projects / mesa.git / blob
? search:


0a75e66351fddc9691576de42d934782dd79a21a
[mesa.git] / src / panfrost / pandecode / decode.c
1 /*
2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
5 *
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:
12 *
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
15 * Software.
16 *
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
23 * SOFTWARE.
24 */
25
26 #include <panfrost-job.h>
27 #include <stdio.h>
28 #include <stdlib.h>
29 #include <memory.h>
30 #include <stdbool.h>
31 #include <stdarg.h>
32 #include <ctype.h>
33 #include "decode.h"
34 #include "util/macros.h"
35 #include "util/u_math.h"
36
37 #include "pan_pretty_print.h"
38 #include "midgard/disassemble.h"
39 #include "bifrost/disassemble.h"
40
41 #include "pan_encoder.h"
42
43 static void pandecode_swizzle(unsigned swizzle, enum mali_format format);
44
45 #define MEMORY_PROP(obj, p) {\
46 if (obj->p) { \
47 char *a = pointer_as_memory_reference(obj->p); \
48 pandecode_prop("%s = %s", #p, a); \
49 free(a); \
50 } \
51 }
52
53 #define MEMORY_PROP_DIR(obj, p) {\
54 if (obj.p) { \
55 char *a = pointer_as_memory_reference(obj.p); \
56 pandecode_prop("%s = %s", #p, a); \
57 free(a); \
58 } \
59 }
60
61 FILE *pandecode_dump_stream;
62
63 /* Semantic logging type.
64 *
65 * Raw: for raw messages to be printed as is.
66 * Message: for helpful information to be commented out in replays.
67 * Property: for properties of a struct
68 *
69 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
70 */
71
72 enum pandecode_log_type {
73 PANDECODE_RAW,
74 PANDECODE_MESSAGE,
75 PANDECODE_PROPERTY
76 };
77
78 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
79 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
80 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
81
82 unsigned pandecode_indent = 0;
83
84 static void
85 pandecode_make_indent(void)
86 {
87 for (unsigned i = 0; i < pandecode_indent; ++i)
88 fprintf(pandecode_dump_stream, " ");
89 }
90
91 static void
92 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
93 {
94 va_list ap;
95
96 pandecode_make_indent();
97
98 if (type == PANDECODE_MESSAGE)
99 fprintf(pandecode_dump_stream, "// ");
100 else if (type == PANDECODE_PROPERTY)
101 fprintf(pandecode_dump_stream, ".");
102
103 va_start(ap, format);
104 vfprintf(pandecode_dump_stream, format, ap);
105 va_end(ap);
106
107 if (type == PANDECODE_PROPERTY)
108 fprintf(pandecode_dump_stream, ",\n");
109 }
110
111 static void
112 pandecode_log_cont(const char *format, ...)
113 {
114 va_list ap;
115
116 va_start(ap, format);
117 vfprintf(pandecode_dump_stream, format, ap);
118 va_end(ap);
119 }
120
121 /* To check for memory safety issues, validates that the given pointer in GPU
122 * memory is valid, containing at least sz bytes. The goal is to eliminate
123 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
124 * overruns) by statically validating pointers.
125 */
126
127 static void
128 pandecode_validate_buffer(mali_ptr addr, size_t sz)
129 {
130 if (!addr) {
131 pandecode_msg("XXX: null pointer deref");
132 return;
133 }
134
135 /* Find a BO */
136
137 struct pandecode_mapped_memory *bo =
138 pandecode_find_mapped_gpu_mem_containing(addr);
139
140 if (!bo) {
141 pandecode_msg("XXX: invalid memory dereference\n");
142 return;
143 }
144
145 /* Bounds check */
146
147 unsigned offset = addr - bo->gpu_va;
148 unsigned total = offset + sz;
149
150 if (total > bo->length) {
151 pandecode_msg("XXX: buffer overrun. "
152 "Chunk of size %zu at offset %d in buffer of size %zu. "
153 "Overrun by %zu bytes. \n",
154 sz, offset, bo->length, total - bo->length);
155 return;
156 }
157 }
158
159 struct pandecode_flag_info {
160 u64 flag;
161 const char *name;
162 };
163
164 static void
165 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
166 u64 flags)
167 {
168 bool decodable_flags_found = false;
169
170 for (int i = 0; flag_info[i].name; i++) {
171 if ((flags & flag_info[i].flag) != flag_info[i].flag)
172 continue;
173
174 if (!decodable_flags_found) {
175 decodable_flags_found = true;
176 } else {
177 pandecode_log_cont(" | ");
178 }
179
180 pandecode_log_cont("%s", flag_info[i].name);
181
182 flags &= ~flag_info[i].flag;
183 }
184
185 if (decodable_flags_found) {
186 if (flags)
187 pandecode_log_cont(" | 0x%" PRIx64, flags);
188 } else {
189 pandecode_log_cont("0x%" PRIx64, flags);
190 }
191 }
192
193 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
194 static const struct pandecode_flag_info gl_enable_flag_info[] = {
195 FLAG_INFO(OCCLUSION_QUERY),
196 FLAG_INFO(OCCLUSION_PRECISE),
197 FLAG_INFO(FRONT_CCW_TOP),
198 FLAG_INFO(CULL_FACE_FRONT),
199 FLAG_INFO(CULL_FACE_BACK),
200 {}
201 };
202 #undef FLAG_INFO
203
204 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
205 static const struct pandecode_flag_info clear_flag_info[] = {
206 FLAG_INFO(FAST),
207 FLAG_INFO(SLOW),
208 FLAG_INFO(SLOW_STENCIL),
209 {}
210 };
211 #undef FLAG_INFO
212
213 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
214 static const struct pandecode_flag_info mask_flag_info[] = {
215 FLAG_INFO(R),
216 FLAG_INFO(G),
217 FLAG_INFO(B),
218 FLAG_INFO(A),
219 {}
220 };
221 #undef FLAG_INFO
222
223 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
224 static const struct pandecode_flag_info u3_flag_info[] = {
225 FLAG_INFO(HAS_MSAA),
226 FLAG_INFO(CAN_DISCARD),
227 FLAG_INFO(HAS_BLEND_SHADER),
228 FLAG_INFO(DEPTH_WRITEMASK),
229 {}
230 };
231
232 static const struct pandecode_flag_info u4_flag_info[] = {
233 FLAG_INFO(NO_MSAA),
234 FLAG_INFO(NO_DITHER),
235 FLAG_INFO(DEPTH_RANGE_A),
236 FLAG_INFO(DEPTH_RANGE_B),
237 FLAG_INFO(STENCIL_TEST),
238 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER),
239 {}
240 };
241 #undef FLAG_INFO
242
243 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
244 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
245 FLAG_INFO(MSAA),
246 FLAG_INFO(SRGB),
247 {}
248 };
249 #undef FLAG_INFO
250
251 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
252 static const struct pandecode_flag_info mfbd_extra_flag_hi_info[] = {
253 FLAG_INFO(PRESENT),
254 {}
255 };
256 #undef FLAG_INFO
257
258 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
259 static const struct pandecode_flag_info mfbd_extra_flag_lo_info[] = {
260 FLAG_INFO(ZS),
261 {}
262 };
263 #undef FLAG_INFO
264
265 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
266 static const struct pandecode_flag_info shader_midgard1_flag_lo_info [] = {
267 FLAG_INFO(WRITES_Z),
268 FLAG_INFO(EARLY_Z),
269 FLAG_INFO(READS_TILEBUFFER),
270 FLAG_INFO(READS_ZS),
271 {}
272 };
273
274 static const struct pandecode_flag_info shader_midgard1_flag_hi_info [] = {
275 FLAG_INFO(WRITES_S),
276 {}
277 };
278 #undef FLAG_INFO
279
280 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
281 static const struct pandecode_flag_info mfbd_flag_info [] = {
282 FLAG_INFO(DEPTH_WRITE),
283 FLAG_INFO(EXTRA),
284 {}
285 };
286 #undef FLAG_INFO
287
288 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
289 static const struct pandecode_flag_info sampler_flag_info [] = {
290 FLAG_INFO(MAG_NEAREST),
291 FLAG_INFO(MIN_NEAREST),
292 FLAG_INFO(MIP_LINEAR_1),
293 FLAG_INFO(MIP_LINEAR_2),
294 FLAG_INFO(NORM_COORDS),
295 {}
296 };
297 #undef FLAG_INFO
298
299 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
300 static const struct pandecode_flag_info sfbd_unk1_info [] = {
301 FLAG_INFO(MSAA_8),
302 FLAG_INFO(MSAA_A),
303 {}
304 };
305 #undef FLAG_INFO
306
307 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
308 static const struct pandecode_flag_info sfbd_unk2_info [] = {
309 FLAG_INFO(MSAA_B),
310 FLAG_INFO(SRGB),
311 {}
312 };
313 #undef FLAG_INFO
314
315 extern char *replace_fragment;
316 extern char *replace_vertex;
317
318 static char *
319 pandecode_job_type(enum mali_job_type type)
320 {
321 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
322
323 switch (type) {
324 DEFINE_CASE(NULL);
325 DEFINE_CASE(WRITE_VALUE);
326 DEFINE_CASE(CACHE_FLUSH);
327 DEFINE_CASE(COMPUTE);
328 DEFINE_CASE(VERTEX);
329 DEFINE_CASE(TILER);
330 DEFINE_CASE(FUSED);
331 DEFINE_CASE(FRAGMENT);
332
333 case JOB_NOT_STARTED:
334 return "NOT_STARTED";
335
336 default:
337 pandecode_log("Warning! Unknown job type %x\n", type);
338 return "!?!?!?";
339 }
340
341 #undef DEFINE_CASE
342 }
343
344 static char *
345 pandecode_draw_mode(enum mali_draw_mode mode)
346 {
347 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
348
349 switch (mode) {
350 DEFINE_CASE(DRAW_NONE);
351 DEFINE_CASE(POINTS);
352 DEFINE_CASE(LINES);
353 DEFINE_CASE(TRIANGLES);
354 DEFINE_CASE(TRIANGLE_STRIP);
355 DEFINE_CASE(TRIANGLE_FAN);
356 DEFINE_CASE(LINE_STRIP);
357 DEFINE_CASE(LINE_LOOP);
358 DEFINE_CASE(POLYGON);
359 DEFINE_CASE(QUADS);
360 DEFINE_CASE(QUAD_STRIP);
361
362 default:
363 pandecode_msg("XXX: invalid draw mode %X\n", mode);
364 return "";
365 }
366
367 #undef DEFINE_CASE
368 }
369
370 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
371 static char *
372 pandecode_func(enum mali_func mode)
373 {
374 switch (mode) {
375 DEFINE_CASE(NEVER);
376 DEFINE_CASE(LESS);
377 DEFINE_CASE(EQUAL);
378 DEFINE_CASE(LEQUAL);
379 DEFINE_CASE(GREATER);
380 DEFINE_CASE(NOTEQUAL);
381 DEFINE_CASE(GEQUAL);
382 DEFINE_CASE(ALWAYS);
383
384 default:
385 pandecode_msg("XXX: invalid func %X\n", mode);
386 return "";
387 }
388 }
389 #undef DEFINE_CASE
390
391 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
392 static char *
393 pandecode_stencil_op(enum mali_stencil_op op)
394 {
395 switch (op) {
396 DEFINE_CASE(KEEP);
397 DEFINE_CASE(REPLACE);
398 DEFINE_CASE(ZERO);
399 DEFINE_CASE(INVERT);
400 DEFINE_CASE(INCR_WRAP);
401 DEFINE_CASE(DECR_WRAP);
402 DEFINE_CASE(INCR);
403 DEFINE_CASE(DECR);
404
405 default:
406 pandecode_msg("XXX: invalid stencil op %X\n", op);
407 return "";
408 }
409 }
410
411 #undef DEFINE_CASE
412
413 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
414 {
415 switch(mode) {
416 /* TODO: Combine to just "instanced" once this can be done
417 * unambiguously in all known cases */
418 case MALI_ATTR_POT_DIVIDE:
419 return "instanced_pot";
420 case MALI_ATTR_MODULO:
421 return "instanced_mod";
422 case MALI_ATTR_NPOT_DIVIDE:
423 return "instanced_npot";
424 case MALI_ATTR_IMAGE:
425 return "image";
426 default:
427 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
428 return "";
429 }
430 }
431
432 static const char *
433 pandecode_special_record(uint64_t v, bool* attribute)
434 {
435 switch(v) {
436 case MALI_ATTR_VERTEXID:
437 *attribute = true;
438 return "gl_VertexID";
439 case MALI_ATTR_INSTANCEID:
440 *attribute = true;
441 return "gl_InstanceID";
442 case MALI_VARYING_FRAG_COORD:
443 return "gl_FragCoord";
444 case MALI_VARYING_FRONT_FACING:
445 return "gl_FrontFacing";
446 case MALI_VARYING_POINT_COORD:
447 return "gl_PointCoord";
448 default:
449 pandecode_msg("XXX: invalid special record %" PRIx64 "\n", v);
450 return "";
451 }
452 }
453
454 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
455 static char *
456 pandecode_wrap_mode(enum mali_wrap_mode op)
457 {
458 switch (op) {
459 DEFINE_CASE(REPEAT);
460 DEFINE_CASE(CLAMP_TO_EDGE);
461 DEFINE_CASE(CLAMP_TO_BORDER);
462 DEFINE_CASE(MIRRORED_REPEAT);
463
464 default:
465 pandecode_msg("XXX: invalid wrap mode %X\n", op);
466 return "";
467 }
468 }
469 #undef DEFINE_CASE
470
471 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
472 static char *
473 pandecode_block_format(enum mali_block_format fmt)
474 {
475 switch (fmt) {
476 DEFINE_CASE(TILED);
477 DEFINE_CASE(UNKNOWN);
478 DEFINE_CASE(LINEAR);
479 DEFINE_CASE(AFBC);
480
481 default:
482 unreachable("Invalid case");
483 }
484 }
485 #undef DEFINE_CASE
486
487 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
488 char *
489 pandecode_exception_access(unsigned access)
490 {
491 switch (access) {
492 DEFINE_CASE(NONE);
493 DEFINE_CASE(EXECUTE);
494 DEFINE_CASE(READ);
495 DEFINE_CASE(WRITE);
496
497 default:
498 unreachable("Invalid case");
499 }
500 }
501 #undef DEFINE_CASE
502
503 /* Midgard's tiler descriptor is embedded within the
504 * larger FBD */
505
506 static void
507 pandecode_midgard_tiler_descriptor(
508 const struct midgard_tiler_descriptor *t,
509 unsigned width,
510 unsigned height,
511 bool is_fragment,
512 bool has_hierarchy)
513 {
514 pandecode_log(".tiler = {\n");
515 pandecode_indent++;
516
517 if (t->hierarchy_mask == MALI_TILER_DISABLED)
518 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
519 else
520 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
521
522 /* We know this name from the kernel, but we never see it nonzero */
523
524 if (t->flags)
525 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
526
527 MEMORY_PROP(t, polygon_list);
528
529 /* The body is offset from the base of the polygon list */
530 assert(t->polygon_list_body > t->polygon_list);
531 unsigned body_offset = t->polygon_list_body - t->polygon_list;
532
533 /* It needs to fit inside the reported size */
534 assert(t->polygon_list_size >= body_offset);
535
536 /* Check that we fit */
537 struct pandecode_mapped_memory *plist =
538 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
539
540 assert(t->polygon_list_size <= plist->length);
541
542 /* Now that we've sanity checked, we'll try to calculate the sizes
543 * ourselves for comparison */
544
545 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
546 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
547
548 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
549 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
550 ref_header, ref_size);
551 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
552 pandecode_msg("body offset %d\n", body_offset);
553 }
554
555 /* The tiler heap has a start and end specified -- it should be
556 * identical to what we have in the BO. The exception is if tiling is
557 * disabled. */
558
559 MEMORY_PROP(t, heap_start);
560 assert(t->heap_end >= t->heap_start);
561
562 struct pandecode_mapped_memory *heap =
563 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
564
565 unsigned heap_size = t->heap_end - t->heap_start;
566
567 /* Tiling is enabled with a special flag */
568 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
569 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
570
571 bool tiling_enabled = hierarchy_mask;
572
573 if (tiling_enabled) {
574 /* When tiling is enabled, the heap should be a tight fit */
575 unsigned heap_offset = t->heap_start - heap->gpu_va;
576 if ((heap_offset + heap_size) != heap->length) {
577 pandecode_msg("XXX: heap size %u (expected %zu)\n",
578 heap_size, heap->length - heap_offset);
579 }
580
581 /* We should also have no other flags */
582 if (tiler_flags)
583 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
584 } else {
585 /* When tiling is disabled, we should have that flag and no others */
586
587 if (tiler_flags != MALI_TILER_DISABLED) {
588 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
589 tiler_flags);
590 }
591
592 /* We should also have an empty heap */
593 if (heap_size) {
594 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
595 heap_size);
596 }
597
598 /* Disabled tiling is used only for clear-only jobs, which are
599 * purely FRAGMENT, so we should never see this for
600 * non-FRAGMENT descriptors. */
601
602 if (!is_fragment)
603 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
604 }
605
606 /* We've never seen weights used in practice, but we know from the
607 * kernel these fields is there */
608
609 bool nonzero_weights = false;
610
611 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
612 nonzero_weights |= t->weights[w] != 0x0;
613 }
614
615 if (nonzero_weights) {
616 pandecode_log(".weights = {");
617
618 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
619 pandecode_log("%d, ", t->weights[w]);
620 }
621
622 pandecode_log("},");
623 }
624
625 pandecode_indent--;
626 pandecode_log("}\n");
627 }
628
629 /* Information about the framebuffer passed back for
630 * additional analysis */
631
632 struct pandecode_fbd {
633 unsigned width;
634 unsigned height;
635 unsigned rt_count;
636 bool has_extra;
637 };
638
639 static void
640 pandecode_sfbd_format(struct mali_sfbd_format format)
641 {
642 pandecode_log(".format = {\n");
643 pandecode_indent++;
644
645 pandecode_log(".unk1 = ");
646 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
647 pandecode_log_cont(",\n");
648
649 /* TODO: Map formats so we can check swizzles and print nicely */
650 pandecode_log("swizzle");
651 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
652 pandecode_log_cont(",\n");
653
654 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
655 (format.nr_channels + 1));
656
657 pandecode_log(".unk2 = ");
658 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
659 pandecode_log_cont(",\n");
660
661 pandecode_prop("block = %s", pandecode_block_format(format.block));
662
663 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
664
665 pandecode_indent--;
666 pandecode_log("},\n");
667 }
668
669 static void
670 pandecode_shared_memory(const struct mali_shared_memory *desc, bool is_compute)
671 {
672 pandecode_prop("stack_shift = 0x%x", desc->stack_shift);
673
674 if (desc->unk0)
675 pandecode_prop("unk0 = 0x%x", desc->unk0);
676
677 if (desc->shared_workgroup_count != 0x1F) {
678 pandecode_prop("shared_workgroup_count = %d", desc->shared_workgroup_count);
679 if (!is_compute)
680 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
681 }
682
683 if (desc->shared_unk1 || desc->shared_shift) {
684 pandecode_prop("shared_unk1 = %X", desc->shared_unk1);
685 pandecode_prop("shared_shift = %X", desc->shared_shift);
686
687 if (!is_compute)
688 pandecode_msg("XXX: shared memory configured in noncompute shader");
689 }
690
691 if (desc->shared_zero) {
692 pandecode_msg("XXX: shared memory zero tripped\n");
693 pandecode_prop("shared_zero = 0x%" PRIx32, desc->shared_zero);
694 }
695
696 if (desc->shared_memory && !is_compute)
697 pandecode_msg("XXX: shared memory used in noncompute shader\n");
698
699 MEMORY_PROP(desc, scratchpad);
700 MEMORY_PROP(desc, shared_memory);
701 MEMORY_PROP(desc, unknown1);
702 }
703
704 static struct pandecode_fbd
705 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
706 {
707 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
708 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
709
710 struct pandecode_fbd info = {
711 .has_extra = false,
712 .rt_count = 1
713 };
714
715 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
716 pandecode_indent++;
717
718 pandecode_log(".shared_memory = {\n");
719 pandecode_indent++;
720 pandecode_shared_memory(&s->shared_memory, false);
721 pandecode_indent--;
722 pandecode_log("},\n");
723
724 pandecode_sfbd_format(s->format);
725
726 info.width = s->width + 1;
727 info.height = s->height + 1;
728
729 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
730 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
731
732 MEMORY_PROP(s, checksum);
733
734 if (s->checksum_stride)
735 pandecode_prop("checksum_stride = %d", s->checksum_stride);
736
737 MEMORY_PROP(s, framebuffer);
738 pandecode_prop("stride = %d", s->stride);
739
740 /* Earlier in the actual commandstream -- right before width -- but we
741 * delay to flow nicer */
742
743 pandecode_log(".clear_flags = ");
744 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
745 pandecode_log_cont(",\n");
746
747 if (s->depth_buffer) {
748 MEMORY_PROP(s, depth_buffer);
749 pandecode_prop("depth_stride = %d", s->depth_stride);
750 }
751
752 if (s->stencil_buffer) {
753 MEMORY_PROP(s, stencil_buffer);
754 pandecode_prop("stencil_stride = %d", s->stencil_stride);
755 }
756
757 if (s->depth_stride_zero ||
758 s->stencil_stride_zero ||
759 s->zero7 || s->zero8) {
760 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
761 pandecode_prop("depth_stride_zero = 0x%x",
762 s->depth_stride_zero);
763 pandecode_prop("stencil_stride_zero = 0x%x",
764 s->stencil_stride_zero);
765 pandecode_prop("zero7 = 0x%" PRIx32,
766 s->zero7);
767 pandecode_prop("zero8 = 0x%" PRIx32,
768 s->zero8);
769 }
770
771 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
772 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
773 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
774 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
775 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
776 }
777
778 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
779 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
780 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
781 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
782 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
783 }
784
785 if (s->clear_stencil) {
786 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
787 }
788
789 const struct midgard_tiler_descriptor t = s->tiler;
790
791 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
792 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
793
794 pandecode_indent--;
795 pandecode_log("};\n");
796
797 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
798 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
799 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
800
801 pandecode_log_cont(".zero3 = {");
802
803 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
804 pandecode_log_cont("%X, ", s->zero3[i]);
805
806 pandecode_log_cont("},\n");
807
808 pandecode_log_cont(".zero6 = {");
809
810 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
811 pandecode_log_cont("%X, ", s->zero6[i]);
812
813 pandecode_log_cont("},\n");
814
815 return info;
816 }
817
818 static void
819 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
820 {
821 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
822 const struct mali_shared_memory *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
823
824 pandecode_log("struct mali_shared_memory shared_%"PRIx64"_%d = {\n", gpu_va, job_no);
825 pandecode_indent++;
826 pandecode_shared_memory(s, true);
827 pandecode_indent--;
828 pandecode_log("},\n");
829 }
830
831 /* Extracts the number of components associated with a Mali format */
832
833 static unsigned
834 pandecode_format_component_count(enum mali_format fmt)
835 {
836 /* Mask out the format class */
837 unsigned top = fmt & 0b11100000;
838
839 switch (top) {
840 case MALI_FORMAT_SNORM:
841 case MALI_FORMAT_UINT:
842 case MALI_FORMAT_UNORM:
843 case MALI_FORMAT_SINT:
844 return ((fmt >> 3) & 3) + 1;
845 default:
846 /* TODO: Validate */
847 return 4;
848 }
849 }
850
851 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
852
853 static unsigned
854 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
855 {
856 unsigned mask = 0;
857 assert(MALI_CHANNEL_RED == 0);
858
859 for (unsigned c = 0; c < 4; ++c) {
860 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
861
862 if (chan <= MALI_CHANNEL_ALPHA)
863 mask |= (1 << chan);
864 }
865
866 return mask;
867 }
868
869 /* Validates that a (format, swizzle) pair is valid, in the sense that the
870 * swizzle doesn't access any components that are undefined in the format.
871 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
872 * omitted */
873
874 static bool
875 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
876 {
877 unsigned nr_comp = pandecode_format_component_count(fmt);
878 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
879 unsigned valid_mask = (1 << nr_comp) - 1;
880 unsigned invalid_mask = ~valid_mask;
881
882 if (access_mask & invalid_mask) {
883 pandecode_msg("XXX: invalid components accessed\n");
884 return false;
885 }
886
887 /* Check for the default non-swizzling swizzle so we can suppress
888 * useless printing for the defaults */
889
890 unsigned default_swizzles[4] = {
891 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
892 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
893 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
894 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
895 };
896
897 return (swizzle == default_swizzles[nr_comp - 1]);
898 }
899
900 /* Maps MALI_RGBA32F to rgba32f, etc */
901
902 static void
903 pandecode_format_short(enum mali_format fmt, bool srgb)
904 {
905 /* We want a type-like format, so cut off the initial MALI_ */
906 char *format = pandecode_format(fmt);
907 format += strlen("MALI_");
908
909 unsigned len = strlen(format);
910 char *lower_format = calloc(1, len + 1);
911
912 for (unsigned i = 0; i < len; ++i)
913 lower_format[i] = tolower(format[i]);
914
915 /* Sanity check sRGB flag is applied to RGB, per the name */
916 if (srgb && lower_format[0] != 'r')
917 pandecode_msg("XXX: sRGB applied to non-colour format\n");
918
919 /* Just prefix with an s, so you get formats like srgba8_unorm */
920 if (srgb)
921 pandecode_log_cont("s");
922
923 pandecode_log_cont("%s", lower_format);
924 free(lower_format);
925 }
926
927 static void
928 pandecode_swizzle(unsigned swizzle, enum mali_format format)
929 {
930 /* First, do some validation */
931 bool trivial_swizzle = pandecode_validate_format_swizzle(
932 format, swizzle);
933
934 if (trivial_swizzle)
935 return;
936
937 /* Next, print the swizzle */
938 pandecode_log_cont(".");
939
940 static const char components[] = "rgba01";
941
942 for (unsigned c = 0; c < 4; ++c) {
943 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
944
945 if (chan >= MALI_CHANNEL_RESERVED_0) {
946 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
947 continue;
948 }
949 pandecode_log_cont("%c", components[chan]);
950 }
951 }
952
953 static void
954 pandecode_rt_format(struct mali_rt_format format)
955 {
956 pandecode_log(".format = {\n");
957 pandecode_indent++;
958
959 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
960 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
961 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
962
963 pandecode_prop("block = %s", pandecode_block_format(format.block));
964
965 /* TODO: Map formats so we can check swizzles and print nicely */
966 pandecode_log("swizzle");
967 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
968 pandecode_log_cont(",\n");
969
970 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
971 (format.nr_channels + 1));
972
973 pandecode_log(".flags = ");
974 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
975 pandecode_log_cont(",\n");
976
977 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
978 * which is a faster hardware-based alternative to the wallpaper method
979 * to preserve framebuffer contents across frames. In practice, MFBD
980 * preload is buggy on Midgard, and so this is a chicken bit. If this
981 * bit isn't set, most likely something broke unrelated to preload */
982
983 if (!format.no_preload) {
984 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
985 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
986 }
987
988 if (format.zero)
989 pandecode_prop("zero = 0x%" PRIx32, format.zero);
990
991 pandecode_indent--;
992 pandecode_log("},\n");
993 }
994
995 static void
996 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct mali_framebuffer *fb)
997 {
998 pandecode_log("struct mali_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
999 pandecode_indent++;
1000
1001 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
1002 mali_ptr rt_va = gpu_va + i * sizeof(struct mali_render_target);
1003 struct pandecode_mapped_memory *mem =
1004 pandecode_find_mapped_gpu_mem_containing(rt_va);
1005 const struct mali_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
1006
1007 pandecode_log("{\n");
1008 pandecode_indent++;
1009
1010 pandecode_rt_format(rt->format);
1011
1012 if (rt->format.block == MALI_BLOCK_AFBC) {
1013 pandecode_log(".afbc = {\n");
1014 pandecode_indent++;
1015
1016 char *a = pointer_as_memory_reference(rt->afbc.metadata);
1017 pandecode_prop("metadata = %s", a);
1018 free(a);
1019
1020 pandecode_prop("stride = %d", rt->afbc.stride);
1021 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
1022
1023 pandecode_indent--;
1024 pandecode_log("},\n");
1025 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
1026 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1027 rt->afbc.metadata,
1028 rt->afbc.stride,
1029 rt->afbc.unk);
1030 }
1031
1032 MEMORY_PROP(rt, framebuffer);
1033 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1034
1035 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1036 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1037 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1038 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1039 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1040 }
1041
1042 if (rt->zero1 || rt->zero2 || rt->zero3) {
1043 pandecode_msg("XXX: render target zeros tripped\n");
1044 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1045 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1046 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
1047 }
1048
1049 pandecode_indent--;
1050 pandecode_log("},\n");
1051 }
1052
1053 pandecode_indent--;
1054 pandecode_log("};\n");
1055 }
1056
1057 static struct pandecode_fbd
1058 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment, bool is_compute)
1059 {
1060 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1061 const struct mali_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1062
1063 struct pandecode_fbd info;
1064
1065 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1066 pandecode_indent++;
1067
1068 pandecode_log(".shared_memory = {\n");
1069 pandecode_indent++;
1070 pandecode_shared_memory(&fb->shared_memory, is_compute);
1071 pandecode_indent--;
1072 pandecode_log("},\n");
1073
1074 info.width = fb->width1 + 1;
1075 info.height = fb->height1 + 1;
1076 info.rt_count = fb->rt_count_1 + 1;
1077
1078 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1079 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1080 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1081 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1082
1083 pandecode_prop("unk1 = 0x%x", fb->unk1);
1084 pandecode_prop("unk2 = 0x%x", fb->unk2);
1085 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1086 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1087
1088 pandecode_log(".mfbd_flags = ");
1089 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1090 pandecode_log_cont(",\n");
1091
1092 if (fb->clear_stencil)
1093 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1094
1095 if (fb->clear_depth)
1096 pandecode_prop("clear_depth = %f", fb->clear_depth);
1097
1098 const struct midgard_tiler_descriptor t = fb->tiler;
1099 if (!is_compute)
1100 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
1101 else
1102 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1103
1104 if (fb->zero3 || fb->zero4) {
1105 pandecode_msg("XXX: framebuffer zeros tripped\n");
1106 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1107 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1108 }
1109
1110 pandecode_indent--;
1111 pandecode_log("};\n");
1112
1113 gpu_va += sizeof(struct mali_framebuffer);
1114
1115 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1116
1117 if (info.has_extra) {
1118 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1119 const struct mali_framebuffer_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1120
1121 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1122 pandecode_indent++;
1123
1124 MEMORY_PROP(fbx, checksum);
1125
1126 if (fbx->checksum_stride)
1127 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1128
1129 pandecode_log(".flags_hi = ");
1130 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_lo);
1131 pandecode_log_cont(",\n");
1132
1133 pandecode_log(".flags_lo = ");
1134 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
1135 pandecode_log_cont(",\n");
1136
1137 pandecode_prop("zs_block = %s", pandecode_block_format(fbx->zs_block));
1138
1139 if (fbx->zs_block == MALI_BLOCK_AFBC) {
1140 pandecode_log(".ds_afbc = {\n");
1141 pandecode_indent++;
1142
1143 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1144 pandecode_prop("depth_stencil_afbc_stride = %d",
1145 fbx->ds_afbc.depth_stencil_afbc_stride);
1146 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1147
1148 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
1149 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1150 pandecode_prop("zero1 = 0x%" PRIx32,
1151 fbx->ds_afbc.zero1);
1152 pandecode_prop("padding = 0x%" PRIx64,
1153 fbx->ds_afbc.padding);
1154 }
1155
1156 pandecode_indent--;
1157 pandecode_log("},\n");
1158 } else {
1159 pandecode_log(".ds_linear = {\n");
1160 pandecode_indent++;
1161
1162 if (fbx->ds_linear.depth) {
1163 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1164 pandecode_prop("depth_stride = %d",
1165 fbx->ds_linear.depth_stride);
1166 } else if (fbx->ds_linear.depth_stride) {
1167 pandecode_msg("XXX: depth stride zero tripped %d\n", fbx->ds_linear.depth_stride);
1168 }
1169
1170 if (fbx->ds_linear.stencil) {
1171 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1172 pandecode_prop("stencil_stride = %d",
1173 fbx->ds_linear.stencil_stride);
1174 } else if (fbx->ds_linear.stencil_stride) {
1175 pandecode_msg("XXX: stencil stride zero tripped %d\n", fbx->ds_linear.stencil_stride);
1176 }
1177
1178 if (fbx->ds_linear.depth_stride_zero ||
1179 fbx->ds_linear.stencil_stride_zero ||
1180 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1181 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1182 pandecode_prop("depth_stride_zero = 0x%x",
1183 fbx->ds_linear.depth_stride_zero);
1184 pandecode_prop("stencil_stride_zero = 0x%x",
1185 fbx->ds_linear.stencil_stride_zero);
1186 pandecode_prop("zero1 = 0x%" PRIx32,
1187 fbx->ds_linear.zero1);
1188 pandecode_prop("zero2 = 0x%" PRIx32,
1189 fbx->ds_linear.zero2);
1190 }
1191
1192 pandecode_indent--;
1193 pandecode_log("},\n");
1194 }
1195
1196 if (fbx->zero3 || fbx->zero4) {
1197 pandecode_msg("XXX: fb_extra zeros tripped\n");
1198 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1199 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1200 }
1201
1202 pandecode_indent--;
1203 pandecode_log("};\n");
1204
1205 gpu_va += sizeof(struct mali_framebuffer_extra);
1206 }
1207
1208 if (is_fragment)
1209 pandecode_render_target(gpu_va, job_no, fb);
1210
1211 return info;
1212 }
1213
1214 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1215 * count */
1216
1217 static void
1218 pandecode_padded_vertices(unsigned shift, unsigned k)
1219 {
1220 unsigned odd = 2*k + 1;
1221 unsigned pot = 1 << shift;
1222 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1223 }
1224
1225 /* Given a magic divisor, recover what we were trying to divide by.
1226 *
1227 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1228 * 0 <= m < 2^N, for N bits in m.
1229 *
1230 * Let q represent the number we would like to divide by.
1231 *
1232 * By definition of a magic divisor for N-bit unsigned integers (a number you
1233 * multiply by to magically get division), m is a number such that:
1234 *
1235 * (m * x) & (2^N - 1) = floor(x/q).
1236 * for all x on Z where 0 <= x < 2^N
1237 *
1238 * Ignore the case where any of the above values equals zero; it is irrelevant
1239 * for our purposes (instanced arrays).
1240 *
1241 * Choose x = q. Then:
1242 *
1243 * (m * x) & (2^N - 1) = floor(x/q).
1244 * (m * q) & (2^N - 1) = floor(q/q).
1245 *
1246 * floor(q/q) = floor(1) = 1, therefore:
1247 *
1248 * (m * q) & (2^N - 1) = 1
1249 *
1250 * Recall the identity that the bitwise AND of one less than a power-of-two
1251 * equals the modulo with that power of two, i.e. for all x:
1252 *
1253 * x & (2^N - 1) = x % N
1254 *
1255 * Therefore:
1256 *
1257 * mq % (2^N) = 1
1258 *
1259 * By definition, a modular multiplicative inverse of a number m is the number
1260 * q such that with respect to a modulos M:
1261 *
1262 * mq % M = 1
1263 *
1264 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1265 *
1266 */
1267
1268 static void
1269 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1270 {
1271 #if 0
1272 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1273 * shift) the most lame way possible... just repeatedly add.
1274 * Asymptoptically slow but nobody cares in practice, unless you have
1275 * massive numbers of vertices or high divisors. */
1276
1277 unsigned inverse = 0;
1278
1279 /* Magic implicitly has the highest bit set */
1280 magic |= (1 << 31);
1281
1282 /* Depending on rounding direction */
1283 if (extra)
1284 magic++;
1285
1286 for (;;) {
1287 uint32_t product = magic * inverse;
1288
1289 if (shift) {
1290 product >>= shift;
1291 }
1292
1293 if (product == 1)
1294 break;
1295
1296 ++inverse;
1297 }
1298
1299 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1300
1301 /* Recall we're supposed to divide by (gl_level_divisor *
1302 * padded_num_vertices) */
1303
1304 unsigned padded_num_vertices = inverse / orig_divisor;
1305
1306 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1307 #endif
1308 }
1309
1310 static void
1311 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1312 mali_ptr addr, int job_no, char *suffix,
1313 int count, bool varying, enum mali_job_type job_type)
1314 {
1315 char *prefix = varying ? "varying" : "attribute";
1316 assert(addr);
1317
1318 if (!count) {
1319 pandecode_msg("warn: No %s records\n", prefix);
1320 return;
1321 }
1322
1323 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1324
1325 for (int i = 0; i < count; ++i) {
1326 /* First, check for special records */
1327 if (attr[i].elements < MALI_RECORD_SPECIAL) {
1328 if (attr[i].size)
1329 pandecode_msg("XXX: tripped size=%d\n", attr[i].size);
1330
1331 if (attr[i].stride) {
1332 /* gl_InstanceID passes a magic divisor in the
1333 * stride field to divide by the padded vertex
1334 * count. No other records should do so, so
1335 * stride should otherwise be zero. Note that
1336 * stride in the usual attribute sense doesn't
1337 * apply to special records. */
1338
1339 bool has_divisor = attr[i].elements == MALI_ATTR_INSTANCEID;
1340
1341 pandecode_log_cont("/* %smagic divisor = %X */ ",
1342 has_divisor ? "" : "XXX: ", attr[i].stride);
1343 }
1344
1345 if (attr[i].shift || attr[i].extra_flags) {
1346 /* Attributes use these fields for
1347 * instancing/padding/etc type issues, but
1348 * varyings don't */
1349
1350 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1351 varying ? "XXX: " : "",
1352 attr[i].shift, attr[i].extra_flags);
1353 }
1354
1355 /* Print the special record name */
1356 bool attribute = false;
1357 pandecode_log("%s_%d = %s;\n", prefix, i, pandecode_special_record(attr[i].elements, &attribute));
1358
1359 /* Sanity check */
1360 if (attribute == varying)
1361 pandecode_msg("XXX: mismatched special record\n");
1362
1363 continue;
1364 }
1365
1366 enum mali_attr_mode mode = attr[i].elements & 7;
1367
1368 if (mode == MALI_ATTR_UNUSED)
1369 pandecode_msg("XXX: unused attribute record\n");
1370
1371 /* For non-linear records, we need to print the type of record */
1372 if (mode != MALI_ATTR_LINEAR)
1373 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1374
1375 /* Print the name to link with attr_meta */
1376 pandecode_log_cont("%s_%d", prefix, i);
1377
1378 /* Print the stride and size */
1379 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1380
1381 /* TODO: Sanity check the quotient itself. It must be equal to
1382 * (or be greater than, if the driver added padding) the padded
1383 * vertex count. */
1384
1385 /* Finally, print the pointer */
1386 mali_ptr raw_elements = attr[i].elements & ~7;
1387 char *a = pointer_as_memory_reference(raw_elements);
1388 pandecode_log_cont(" = (%s);\n", a);
1389 free(a);
1390
1391 /* Check the pointer */
1392 pandecode_validate_buffer(raw_elements, attr[i].size);
1393
1394 /* shift/extra_flags exist only for instanced */
1395 if (attr[i].shift | attr[i].extra_flags) {
1396 /* These are set to random values by the blob for
1397 * varyings, most likely a symptom of uninitialized
1398 * memory where the hardware masked the bug. As such we
1399 * put this at a warning, not an error. */
1400
1401 if (mode == MALI_ATTR_LINEAR)
1402 pandecode_msg("warn: instancing fields set for linear\n");
1403
1404 pandecode_prop("shift = %d", attr[i].shift);
1405 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1406 }
1407
1408 /* Decode further where possible */
1409
1410 if (mode == MALI_ATTR_MODULO) {
1411 pandecode_padded_vertices(
1412 attr[i].shift,
1413 attr[i].extra_flags);
1414 }
1415
1416 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1417 i++;
1418 pandecode_log("{\n");
1419 pandecode_indent++;
1420 pandecode_prop("unk = 0x%x", attr[i].unk);
1421 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1422 if (attr[i].zero != 0)
1423 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1424 pandecode_prop("divisor = %d", attr[i].divisor);
1425 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1426 pandecode_indent--;
1427 pandecode_log("}, \n");
1428 }
1429
1430 }
1431
1432 pandecode_log("\n");
1433 }
1434
1435 static mali_ptr
1436 pandecode_shader_address(const char *name, mali_ptr ptr)
1437 {
1438 /* TODO: Decode flags */
1439 mali_ptr shader_ptr = ptr & ~15;
1440
1441 char *a = pointer_as_memory_reference(shader_ptr);
1442 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1443 free(a);
1444
1445 return shader_ptr;
1446 }
1447
1448 static void
1449 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1450 {
1451 unsigned any_nonzero =
1452 stencil->ref | stencil->mask | stencil->func |
1453 stencil->sfail | stencil->dpfail | stencil->dppass;
1454
1455 if (any_nonzero == 0)
1456 return;
1457
1458 const char *func = pandecode_func(stencil->func);
1459 const char *sfail = pandecode_stencil_op(stencil->sfail);
1460 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1461 const char *dppass = pandecode_stencil_op(stencil->dppass);
1462
1463 if (stencil->zero)
1464 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1465
1466 pandecode_log(".stencil_%s = {\n", name);
1467 pandecode_indent++;
1468 pandecode_prop("ref = %d", stencil->ref);
1469 pandecode_prop("mask = 0x%02X", stencil->mask);
1470 pandecode_prop("func = %s", func);
1471 pandecode_prop("sfail = %s", sfail);
1472 pandecode_prop("dpfail = %s", dpfail);
1473 pandecode_prop("dppass = %s", dppass);
1474 pandecode_indent--;
1475 pandecode_log("},\n");
1476 }
1477
1478 static void
1479 pandecode_blend_equation(const struct mali_blend_equation *blend)
1480 {
1481 if (blend->zero1)
1482 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1483
1484 pandecode_log(".equation = {\n");
1485 pandecode_indent++;
1486
1487 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1488 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1489
1490 pandecode_log(".color_mask = ");
1491 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1492 pandecode_log_cont(",\n");
1493
1494 pandecode_indent--;
1495 pandecode_log("},\n");
1496 }
1497
1498 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1499
1500 static unsigned
1501 decode_bifrost_constant(u16 constant)
1502 {
1503 float lo = (float) (constant & 0xFF);
1504 float hi = (float) (constant >> 8);
1505
1506 return (hi / 255.0) + (lo / 65535.0);
1507 }
1508
1509 static mali_ptr
1510 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1511 {
1512 struct bifrost_blend_rt *b =
1513 ((struct bifrost_blend_rt *) descs) + rt_no;
1514
1515 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1516 pandecode_indent++;
1517
1518 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1519 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1520 b->constant, decode_bifrost_constant(b->constant));
1521
1522 /* TODO figure out blend shader enable bit */
1523 pandecode_blend_equation(&b->equation);
1524 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1525 pandecode_prop("index = 0x%" PRIx16, b->index);
1526 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1527
1528 pandecode_indent--;
1529 pandecode_log("},\n");
1530
1531 return 0;
1532 }
1533
1534 static mali_ptr
1535 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1536 {
1537 /* constant/equation is in a union */
1538 if (!blend->shader)
1539 return 0;
1540
1541 pandecode_log(".blend = {\n");
1542 pandecode_indent++;
1543
1544 if (is_shader) {
1545 pandecode_shader_address("shader", blend->shader);
1546 } else {
1547 pandecode_blend_equation(&blend->equation);
1548 pandecode_prop("constant = %f", blend->constant);
1549 }
1550
1551 pandecode_indent--;
1552 pandecode_log("},\n");
1553
1554 /* Return blend shader to disassemble if present */
1555 return is_shader ? (blend->shader & ~0xF) : 0;
1556 }
1557
1558 static mali_ptr
1559 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1560 {
1561 struct midgard_blend_rt *b =
1562 ((struct midgard_blend_rt *) descs) + rt_no;
1563
1564 /* Flags determine presence of blend shader */
1565 bool is_shader = (b->flags & 0xF) >= 0x2;
1566
1567 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1568 pandecode_indent++;
1569
1570 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1571
1572 union midgard_blend blend = b->blend;
1573 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1574
1575 pandecode_indent--;
1576 pandecode_log("};\n");
1577
1578 return shader;
1579 }
1580
1581 /* Attributes and varyings have descriptor records, which contain information
1582 * about their format and ordering with the attribute/varying buffers. We'll
1583 * want to validate that the combinations specified are self-consistent.
1584 */
1585
1586 static int
1587 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1588 {
1589 char base[128];
1590 char *prefix = varying ? "varying" : "attribute";
1591 unsigned max_index = 0;
1592 snprintf(base, sizeof(base), "%s_meta", prefix);
1593
1594 struct mali_attr_meta *attr_meta;
1595 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1596
1597 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1598
1599 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1600 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1601 sizeof(*attr_mem));
1602
1603 /* If the record is discard, it should be zero for everything else */
1604
1605 if (attr_meta->format == MALI_VARYING_DISCARD) {
1606 uint64_t zero =
1607 attr_meta->index |
1608 attr_meta->unknown1 |
1609 attr_meta->unknown3 |
1610 attr_meta->src_offset;
1611
1612 if (zero)
1613 pandecode_msg("XXX: expected empty record for varying discard\n");
1614
1615 /* We want to look for a literal 0000 swizzle -- this
1616 * is not encoded with all zeroes, however */
1617
1618 enum mali_channel z = MALI_CHANNEL_ZERO;
1619 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1620 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1621
1622 if (!good_swizzle)
1623 pandecode_msg("XXX: expected zero swizzle for discard\n");
1624
1625 if (!varying)
1626 pandecode_msg("XXX: cannot discard attribute\n");
1627
1628 /* If we're all good, omit the record */
1629 if (!zero && varying && good_swizzle) {
1630 pandecode_log("/* discarded varying */\n");
1631 continue;
1632 }
1633 }
1634
1635 if (attr_meta->index > max_index)
1636 max_index = attr_meta->index;
1637
1638 if (attr_meta->unknown1 != 0x2) {
1639 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1640 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1641 }
1642
1643 if (attr_meta->unknown3) {
1644 pandecode_msg("XXX: unexpected unknown3 set\n");
1645 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1646 }
1647
1648 pandecode_format_short(attr_meta->format, false);
1649 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1650
1651 if (attr_meta->src_offset)
1652 pandecode_log_cont("[%u]", attr_meta->src_offset);
1653
1654 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1655
1656 pandecode_log_cont(";\n");
1657 }
1658
1659 pandecode_log("\n");
1660
1661 return count ? (max_index + 1) : 0;
1662 }
1663
1664 /* return bits [lo, hi) of word */
1665 static u32
1666 bits(u32 word, u32 lo, u32 hi)
1667 {
1668 if (hi - lo >= 32)
1669 return word; // avoid undefined behavior with the shift
1670
1671 return (word >> lo) & ((1 << (hi - lo)) - 1);
1672 }
1673
1674 static void
1675 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1676 {
1677 pandecode_log_cont("{\n");
1678 pandecode_indent++;
1679
1680 /* Decode invocation_count. See the comment before the definition of
1681 * invocation_count for an explanation.
1682 */
1683
1684 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1685 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1686 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1687 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1688 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1689 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1690
1691 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1692 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1693 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1694
1695 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1696 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1697 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1698
1699 /* Even though we have this decoded, we want to ensure that the
1700 * representation is "unique" so we don't lose anything by printing only
1701 * the final result. More specifically, we need to check that we were
1702 * passed something in canonical form, since the definition per the
1703 * hardware is inherently not unique. How? Well, take the resulting
1704 * decode and pack it ourselves! If it is bit exact with what we
1705 * decoded, we're good to go. */
1706
1707 struct mali_vertex_tiler_prefix ref;
1708 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1709
1710 bool canonical =
1711 (p->invocation_count == ref.invocation_count) &&
1712 (p->invocation_shifts == ref.invocation_shifts);
1713
1714 if (!canonical) {
1715 pandecode_msg("XXX: non-canonical workgroups packing\n");
1716 pandecode_msg("expected: %X, %X",
1717 ref.invocation_count,
1718 ref.invocation_shifts);
1719
1720 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1721 pandecode_prop("size_y_shift = %d", size_y_shift);
1722 pandecode_prop("size_z_shift = %d", size_z_shift);
1723 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
1724 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
1725 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
1726 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
1727 }
1728
1729 /* Regardless, print the decode */
1730 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1731 size_x, size_y, size_z,
1732 groups_x, groups_y, groups_z);
1733
1734 /* TODO: Decode */
1735 if (p->unknown_draw)
1736 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1737
1738 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1739
1740 if (p->draw_mode != MALI_DRAW_NONE)
1741 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1742
1743 /* Index count only exists for tiler jobs anyway */
1744
1745 if (p->index_count)
1746 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1747
1748
1749 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1750 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1751
1752 /* Validate an index buffer is present if we need one. TODO: verify
1753 * relationship between invocation_count and index_count */
1754
1755 if (p->indices) {
1756 unsigned count = p->index_count;
1757
1758 /* Grab the size */
1759 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1760
1761 /* Ensure we got a size, and if so, validate the index buffer
1762 * is large enough to hold a full set of indices of the given
1763 * size */
1764
1765 if (!index_raw_size)
1766 pandecode_msg("XXX: index size missing\n");
1767 else
1768 pandecode_validate_buffer(p->indices, count * size);
1769 } else if (index_raw_size)
1770 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1771
1772 if (p->offset_bias_correction)
1773 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1774
1775 /* TODO: Figure out what this is. It's not zero */
1776 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1777
1778 pandecode_indent--;
1779 pandecode_log("},\n");
1780 }
1781
1782 static void
1783 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1784 {
1785 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1786 uint64_t *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1787
1788 for (int i = 0; i < ubufs_count; i++) {
1789 unsigned size = (ubufs[i] & ((1 << 10) - 1)) * 16;
1790 mali_ptr addr = (ubufs[i] >> 10) << 2;
1791
1792 pandecode_validate_buffer(addr, size);
1793
1794 char *ptr = pointer_as_memory_reference(addr);
1795 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1796 free(ptr);
1797 }
1798
1799 pandecode_log("\n");
1800 }
1801
1802 static void
1803 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1804 {
1805 pandecode_validate_buffer(uniforms, uniform_count * 16);
1806
1807 char *ptr = pointer_as_memory_reference(uniforms);
1808 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1809 free(ptr);
1810 }
1811
1812 static const char *
1813 shader_type_for_job(unsigned type)
1814 {
1815 switch (type) {
1816 case JOB_TYPE_VERTEX: return "VERTEX";
1817 case JOB_TYPE_TILER: return "FRAGMENT";
1818 case JOB_TYPE_COMPUTE: return "COMPUTE";
1819 default:
1820 return "UNKNOWN";
1821 }
1822 }
1823
1824 static unsigned shader_id = 0;
1825
1826 static struct midgard_disasm_stats
1827 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1828 bool is_bifrost, unsigned gpu_id)
1829 {
1830 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1831 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1832
1833 /* Compute maximum possible size */
1834 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1835
1836 /* Print some boilerplate to clearly denote the assembly (which doesn't
1837 * obey indentation rules), and actually do the disassembly! */
1838
1839 pandecode_log_cont("\n\n");
1840
1841 struct midgard_disasm_stats stats;
1842
1843 if (is_bifrost) {
1844 disassemble_bifrost(pandecode_dump_stream, code, sz, false);
1845
1846 /* TODO: Extend stats to Bifrost */
1847 stats.texture_count = -128;
1848 stats.sampler_count = -128;
1849 stats.attribute_count = -128;
1850 stats.varying_count = -128;
1851 stats.uniform_count = -128;
1852 stats.uniform_buffer_count = -128;
1853 stats.work_count = -128;
1854
1855 stats.instruction_count = 0;
1856 stats.bundle_count = 0;
1857 stats.quadword_count = 0;
1858 stats.helper_invocations = false;
1859 } else {
1860 stats = disassemble_midgard(pandecode_dump_stream,
1861 code, sz, gpu_id,
1862 type == JOB_TYPE_TILER ?
1863 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
1864 }
1865
1866 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1867 * driver-internal purposes with the blob and interfere */
1868
1869 bool should_shaderdb = type != JOB_TYPE_COMPUTE;
1870
1871 if (should_shaderdb) {
1872 unsigned nr_threads =
1873 (stats.work_count <= 4) ? 4 :
1874 (stats.work_count <= 8) ? 2 :
1875 1;
1876
1877 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
1878 "%u inst, %u bundles, %u quadwords, "
1879 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1880 shader_id++,
1881 shader_type_for_job(type),
1882 stats.instruction_count, stats.bundle_count, stats.quadword_count,
1883 stats.work_count, nr_threads);
1884 }
1885
1886
1887 return stats;
1888 }
1889
1890 static void
1891 pandecode_texture(mali_ptr u,
1892 struct pandecode_mapped_memory *tmem,
1893 unsigned job_no, unsigned tex)
1894 {
1895 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
1896
1897 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
1898 pandecode_indent++;
1899
1900 struct mali_texture_format f = t->format;
1901
1902 /* See the definiton of enum mali_texture_type */
1903
1904 bool is_cube = f.type == MALI_TEX_CUBE;
1905 unsigned dimension = is_cube ? 2 : f.type;
1906
1907 pandecode_make_indent();
1908
1909 /* TODO: Are there others? */
1910 bool is_zs = f.format == MALI_Z32_UNORM;
1911
1912 /* Recall Z/S switched the meaning of linear/tiled .. */
1913 if (is_zs && f.layout == MALI_TEXTURE_LINEAR)
1914 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1915
1916 /* Print the layout. Default is linear; a modifier can denote AFBC or
1917 * u-interleaved/tiled modes */
1918
1919 if (f.layout == MALI_TEXTURE_AFBC)
1920 pandecode_log_cont("afbc");
1921 else if (f.layout == MALI_TEXTURE_TILED)
1922 pandecode_log_cont("tiled");
1923 else if (f.layout == MALI_TEXTURE_LINEAR)
1924 pandecode_log_cont("linear");
1925 else
1926 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
1927
1928 pandecode_swizzle(t->swizzle, f.format);
1929 pandecode_log_cont(" ");
1930
1931 /* Distinguish cube/2D with modifier */
1932
1933 if (is_cube)
1934 pandecode_log_cont("cube ");
1935
1936 pandecode_format_short(f.format, f.srgb);
1937 pandecode_swizzle(f.swizzle, f.format);
1938
1939 /* All four width/height/depth/array_size dimensions are present
1940 * regardless of the type of texture, but it is an error to have
1941 * non-zero dimensions for unused dimensions. Verify this. array_size
1942 * can always be set, as can width. */
1943
1944 if (t->height && dimension < 2)
1945 pandecode_msg("XXX: nonzero height for <2D texture\n");
1946
1947 if (t->depth && dimension < 3)
1948 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1949
1950 /* Print only the dimensions that are actually there */
1951
1952 pandecode_log_cont(": %d", t->width + 1);
1953
1954 if (dimension >= 2)
1955 pandecode_log_cont("x%u", t->height + 1);
1956
1957 if (dimension >= 3)
1958 pandecode_log_cont("x%u", t->depth + 1);
1959
1960 if (t->array_size)
1961 pandecode_log_cont("[%u]", t->array_size + 1);
1962
1963 if (t->levels)
1964 pandecode_log_cont(" mip %u", t->levels);
1965
1966 pandecode_log_cont("\n");
1967
1968 if (f.unknown1 | f.zero) {
1969 pandecode_msg("XXX: texture format zero tripped\n");
1970 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
1971 pandecode_prop("zero = %" PRId32, f.zero);
1972 }
1973
1974 if (!f.unknown2) {
1975 pandecode_msg("XXX: expected unknown texture bit set\n");
1976 pandecode_prop("unknown2 = %" PRId32, f.unknown2);
1977 }
1978
1979 if (t->swizzle_zero) {
1980 pandecode_msg("XXX: swizzle zero tripped\n");
1981 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
1982 }
1983
1984 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
1985 pandecode_msg("XXX: texture zero tripped\n");
1986 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
1987 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
1988 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
1989 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
1990 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
1991 }
1992
1993 pandecode_log(".payload = {\n");
1994 pandecode_indent++;
1995
1996 /* A bunch of bitmap pointers follow.
1997 * We work out the correct number,
1998 * based on the mipmap/cubemap
1999 * properties, but dump extra
2000 * possibilities to futureproof */
2001
2002 int bitmap_count = t->levels + 1;
2003
2004 /* Miptree for each face */
2005 if (f.type == MALI_TEX_CUBE)
2006 bitmap_count *= 6;
2007 else if (f.type == MALI_TEX_3D && f.layout == MALI_TEXTURE_LINEAR)
2008 bitmap_count *= (t->depth + 1);
2009
2010 /* Array of textures */
2011 bitmap_count *= (t->array_size + 1);
2012
2013 /* Stride for each element */
2014 if (f.manual_stride)
2015 bitmap_count *= 2;
2016
2017 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
2018 u + sizeof(*t), sizeof(mali_ptr) * bitmap_count);
2019 for (int i = 0; i < bitmap_count; ++i) {
2020 /* How we dump depends if this is a stride or a pointer */
2021
2022 if (f.manual_stride && (i & 1)) {
2023 /* signed 32-bit snuck in as a 64-bit pointer */
2024 uint64_t stride_set = pointers_and_strides[i];
2025 uint32_t clamped_stride = stride_set;
2026 int32_t stride = clamped_stride;
2027 assert(stride_set == clamped_stride);
2028 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2029 } else {
2030 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
2031 pandecode_log("%s, \n", a);
2032 free(a);
2033 }
2034 }
2035
2036 pandecode_indent--;
2037 pandecode_log("},\n");
2038
2039 pandecode_indent--;
2040 pandecode_log("};\n");
2041 }
2042
2043 /* 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 */
2044
2045 static void
2046 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2047 {
2048 /* Nothing to do */
2049 if (claim == truth)
2050 return;
2051
2052 if (fuzzy)
2053 assert(truth >= 0);
2054
2055 if ((truth >= 0) && !fuzzy) {
2056 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2057 (truth < claim) ? "warn" : "XXX",
2058 name, truth, claim);
2059 } else if ((claim > -truth) && !fuzzy) {
2060 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2061 name, -truth, claim);
2062 } else if (fuzzy && (claim < truth))
2063 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2064 name, truth, claim);
2065
2066 pandecode_log(".%s = %" PRId16, name, claim);
2067
2068 if (fuzzy)
2069 pandecode_log_cont(" /* %u used */", truth);
2070
2071 pandecode_log_cont(",\n");
2072 }
2073
2074 static void
2075 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2076 bool is_bifrost, unsigned gpu_id)
2077 {
2078 struct midgard_disasm_stats stats =
2079 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
2080
2081 bool has_texture = (stats.texture_count > 0);
2082 bool has_sampler = (stats.sampler_count > 0);
2083 bool has_attribute = (stats.attribute_count > 0);
2084 bool has_varying = (stats.varying_count > 0);
2085 bool has_uniform = (stats.uniform_count > 0);
2086 bool has_ubo = (stats.uniform_buffer_count > 0);
2087
2088 if (has_texture || has_sampler)
2089 pandecode_msg("XXX: blend shader accessing textures\n");
2090
2091 if (has_attribute || has_varying)
2092 pandecode_msg("XXX: blend shader accessing interstage\n");
2093
2094 if (has_uniform || has_ubo)
2095 pandecode_msg("XXX: blend shader accessing uniforms\n");
2096 }
2097
2098 static void
2099 pandecode_vertex_tiler_postfix_pre(
2100 const struct mali_vertex_tiler_postfix *p,
2101 int job_no, enum mali_job_type job_type,
2102 char *suffix, bool is_bifrost, unsigned gpu_id)
2103 {
2104 struct pandecode_mapped_memory *attr_mem;
2105
2106 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2107 * are the only things actually needed from the FBD, vertex/tiler jobs
2108 * no longer reference the FBD -- instead, this field points to some
2109 * info about the scratchpad.
2110 */
2111
2112 struct pandecode_fbd fbd_info = {
2113 /* Default for Bifrost */
2114 .rt_count = 1
2115 };
2116
2117 if (is_bifrost) {
2118 pandecode_log_cont("\t/* %X %/\n", p->shared_memory & 1);
2119 pandecode_compute_fbd(p->shared_memory & ~1, job_no);
2120 } else if (p->shared_memory & MALI_MFBD)
2121 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->shared_memory) & FBD_MASK, job_no, false, job_type == JOB_TYPE_COMPUTE);
2122 else if (job_type == JOB_TYPE_COMPUTE)
2123 pandecode_compute_fbd((u64) (uintptr_t) p->shared_memory, job_no);
2124 else
2125 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->shared_memory, job_no, false, gpu_id);
2126
2127 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2128 int texture_count = 0, sampler_count = 0;
2129
2130 if (p->shader) {
2131 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2132 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2133
2134 /* Disassemble ahead-of-time to get stats. Initialize with
2135 * stats for the missing-shader case so we get validation
2136 * there, too */
2137
2138 struct midgard_disasm_stats info = {
2139 .texture_count = 0,
2140 .sampler_count = 0,
2141 .attribute_count = 0,
2142 .varying_count = 0,
2143 .work_count = 1,
2144
2145 .uniform_count = -128,
2146 .uniform_buffer_count = 0
2147 };
2148
2149 if (s->shader & ~0xF)
2150 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
2151
2152 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2153 pandecode_indent++;
2154
2155 /* Save for dumps */
2156 attribute_count = s->attribute_count;
2157 varying_count = s->varying_count;
2158 texture_count = s->texture_count;
2159 sampler_count = s->sampler_count;
2160
2161 if (is_bifrost) {
2162 uniform_count = s->bifrost2.uniform_count;
2163 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2164 } else {
2165 uniform_count = s->midgard1.uniform_count;
2166 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2167 }
2168
2169 pandecode_shader_address("shader", s->shader);
2170
2171 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2172 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2173 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2174 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2175 pandecode_shader_prop("uniform_buffer_count",
2176 uniform_buffer_count,
2177 info.uniform_buffer_count, true);
2178
2179 if (!is_bifrost) {
2180 pandecode_shader_prop("uniform_count",
2181 uniform_count,
2182 info.uniform_count, false);
2183
2184 pandecode_shader_prop("work_count",
2185 s->midgard1.work_count, info.work_count, false);
2186 }
2187
2188 if (is_bifrost) {
2189 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32, s->bifrost1.unk1);
2190 } else {
2191 bool helpers = s->midgard1.flags_lo & MALI_HELPER_INVOCATIONS;
2192 s->midgard1.flags_lo &= ~MALI_HELPER_INVOCATIONS;
2193
2194 if (helpers != info.helper_invocations) {
2195 pandecode_msg("XXX: expected helpers %u but got %u\n",
2196 info.helper_invocations, helpers);
2197 }
2198
2199 pandecode_log(".midgard1.flags_lo = ");
2200 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info, s->midgard1.flags_lo);
2201 pandecode_log_cont(",\n");
2202
2203 pandecode_log(".midgard1.flags_hi = ");
2204 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info, s->midgard1.flags_hi);
2205 pandecode_log_cont(",\n");
2206 }
2207
2208 if (s->depth_units || s->depth_factor) {
2209 pandecode_prop("depth_factor = %f", s->depth_factor);
2210 pandecode_prop("depth_units = %f", s->depth_units);
2211 }
2212
2213 if (s->alpha_coverage) {
2214 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
2215 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
2216
2217 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2218 invert_alpha_coverage ? "~" : "",
2219 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
2220 }
2221
2222 if (s->unknown2_3 || s->unknown2_4) {
2223 pandecode_log(".unknown2_3 = ");
2224
2225 int unknown2_3 = s->unknown2_3;
2226 int unknown2_4 = s->unknown2_4;
2227
2228 /* We're not quite sure what these flags mean without the depth test, if anything */
2229
2230 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2231 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2232 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2233
2234 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2235 }
2236
2237 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2238 pandecode_log_cont(",\n");
2239
2240 pandecode_log(".unknown2_4 = ");
2241 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2242 pandecode_log_cont(",\n");
2243 }
2244
2245 if (s->stencil_mask_front || s->stencil_mask_back) {
2246 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2247 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2248 }
2249
2250 pandecode_stencil("front", &s->stencil_front);
2251 pandecode_stencil("back", &s->stencil_back);
2252
2253 if (is_bifrost) {
2254 pandecode_log(".bifrost2 = {\n");
2255 pandecode_indent++;
2256
2257 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2258 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2259 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2260 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2261
2262 pandecode_indent--;
2263 pandecode_log("},\n");
2264 } else if (s->midgard2.unknown2_7) {
2265 pandecode_log(".midgard2 = {\n");
2266 pandecode_indent++;
2267
2268 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2269 pandecode_indent--;
2270 pandecode_log("},\n");
2271 }
2272
2273 if (s->padding) {
2274 pandecode_msg("XXX: shader padding tripped\n");
2275 pandecode_prop("padding = 0x%" PRIx32, s->padding);
2276 }
2277
2278 if (!is_bifrost) {
2279 /* TODO: Blend shaders routing/disasm */
2280 union midgard_blend blend = s->blend;
2281 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2282 if (shader & ~0xF)
2283 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2284 }
2285
2286 pandecode_indent--;
2287 pandecode_log("};\n");
2288
2289 /* MRT blend fields are used whenever MFBD is used, with
2290 * per-RT descriptors */
2291
2292 if (job_type == JOB_TYPE_TILER && p->shared_memory & MALI_MFBD) {
2293 void* blend_base = (void *) (s + 1);
2294
2295 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2296 mali_ptr shader = 0;
2297
2298 if (is_bifrost)
2299 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2300 else
2301 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2302
2303 if (shader & ~0xF)
2304 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2305
2306 }
2307 }
2308 } else
2309 pandecode_msg("XXX: missing shader descriptor\n");
2310
2311 if (p->viewport) {
2312 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2313 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2314
2315 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2316 pandecode_indent++;
2317
2318 pandecode_prop("clip_minx = %f", f->clip_minx);
2319 pandecode_prop("clip_miny = %f", f->clip_miny);
2320 pandecode_prop("clip_minz = %f", f->clip_minz);
2321 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2322 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2323 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2324
2325 /* Only the higher coordinates are MALI_POSITIVE scaled */
2326
2327 pandecode_prop("viewport0 = { %d, %d }",
2328 f->viewport0[0], f->viewport0[1]);
2329
2330 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2331 f->viewport1[0] + 1, f->viewport1[1] + 1);
2332
2333 pandecode_indent--;
2334 pandecode_log("};\n");
2335 }
2336
2337 unsigned max_attr_index = 0;
2338
2339 if (p->attribute_meta)
2340 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2341
2342 if (p->attributes) {
2343 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2344 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2345 }
2346
2347 /* Varyings are encoded like attributes but not actually sent; we just
2348 * pass a zero buffer with the right stride/size set, (or whatever)
2349 * since the GPU will write to it itself */
2350
2351 if (p->varying_meta) {
2352 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2353 }
2354
2355 if (p->varyings) {
2356 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2357
2358 /* Number of descriptors depends on whether there are
2359 * non-internal varyings */
2360
2361 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2362 }
2363
2364 if (p->uniform_buffers) {
2365 if (uniform_buffer_count)
2366 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2367 else
2368 pandecode_msg("warn: UBOs specified but not referenced\n");
2369 } else if (uniform_buffer_count)
2370 pandecode_msg("XXX: UBOs referenced but not specified\n");
2371
2372 /* We don't want to actually dump uniforms, but we do need to validate
2373 * that the counts we were given are sane */
2374
2375 if (p->uniforms) {
2376 if (uniform_count)
2377 pandecode_uniforms(p->uniforms, uniform_count);
2378 else
2379 pandecode_msg("warn: Uniforms specified but not referenced\n");
2380 } else if (uniform_count)
2381 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2382
2383 if (p->texture_trampoline) {
2384 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
2385
2386 if (mmem) {
2387 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
2388
2389 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
2390 pandecode_indent++;
2391
2392 for (int tex = 0; tex < texture_count; ++tex) {
2393 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2394 char *a = pointer_as_memory_reference(*u);
2395 pandecode_log("%s,\n", a);
2396 free(a);
2397 }
2398
2399 pandecode_indent--;
2400 pandecode_log("};\n");
2401
2402 /* Now, finally, descend down into the texture descriptor */
2403 for (unsigned tex = 0; tex < texture_count; ++tex) {
2404 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2405 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2406 if (tmem)
2407 pandecode_texture(*u, tmem, job_no, tex);
2408 }
2409 }
2410 }
2411
2412 if (p->sampler_descriptor) {
2413 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2414
2415 if (smem) {
2416 struct mali_sampler_descriptor *s;
2417
2418 mali_ptr d = p->sampler_descriptor;
2419
2420 for (int i = 0; i < sampler_count; ++i) {
2421 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2422
2423 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2424 pandecode_indent++;
2425
2426 pandecode_log(".filter_mode = ");
2427 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2428 pandecode_log_cont(",\n");
2429
2430 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2431 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2432
2433 if (s->lod_bias)
2434 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s->lod_bias));
2435
2436 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2437 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2438 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2439
2440 pandecode_prop("compare_func = %s", pandecode_func(s->compare_func));
2441
2442 if (s->zero || s->zero2) {
2443 pandecode_msg("XXX: sampler zero tripped\n");
2444 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2445 }
2446
2447 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2448
2449 pandecode_prop("border_color = { %f, %f, %f, %f }",
2450 s->border_color[0],
2451 s->border_color[1],
2452 s->border_color[2],
2453 s->border_color[3]);
2454
2455 pandecode_indent--;
2456 pandecode_log("};\n");
2457 }
2458 }
2459 }
2460 }
2461
2462 static void
2463 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2464 {
2465 if (p->shader & 0xF)
2466 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2467
2468 if (!(p->position_varying || p->occlusion_counter))
2469 return;
2470
2471 pandecode_log(".postfix = {\n");
2472 pandecode_indent++;
2473
2474 MEMORY_PROP(p, position_varying);
2475 MEMORY_PROP(p, occlusion_counter);
2476
2477 pandecode_indent--;
2478 pandecode_log("},\n");
2479 }
2480
2481 static void
2482 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2483 {
2484 pandecode_log_cont("{\n");
2485 pandecode_indent++;
2486
2487 pandecode_prop("unk2 = 0x%x", v->unk2);
2488
2489 if (v->zero0 || v->zero1) {
2490 pandecode_msg("XXX: vertex only zero tripped");
2491 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2492 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2493 }
2494
2495 pandecode_indent--;
2496 pandecode_log("}\n");
2497 }
2498
2499 static void
2500 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2501 {
2502
2503 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2504 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2505
2506 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2507 pandecode_indent++;
2508
2509 if (h->zero) {
2510 pandecode_msg("XXX: tiler heap zero tripped\n");
2511 pandecode_prop("zero = 0x%x", h->zero);
2512 }
2513
2514 for (int i = 0; i < 12; i++) {
2515 if (h->zeros[i] != 0) {
2516 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2517 i, h->zeros[i]);
2518 }
2519 }
2520
2521 pandecode_prop("heap_size = 0x%x", h->heap_size);
2522 MEMORY_PROP(h, tiler_heap_start);
2523 MEMORY_PROP(h, tiler_heap_free);
2524
2525 /* this might point to the beginning of another buffer, when it's
2526 * really the end of the tiler heap buffer, so we have to be careful
2527 * here. but for zero length, we need the same pointer.
2528 */
2529
2530 if (h->tiler_heap_end == h->tiler_heap_start) {
2531 MEMORY_PROP(h, tiler_heap_start);
2532 } else {
2533 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2534 pandecode_prop("tiler_heap_end = %s + 1", a);
2535 free(a);
2536 }
2537
2538 pandecode_indent--;
2539 pandecode_log("};\n");
2540 }
2541
2542 static void
2543 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2544 {
2545 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2546 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2547
2548 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2549
2550 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2551 pandecode_indent++;
2552
2553 if (t->zero0 || t->zero1) {
2554 pandecode_msg("XXX: tiler meta zero tripped\n");
2555 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2556 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2557 }
2558
2559 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2560 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2561
2562 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2563 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2564
2565 for (int i = 0; i < 12; i++) {
2566 if (t->zeros[i] != 0) {
2567 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2568 i, t->zeros[i]);
2569 }
2570 }
2571
2572 pandecode_indent--;
2573 pandecode_log("};\n");
2574 }
2575
2576 static void
2577 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2578 {
2579 pandecode_log(".gl_enables = ");
2580
2581 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2582
2583 pandecode_log_cont(",\n");
2584 }
2585
2586 static void
2587 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2588 {
2589 if (u.pointer == 0x0)
2590 return;
2591
2592 pandecode_log(".primitive_size = {\n");
2593 pandecode_indent++;
2594
2595 if (constant) {
2596 pandecode_prop("constant = %f", u.constant);
2597 } else {
2598 MEMORY_PROP((&u), pointer);
2599 }
2600
2601 pandecode_indent--;
2602 pandecode_log("},\n");
2603 }
2604
2605 static void
2606 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2607 {
2608 pandecode_log_cont("{\n");
2609 pandecode_indent++;
2610
2611 /* TODO: gl_PointSize on Bifrost */
2612 pandecode_primitive_size(t->primitive_size, true);
2613
2614 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2615
2616 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2617 || t->zero6 || t->zero7 || t->zero8) {
2618 pandecode_msg("XXX: tiler only zero tripped\n");
2619 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2620 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2621 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2622 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2623 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2624 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2625 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2626 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2627 }
2628
2629 pandecode_indent--;
2630 pandecode_log("},\n");
2631 }
2632
2633 static int
2634 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2635 const struct pandecode_mapped_memory *mem,
2636 mali_ptr payload, int job_no, unsigned gpu_id)
2637 {
2638 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2639
2640 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
2641
2642 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2643 pandecode_indent++;
2644
2645 pandecode_log(".prefix = ");
2646 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2647
2648 pandecode_log(".vertex = ");
2649 pandecode_vertex_only_bfr(&v->vertex);
2650
2651 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2652
2653 pandecode_indent--;
2654 pandecode_log("};\n");
2655
2656 return sizeof(*v);
2657 }
2658
2659 static int
2660 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2661 const struct pandecode_mapped_memory *mem,
2662 mali_ptr payload, int job_no, unsigned gpu_id)
2663 {
2664 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2665
2666 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
2667 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2668
2669 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2670 pandecode_indent++;
2671
2672 pandecode_log(".prefix = ");
2673 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2674
2675 pandecode_log(".tiler = ");
2676 pandecode_tiler_only_bfr(&t->tiler, job_no);
2677
2678 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2679
2680 pandecode_indent--;
2681 pandecode_log("};\n");
2682
2683 return sizeof(*t);
2684 }
2685
2686 static int
2687 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2688 const struct pandecode_mapped_memory *mem,
2689 mali_ptr payload, int job_no, unsigned gpu_id)
2690 {
2691 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2692
2693 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
2694
2695 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2696 pandecode_indent++;
2697
2698 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2699 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2700
2701 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2702
2703 pandecode_log(".prefix = ");
2704 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
2705
2706 pandecode_gl_enables(v->gl_enables, h->job_type);
2707
2708 if (v->instance_shift || v->instance_odd) {
2709 pandecode_prop("instance_shift = 0x%d /* %d */",
2710 v->instance_shift, 1 << v->instance_shift);
2711 pandecode_prop("instance_odd = 0x%X /* %d */",
2712 v->instance_odd, (2 * v->instance_odd) + 1);
2713
2714 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2715 }
2716
2717 if (v->offset_start)
2718 pandecode_prop("offset_start = %d", v->offset_start);
2719
2720 if (v->zero5) {
2721 pandecode_msg("XXX: midgard payload zero tripped\n");
2722 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2723 }
2724
2725 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2726
2727 pandecode_indent--;
2728 pandecode_log("};\n");
2729
2730 return sizeof(*v);
2731 }
2732
2733 static int
2734 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2735 mali_ptr payload, int job_no,
2736 bool is_bifrost, unsigned gpu_id)
2737 {
2738 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2739
2740 bool is_mfbd = s->framebuffer & MALI_MFBD;
2741
2742 if (!is_mfbd && is_bifrost)
2743 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2744
2745 struct pandecode_fbd info;
2746
2747 if (is_mfbd)
2748 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true, false);
2749 else
2750 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
2751
2752 /* Compute the tag for the tagged pointer. This contains the type of
2753 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2754 * additional structures follow the MFBD header (an extra payload or
2755 * not, as well as a count of render targets) */
2756
2757 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
2758
2759 if (is_mfbd) {
2760 if (info.has_extra)
2761 expected_tag |= MALI_MFBD_TAG_EXTRA;
2762
2763 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
2764 }
2765
2766 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
2767 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2768 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
2769 pandecode_prop("max_tile_coord = 0x%X\n", s->min_tile_coord);
2770 }
2771
2772 /* Extract tile coordinates */
2773
2774 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
2775 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
2776
2777 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2778 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2779
2780 /* For the max, we also want the floored (rather than ceiled) version for checking */
2781
2782 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
2783 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
2784
2785 /* Validate the coordinates are well-ordered */
2786
2787 if (min_x == max_x)
2788 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2789 else if (min_x > max_x)
2790 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2791
2792 if (min_y == max_y)
2793 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2794 else if (min_y > max_y)
2795 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2796
2797 /* Validate the coordinates fit inside the framebuffer. We use floor,
2798 * rather than ceil, for the max coordinates, since the tile
2799 * coordinates for something like an 800x600 framebuffer will actually
2800 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2801
2802 if ((min_x > info.width) || (max_x_f > info.width))
2803 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2804
2805 if ((min_y > info.height) || (max_y_f > info.height))
2806 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2807
2808 /* After validation, we print */
2809
2810 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
2811
2812 /* The FBD is a tagged pointer */
2813
2814 unsigned tag = (s->framebuffer & ~FBD_MASK);
2815
2816 if (tag != expected_tag)
2817 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
2818
2819 return sizeof(*s);
2820 }
2821
2822 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
2823 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
2824 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
2825 * might wish to combine this with the bifrost parameter) because some details
2826 * are model-specific even within a particular architecture. Minimal traces
2827 * *only* examine the job descriptors, skipping printing entirely if there is
2828 * no faults, and only descends into the payload if there are faults. This is
2829 * useful for looking for faults without the overhead of invasive traces. */
2830
2831 void
2832 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id, bool minimal)
2833 {
2834 struct mali_job_descriptor_header *h;
2835 unsigned job_descriptor_number = 0;
2836
2837 do {
2838 struct pandecode_mapped_memory *mem =
2839 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2840
2841 void *payload;
2842
2843 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2844
2845 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2846 * high 32-bits of the 64-bit pointer are reused to store
2847 * something else.
2848 */
2849 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2850 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2851 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2852
2853 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
2854
2855 int job_no = job_descriptor_number++;
2856
2857 /* If the job is good to go, skip it in minimal mode */
2858 if (minimal && (h->exception_status == 0x0 || h->exception_status == 0x1))
2859 continue;
2860
2861 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2862 pandecode_indent++;
2863
2864 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2865
2866 if (h->job_descriptor_size)
2867 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2868
2869 if (h->exception_status && h->exception_status != 0x1)
2870 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2871 h->exception_status,
2872 (h->exception_status >> 16) & 0xFFFF,
2873 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2874 h->exception_status & 0xFF);
2875
2876 if (h->first_incomplete_task)
2877 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2878
2879 if (h->fault_pointer)
2880 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2881
2882 if (h->job_barrier)
2883 pandecode_prop("job_barrier = %d", h->job_barrier);
2884
2885 pandecode_prop("job_index = %d", h->job_index);
2886
2887 if (h->unknown_flags)
2888 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2889
2890 if (h->job_dependency_index_1)
2891 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2892
2893 if (h->job_dependency_index_2)
2894 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2895
2896 pandecode_indent--;
2897 pandecode_log("};\n");
2898
2899 switch (h->job_type) {
2900 case JOB_TYPE_WRITE_VALUE: {
2901 struct mali_payload_write_value *s = payload;
2902 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2903 pandecode_indent++;
2904 MEMORY_PROP(s, address);
2905
2906 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
2907 pandecode_msg("XXX: unknown value descriptor\n");
2908 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
2909 }
2910
2911 if (s->reserved) {
2912 pandecode_msg("XXX: set value tripped\n");
2913 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
2914 }
2915
2916 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
2917 pandecode_indent--;
2918 pandecode_log("};\n");
2919
2920 break;
2921 }
2922
2923 case JOB_TYPE_TILER:
2924 case JOB_TYPE_VERTEX:
2925 case JOB_TYPE_COMPUTE:
2926 if (bifrost) {
2927 if (h->job_type == JOB_TYPE_TILER)
2928 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2929 else
2930 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2931 } else
2932 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
2933
2934 break;
2935
2936 case JOB_TYPE_FRAGMENT:
2937 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
2938 break;
2939
2940 default:
2941 break;
2942 }
2943 } while ((jc_gpu_va = h->next_job));
2944 }