15;
1573 char *a
= pointer_as_memory_reference(shader_ptr
);
1574 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1581 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1584 pandecode_msg("XXX: blend zero tripped: %X\n", blend
->zero1
);
1586 pandecode_log(".equation = {\n");
1589 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1590 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1592 pandecode_log(".color_mask = ");
1593 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1594 pandecode_log_cont(",\n");
1597 pandecode_log("},\n");
1600 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1603 decode_bifrost_constant(u16 constant
)
1605 float lo
= (float) (constant
& 0xFF);
1606 float hi
= (float) (constant
>> 8);
1608 return (hi
/ 255.0) + (lo
/ 65535.0);
1612 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1614 struct bifrost_blend_rt
*b
=
1615 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1617 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1620 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1621 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1622 b
->constant
, decode_bifrost_constant(b
->constant
));
1624 /* TODO figure out blend shader enable bit */
1625 pandecode_blend_equation(&b
->equation
);
1627 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1628 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1630 pandecode_log(".format = ");
1631 pandecode_format_short(b
->format
, false);
1632 pandecode_swizzle(b
->swizzle
, b
->format
);
1633 pandecode_log_cont(",\n");
1635 pandecode_prop("swizzle = 0x%" PRIx32
, b
->swizzle
);
1636 pandecode_prop("format = 0x%" PRIx32
, b
->format
);
1639 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1640 pandecode_prop("zero1 = 0x%" PRIx32
, b
->zero1
);
1643 pandecode_log(".shader_type = ");
1644 switch(b
->shader_type
) {
1645 case BIFROST_BLEND_F16
:
1646 pandecode_log_cont("BIFROST_BLEND_F16");
1648 case BIFROST_BLEND_F32
:
1649 pandecode_log_cont("BIFROST_BLEND_F32");
1651 case BIFROST_BLEND_I32
:
1652 pandecode_log_cont("BIFROST_BLEND_I32");
1654 case BIFROST_BLEND_U32
:
1655 pandecode_log_cont("BIFROST_BLEND_U32");
1657 case BIFROST_BLEND_I16
:
1658 pandecode_log_cont("BIFROST_BLEND_I16");
1660 case BIFROST_BLEND_U16
:
1661 pandecode_log_cont("BIFROST_BLEND_U16");
1664 pandecode_log_cont(",\n");
1667 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1668 pandecode_prop("zero2 = 0x%" PRIx32
, b
->zero2
);
1671 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1674 pandecode_log("},\n");
1680 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1682 /* constant/equation is in a union */
1686 pandecode_log(".blend = {\n");
1690 pandecode_shader_address("shader", blend
->shader
);
1692 pandecode_blend_equation(&blend
->equation
);
1693 pandecode_prop("constant = %f", blend
->constant
);
1697 pandecode_log("},\n");
1699 /* Return blend shader to disassemble if present */
1700 return is_shader
? (blend
->shader
& ~0xF) : 0;
1704 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1706 struct midgard_blend_rt
*b
=
1707 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1709 /* Flags determine presence of blend shader */
1710 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1712 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1715 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1717 union midgard_blend blend
= b
->blend
;
1718 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1721 pandecode_log("};\n");
1726 /* Attributes and varyings have descriptor records, which contain information
1727 * about their format and ordering with the attribute/varying buffers. We'll
1728 * want to validate that the combinations specified are self-consistent.
1732 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1735 char *prefix
= varying
? "varying" : "attribute";
1736 unsigned max_index
= 0;
1737 snprintf(base
, sizeof(base
), "%s_meta", prefix
);
1739 struct mali_attr_meta
*attr_meta
;
1740 mali_ptr p
= varying
? v
->varying_meta
: v
->attribute_meta
;
1742 struct pandecode_mapped_memory
*attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
);
1744 for (int i
= 0; i
< count
; ++i
, p
+= sizeof(struct mali_attr_meta
)) {
1745 attr_meta
= pandecode_fetch_gpu_mem(attr_mem
, p
,
1748 /* If the record is discard, it should be zero for everything else */
1750 if (attr_meta
->format
== MALI_VARYING_DISCARD
) {
1753 attr_meta
->unknown1
|
1754 attr_meta
->unknown3
|
1755 attr_meta
->src_offset
;
1758 pandecode_msg("XXX: expected empty record for varying discard\n");
1760 /* We want to look for a literal 0000 swizzle -- this
1761 * is not encoded with all zeroes, however */
1763 enum mali_channel z
= MALI_CHANNEL_ZERO
;
1764 unsigned zero_swizzle
= z
| (z
<< 3) | (z
<< 6) | (z
<< 9);
1765 bool good_swizzle
= attr_meta
->swizzle
== zero_swizzle
;
1768 pandecode_msg("XXX: expected zero swizzle for discard\n");
1771 pandecode_msg("XXX: cannot discard attribute\n");
1773 /* If we're all good, omit the record */
1774 if (!zero
&& varying
&& good_swizzle
) {
1775 pandecode_log("/* discarded varying */\n");
1780 if (attr_meta
->index
> max_index
)
1781 max_index
= attr_meta
->index
;
1783 if (attr_meta
->unknown1
!= 0x2) {
1784 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1785 pandecode_prop("unknown1 = 0x%" PRIx64
, (u64
) attr_meta
->unknown1
);
1788 if (attr_meta
->unknown3
) {
1789 pandecode_msg("XXX: unexpected unknown3 set\n");
1790 pandecode_prop("unknown3 = 0x%" PRIx64
, (u64
) attr_meta
->unknown3
);
1793 pandecode_format_short(attr_meta
->format
, false);
1794 pandecode_log_cont(" %s_%u", prefix
, attr_meta
->index
);
1796 if (attr_meta
->src_offset
)
1797 pandecode_log_cont("[%u]", attr_meta
->src_offset
);
1799 pandecode_swizzle(attr_meta
->swizzle
, attr_meta
->format
);
1801 pandecode_log_cont(";\n");
1804 pandecode_log("\n");
1806 return count
? (max_index
+ 1) : 0;
1809 /* return bits [lo, hi) of word */
1811 bits(u32 word
, u32 lo
, u32 hi
)
1814 return word
; // avoid undefined behavior with the shift
1816 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1820 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool graphics
)
1822 pandecode_log(".prefix = {\n");
1825 /* Decode invocation_count. See the comment before the definition of
1826 * invocation_count for an explanation.
1829 unsigned size_y_shift
= bits(p
->invocation_shifts
, 0, 5);
1830 unsigned size_z_shift
= bits(p
->invocation_shifts
, 5, 10);
1831 unsigned workgroups_x_shift
= bits(p
->invocation_shifts
, 10, 16);
1832 unsigned workgroups_y_shift
= bits(p
->invocation_shifts
, 16, 22);
1833 unsigned workgroups_z_shift
= bits(p
->invocation_shifts
, 22, 28);
1834 unsigned workgroups_x_shift_2
= bits(p
->invocation_shifts
, 28, 32);
1836 unsigned size_x
= bits(p
->invocation_count
, 0, size_y_shift
) + 1;
1837 unsigned size_y
= bits(p
->invocation_count
, size_y_shift
, size_z_shift
) + 1;
1838 unsigned size_z
= bits(p
->invocation_count
, size_z_shift
, workgroups_x_shift
) + 1;
1840 unsigned groups_x
= bits(p
->invocation_count
, workgroups_x_shift
, workgroups_y_shift
) + 1;
1841 unsigned groups_y
= bits(p
->invocation_count
, workgroups_y_shift
, workgroups_z_shift
) + 1;
1842 unsigned groups_z
= bits(p
->invocation_count
, workgroups_z_shift
, 32) + 1;
1844 /* Even though we have this decoded, we want to ensure that the
1845 * representation is "unique" so we don't lose anything by printing only
1846 * the final result. More specifically, we need to check that we were
1847 * passed something in canonical form, since the definition per the
1848 * hardware is inherently not unique. How? Well, take the resulting
1849 * decode and pack it ourselves! If it is bit exact with what we
1850 * decoded, we're good to go. */
1852 struct mali_vertex_tiler_prefix ref
;
1853 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, graphics
);
1856 (p
->invocation_count
== ref
.invocation_count
) &&
1857 (p
->invocation_shifts
== ref
.invocation_shifts
);
1860 pandecode_msg("XXX: non-canonical workgroups packing\n");
1861 pandecode_msg("expected: %X, %X",
1862 ref
.invocation_count
,
1863 ref
.invocation_shifts
);
1865 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1866 pandecode_prop("size_y_shift = %d", size_y_shift
);
1867 pandecode_prop("size_z_shift = %d", size_z_shift
);
1868 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift
);
1869 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift
);
1870 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift
);
1871 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2
);
1874 /* Regardless, print the decode */
1875 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1876 size_x
, size_y
, size_z
,
1877 groups_x
, groups_y
, groups_z
);
1880 if (p
->unknown_draw
)
1881 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1883 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1885 if (p
->draw_mode
!= MALI_DRAW_MODE_NONE
)
1886 pandecode_prop("draw_mode = %s", mali_draw_mode_as_str(p
->draw_mode
));
1888 /* Index count only exists for tiler jobs anyway */
1891 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1894 unsigned index_raw_size
= (p
->unknown_draw
& MALI_DRAW_INDEXED_SIZE
);
1895 index_raw_size
>>= MALI_DRAW_INDEXED_SHIFT
;
1897 /* Validate an index buffer is present if we need one. TODO: verify
1898 * relationship between invocation_count and index_count */
1901 unsigned count
= p
->index_count
;
1904 unsigned size
= (index_raw_size
== 0x3) ? 4 : index_raw_size
;
1906 /* Ensure we got a size, and if so, validate the index buffer
1907 * is large enough to hold a full set of indices of the given
1910 if (!index_raw_size
)
1911 pandecode_msg("XXX: index size missing\n");
1913 pandecode_validate_buffer(p
->indices
, count
* size
);
1914 } else if (index_raw_size
)
1915 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size
);
1917 if (p
->offset_bias_correction
)
1918 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1920 /* TODO: Figure out what this is. It's not zero */
1921 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1924 pandecode_log("},\n");
1928 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1930 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1931 uint64_t *PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1933 for (int i
= 0; i
< ubufs_count
; i
++) {
1934 unsigned size
= (ubufs
[i
] & ((1 << 10) - 1)) * 16;
1935 mali_ptr addr
= (ubufs
[i
] >> 10) << 2;
1937 pandecode_validate_buffer(addr
, size
);
1939 char *ptr
= pointer_as_memory_reference(addr
);
1940 pandecode_log("ubuf_%d[%u] = %s;\n", i
, size
, ptr
);
1944 pandecode_log("\n");
1948 pandecode_uniforms(mali_ptr uniforms
, unsigned uniform_count
)
1950 pandecode_validate_buffer(uniforms
, uniform_count
* 16);
1952 char *ptr
= pointer_as_memory_reference(uniforms
);
1953 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count
, ptr
);
1958 shader_type_for_job(unsigned type
)
1961 case MALI_JOB_TYPE_VERTEX
: return "VERTEX";
1962 case MALI_JOB_TYPE_TILER
: return "FRAGMENT";
1963 case MALI_JOB_TYPE_COMPUTE
: return "COMPUTE";
1969 static unsigned shader_id
= 0;
1971 static struct midgard_disasm_stats
1972 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1973 bool is_bifrost
, unsigned gpu_id
)
1975 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1976 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1978 /* Compute maximum possible size */
1979 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1981 /* Print some boilerplate to clearly denote the assembly (which doesn't
1982 * obey indentation rules), and actually do the disassembly! */
1984 pandecode_log_cont("\n\n");
1986 struct midgard_disasm_stats stats
;
1989 disassemble_bifrost(pandecode_dump_stream
, code
, sz
, true);
1991 /* TODO: Extend stats to Bifrost */
1992 stats
.texture_count
= -128;
1993 stats
.sampler_count
= -128;
1994 stats
.attribute_count
= -128;
1995 stats
.varying_count
= -128;
1996 stats
.uniform_count
= -128;
1997 stats
.uniform_buffer_count
= -128;
1998 stats
.work_count
= -128;
2000 stats
.instruction_count
= 0;
2001 stats
.bundle_count
= 0;
2002 stats
.quadword_count
= 0;
2003 stats
.helper_invocations
= false;
2005 stats
= disassemble_midgard(pandecode_dump_stream
,
2007 type
== MALI_JOB_TYPE_TILER
?
2008 MESA_SHADER_FRAGMENT
: MESA_SHADER_VERTEX
);
2011 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
2012 * driver-internal purposes with the blob and interfere */
2014 bool should_shaderdb
= type
!= MALI_JOB_TYPE_COMPUTE
;
2016 if (should_shaderdb
) {
2017 unsigned nr_threads
=
2018 (stats
.work_count
<= 4) ? 4 :
2019 (stats
.work_count
<= 8) ? 2 :
2022 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
2023 "%u inst, %u bundles, %u quadwords, "
2024 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
2026 shader_type_for_job(type
),
2027 stats
.instruction_count
, stats
.bundle_count
, stats
.quadword_count
,
2028 stats
.work_count
, nr_threads
);
2036 pandecode_texture_payload(mali_ptr payload
,
2037 enum mali_texture_type type
,
2038 enum mali_texture_layout layout
,
2042 uint16_t array_size
,
2043 struct pandecode_mapped_memory
*tmem
)
2045 pandecode_log(".payload = {\n");
2048 /* A bunch of bitmap pointers follow.
2049 * We work out the correct number,
2050 * based on the mipmap/cubemap
2051 * properties, but dump extra
2052 * possibilities to futureproof */
2054 int bitmap_count
= levels
+ 1;
2056 /* Miptree for each face */
2057 if (type
== MALI_TEX_CUBE
)
2060 /* Array of layers */
2061 bitmap_count
*= (depth
+ 1);
2063 /* Array of textures */
2064 bitmap_count
*= (array_size
+ 1);
2066 /* Stride for each element */
2070 mali_ptr
*pointers_and_strides
= pandecode_fetch_gpu_mem(tmem
,
2071 payload
, sizeof(mali_ptr
) * bitmap_count
);
2072 for (int i
= 0; i
< bitmap_count
; ++i
) {
2073 /* How we dump depends if this is a stride or a pointer */
2075 if (manual_stride
&& (i
& 1)) {
2076 /* signed 32-bit snuck in as a 64-bit pointer */
2077 uint64_t stride_set
= pointers_and_strides
[i
];
2078 uint32_t clamped_stride
= stride_set
;
2079 int32_t stride
= clamped_stride
;
2080 assert(stride_set
== clamped_stride
);
2081 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
2083 char *a
= pointer_as_memory_reference(pointers_and_strides
[i
]);
2084 pandecode_log("%s, \n", a
);
2090 pandecode_log("},\n");
2094 pandecode_texture(mali_ptr u
,
2095 struct pandecode_mapped_memory
*tmem
,
2096 unsigned job_no
, unsigned tex
)
2098 struct mali_texture_descriptor
*PANDECODE_PTR_VAR(t
, tmem
, u
);
2100 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64
"_%d_%d = {\n", u
, job_no
, tex
);
2103 pandecode_prop("width = %" PRId32
, t
->width
);
2104 pandecode_prop("height = %" PRId32
, t
->height
);
2105 pandecode_prop("depth = %" PRId32
, t
->depth
);
2106 pandecode_prop("array_size = %" PRId32
, t
->array_size
);
2108 pandecode_log("\n");
2109 pandecode_prop("f.swizzle = 0x%" PRIx32
, t
->format
.swizzle
);
2110 pandecode_prop("f.format = 0x%" PRIx32
, t
->format
.format
);
2111 pandecode_prop("f.srgb = 0x%" PRIx32
, t
->format
.srgb
);
2112 pandecode_prop("f.unknown1 = 0x%" PRIx32
, t
->format
.unknown1
);
2113 pandecode_prop("f.type = %" PRId32
, t
->format
.type
);
2114 pandecode_prop("f.layout = %" PRId32
, t
->format
.layout
);
2115 pandecode_prop("f.unknown2 = 0x%" PRIx32
, t
->format
.unknown2
);
2116 pandecode_prop("f.manual_stride = %" PRId32
, t
->format
.manual_stride
);
2117 pandecode_prop("f.zero = 0x%" PRIx32
, t
->format
.zero
);
2118 pandecode_log("\n");
2120 pandecode_prop("unknown3 = 0x%" PRIx32
, t
->unknown3
);
2121 pandecode_prop("unknown3A = 0x%" PRIx32
, t
->unknown3A
);
2122 pandecode_prop("levels = %" PRId32
, t
->levels
);
2123 pandecode_prop("swizzle = 0x%" PRIx32
, t
->swizzle
);
2124 pandecode_prop("swizzle_zero = 0x%" PRIx32
, t
->swizzle_zero
);
2126 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
2127 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
2128 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
2129 pandecode_log("\n");
2131 struct mali_texture_format f
= t
->format
;
2133 /* See the definiton of enum mali_texture_type */
2135 bool is_cube
= f
.type
== MALI_TEX_CUBE
;
2136 unsigned dimension
= is_cube
? 2 : f
.type
;
2138 pandecode_make_indent();
2140 /* Print the layout. Default is linear; a modifier can denote AFBC or
2141 * u-interleaved/tiled modes */
2143 if (f
.layout
== MALI_TEXTURE_AFBC
)
2144 pandecode_log_cont("afbc");
2145 else if (f
.layout
== MALI_TEXTURE_TILED
)
2146 pandecode_log_cont("tiled");
2147 else if (f
.layout
== MALI_TEXTURE_LINEAR
)
2148 pandecode_log_cont("linear");
2150 pandecode_msg("XXX: invalid texture layout 0x%X\n", f
.layout
);
2152 pandecode_swizzle(t
->swizzle
, f
.format
);
2153 pandecode_log_cont(" ");
2155 /* Distinguish cube/2D with modifier */
2158 pandecode_log_cont("cube ");
2160 pandecode_format_short(f
.format
, f
.srgb
);
2161 pandecode_swizzle(f
.swizzle
, f
.format
);
2163 /* All four width/height/depth/array_size dimensions are present
2164 * regardless of the type of texture, but it is an error to have
2165 * non-zero dimensions for unused dimensions. Verify this. array_size
2166 * can always be set, as can width. Depth used for MSAA. */
2168 if (t
->height
&& dimension
< 2)
2169 pandecode_msg("XXX: nonzero height for <2D texture\n");
2171 /* Print only the dimensions that are actually there */
2173 pandecode_log_cont(": %d", t
->width
+ 1);
2175 if (t
->height
|| t
->depth
)
2176 pandecode_log_cont("x%u", t
->height
+ 1);
2179 pandecode_log_cont("x%u", t
->depth
+ 1);
2182 pandecode_log_cont("[%u]", t
->array_size
+ 1);
2185 pandecode_log_cont(" mip %u", t
->levels
);
2187 pandecode_log_cont("\n");
2189 if (f
.unknown1
| f
.zero
) {
2190 pandecode_msg("XXX: texture format zero tripped\n");
2191 pandecode_prop("unknown1 = %" PRId32
, f
.unknown1
);
2192 pandecode_prop("zero = %" PRId32
, f
.zero
);
2196 pandecode_msg("XXX: expected unknown texture bit set\n");
2197 pandecode_prop("unknown2 = %" PRId32
, f
.unknown2
);
2200 if (t
->swizzle_zero
) {
2201 pandecode_msg("XXX: swizzle zero tripped\n");
2202 pandecode_prop("swizzle_zero = %d", t
->swizzle_zero
);
2205 if (t
->unknown3
| t
->unknown3A
| t
->unknown5
| t
->unknown6
| t
->unknown7
) {
2206 pandecode_msg("XXX: texture zero tripped\n");
2207 pandecode_prop("unknown3 = %" PRId16
, t
->unknown3
);
2208 pandecode_prop("unknown3A = %" PRId8
, t
->unknown3A
);
2209 pandecode_prop("unknown5 = 0x%" PRIx32
, t
->unknown5
);
2210 pandecode_prop("unknown6 = 0x%" PRIx32
, t
->unknown6
);
2211 pandecode_prop("unknown7 = 0x%" PRIx32
, t
->unknown7
);
2214 pandecode_texture_payload(u
+ sizeof(*t
), f
.type
, f
.layout
, f
.manual_stride
, t
->levels
, t
->depth
, t
->array_size
, tmem
);
2217 pandecode_log("};\n");
2221 pandecode_bifrost_texture(
2222 const struct bifrost_texture_descriptor
*t
,
2226 pandecode_log("struct bifrost_texture_descriptor texture_descriptor_%d_%d = {\n", job_no
, tex
);
2229 pandecode_prop("format_unk = 0x%" PRIx32
, t
->format_unk
);
2230 pandecode_prop("type = %" PRId32
, t
->type
);
2233 pandecode_msg("XXX: zero tripped\n");
2234 pandecode_prop("zero = 0x%" PRIx32
, t
->zero
);
2237 pandecode_prop("format_swizzle = 0x%" PRIx32
, t
->format_swizzle
);
2238 pandecode_prop("format = 0x%" PRIx32
, t
->format
);
2239 pandecode_prop("srgb = 0x%" PRIx32
, t
->srgb
);
2240 pandecode_prop("format_unk3 = 0x%" PRIx32
, t
->format_unk3
);
2241 pandecode_prop("width = %" PRId32
, t
->width
);
2242 pandecode_prop("height = %" PRId32
, t
->height
);
2243 pandecode_prop("swizzle = 0x%" PRIx32
, t
->swizzle
);
2244 pandecode_prop("levels = %" PRId32
, t
->levels
);
2245 pandecode_prop("unk1 = 0x%" PRIx32
, t
->unk1
);
2246 pandecode_prop("levels_unk = %" PRId32
, t
->levels_unk
);
2247 pandecode_prop("level_2 = %" PRId32
, t
->level_2
);
2248 pandecode_prop("payload = 0x%" PRIx64
, t
->payload
);
2249 pandecode_prop("array_size = %" PRId32
, t
->array_size
);
2250 pandecode_prop("unk4 = 0x%" PRIx32
, t
->unk4
);
2251 pandecode_prop("depth = %" PRId32
, t
->depth
);
2252 pandecode_prop("unk5 = 0x%" PRIx32
, t
->unk5
);
2253 pandecode_log("\n");
2255 /* See the definiton of enum mali_texture_type */
2257 bool is_cube
= t
->type
== MALI_TEX_CUBE
;
2258 unsigned dimension
= is_cube
? 2 : t
->type
;
2260 /* Print the layout. Default is linear; a modifier can denote AFBC or
2261 * u-interleaved/tiled modes */
2263 if (t
->layout
== MALI_TEXTURE_AFBC
)
2264 pandecode_log_cont("afbc");
2265 else if (t
->layout
== MALI_TEXTURE_TILED
)
2266 pandecode_log_cont("tiled");
2267 else if (t
->layout
== MALI_TEXTURE_LINEAR
)
2268 pandecode_log_cont("linear");
2270 pandecode_msg("XXX: invalid texture layout 0x%X\n", t
->layout
);
2272 pandecode_swizzle(t
->swizzle
, t
->format
);
2273 pandecode_log_cont(" ");
2275 /* Distinguish cube/2D with modifier */
2278 pandecode_log_cont("cube ");
2280 pandecode_format_short(t
->format
, t
->srgb
);
2282 /* All four width/height/depth/array_size dimensions are present
2283 * regardless of the type of texture, but it is an error to have
2284 * non-zero dimensions for unused dimensions. Verify this. array_size
2285 * can always be set, as can width. */
2287 if (t
->height
&& dimension
< 2)
2288 pandecode_msg("XXX: nonzero height for <2D texture\n");
2290 if (t
->depth
&& dimension
< 3)
2291 pandecode_msg("XXX: nonzero depth for <2D texture\n");
2293 /* Print only the dimensions that are actually there */
2295 pandecode_log_cont(": %d", t
->width
+ 1);
2298 pandecode_log_cont("x%u", t
->height
+ 1);
2301 pandecode_log_cont("x%u", t
->depth
+ 1);
2304 pandecode_log_cont("[%u]", t
->array_size
+ 1);
2307 pandecode_log_cont(" mip %u", t
->levels
);
2309 pandecode_log_cont("\n");
2311 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(t
->payload
);
2313 pandecode_texture_payload(t
->payload
, t
->type
, t
->layout
,
2314 true, t
->levels
, t
->depth
,
2315 t
->array_size
, tmem
);
2319 pandecode_log("};\n");
2322 /* 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 */
2325 pandecode_shader_prop(const char *name
, unsigned claim
, signed truth
, bool fuzzy
)
2331 if (fuzzy
&& (truth
< 0))
2332 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name
, claim
, truth
);
2334 if ((truth
>= 0) && !fuzzy
) {
2335 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2336 (truth
< claim
) ? "warn" : "XXX",
2337 name
, truth
, claim
);
2338 } else if ((claim
> -truth
) && !fuzzy
) {
2339 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2340 name
, -truth
, claim
);
2341 } else if (fuzzy
&& (claim
< truth
))
2342 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2343 name
, truth
, claim
);
2345 pandecode_log(".%s = %" PRId16
, name
, claim
);
2348 pandecode_log_cont(" /* %u used */", truth
);
2350 pandecode_log_cont(",\n");
2354 pandecode_blend_shader_disassemble(mali_ptr shader
, int job_no
, int job_type
,
2355 bool is_bifrost
, unsigned gpu_id
)
2357 struct midgard_disasm_stats stats
=
2358 pandecode_shader_disassemble(shader
, job_no
, job_type
, is_bifrost
, gpu_id
);
2360 bool has_texture
= (stats
.texture_count
> 0);
2361 bool has_sampler
= (stats
.sampler_count
> 0);
2362 bool has_attribute
= (stats
.attribute_count
> 0);
2363 bool has_varying
= (stats
.varying_count
> 0);
2364 bool has_uniform
= (stats
.uniform_count
> 0);
2365 bool has_ubo
= (stats
.uniform_buffer_count
> 0);
2367 if (has_texture
|| has_sampler
)
2368 pandecode_msg("XXX: blend shader accessing textures\n");
2370 if (has_attribute
|| has_varying
)
2371 pandecode_msg("XXX: blend shader accessing interstage\n");
2373 if (has_uniform
|| has_ubo
)
2374 pandecode_msg("XXX: blend shader accessing uniforms\n");
2378 pandecode_textures(mali_ptr textures
, unsigned texture_count
, int job_no
, bool is_bifrost
)
2380 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(textures
);
2386 const struct bifrost_texture_descriptor
*PANDECODE_PTR_VAR(t
, mmem
, textures
);
2388 pandecode_log("uint64_t textures_%"PRIx64
"_%d[] = {\n", textures
, job_no
);
2391 for (unsigned tex
= 0; tex
< texture_count
; ++tex
)
2392 pandecode_bifrost_texture(&t
[tex
], job_no
, tex
);
2395 pandecode_log("};\n");
2397 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
);
2399 pandecode_log("uint64_t textures_%"PRIx64
"_%d[] = {\n", textures
, job_no
);
2402 for (int tex
= 0; tex
< texture_count
; ++tex
) {
2403 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
2404 char *a
= pointer_as_memory_reference(*u
);
2405 pandecode_log("%s,\n", a
);
2410 pandecode_log("};\n");
2412 /* Now, finally, descend down into the texture descriptor */
2413 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
2414 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
2415 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
2417 pandecode_texture(*u
, tmem
, job_no
, tex
);
2423 pandecode_samplers(mali_ptr samplers
, unsigned sampler_count
, int job_no
, bool is_bifrost
)
2425 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(samplers
);
2431 struct bifrost_sampler_descriptor
*s
;
2433 for (int i
= 0; i
< sampler_count
; ++i
) {
2434 s
= pandecode_fetch_gpu_mem(smem
, samplers
+ sizeof(*s
) * i
, sizeof(*s
));
2436 pandecode_log("struct bifrost_sampler_descriptor sampler_descriptor_%"PRIx64
"_%d_%d = {\n", samplers
+ sizeof(*s
) * i
, job_no
, i
);
2440 pandecode_msg("XXX: unk1 tripped\n");
2441 pandecode_prop("unk1 = 0x%x", s
->unk1
);
2444 pandecode_prop("wrap_s = %s", mali_wrap_mode_as_str(s
->wrap_s
));
2445 pandecode_prop("wrap_t = %s", mali_wrap_mode_as_str(s
->wrap_t
));
2446 pandecode_prop("wrap_r = %s", mali_wrap_mode_as_str(s
->wrap_r
));
2448 if (s
->unk8
!= 0x8) {
2449 pandecode_msg("XXX: unk8 tripped\n");
2450 pandecode_prop("unk8 = 0x%x", s
->unk8
);
2453 pandecode_prop("unk2 = 0x%x", s
->unk2
);
2454 pandecode_prop("unk3 = 0x%x", s
->unk3
);
2455 pandecode_prop("min_filter = %s", s
->min_filter
? "nearest" : "linear");
2456 pandecode_prop("norm_coords = 0x%x", s
->norm_coords
& 0x1);
2457 pandecode_prop("zero1 = 0x%x", s
->zero1
& 0x1);
2458 pandecode_prop("mip_filter = %s", s
->mip_filter
? "linear" : "nearest");
2459 pandecode_prop("mag_filter = %s", s
->mag_filter
? "linear" : "nearest");
2461 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->min_lod
));
2462 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s
->max_lod
));
2464 if (s
->zero1
|| s
->zero2
|| s
->zero3
|| s
->zero4
) {
2465 pandecode_msg("XXX: sampler zero tripped\n");
2466 pandecode_prop("zero = 0x%" PRIx8
", 0x%" PRIx64
", 0x%" PRIx64
", 0x%" PRIx64
"\n", s
->zero1
, s
->zero2
, s
->zero3
, s
->zero4
);
2470 pandecode_log("};\n");
2473 for (int i
= 0; i
< sampler_count
; ++i
)
2474 DUMP_ADDR("Sampler", MIDGARD_SAMPLER
, samplers
+ (MALI_MIDGARD_SAMPLER_LENGTH
* i
), 1);
2479 pandecode_vertex_tiler_postfix_pre(
2480 const struct mali_vertex_tiler_postfix
*p
,
2481 int job_no
, enum mali_job_type job_type
,
2482 char *suffix
, bool is_bifrost
, unsigned gpu_id
)
2484 struct pandecode_mapped_memory
*attr_mem
;
2486 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2487 * are the only things actually needed from the FBD, vertex/tiler jobs
2488 * no longer reference the FBD -- instead, this field points to some
2489 * info about the scratchpad.
2492 struct pandecode_fbd fbd_info
= {
2493 /* Default for Bifrost */
2498 pandecode_log_cont("\t/* %X %/\n", p
->shared_memory
& 1);
2499 pandecode_compute_fbd(p
->shared_memory
& ~1, job_no
);
2500 } else if (p
->shared_memory
& MALI_MFBD
)
2501 fbd_info
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->shared_memory
) & FBD_MASK
, job_no
, false, job_type
== MALI_JOB_TYPE_COMPUTE
, false);
2502 else if (job_type
== MALI_JOB_TYPE_COMPUTE
)
2503 pandecode_compute_fbd((u64
) (uintptr_t) p
->shared_memory
, job_no
);
2505 fbd_info
= pandecode_sfbd((u64
) (uintptr_t) p
->shared_memory
, job_no
, false, gpu_id
);
2507 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
2508 int texture_count
= 0, sampler_count
= 0;
2511 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->shader
);
2512 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, p
->shader
);
2514 /* Disassemble ahead-of-time to get stats. Initialize with
2515 * stats for the missing-shader case so we get validation
2518 struct midgard_disasm_stats info
= {
2521 .attribute_count
= 0,
2525 .uniform_count
= -128,
2526 .uniform_buffer_count
= 0
2529 if (s
->shader
& ~0xF)
2530 info
= pandecode_shader_disassemble(s
->shader
& ~0xF, job_no
, job_type
, is_bifrost
, gpu_id
);
2532 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", p
->shader
, job_no
, suffix
);
2535 /* Save for dumps */
2536 attribute_count
= s
->attribute_count
;
2537 varying_count
= s
->varying_count
;
2538 texture_count
= s
->texture_count
;
2539 sampler_count
= s
->sampler_count
;
2542 uniform_count
= s
->bifrost2
.uniform_count
;
2543 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
2545 uniform_count
= s
->midgard1
.uniform_count
;
2546 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
2549 pandecode_shader_address("shader", s
->shader
);
2551 pandecode_shader_prop("texture_count", s
->texture_count
, info
.texture_count
, false);
2552 pandecode_shader_prop("sampler_count", s
->sampler_count
, info
.sampler_count
, false);
2553 pandecode_shader_prop("attribute_count", s
->attribute_count
, info
.attribute_count
, false);
2554 pandecode_shader_prop("varying_count", s
->varying_count
, info
.varying_count
, false);
2555 pandecode_shader_prop("uniform_buffer_count",
2556 uniform_buffer_count
,
2557 info
.uniform_buffer_count
, true);
2560 pandecode_shader_prop("uniform_count",
2562 info
.uniform_count
, false);
2564 pandecode_shader_prop("work_count",
2565 s
->midgard1
.work_count
, info
.work_count
, false);
2569 pandecode_log("bifrost1.unk1 = ");
2570 pandecode_log_decoded_flags(shader_bifrost_info
, s
->bifrost1
.unk1
);
2571 pandecode_log_cont(",\n");
2573 bool helpers
= s
->midgard1
.flags_lo
& MALI_HELPER_INVOCATIONS
;
2575 if (helpers
!= info
.helper_invocations
) {
2576 pandecode_msg("XXX: expected helpers %u but got %u\n",
2577 info
.helper_invocations
, helpers
);
2580 pandecode_log(".midgard1.flags_lo = ");
2581 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info
,
2582 s
->midgard1
.flags_lo
& ~MALI_HELPER_INVOCATIONS
);
2583 pandecode_log_cont(",\n");
2585 pandecode_log(".midgard1.flags_hi = ");
2586 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info
, s
->midgard1
.flags_hi
);
2587 pandecode_log_cont(",\n");
2590 if (s
->depth_units
|| s
->depth_factor
) {
2591 pandecode_prop("depth_factor = %f", s
->depth_factor
);
2592 pandecode_prop("depth_units = %f", s
->depth_units
);
2595 if (s
->coverage_mask
)
2596 pandecode_prop("coverage_mask = 0x%X", s
->coverage_mask
);
2599 pandecode_prop(".unknown2_2 = %X", s
->unknown2_2
);
2601 if (s
->unknown2_3
|| s
->unknown2_4
) {
2602 pandecode_log(".unknown2_3 = ");
2604 int unknown2_3
= s
->unknown2_3
;
2605 int unknown2_4
= s
->unknown2_4
;
2607 /* We're not quite sure what these flags mean without the depth test, if anything */
2609 if (unknown2_3
& (MALI_DEPTH_WRITEMASK
| MALI_DEPTH_FUNC_MASK
)) {
2610 const char *func
= mali_func_as_str(MALI_GET_DEPTH_FUNC(unknown2_3
));
2611 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
2613 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
2616 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
2617 pandecode_log_cont(",\n");
2619 pandecode_log(".unknown2_4 = ");
2620 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
2621 pandecode_log_cont(",\n");
2624 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
2625 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
2626 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
2629 DUMP_CL("Stencil front", STENCIL
, &s
->stencil_front
, 1);
2630 DUMP_CL("Stencil back", STENCIL
, &s
->stencil_back
, 1);
2633 pandecode_log(".bifrost2 = {\n");
2636 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
2637 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
2638 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
2639 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
2642 pandecode_log("},\n");
2643 } else if (s
->midgard2
.unknown2_7
) {
2644 pandecode_log(".midgard2 = {\n");
2647 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
2649 pandecode_log("},\n");
2653 pandecode_msg("XXX: shader padding tripped\n");
2654 pandecode_prop("padding = 0x%" PRIx32
, s
->padding
);
2658 /* TODO: Blend shaders routing/disasm */
2659 union midgard_blend blend
= s
->blend
;
2660 mali_ptr shader
= pandecode_midgard_blend(&blend
, s
->unknown2_3
& MALI_HAS_BLEND_SHADER
);
2662 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
2664 pandecode_msg("mdg_blend = %" PRIx64
"\n", s
->blend
.shader
);
2668 pandecode_log("};\n");
2670 /* MRT blend fields are used whenever MFBD is used, with
2671 * per-RT descriptors */
2673 if (job_type
== MALI_JOB_TYPE_TILER
&& (is_bifrost
|| p
->shared_memory
& MALI_MFBD
)) {
2674 void* blend_base
= (void *) (s
+ 1);
2676 for (unsigned i
= 0; i
< fbd_info
.rt_count
; i
++) {
2677 mali_ptr shader
= 0;
2680 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
2682 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
2685 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
2690 pandecode_msg("XXX: missing shader descriptor\n");
2693 DUMP_ADDR("Viewport", VIEWPORT
, p
->viewport
, 1);
2695 unsigned max_attr_index
= 0;
2697 if (p
->attribute_meta
)
2698 max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
2700 if (p
->attributes
) {
2701 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
2702 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false, job_type
);
2705 /* Varyings are encoded like attributes but not actually sent; we just
2706 * pass a zero buffer with the right stride/size set, (or whatever)
2707 * since the GPU will write to it itself */
2709 if (p
->varying_meta
) {
2710 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
2714 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
2716 /* Number of descriptors depends on whether there are
2717 * non-internal varyings */
2719 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true, job_type
);
2722 if (p
->uniform_buffers
) {
2723 if (uniform_buffer_count
)
2724 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
2726 pandecode_msg("warn: UBOs specified but not referenced\n");
2727 } else if (uniform_buffer_count
)
2728 pandecode_msg("XXX: UBOs referenced but not specified\n");
2730 /* We don't want to actually dump uniforms, but we do need to validate
2731 * that the counts we were given are sane */
2735 pandecode_uniforms(p
->uniforms
, uniform_count
);
2737 pandecode_msg("warn: Uniforms specified but not referenced\n");
2738 } else if (uniform_count
)
2739 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2742 pandecode_textures(p
->textures
, texture_count
, job_no
, is_bifrost
);
2744 if (p
->sampler_descriptor
)
2745 pandecode_samplers(p
->sampler_descriptor
, sampler_count
, job_no
, is_bifrost
);
2749 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2751 pandecode_log(".gl_enables = ");
2753 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2755 pandecode_log_cont(",\n");
2759 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2761 if (p
->shader
& 0xF)
2762 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p
->shader
& 0xF));
2764 pandecode_log(".postfix = {\n");
2767 pandecode_gl_enables(p
->gl_enables
, MALI_JOB_TYPE_TILER
);
2768 pandecode_prop("instance_shift = 0x%x", p
->instance_shift
);
2769 pandecode_prop("instance_odd = 0x%x", p
->instance_odd
);
2772 pandecode_msg("XXX: vertex only zero tripped");
2773 pandecode_prop("zero4 = 0x%" PRIx32
, p
->zero4
);
2776 pandecode_prop("offset_start = 0x%x", p
->offset_start
);
2779 pandecode_msg("XXX: vertex only zero tripped");
2780 pandecode_prop("zero5 = 0x%" PRIx64
, p
->zero5
);
2783 MEMORY_PROP(p
, position_varying
);
2784 MEMORY_PROP(p
, occlusion_counter
);
2787 pandecode_log("},\n");
2791 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2793 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2794 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2796 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2800 pandecode_msg("XXX: tiler heap zero tripped\n");
2801 pandecode_prop("zero = 0x%x", h
->zero
);
2804 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2805 MEMORY_PROP(h
, tiler_heap_start
);
2806 MEMORY_PROP(h
, tiler_heap_free
);
2808 /* this might point to the beginning of another buffer, when it's
2809 * really the end of the tiler heap buffer, so we have to be careful
2810 * here. but for zero length, we need the same pointer.
2813 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2814 MEMORY_PROP(h
, tiler_heap_start
);
2816 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2817 pandecode_prop("tiler_heap_end = %s + 1", a
);
2821 for (int i
= 0; i
< 10; i
++) {
2822 if (h
->zeros
[i
] != 0) {
2823 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2828 if (h
->unk1
!= 0x1) {
2829 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2830 pandecode_prop("unk1 = 0x%x", h
->unk1
);
2833 if (h
->unk7e007e
!= 0x7e007e) {
2834 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
2835 pandecode_prop("unk7e007e = 0x%x", h
->unk7e007e
);
2839 pandecode_log("};\n");
2843 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2845 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2846 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2848 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2850 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2853 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32
, t
->tiler_heap_next_start
);
2854 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32
, t
->used_hierarchy_mask
);
2856 if (t
->hierarchy_mask
!= 0xa &&
2857 t
->hierarchy_mask
!= 0x14 &&
2858 t
->hierarchy_mask
!= 0x28 &&
2859 t
->hierarchy_mask
!= 0x50 &&
2860 t
->hierarchy_mask
!= 0xa0)
2861 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
2863 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2865 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2867 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2868 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2871 pandecode_msg("XXX: tiler meta zero tripped\n");
2872 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2875 for (int i
= 0; i
< 12; i
++) {
2876 if (t
->zeros
[i
] != 0) {
2877 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2883 pandecode_log("};\n");
2887 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2889 if (u
.pointer
== 0x0)
2892 pandecode_log(".primitive_size = {\n");
2896 pandecode_prop("constant = %f", u
.constant
);
2898 MEMORY_PROP((&u
), pointer
);
2902 pandecode_log("},\n");
2906 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2908 pandecode_log_cont("{\n");
2911 /* TODO: gl_PointSize on Bifrost */
2912 pandecode_primitive_size(t
->primitive_size
, true);
2914 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2916 pandecode_msg("XXX: tiler only zero tripped\n");
2917 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2918 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2919 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2920 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2921 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2922 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2926 pandecode_log("},\n");
2930 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2931 const struct pandecode_mapped_memory
*mem
,
2932 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2934 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2936 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2938 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2941 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2942 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2945 pandecode_log("};\n");
2951 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2952 const struct pandecode_mapped_memory
*mem
,
2953 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2955 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2957 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2958 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2960 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2963 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2965 pandecode_log(".tiler = ");
2966 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2968 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2971 pandecode_log("};\n");
2977 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2978 const struct pandecode_mapped_memory
*mem
,
2979 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2981 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2982 bool is_graphics
= (h
->job_type
== MALI_JOB_TYPE_VERTEX
) || (h
->job_type
== MALI_JOB_TYPE_TILER
);
2984 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false, gpu_id
);
2986 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2989 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, is_graphics
);
2990 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2992 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2993 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2996 pandecode_log("};\n");
3002 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
3003 mali_ptr payload
, int job_no
,
3004 bool is_bifrost
, unsigned gpu_id
)
3006 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
3008 bool is_mfbd
= s
->framebuffer
& MALI_MFBD
;
3010 if (!is_mfbd
&& is_bifrost
)
3011 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
3013 struct pandecode_fbd info
;
3016 info
= pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true, false, is_bifrost
);
3018 info
= pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true, gpu_id
);
3020 /* Compute the tag for the tagged pointer. This contains the type of
3021 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
3022 * additional structures follow the MFBD header (an extra payload or
3023 * not, as well as a count of render targets) */
3025 unsigned expected_tag
= is_mfbd
? MALI_MFBD
: 0;
3029 expected_tag
|= MALI_MFBD_TAG_EXTRA
;
3031 expected_tag
|= (MALI_POSITIVE(info
.rt_count
) << 2);
3034 if ((s
->min_tile_coord
| s
->max_tile_coord
) & ~(MALI_X_COORD_MASK
| MALI_Y_COORD_MASK
)) {
3035 pandecode_msg("XXX: unexpected tile coordinate bits\n");
3036 pandecode_prop("min_tile_coord = 0x%X\n", s
->min_tile_coord
);
3037 pandecode_prop("max_tile_coord = 0x%X\n", s
->max_tile_coord
);
3040 /* Extract tile coordinates */
3042 unsigned min_x
= MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
3043 unsigned min_y
= MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
3045 unsigned max_x
= (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
3046 unsigned max_y
= (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
3048 /* For the max, we also want the floored (rather than ceiled) version for checking */
3050 unsigned max_x_f
= (MALI_TILE_COORD_X(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
3051 unsigned max_y_f
= (MALI_TILE_COORD_Y(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
3053 /* Validate the coordinates are well-ordered */
3056 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
3057 else if (min_x
> max_x
)
3058 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
3061 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
3062 else if (min_y
> max_y
)
3063 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
3065 /* Validate the coordinates fit inside the framebuffer. We use floor,
3066 * rather than ceil, for the max coordinates, since the tile
3067 * coordinates for something like an 800x600 framebuffer will actually
3068 * resolve to 800x608, which would otherwise trigger a Y-overflow */
3070 if ((min_x
> info
.width
) || (max_x_f
> info
.width
))
3071 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
3073 if ((min_y
> info
.height
) || (max_y_f
> info
.height
))
3074 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
3076 /* After validation, we print */
3078 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x
, min_y
, max_x
, max_y
);
3080 /* The FBD is a tagged pointer */
3082 unsigned tag
= (s
->framebuffer
& ~FBD_MASK
);
3084 if (tag
!= expected_tag
)
3085 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag
, tag
);
3090 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
3091 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
3092 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
3093 * might wish to combine this with the bifrost parameter) because some details
3094 * are model-specific even within a particular architecture. Minimal traces
3095 * *only* examine the job descriptors, skipping printing entirely if there is
3096 * no faults, and only descends into the payload if there are faults. This is
3097 * useful for looking for faults without the overhead of invasive traces. */
3100 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
, unsigned gpu_id
, bool minimal
)
3102 pandecode_dump_file_open();
3104 struct mali_job_descriptor_header
*h
;
3105 unsigned job_descriptor_number
= 0;
3108 struct pandecode_mapped_memory
*mem
=
3109 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
3113 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
3115 /* On Midgard, for 32-bit jobs except for fragment jobs, the
3116 * high 32-bits of the 64-bit pointer are reused to store
3119 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
3120 h
->job_type
!= MALI_JOB_TYPE_FRAGMENT
? 4 : 0;
3121 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
3123 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
, 256);
3125 int job_no
= job_descriptor_number
++;
3127 /* If the job is good to go, skip it in minimal mode */
3128 if (minimal
&& (h
->exception_status
== 0x0 || h
->exception_status
== 0x1))
3131 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
3134 pandecode_prop("job_type = %s", mali_job_type_as_str(h
->job_type
));
3136 if (h
->job_descriptor_size
)
3137 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
3139 if (h
->exception_status
&& h
->exception_status
!= 0x1)
3140 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
3141 h
->exception_status
,
3142 (h
->exception_status
>> 16) & 0xFFFF,
3143 pandecode_exception_access((h
->exception_status
>> 8) & 0x3),
3144 h
->exception_status
& 0xFF);
3146 if (h
->first_incomplete_task
)
3147 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
3149 if (h
->fault_pointer
)
3150 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
3153 pandecode_prop("job_barrier = %d", h
->job_barrier
);
3155 pandecode_prop("job_index = %d", h
->job_index
);
3157 if (h
->unknown_flags
)
3158 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
3160 if (h
->job_dependency_index_1
)
3161 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
3163 if (h
->job_dependency_index_2
)
3164 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
3167 pandecode_log("};\n");
3169 switch (h
->job_type
) {
3170 case MALI_JOB_TYPE_WRITE_VALUE
: {
3171 struct mali_payload_write_value
*s
= payload
;
3172 pandecode_log("struct mali_payload_write_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
3174 MEMORY_PROP(s
, address
);
3176 if (s
->value_descriptor
!= MALI_WRITE_VALUE_ZERO
) {
3177 pandecode_msg("XXX: unknown value descriptor\n");
3178 pandecode_prop("value_descriptor = 0x%" PRIX32
, s
->value_descriptor
);
3182 pandecode_msg("XXX: set value tripped\n");
3183 pandecode_prop("reserved = 0x%" PRIX32
, s
->reserved
);
3186 pandecode_prop("immediate = 0x%" PRIX64
, s
->immediate
);
3188 pandecode_log("};\n");
3193 case MALI_JOB_TYPE_TILER
:
3194 case MALI_JOB_TYPE_VERTEX
:
3195 case MALI_JOB_TYPE_COMPUTE
:
3197 if (h
->job_type
== MALI_JOB_TYPE_TILER
)
3198 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
3200 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
3202 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
, gpu_id
);
3206 case MALI_JOB_TYPE_FRAGMENT
:
3207 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
, gpu_id
);
3213 } while ((jc_gpu_va
= h
->next_job
));
3215 pandecode_map_read_write();