projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
panfrost: Document Midgard Inf/NaN suppress bit
[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 FLAG_INFO(SUPPRESS_INF_NAN),
277 {}
278 };
279 #undef FLAG_INFO
280
281 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
282 static const struct pandecode_flag_info mfbd_flag_info [] = {
283 FLAG_INFO(DEPTH_WRITE),
284 FLAG_INFO(EXTRA),
285 {}
286 };
287 #undef FLAG_INFO
288
289 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
290 static const struct pandecode_flag_info sampler_flag_info [] = {
291 FLAG_INFO(MAG_NEAREST),
292 FLAG_INFO(MIN_NEAREST),
293 FLAG_INFO(MIP_LINEAR_1),
294 FLAG_INFO(MIP_LINEAR_2),
295 FLAG_INFO(NORM_COORDS),
296 {}
297 };
298 #undef FLAG_INFO
299
300 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
301 static const struct pandecode_flag_info sfbd_unk1_info [] = {
302 FLAG_INFO(MSAA_8),
303 FLAG_INFO(MSAA_A),
304 {}
305 };
306 #undef FLAG_INFO
307
308 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
309 static const struct pandecode_flag_info sfbd_unk2_info [] = {
310 FLAG_INFO(MSAA_B),
311 FLAG_INFO(SRGB),
312 {}
313 };
314 #undef FLAG_INFO
315
316 extern char *replace_fragment;
317 extern char *replace_vertex;
318
319 static char *
320 pandecode_job_type(enum mali_job_type type)
321 {
322 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
323
324 switch (type) {
325 DEFINE_CASE(NULL);
326 DEFINE_CASE(WRITE_VALUE);
327 DEFINE_CASE(CACHE_FLUSH);
328 DEFINE_CASE(COMPUTE);
329 DEFINE_CASE(VERTEX);
330 DEFINE_CASE(TILER);
331 DEFINE_CASE(FUSED);
332 DEFINE_CASE(FRAGMENT);
333
334 case JOB_NOT_STARTED:
335 return "NOT_STARTED";
336
337 default:
338 pandecode_log("Warning! Unknown job type %x\n", type);
339 return "!?!?!?";
340 }
341
342 #undef DEFINE_CASE
343 }
344
345 static char *
346 pandecode_draw_mode(enum mali_draw_mode mode)
347 {
348 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
349
350 switch (mode) {
351 DEFINE_CASE(DRAW_NONE);
352 DEFINE_CASE(POINTS);
353 DEFINE_CASE(LINES);
354 DEFINE_CASE(TRIANGLES);
355 DEFINE_CASE(TRIANGLE_STRIP);
356 DEFINE_CASE(TRIANGLE_FAN);
357 DEFINE_CASE(LINE_STRIP);
358 DEFINE_CASE(LINE_LOOP);
359 DEFINE_CASE(POLYGON);
360 DEFINE_CASE(QUADS);
361 DEFINE_CASE(QUAD_STRIP);
362
363 default:
364 pandecode_msg("XXX: invalid draw mode %X\n", mode);
365 return "";
366 }
367
368 #undef DEFINE_CASE
369 }
370
371 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
372 static char *
373 pandecode_func(enum mali_func mode)
374 {
375 switch (mode) {
376 DEFINE_CASE(NEVER);
377 DEFINE_CASE(LESS);
378 DEFINE_CASE(EQUAL);
379 DEFINE_CASE(LEQUAL);
380 DEFINE_CASE(GREATER);
381 DEFINE_CASE(NOTEQUAL);
382 DEFINE_CASE(GEQUAL);
383 DEFINE_CASE(ALWAYS);
384
385 default:
386 pandecode_msg("XXX: invalid func %X\n", mode);
387 return "";
388 }
389 }
390 #undef DEFINE_CASE
391
392 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
393 static char *
394 pandecode_stencil_op(enum mali_stencil_op op)
395 {
396 switch (op) {
397 DEFINE_CASE(KEEP);
398 DEFINE_CASE(REPLACE);
399 DEFINE_CASE(ZERO);
400 DEFINE_CASE(INVERT);
401 DEFINE_CASE(INCR_WRAP);
402 DEFINE_CASE(DECR_WRAP);
403 DEFINE_CASE(INCR);
404 DEFINE_CASE(DECR);
405
406 default:
407 pandecode_msg("XXX: invalid stencil op %X\n", op);
408 return "";
409 }
410 }
411
412 #undef DEFINE_CASE
413
414 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
415 {
416 switch(mode) {
417 /* TODO: Combine to just "instanced" once this can be done
418 * unambiguously in all known cases */
419 case MALI_ATTR_POT_DIVIDE:
420 return "instanced_pot";
421 case MALI_ATTR_MODULO:
422 return "instanced_mod";
423 case MALI_ATTR_NPOT_DIVIDE:
424 return "instanced_npot";
425 case MALI_ATTR_IMAGE:
426 return "image";
427 default:
428 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
429 return "";
430 }
431 }
432
433 static const char *
434 pandecode_special_record(uint64_t v, bool* attribute)
435 {
436 switch(v) {
437 case MALI_ATTR_VERTEXID:
438 *attribute = true;
439 return "gl_VertexID";
440 case MALI_ATTR_INSTANCEID:
441 *attribute = true;
442 return "gl_InstanceID";
443 case MALI_VARYING_FRAG_COORD:
444 return "gl_FragCoord";
445 case MALI_VARYING_FRONT_FACING:
446 return "gl_FrontFacing";
447 case MALI_VARYING_POINT_COORD:
448 return "gl_PointCoord";
449 default:
450 pandecode_msg("XXX: invalid special record %" PRIx64 "\n", v);
451 return "";
452 }
453 }
454
455 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
456 static char *
457 pandecode_wrap_mode(enum mali_wrap_mode op)
458 {
459 switch (op) {
460 DEFINE_CASE(REPEAT);
461 DEFINE_CASE(CLAMP_TO_EDGE);
462 DEFINE_CASE(CLAMP_TO_BORDER);
463 DEFINE_CASE(MIRRORED_REPEAT);
464
465 default:
466 pandecode_msg("XXX: invalid wrap mode %X\n", op);
467 return "";
468 }
469 }
470 #undef DEFINE_CASE
471
472 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
473 static char *
474 pandecode_block_format(enum mali_block_format fmt)
475 {
476 switch (fmt) {
477 DEFINE_CASE(TILED);
478 DEFINE_CASE(UNKNOWN);
479 DEFINE_CASE(LINEAR);
480 DEFINE_CASE(AFBC);
481
482 default:
483 unreachable("Invalid case");
484 }
485 }
486 #undef DEFINE_CASE
487
488 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
489 char *
490 pandecode_exception_access(unsigned access)
491 {
492 switch (access) {
493 DEFINE_CASE(NONE);
494 DEFINE_CASE(EXECUTE);
495 DEFINE_CASE(READ);
496 DEFINE_CASE(WRITE);
497
498 default:
499 unreachable("Invalid case");
500 }
501 }
502 #undef DEFINE_CASE
503
504 /* Midgard's tiler descriptor is embedded within the
505 * larger FBD */
506
507 static void
508 pandecode_midgard_tiler_descriptor(
509 const struct midgard_tiler_descriptor *t,
510 unsigned width,
511 unsigned height,
512 bool is_fragment,
513 bool has_hierarchy)
514 {
515 pandecode_log(".tiler = {\n");
516 pandecode_indent++;
517
518 if (t->hierarchy_mask == MALI_TILER_DISABLED)
519 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
520 else
521 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
522
523 /* We know this name from the kernel, but we never see it nonzero */
524
525 if (t->flags)
526 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
527
528 MEMORY_PROP(t, polygon_list);
529
530 /* The body is offset from the base of the polygon list */
531 //assert(t->polygon_list_body > t->polygon_list);
532 unsigned body_offset = t->polygon_list_body - t->polygon_list;
533
534 /* It needs to fit inside the reported size */
535 //assert(t->polygon_list_size >= body_offset);
536
537 /* Now that we've sanity checked, we'll try to calculate the sizes
538 * ourselves for comparison */
539
540 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
541 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
542
543 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
544 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
545 ref_header, ref_size);
546 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
547 pandecode_msg("body offset %d\n", body_offset);
548 }
549
550 /* The tiler heap has a start and end specified -- it should be
551 * identical to what we have in the BO. The exception is if tiling is
552 * disabled. */
553
554 MEMORY_PROP(t, heap_start);
555 assert(t->heap_end >= t->heap_start);
556
557 struct pandecode_mapped_memory *heap =
558 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
559
560 unsigned heap_size = t->heap_end - t->heap_start;
561
562 /* Tiling is enabled with a special flag */
563 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
564 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
565
566 bool tiling_enabled = hierarchy_mask;
567
568 if (tiling_enabled) {
569 /* When tiling is enabled, the heap should be a tight fit */
570 unsigned heap_offset = t->heap_start - heap->gpu_va;
571 if ((heap_offset + heap_size) != heap->length) {
572 pandecode_msg("XXX: heap size %u (expected %zu)\n",
573 heap_size, heap->length - heap_offset);
574 }
575
576 /* We should also have no other flags */
577 if (tiler_flags)
578 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
579 } else {
580 /* When tiling is disabled, we should have that flag and no others */
581
582 if (tiler_flags != MALI_TILER_DISABLED) {
583 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
584 tiler_flags);
585 }
586
587 /* We should also have an empty heap */
588 if (heap_size) {
589 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
590 heap_size);
591 }
592
593 /* Disabled tiling is used only for clear-only jobs, which are
594 * purely FRAGMENT, so we should never see this for
595 * non-FRAGMENT descriptors. */
596
597 if (!is_fragment)
598 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
599 }
600
601 /* We've never seen weights used in practice, but we know from the
602 * kernel these fields is there */
603
604 bool nonzero_weights = false;
605
606 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
607 nonzero_weights |= t->weights[w] != 0x0;
608 }
609
610 if (nonzero_weights) {
611 pandecode_log(".weights = { ");
612
613 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
614 pandecode_log_cont("%d, ", t->weights[w]);
615 }
616
617 pandecode_log("},");
618 }
619
620 pandecode_indent--;
621 pandecode_log("}\n");
622 }
623
624 /* TODO: The Bifrost tiler is not understood at all yet */
625
626 static void
627 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer *fb)
628 {
629 pandecode_log(".tiler = {\n");
630 pandecode_indent++;
631
632 MEMORY_PROP(fb, tiler_meta);
633
634 for (int i = 0; i < 16; i++) {
635 if (fb->zeros[i] != 0) {
636 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
637 i, fb->zeros[i]);
638 }
639 }
640
641 pandecode_log("},\n");
642
643 pandecode_indent--;
644 pandecode_log("}\n");
645
646 }
647
648 /* Information about the framebuffer passed back for
649 * additional analysis */
650
651 struct pandecode_fbd {
652 unsigned width;
653 unsigned height;
654 unsigned rt_count;
655 bool has_extra;
656 };
657
658 static void
659 pandecode_sfbd_format(struct mali_sfbd_format format)
660 {
661 pandecode_log(".format = {\n");
662 pandecode_indent++;
663
664 pandecode_log(".unk1 = ");
665 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
666 pandecode_log_cont(",\n");
667
668 /* TODO: Map formats so we can check swizzles and print nicely */
669 pandecode_log("swizzle");
670 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
671 pandecode_log_cont(",\n");
672
673 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
674 (format.nr_channels + 1));
675
676 pandecode_log(".unk2 = ");
677 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
678 pandecode_log_cont(",\n");
679
680 pandecode_prop("block = %s", pandecode_block_format(format.block));
681
682 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
683
684 pandecode_indent--;
685 pandecode_log("},\n");
686 }
687
688 static void
689 pandecode_shared_memory(const struct mali_shared_memory *desc, bool is_compute)
690 {
691 pandecode_prop("stack_shift = 0x%x", desc->stack_shift);
692
693 if (desc->unk0)
694 pandecode_prop("unk0 = 0x%x", desc->unk0);
695
696 if (desc->shared_workgroup_count != 0x1F) {
697 pandecode_prop("shared_workgroup_count = %d", desc->shared_workgroup_count);
698 if (!is_compute)
699 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
700 }
701
702 if (desc->shared_unk1 || desc->shared_shift) {
703 pandecode_prop("shared_unk1 = %X", desc->shared_unk1);
704 pandecode_prop("shared_shift = %X", desc->shared_shift);
705
706 if (!is_compute)
707 pandecode_msg("XXX: shared memory configured in noncompute shader");
708 }
709
710 if (desc->shared_zero) {
711 pandecode_msg("XXX: shared memory zero tripped\n");
712 pandecode_prop("shared_zero = 0x%" PRIx32, desc->shared_zero);
713 }
714
715 if (desc->shared_memory && !is_compute)
716 pandecode_msg("XXX: shared memory used in noncompute shader\n");
717
718 MEMORY_PROP(desc, scratchpad);
719 MEMORY_PROP(desc, shared_memory);
720 MEMORY_PROP(desc, unknown1);
721
722 if (desc->scratchpad) {
723 struct pandecode_mapped_memory *smem =
724 pandecode_find_mapped_gpu_mem_containing(desc->scratchpad);
725
726 pandecode_msg("scratchpad size %u\n", smem->length);
727 }
728
729 }
730
731 static struct pandecode_fbd
732 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
733 {
734 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
735 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
736
737 struct pandecode_fbd info = {
738 .has_extra = false,
739 .rt_count = 1
740 };
741
742 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
743 pandecode_indent++;
744
745 pandecode_log(".shared_memory = {\n");
746 pandecode_indent++;
747 pandecode_shared_memory(&s->shared_memory, false);
748 pandecode_indent--;
749 pandecode_log("},\n");
750
751 pandecode_sfbd_format(s->format);
752
753 info.width = s->width + 1;
754 info.height = s->height + 1;
755
756 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
757 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
758
759 MEMORY_PROP(s, checksum);
760
761 if (s->checksum_stride)
762 pandecode_prop("checksum_stride = %d", s->checksum_stride);
763
764 MEMORY_PROP(s, framebuffer);
765 pandecode_prop("stride = %d", s->stride);
766
767 /* Earlier in the actual commandstream -- right before width -- but we
768 * delay to flow nicer */
769
770 pandecode_log(".clear_flags = ");
771 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
772 pandecode_log_cont(",\n");
773
774 if (s->depth_buffer) {
775 MEMORY_PROP(s, depth_buffer);
776 pandecode_prop("depth_stride = %d", s->depth_stride);
777 }
778
779 if (s->stencil_buffer) {
780 MEMORY_PROP(s, stencil_buffer);
781 pandecode_prop("stencil_stride = %d", s->stencil_stride);
782 }
783
784 if (s->depth_stride_zero ||
785 s->stencil_stride_zero ||
786 s->zero7 || s->zero8) {
787 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
788 pandecode_prop("depth_stride_zero = 0x%x",
789 s->depth_stride_zero);
790 pandecode_prop("stencil_stride_zero = 0x%x",
791 s->stencil_stride_zero);
792 pandecode_prop("zero7 = 0x%" PRIx32,
793 s->zero7);
794 pandecode_prop("zero8 = 0x%" PRIx32,
795 s->zero8);
796 }
797
798 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
799 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
800 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
801 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
802 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
803 }
804
805 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
806 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
807 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
808 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
809 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
810 }
811
812 if (s->clear_stencil) {
813 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
814 }
815
816 const struct midgard_tiler_descriptor t = s->tiler;
817
818 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
819 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
820
821 pandecode_indent--;
822 pandecode_log("};\n");
823
824 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
825 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
826 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
827
828 pandecode_log_cont(".zero3 = {");
829
830 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
831 pandecode_log_cont("%X, ", s->zero3[i]);
832
833 pandecode_log_cont("},\n");
834
835 pandecode_log_cont(".zero6 = {");
836
837 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
838 pandecode_log_cont("%X, ", s->zero6[i]);
839
840 pandecode_log_cont("},\n");
841
842 return info;
843 }
844
845 static void
846 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
847 {
848 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
849 const struct mali_shared_memory *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
850
851 pandecode_log("struct mali_shared_memory shared_%"PRIx64"_%d = {\n", gpu_va, job_no);
852 pandecode_indent++;
853 pandecode_shared_memory(s, true);
854 pandecode_indent--;
855 pandecode_log("},\n");
856 }
857
858 /* Extracts the number of components associated with a Mali format */
859
860 static unsigned
861 pandecode_format_component_count(enum mali_format fmt)
862 {
863 /* Mask out the format class */
864 unsigned top = fmt & 0b11100000;
865
866 switch (top) {
867 case MALI_FORMAT_SNORM:
868 case MALI_FORMAT_UINT:
869 case MALI_FORMAT_UNORM:
870 case MALI_FORMAT_SINT:
871 return ((fmt >> 3) & 3) + 1;
872 default:
873 /* TODO: Validate */
874 return 4;
875 }
876 }
877
878 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
879
880 static unsigned
881 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
882 {
883 unsigned mask = 0;
884 assert(MALI_CHANNEL_RED == 0);
885
886 for (unsigned c = 0; c < 4; ++c) {
887 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
888
889 if (chan <= MALI_CHANNEL_ALPHA)
890 mask |= (1 << chan);
891 }
892
893 return mask;
894 }
895
896 /* Validates that a (format, swizzle) pair is valid, in the sense that the
897 * swizzle doesn't access any components that are undefined in the format.
898 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
899 * omitted */
900
901 static bool
902 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
903 {
904 unsigned nr_comp = pandecode_format_component_count(fmt);
905 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
906 unsigned valid_mask = (1 << nr_comp) - 1;
907 unsigned invalid_mask = ~valid_mask;
908
909 if (access_mask & invalid_mask) {
910 pandecode_msg("XXX: invalid components accessed\n");
911 return false;
912 }
913
914 /* Check for the default non-swizzling swizzle so we can suppress
915 * useless printing for the defaults */
916
917 unsigned default_swizzles[4] = {
918 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
919 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
920 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
921 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
922 };
923
924 return (swizzle == default_swizzles[nr_comp - 1]);
925 }
926
927 /* Maps MALI_RGBA32F to rgba32f, etc */
928
929 static void
930 pandecode_format_short(enum mali_format fmt, bool srgb)
931 {
932 /* We want a type-like format, so cut off the initial MALI_ */
933 char *format = pandecode_format(fmt);
934 format += strlen("MALI_");
935
936 unsigned len = strlen(format);
937 char *lower_format = calloc(1, len + 1);
938
939 for (unsigned i = 0; i < len; ++i)
940 lower_format[i] = tolower(format[i]);
941
942 /* Sanity check sRGB flag is applied to RGB, per the name */
943 if (srgb && lower_format[0] != 'r')
944 pandecode_msg("XXX: sRGB applied to non-colour format\n");
945
946 /* Just prefix with an s, so you get formats like srgba8_unorm */
947 if (srgb)
948 pandecode_log_cont("s");
949
950 pandecode_log_cont("%s", lower_format);
951 free(lower_format);
952 }
953
954 static void
955 pandecode_swizzle(unsigned swizzle, enum mali_format format)
956 {
957 /* First, do some validation */
958 bool trivial_swizzle = pandecode_validate_format_swizzle(
959 format, swizzle);
960
961 if (trivial_swizzle)
962 return;
963
964 /* Next, print the swizzle */
965 pandecode_log_cont(".");
966
967 static const char components[] = "rgba01";
968
969 for (unsigned c = 0; c < 4; ++c) {
970 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
971
972 if (chan >= MALI_CHANNEL_RESERVED_0) {
973 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
974 continue;
975 }
976 pandecode_log_cont("%c", components[chan]);
977 }
978 }
979
980 static void
981 pandecode_rt_format(struct mali_rt_format format)
982 {
983 pandecode_log(".format = {\n");
984 pandecode_indent++;
985
986 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
987 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
988 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
989 pandecode_prop("unk4 = 0x%" PRIx32, format.unk4);
990
991 pandecode_prop("block = %s", pandecode_block_format(format.block));
992
993 /* TODO: Map formats so we can check swizzles and print nicely */
994 pandecode_log("swizzle");
995 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
996 pandecode_log_cont(",\n");
997
998 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
999 (format.nr_channels + 1));
1000
1001 pandecode_log(".flags = ");
1002 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
1003 pandecode_log_cont(",\n");
1004
1005 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
1006 * which is a faster hardware-based alternative to the wallpaper method
1007 * to preserve framebuffer contents across frames. In practice, MFBD
1008 * preload is buggy on Midgard, and so this is a chicken bit. If this
1009 * bit isn't set, most likely something broke unrelated to preload */
1010
1011 if (!format.no_preload) {
1012 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
1013 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
1014 }
1015
1016 if (format.zero)
1017 pandecode_prop("zero = 0x%" PRIx32, format.zero);
1018
1019 pandecode_indent--;
1020 pandecode_log("},\n");
1021 }
1022
1023 static void
1024 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct mali_framebuffer *fb)
1025 {
1026 pandecode_log("struct mali_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
1027 pandecode_indent++;
1028
1029 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
1030 mali_ptr rt_va = gpu_va + i * sizeof(struct mali_render_target);
1031 struct pandecode_mapped_memory *mem =
1032 pandecode_find_mapped_gpu_mem_containing(rt_va);
1033 const struct mali_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
1034
1035 pandecode_log("{\n");
1036 pandecode_indent++;
1037
1038 pandecode_rt_format(rt->format);
1039
1040 if (rt->format.block == MALI_BLOCK_AFBC) {
1041 pandecode_log(".afbc = {\n");
1042 pandecode_indent++;
1043
1044 char *a = pointer_as_memory_reference(rt->afbc.metadata);
1045 pandecode_prop("metadata = %s", a);
1046 free(a);
1047
1048 pandecode_prop("stride = %d", rt->afbc.stride);
1049 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
1050
1051 pandecode_indent--;
1052 pandecode_log("},\n");
1053 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
1054 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1055 rt->afbc.metadata,
1056 rt->afbc.stride,
1057 rt->afbc.unk);
1058 }
1059
1060 MEMORY_PROP(rt, framebuffer);
1061 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1062
1063 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1064 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1065 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1066 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1067 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1068 }
1069
1070 if (rt->zero1 || rt->zero2 || rt->zero3) {
1071 pandecode_msg("XXX: render target zeros tripped\n");
1072 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1073 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1074 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
1075 }
1076
1077 pandecode_indent--;
1078 pandecode_log("},\n");
1079 }
1080
1081 pandecode_indent--;
1082 pandecode_log("};\n");
1083 }
1084
1085 static struct pandecode_fbd
1086 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment, bool is_compute, bool is_bifrost)
1087 {
1088 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1089 const struct mali_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1090
1091 struct pandecode_fbd info;
1092
1093 if (is_bifrost && fb->msaa.sample_locations) {
1094 /* The blob stores all possible sample locations in a single buffer
1095 * allocated on startup, and just switches the pointer when switching
1096 * MSAA state. For now, we just put the data into the cmdstream, but we
1097 * should do something like what the blob does with a real driver.
1098 *
1099 * There seem to be 32 slots for sample locations, followed by another
1100 * 16. The second 16 is just the center location followed by 15 zeros
1101 * in all the cases I've identified (maybe shader vs. depth/color
1102 * samples?).
1103 */
1104
1105 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->msaa.sample_locations);
1106
1107 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->msaa.sample_locations);
1108
1109 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
1110 pandecode_indent++;
1111
1112 for (int i = 0; i < 32 + 16; i++) {
1113 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1114 }
1115
1116 pandecode_indent--;
1117 pandecode_log("};\n");
1118 }
1119
1120 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1121 pandecode_indent++;
1122
1123 if (is_bifrost) {
1124 pandecode_log(".msaa = {\n");
1125 pandecode_indent++;
1126
1127 if (fb->msaa.sample_locations)
1128 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1129 else
1130 pandecode_msg("XXX: sample_locations missing\n");
1131
1132 if (fb->msaa.zero1 || fb->msaa.zero2 || fb->msaa.zero4) {
1133 pandecode_msg("XXX: multisampling zero tripped\n");
1134 pandecode_prop("zero1 = %" PRIx64, fb->msaa.zero1);
1135 pandecode_prop("zero2 = %" PRIx64, fb->msaa.zero2);
1136 pandecode_prop("zero4 = %" PRIx64, fb->msaa.zero4);
1137 }
1138
1139 pandecode_indent--;
1140 pandecode_log("},\n");
1141 } else {
1142 pandecode_log(".shared_memory = {\n");
1143 pandecode_indent++;
1144 pandecode_shared_memory(&fb->shared_memory, is_compute);
1145 pandecode_indent--;
1146 pandecode_log("},\n");
1147 }
1148
1149 info.width = fb->width1 + 1;
1150 info.height = fb->height1 + 1;
1151 info.rt_count = fb->rt_count_1 + 1;
1152
1153 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1154 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1155 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1156 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1157
1158 pandecode_prop("unk1 = 0x%x", fb->unk1);
1159 pandecode_prop("unk2 = 0x%x", fb->unk2);
1160 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1161 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1162
1163 pandecode_log(".mfbd_flags = ");
1164 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1165 pandecode_log_cont(",\n");
1166
1167 if (fb->clear_stencil)
1168 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1169
1170 if (fb->clear_depth)
1171 pandecode_prop("clear_depth = %f", fb->clear_depth);
1172
1173 if (!is_compute)
1174 if (is_bifrost)
1175 pandecode_bifrost_tiler_descriptor(fb);
1176 else {
1177 const struct midgard_tiler_descriptor t = fb->tiler;
1178 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
1179 }
1180 else
1181 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1182
1183 if (fb->zero3 || fb->zero4) {
1184 pandecode_msg("XXX: framebuffer zeros tripped\n");
1185 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1186 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1187 }
1188
1189 pandecode_indent--;
1190 pandecode_log("};\n");
1191
1192 gpu_va += sizeof(struct mali_framebuffer);
1193
1194 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1195
1196 if (info.has_extra) {
1197 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1198 const struct mali_framebuffer_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1199
1200 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1201 pandecode_indent++;
1202
1203 MEMORY_PROP(fbx, checksum);
1204
1205 if (fbx->checksum_stride)
1206 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1207
1208 pandecode_log(".flags_hi = ");
1209 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_hi);
1210 pandecode_log_cont(",\n");
1211
1212 pandecode_log(".flags_lo = ");
1213 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
1214 pandecode_log_cont(",\n");
1215
1216 pandecode_prop("zs_block = %s", pandecode_block_format(fbx->zs_block));
1217
1218 if (fbx->zs_block == MALI_BLOCK_AFBC) {
1219 pandecode_log(".ds_afbc = {\n");
1220 pandecode_indent++;
1221
1222 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1223 pandecode_prop("depth_stencil_afbc_stride = %d",
1224 fbx->ds_afbc.depth_stencil_afbc_stride);
1225 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1226
1227 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
1228 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1229 pandecode_prop("zero1 = 0x%" PRIx32,
1230 fbx->ds_afbc.zero1);
1231 pandecode_prop("padding = 0x%" PRIx64,
1232 fbx->ds_afbc.padding);
1233 }
1234
1235 pandecode_indent--;
1236 pandecode_log("},\n");
1237 } else {
1238 pandecode_log(".ds_linear = {\n");
1239 pandecode_indent++;
1240
1241 if (fbx->ds_linear.depth) {
1242 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1243 pandecode_prop("depth_stride = %d",
1244 fbx->ds_linear.depth_stride);
1245 } else if (fbx->ds_linear.depth_stride) {
1246 pandecode_msg("XXX: depth stride zero tripped %d\n", fbx->ds_linear.depth_stride);
1247 }
1248
1249 if (fbx->ds_linear.stencil) {
1250 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1251 pandecode_prop("stencil_stride = %d",
1252 fbx->ds_linear.stencil_stride);
1253 } else if (fbx->ds_linear.stencil_stride) {
1254 pandecode_msg("XXX: stencil stride zero tripped %d\n", fbx->ds_linear.stencil_stride);
1255 }
1256
1257 if (fbx->ds_linear.depth_stride_zero ||
1258 fbx->ds_linear.stencil_stride_zero ||
1259 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1260 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1261 pandecode_prop("depth_stride_zero = 0x%x",
1262 fbx->ds_linear.depth_stride_zero);
1263 pandecode_prop("stencil_stride_zero = 0x%x",
1264 fbx->ds_linear.stencil_stride_zero);
1265 pandecode_prop("zero1 = 0x%" PRIx32,
1266 fbx->ds_linear.zero1);
1267 pandecode_prop("zero2 = 0x%" PRIx32,
1268 fbx->ds_linear.zero2);
1269 }
1270
1271 pandecode_indent--;
1272 pandecode_log("},\n");
1273 }
1274
1275 if (fbx->clear_color_1 | fbx->clear_color_2) {
1276 pandecode_prop("clear_color_1 = 0x%" PRIx32, fbx->clear_color_1);
1277 pandecode_prop("clear_color_2 = 0x%" PRIx32, fbx->clear_color_2);
1278 }
1279
1280 if (fbx->zero3) {
1281 pandecode_msg("XXX: fb_extra zeros tripped\n");
1282 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1283 }
1284
1285 pandecode_indent--;
1286 pandecode_log("};\n");
1287
1288 gpu_va += sizeof(struct mali_framebuffer_extra);
1289 }
1290
1291 if (is_fragment)
1292 pandecode_render_target(gpu_va, job_no, fb);
1293
1294 return info;
1295 }
1296
1297 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1298 * count */
1299
1300 static void
1301 pandecode_padded_vertices(unsigned shift, unsigned k)
1302 {
1303 unsigned odd = 2*k + 1;
1304 unsigned pot = 1 << shift;
1305 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1306 }
1307
1308 /* Given a magic divisor, recover what we were trying to divide by.
1309 *
1310 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1311 * 0 <= m < 2^N, for N bits in m.
1312 *
1313 * Let q represent the number we would like to divide by.
1314 *
1315 * By definition of a magic divisor for N-bit unsigned integers (a number you
1316 * multiply by to magically get division), m is a number such that:
1317 *
1318 * (m * x) & (2^N - 1) = floor(x/q).
1319 * for all x on Z where 0 <= x < 2^N
1320 *
1321 * Ignore the case where any of the above values equals zero; it is irrelevant
1322 * for our purposes (instanced arrays).
1323 *
1324 * Choose x = q. Then:
1325 *
1326 * (m * x) & (2^N - 1) = floor(x/q).
1327 * (m * q) & (2^N - 1) = floor(q/q).
1328 *
1329 * floor(q/q) = floor(1) = 1, therefore:
1330 *
1331 * (m * q) & (2^N - 1) = 1
1332 *
1333 * Recall the identity that the bitwise AND of one less than a power-of-two
1334 * equals the modulo with that power of two, i.e. for all x:
1335 *
1336 * x & (2^N - 1) = x % N
1337 *
1338 * Therefore:
1339 *
1340 * mq % (2^N) = 1
1341 *
1342 * By definition, a modular multiplicative inverse of a number m is the number
1343 * q such that with respect to a modulos M:
1344 *
1345 * mq % M = 1
1346 *
1347 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1348 *
1349 */
1350
1351 static void
1352 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1353 {
1354 #if 0
1355 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1356 * shift) the most lame way possible... just repeatedly add.
1357 * Asymptoptically slow but nobody cares in practice, unless you have
1358 * massive numbers of vertices or high divisors. */
1359
1360 unsigned inverse = 0;
1361
1362 /* Magic implicitly has the highest bit set */
1363 magic |= (1 << 31);
1364
1365 /* Depending on rounding direction */
1366 if (extra)
1367 magic++;
1368
1369 for (;;) {
1370 uint32_t product = magic * inverse;
1371
1372 if (shift) {
1373 product >>= shift;
1374 }
1375
1376 if (product == 1)
1377 break;
1378
1379 ++inverse;
1380 }
1381
1382 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1383
1384 /* Recall we're supposed to divide by (gl_level_divisor *
1385 * padded_num_vertices) */
1386
1387 unsigned padded_num_vertices = inverse / orig_divisor;
1388
1389 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1390 #endif
1391 }
1392
1393 static void
1394 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1395 mali_ptr addr, int job_no, char *suffix,
1396 int count, bool varying, enum mali_job_type job_type)
1397 {
1398 char *prefix = varying ? "varying" : "attribute";
1399 assert(addr);
1400
1401 if (!count) {
1402 pandecode_msg("warn: No %s records\n", prefix);
1403 return;
1404 }
1405
1406 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1407
1408 for (int i = 0; i < count; ++i) {
1409 /* First, check for special records */
1410 if (attr[i].elements < MALI_RECORD_SPECIAL) {
1411 if (attr[i].size)
1412 pandecode_msg("XXX: tripped size=%d\n", attr[i].size);
1413
1414 if (attr[i].stride) {
1415 /* gl_InstanceID passes a magic divisor in the
1416 * stride field to divide by the padded vertex
1417 * count. No other records should do so, so
1418 * stride should otherwise be zero. Note that
1419 * stride in the usual attribute sense doesn't
1420 * apply to special records. */
1421
1422 bool has_divisor = attr[i].elements == MALI_ATTR_INSTANCEID;
1423
1424 pandecode_log_cont("/* %smagic divisor = %X */ ",
1425 has_divisor ? "" : "XXX: ", attr[i].stride);
1426 }
1427
1428 if (attr[i].shift || attr[i].extra_flags) {
1429 /* Attributes use these fields for
1430 * instancing/padding/etc type issues, but
1431 * varyings don't */
1432
1433 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1434 varying ? "XXX: " : "",
1435 attr[i].shift, attr[i].extra_flags);
1436 }
1437
1438 /* Print the special record name */
1439 bool attribute = false;
1440 pandecode_log("%s_%d = %s;\n", prefix, i, pandecode_special_record(attr[i].elements, &attribute));
1441
1442 /* Sanity check */
1443 if (attribute == varying)
1444 pandecode_msg("XXX: mismatched special record\n");
1445
1446 continue;
1447 }
1448
1449 enum mali_attr_mode mode = attr[i].elements & 7;
1450
1451 if (mode == MALI_ATTR_UNUSED)
1452 pandecode_msg("XXX: unused attribute record\n");
1453
1454 /* For non-linear records, we need to print the type of record */
1455 if (mode != MALI_ATTR_LINEAR)
1456 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1457
1458 /* Print the name to link with attr_meta */
1459 pandecode_log_cont("%s_%d", prefix, i);
1460
1461 /* Print the stride and size */
1462 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1463
1464 /* TODO: Sanity check the quotient itself. It must be equal to
1465 * (or be greater than, if the driver added padding) the padded
1466 * vertex count. */
1467
1468 /* Finally, print the pointer */
1469 mali_ptr raw_elements = attr[i].elements & ~7;
1470 char *a = pointer_as_memory_reference(raw_elements);
1471 pandecode_log_cont(" = (%s);\n", a);
1472 free(a);
1473
1474 /* Check the pointer */
1475 pandecode_validate_buffer(raw_elements, attr[i].size);
1476
1477 /* shift/extra_flags exist only for instanced */
1478 if (attr[i].shift | attr[i].extra_flags) {
1479 /* These are set to random values by the blob for
1480 * varyings, most likely a symptom of uninitialized
1481 * memory where the hardware masked the bug. As such we
1482 * put this at a warning, not an error. */
1483
1484 if (mode == MALI_ATTR_LINEAR)
1485 pandecode_msg("warn: instancing fields set for linear\n");
1486
1487 pandecode_prop("shift = %d", attr[i].shift);
1488 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1489 }
1490
1491 /* Decode further where possible */
1492
1493 if (mode == MALI_ATTR_MODULO) {
1494 pandecode_padded_vertices(
1495 attr[i].shift,
1496 attr[i].extra_flags);
1497 }
1498
1499 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1500 i++;
1501 pandecode_log("{\n");
1502 pandecode_indent++;
1503 pandecode_prop("unk = 0x%x", attr[i].unk);
1504 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1505 if (attr[i].zero != 0)
1506 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1507 pandecode_prop("divisor = %d", attr[i].divisor);
1508 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1509 pandecode_indent--;
1510 pandecode_log("}, \n");
1511 }
1512
1513 }
1514
1515 pandecode_log("\n");
1516 }
1517
1518 static mali_ptr
1519 pandecode_shader_address(const char *name, mali_ptr ptr)
1520 {
1521 /* TODO: Decode flags */
1522 mali_ptr shader_ptr = ptr & ~15;
1523
1524 char *a = pointer_as_memory_reference(shader_ptr);
1525 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1526 free(a);
1527
1528 return shader_ptr;
1529 }
1530
1531 static void
1532 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1533 {
1534 unsigned any_nonzero =
1535 stencil->ref | stencil->mask | stencil->func |
1536 stencil->sfail | stencil->dpfail | stencil->dppass;
1537
1538 if (any_nonzero == 0)
1539 return;
1540
1541 const char *func = pandecode_func(stencil->func);
1542 const char *sfail = pandecode_stencil_op(stencil->sfail);
1543 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1544 const char *dppass = pandecode_stencil_op(stencil->dppass);
1545
1546 if (stencil->zero)
1547 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1548
1549 pandecode_log(".stencil_%s = {\n", name);
1550 pandecode_indent++;
1551 pandecode_prop("ref = %d", stencil->ref);
1552 pandecode_prop("mask = 0x%02X", stencil->mask);
1553 pandecode_prop("func = %s", func);
1554 pandecode_prop("sfail = %s", sfail);
1555 pandecode_prop("dpfail = %s", dpfail);
1556 pandecode_prop("dppass = %s", dppass);
1557 pandecode_indent--;
1558 pandecode_log("},\n");
1559 }
1560
1561 static void
1562 pandecode_blend_equation(const struct mali_blend_equation *blend)
1563 {
1564 if (blend->zero1)
1565 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1566
1567 pandecode_log(".equation = {\n");
1568 pandecode_indent++;
1569
1570 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1571 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1572
1573 pandecode_log(".color_mask = ");
1574 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1575 pandecode_log_cont(",\n");
1576
1577 pandecode_indent--;
1578 pandecode_log("},\n");
1579 }
1580
1581 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1582
1583 static unsigned
1584 decode_bifrost_constant(u16 constant)
1585 {
1586 float lo = (float) (constant & 0xFF);
1587 float hi = (float) (constant >> 8);
1588
1589 return (hi / 255.0) + (lo / 65535.0);
1590 }
1591
1592 static mali_ptr
1593 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1594 {
1595 struct bifrost_blend_rt *b =
1596 ((struct bifrost_blend_rt *) descs) + rt_no;
1597
1598 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1599 pandecode_indent++;
1600
1601 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1602 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1603 b->constant, decode_bifrost_constant(b->constant));
1604
1605 /* TODO figure out blend shader enable bit */
1606 pandecode_blend_equation(&b->equation);
1607
1608 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1609 pandecode_prop("index = 0x%" PRIx16, b->index);
1610
1611 pandecode_log(".format = ");
1612 pandecode_format_short(b->format, false);
1613 pandecode_swizzle(b->swizzle, b->format);
1614 pandecode_log_cont(",\n");
1615
1616 pandecode_prop("swizzle = 0x%" PRIx32, b->swizzle);
1617 pandecode_prop("format = 0x%" PRIx32, b->format);
1618
1619 if (b->zero1) {
1620 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1621 pandecode_prop("zero1 = 0x%" PRIx32, b->zero1);
1622 }
1623
1624 pandecode_log(".shader_type = ");
1625 switch(b->shader_type) {
1626 case BIFROST_BLEND_F16:
1627 pandecode_log_cont("BIFROST_BLEND_F16");
1628 break;
1629 case BIFROST_BLEND_F32:
1630 pandecode_log_cont("BIFROST_BLEND_F32");
1631 break;
1632 case BIFROST_BLEND_I32:
1633 pandecode_log_cont("BIFROST_BLEND_I32");
1634 break;
1635 case BIFROST_BLEND_U32:
1636 pandecode_log_cont("BIFROST_BLEND_U32");
1637 break;
1638 case BIFROST_BLEND_I16:
1639 pandecode_log_cont("BIFROST_BLEND_I16");
1640 break;
1641 case BIFROST_BLEND_U16:
1642 pandecode_log_cont("BIFROST_BLEND_U16");
1643 break;
1644 }
1645 pandecode_log_cont(",\n");
1646
1647 if (b->zero2) {
1648 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1649 pandecode_prop("zero2 = 0x%" PRIx32, b->zero2);
1650 }
1651
1652 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1653
1654 pandecode_indent--;
1655 pandecode_log("},\n");
1656
1657 return 0;
1658 }
1659
1660 static mali_ptr
1661 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1662 {
1663 /* constant/equation is in a union */
1664 if (!blend->shader)
1665 return 0;
1666
1667 pandecode_log(".blend = {\n");
1668 pandecode_indent++;
1669
1670 if (is_shader) {
1671 pandecode_shader_address("shader", blend->shader);
1672 } else {
1673 pandecode_blend_equation(&blend->equation);
1674 pandecode_prop("constant = %f", blend->constant);
1675 }
1676
1677 pandecode_indent--;
1678 pandecode_log("},\n");
1679
1680 /* Return blend shader to disassemble if present */
1681 return is_shader ? (blend->shader & ~0xF) : 0;
1682 }
1683
1684 static mali_ptr
1685 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1686 {
1687 struct midgard_blend_rt *b =
1688 ((struct midgard_blend_rt *) descs) + rt_no;
1689
1690 /* Flags determine presence of blend shader */
1691 bool is_shader = (b->flags & 0xF) >= 0x2;
1692
1693 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1694 pandecode_indent++;
1695
1696 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1697
1698 union midgard_blend blend = b->blend;
1699 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1700
1701 pandecode_indent--;
1702 pandecode_log("};\n");
1703
1704 return shader;
1705 }
1706
1707 /* Attributes and varyings have descriptor records, which contain information
1708 * about their format and ordering with the attribute/varying buffers. We'll
1709 * want to validate that the combinations specified are self-consistent.
1710 */
1711
1712 static int
1713 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1714 {
1715 char base[128];
1716 char *prefix = varying ? "varying" : "attribute";
1717 unsigned max_index = 0;
1718 snprintf(base, sizeof(base), "%s_meta", prefix);
1719
1720 struct mali_attr_meta *attr_meta;
1721 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1722
1723 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1724
1725 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1726 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1727 sizeof(*attr_mem));
1728
1729 /* If the record is discard, it should be zero for everything else */
1730
1731 if (attr_meta->format == MALI_VARYING_DISCARD) {
1732 uint64_t zero =
1733 attr_meta->index |
1734 attr_meta->unknown1 |
1735 attr_meta->unknown3 |
1736 attr_meta->src_offset;
1737
1738 if (zero)
1739 pandecode_msg("XXX: expected empty record for varying discard\n");
1740
1741 /* We want to look for a literal 0000 swizzle -- this
1742 * is not encoded with all zeroes, however */
1743
1744 enum mali_channel z = MALI_CHANNEL_ZERO;
1745 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1746 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1747
1748 if (!good_swizzle)
1749 pandecode_msg("XXX: expected zero swizzle for discard\n");
1750
1751 if (!varying)
1752 pandecode_msg("XXX: cannot discard attribute\n");
1753
1754 /* If we're all good, omit the record */
1755 if (!zero && varying && good_swizzle) {
1756 pandecode_log("/* discarded varying */\n");
1757 continue;
1758 }
1759 }
1760
1761 if (attr_meta->index > max_index)
1762 max_index = attr_meta->index;
1763
1764 if (attr_meta->unknown1 != 0x2) {
1765 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1766 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1767 }
1768
1769 if (attr_meta->unknown3) {
1770 pandecode_msg("XXX: unexpected unknown3 set\n");
1771 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1772 }
1773
1774 pandecode_format_short(attr_meta->format, false);
1775 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1776
1777 if (attr_meta->src_offset)
1778 pandecode_log_cont("[%u]", attr_meta->src_offset);
1779
1780 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1781
1782 pandecode_log_cont(";\n");
1783 }
1784
1785 pandecode_log("\n");
1786
1787 return count ? (max_index + 1) : 0;
1788 }
1789
1790 /* return bits [lo, hi) of word */
1791 static u32
1792 bits(u32 word, u32 lo, u32 hi)
1793 {
1794 if (hi - lo >= 32)
1795 return word; // avoid undefined behavior with the shift
1796
1797 return (word >> lo) & ((1 << (hi - lo)) - 1);
1798 }
1799
1800 static void
1801 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1802 {
1803 pandecode_log(".prefix = {\n");
1804 pandecode_indent++;
1805
1806 /* Decode invocation_count. See the comment before the definition of
1807 * invocation_count for an explanation.
1808 */
1809
1810 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1811 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1812 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1813 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1814 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1815 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1816
1817 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1818 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1819 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1820
1821 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1822 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1823 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1824
1825 /* Even though we have this decoded, we want to ensure that the
1826 * representation is "unique" so we don't lose anything by printing only
1827 * the final result. More specifically, we need to check that we were
1828 * passed something in canonical form, since the definition per the
1829 * hardware is inherently not unique. How? Well, take the resulting
1830 * decode and pack it ourselves! If it is bit exact with what we
1831 * decoded, we're good to go. */
1832
1833 struct mali_vertex_tiler_prefix ref;
1834 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1835
1836 bool canonical =
1837 (p->invocation_count == ref.invocation_count) &&
1838 (p->invocation_shifts == ref.invocation_shifts);
1839
1840 if (!canonical) {
1841 pandecode_msg("XXX: non-canonical workgroups packing\n");
1842 pandecode_msg("expected: %X, %X",
1843 ref.invocation_count,
1844 ref.invocation_shifts);
1845
1846 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1847 pandecode_prop("size_y_shift = %d", size_y_shift);
1848 pandecode_prop("size_z_shift = %d", size_z_shift);
1849 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
1850 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
1851 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
1852 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
1853 }
1854
1855 /* Regardless, print the decode */
1856 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1857 size_x, size_y, size_z,
1858 groups_x, groups_y, groups_z);
1859
1860 /* TODO: Decode */
1861 if (p->unknown_draw)
1862 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1863
1864 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1865
1866 if (p->draw_mode != MALI_DRAW_NONE)
1867 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1868
1869 /* Index count only exists for tiler jobs anyway */
1870
1871 if (p->index_count)
1872 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1873
1874
1875 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1876 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1877
1878 /* Validate an index buffer is present if we need one. TODO: verify
1879 * relationship between invocation_count and index_count */
1880
1881 if (p->indices) {
1882 unsigned count = p->index_count;
1883
1884 /* Grab the size */
1885 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1886
1887 /* Ensure we got a size, and if so, validate the index buffer
1888 * is large enough to hold a full set of indices of the given
1889 * size */
1890
1891 if (!index_raw_size)
1892 pandecode_msg("XXX: index size missing\n");
1893 else
1894 pandecode_validate_buffer(p->indices, count * size);
1895 } else if (index_raw_size)
1896 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1897
1898 if (p->offset_bias_correction)
1899 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1900
1901 /* TODO: Figure out what this is. It's not zero */
1902 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1903
1904 pandecode_indent--;
1905 pandecode_log("},\n");
1906 }
1907
1908 static void
1909 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1910 {
1911 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1912 uint64_t *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1913
1914 for (int i = 0; i < ubufs_count; i++) {
1915 unsigned size = (ubufs[i] & ((1 << 10) - 1)) * 16;
1916 mali_ptr addr = (ubufs[i] >> 10) << 2;
1917
1918 pandecode_validate_buffer(addr, size);
1919
1920 char *ptr = pointer_as_memory_reference(addr);
1921 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1922 free(ptr);
1923 }
1924
1925 pandecode_log("\n");
1926 }
1927
1928 static void
1929 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1930 {
1931 pandecode_validate_buffer(uniforms, uniform_count * 16);
1932
1933 char *ptr = pointer_as_memory_reference(uniforms);
1934 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1935 free(ptr);
1936 }
1937
1938 static const char *
1939 shader_type_for_job(unsigned type)
1940 {
1941 switch (type) {
1942 case JOB_TYPE_VERTEX: return "VERTEX";
1943 case JOB_TYPE_TILER: return "FRAGMENT";
1944 case JOB_TYPE_COMPUTE: return "COMPUTE";
1945 default:
1946 return "UNKNOWN";
1947 }
1948 }
1949
1950 static unsigned shader_id = 0;
1951
1952 static struct midgard_disasm_stats
1953 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1954 bool is_bifrost, unsigned gpu_id)
1955 {
1956 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1957 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1958
1959 /* Compute maximum possible size */
1960 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1961
1962 /* Print some boilerplate to clearly denote the assembly (which doesn't
1963 * obey indentation rules), and actually do the disassembly! */
1964
1965 pandecode_log_cont("\n\n");
1966
1967 struct midgard_disasm_stats stats;
1968
1969 if (is_bifrost) {
1970 disassemble_bifrost(pandecode_dump_stream, code, sz, true);
1971
1972 /* TODO: Extend stats to Bifrost */
1973 stats.texture_count = -128;
1974 stats.sampler_count = -128;
1975 stats.attribute_count = -128;
1976 stats.varying_count = -128;
1977 stats.uniform_count = -128;
1978 stats.uniform_buffer_count = -128;
1979 stats.work_count = -128;
1980
1981 stats.instruction_count = 0;
1982 stats.bundle_count = 0;
1983 stats.quadword_count = 0;
1984 stats.helper_invocations = false;
1985 } else {
1986 stats = disassemble_midgard(pandecode_dump_stream,
1987 code, sz, gpu_id,
1988 type == JOB_TYPE_TILER ?
1989 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
1990 }
1991
1992 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1993 * driver-internal purposes with the blob and interfere */
1994
1995 bool should_shaderdb = type != JOB_TYPE_COMPUTE;
1996
1997 if (should_shaderdb) {
1998 unsigned nr_threads =
1999 (stats.work_count <= 4) ? 4 :
2000 (stats.work_count <= 8) ? 2 :
2001 1;
2002
2003 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
2004 "%u inst, %u bundles, %u quadwords, "
2005 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
2006 shader_id++,
2007 shader_type_for_job(type),
2008 stats.instruction_count, stats.bundle_count, stats.quadword_count,
2009 stats.work_count, nr_threads);
2010 }
2011
2012
2013 return stats;
2014 }
2015
2016 static void
2017 pandecode_texture_payload(mali_ptr payload,
2018 enum mali_texture_type type,
2019 enum mali_texture_layout layout,
2020 bool manual_stride,
2021 uint8_t levels,
2022 uint16_t depth,
2023 uint16_t array_size,
2024 struct pandecode_mapped_memory *tmem)
2025 {
2026 pandecode_log(".payload = {\n");
2027 pandecode_indent++;
2028
2029 /* A bunch of bitmap pointers follow.
2030 * We work out the correct number,
2031 * based on the mipmap/cubemap
2032 * properties, but dump extra
2033 * possibilities to futureproof */
2034
2035 int bitmap_count = levels + 1;
2036
2037 /* Miptree for each face */
2038 if (type == MALI_TEX_CUBE)
2039 bitmap_count *= 6;
2040 else if (type == MALI_TEX_3D && layout == MALI_TEXTURE_LINEAR)
2041 bitmap_count *= (depth + 1);
2042
2043 /* Array of textures */
2044 bitmap_count *= (array_size + 1);
2045
2046 /* Stride for each element */
2047 if (manual_stride)
2048 bitmap_count *= 2;
2049
2050 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
2051 payload, sizeof(mali_ptr) * bitmap_count);
2052 for (int i = 0; i < bitmap_count; ++i) {
2053 /* How we dump depends if this is a stride or a pointer */
2054
2055 if (manual_stride && (i & 1)) {
2056 /* signed 32-bit snuck in as a 64-bit pointer */
2057 uint64_t stride_set = pointers_and_strides[i];
2058 uint32_t clamped_stride = stride_set;
2059 int32_t stride = clamped_stride;
2060 assert(stride_set == clamped_stride);
2061 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2062 } else {
2063 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
2064 pandecode_log("%s, \n", a);
2065 free(a);
2066 }
2067 }
2068
2069 pandecode_indent--;
2070 pandecode_log("},\n");
2071 }
2072
2073 static void
2074 pandecode_texture(mali_ptr u,
2075 struct pandecode_mapped_memory *tmem,
2076 unsigned job_no, unsigned tex)
2077 {
2078 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
2079
2080 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
2081 pandecode_indent++;
2082
2083 pandecode_prop("width = %" PRId32, t->width);
2084 pandecode_prop("height = %" PRId32, t->height);
2085 pandecode_prop("depth = %" PRId32, t->depth);
2086 pandecode_prop("array_size = %" PRId32, t->array_size);
2087
2088 pandecode_log("\n");
2089 pandecode_prop("f.swizzle = 0x%" PRIx32, t->format.swizzle);
2090 pandecode_prop("f.format = 0x%" PRIx32, t->format.format);
2091 pandecode_prop("f.srgb = 0x%" PRIx32, t->format.srgb);
2092 pandecode_prop("f.unknown1 = 0x%" PRIx32, t->format.unknown1);
2093 pandecode_prop("f.type = %" PRId32, t->format.type);
2094 pandecode_prop("f.layout = %" PRId32, t->format.layout);
2095 pandecode_prop("f.unknown2 = 0x%" PRIx32, t->format.unknown2);
2096 pandecode_prop("f.manual_stride = %" PRId32, t->format.manual_stride);
2097 pandecode_prop("f.zero = 0x%" PRIx32, t->format.zero);
2098 pandecode_log("\n");
2099
2100 pandecode_prop("unknown3 = 0x%" PRIx32, t->unknown3);
2101 pandecode_prop("unknown3A = 0x%" PRIx32, t->unknown3A);
2102 pandecode_prop("levels = %" PRId32, t->levels);
2103 pandecode_prop("swizzle = 0x%" PRIx32, t->swizzle);
2104 pandecode_prop("swizzle_zero = 0x%" PRIx32, t->swizzle_zero);
2105
2106 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2107 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2108 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2109 pandecode_log("\n");
2110
2111 struct mali_texture_format f = t->format;
2112
2113 /* See the definiton of enum mali_texture_type */
2114
2115 bool is_cube = f.type == MALI_TEX_CUBE;
2116 unsigned dimension = is_cube ? 2 : f.type;
2117
2118 pandecode_make_indent();
2119
2120 /* Print the layout. Default is linear; a modifier can denote AFBC or
2121 * u-interleaved/tiled modes */
2122
2123 if (f.layout == MALI_TEXTURE_AFBC)
2124 pandecode_log_cont("afbc");
2125 else if (f.layout == MALI_TEXTURE_TILED)
2126 pandecode_log_cont("tiled");
2127 else if (f.layout == MALI_TEXTURE_LINEAR)
2128 pandecode_log_cont("linear");
2129 else
2130 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
2131
2132 pandecode_swizzle(t->swizzle, f.format);
2133 pandecode_log_cont(" ");
2134
2135 /* Distinguish cube/2D with modifier */
2136
2137 if (is_cube)
2138 pandecode_log_cont("cube ");
2139
2140 pandecode_format_short(f.format, f.srgb);
2141 pandecode_swizzle(f.swizzle, f.format);
2142
2143 /* All four width/height/depth/array_size dimensions are present
2144 * regardless of the type of texture, but it is an error to have
2145 * non-zero dimensions for unused dimensions. Verify this. array_size
2146 * can always be set, as can width. */
2147
2148 if (t->height && dimension < 2)
2149 pandecode_msg("XXX: nonzero height for <2D texture\n");
2150
2151 if (t->depth && dimension < 3)
2152 pandecode_msg("XXX: nonzero depth for <2D texture\n");
2153
2154 /* Print only the dimensions that are actually there */
2155
2156 pandecode_log_cont(": %d", t->width + 1);
2157
2158 if (dimension >= 2)
2159 pandecode_log_cont("x%u", t->height + 1);
2160
2161 if (dimension >= 3)
2162 pandecode_log_cont("x%u", t->depth + 1);
2163
2164 if (t->array_size)
2165 pandecode_log_cont("[%u]", t->array_size + 1);
2166
2167 if (t->levels)
2168 pandecode_log_cont(" mip %u", t->levels);
2169
2170 pandecode_log_cont("\n");
2171
2172 if (f.unknown1 | f.zero) {
2173 pandecode_msg("XXX: texture format zero tripped\n");
2174 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
2175 pandecode_prop("zero = %" PRId32, f.zero);
2176 }
2177
2178 if (!f.unknown2) {
2179 pandecode_msg("XXX: expected unknown texture bit set\n");
2180 pandecode_prop("unknown2 = %" PRId32, f.unknown2);
2181 }
2182
2183 if (t->swizzle_zero) {
2184 pandecode_msg("XXX: swizzle zero tripped\n");
2185 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
2186 }
2187
2188 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
2189 pandecode_msg("XXX: texture zero tripped\n");
2190 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
2191 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
2192 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2193 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2194 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2195 }
2196
2197 pandecode_texture_payload(u + sizeof(*t), f.type, f.layout, f.manual_stride, t->levels, t->depth, t->array_size, tmem);
2198
2199 pandecode_indent--;
2200 pandecode_log("};\n");
2201 }
2202
2203 static void
2204 pandecode_bifrost_texture(
2205 const struct bifrost_texture_descriptor *t,
2206 unsigned job_no,
2207 unsigned tex)
2208 {
2209 pandecode_log("struct bifrost_texture_descriptor texture_descriptor_%d_%d = {\n", job_no, tex);
2210 pandecode_indent++;
2211
2212 pandecode_prop("format_unk = 0x%" PRIx32, t->format_unk);
2213 pandecode_prop("type = %" PRId32, t->type);
2214
2215 if (t->zero) {
2216 pandecode_msg("XXX: zero tripped\n");
2217 pandecode_prop("zero = 0x%" PRIx32, t->zero);
2218 }
2219
2220 pandecode_prop("format_swizzle = 0x%" PRIx32, t->format_swizzle);
2221 pandecode_prop("format = 0x%" PRIx32, t->format);
2222 pandecode_prop("srgb = 0x%" PRIx32, t->srgb);
2223 pandecode_prop("format_unk3 = 0x%" PRIx32, t->format_unk3);
2224 pandecode_prop("width = %" PRId32, t->width);
2225 pandecode_prop("height = %" PRId32, t->height);
2226 pandecode_prop("swizzle = 0x%" PRIx32, t->swizzle);
2227 pandecode_prop("levels = %" PRId32, t->levels);
2228 pandecode_prop("unk1 = 0x%" PRIx32, t->unk1);
2229 pandecode_prop("levels_unk = %" PRId32, t->levels_unk);
2230 pandecode_prop("level_2 = %" PRId32, t->level_2);
2231 pandecode_prop("payload = 0x%" PRIx64, t->payload);
2232 pandecode_prop("array_size = %" PRId32, t->array_size);
2233 pandecode_prop("unk4 = 0x%" PRIx32, t->unk4);
2234 pandecode_prop("depth = %" PRId32, t->depth);
2235 pandecode_prop("unk5 = 0x%" PRIx32, t->unk5);
2236 pandecode_log("\n");
2237
2238 /* See the definiton of enum mali_texture_type */
2239
2240 bool is_cube = t->type == MALI_TEX_CUBE;
2241 unsigned dimension = is_cube ? 2 : t->type;
2242
2243 /* Print the layout. Default is linear; a modifier can denote AFBC or
2244 * u-interleaved/tiled modes */
2245
2246 if (t->layout == MALI_TEXTURE_AFBC)
2247 pandecode_log_cont("afbc");
2248 else if (t->layout == MALI_TEXTURE_TILED)
2249 pandecode_log_cont("tiled");
2250 else if (t->layout == MALI_TEXTURE_LINEAR)
2251 pandecode_log_cont("linear");
2252 else
2253 pandecode_msg("XXX: invalid texture layout 0x%X\n", t->layout);
2254
2255 pandecode_swizzle(t->swizzle, t->format);
2256 pandecode_log_cont(" ");
2257
2258 /* Distinguish cube/2D with modifier */
2259
2260 if (is_cube)
2261 pandecode_log_cont("cube ");
2262
2263 pandecode_format_short(t->format, t->srgb);
2264
2265 /* All four width/height/depth/array_size dimensions are present
2266 * regardless of the type of texture, but it is an error to have
2267 * non-zero dimensions for unused dimensions. Verify this. array_size
2268 * can always be set, as can width. */
2269
2270 if (t->height && dimension < 2)
2271 pandecode_msg("XXX: nonzero height for <2D texture\n");
2272
2273 if (t->depth && dimension < 3)
2274 pandecode_msg("XXX: nonzero depth for <2D texture\n");
2275
2276 /* Print only the dimensions that are actually there */
2277
2278 pandecode_log_cont(": %d", t->width + 1);
2279
2280 if (dimension >= 2)
2281 pandecode_log_cont("x%u", t->height + 1);
2282
2283 if (dimension >= 3)
2284 pandecode_log_cont("x%u", t->depth + 1);
2285
2286 if (t->array_size)
2287 pandecode_log_cont("[%u]", t->array_size + 1);
2288
2289 if (t->levels)
2290 pandecode_log_cont(" mip %u", t->levels);
2291
2292 pandecode_log_cont("\n");
2293
2294 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(t->payload);
2295 if (t->payload) {
2296 pandecode_texture_payload(t->payload, t->type, t->layout,
2297 true, t->levels, t->depth,
2298 t->array_size, tmem);
2299 }
2300
2301 pandecode_indent--;
2302 pandecode_log("};\n");
2303 }
2304
2305 /* 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 */
2306
2307 static void
2308 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2309 {
2310 /* Nothing to do */
2311 if (claim == truth)
2312 return;
2313
2314 if (fuzzy && (truth < 0))
2315 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name, claim, truth);
2316
2317 if ((truth >= 0) && !fuzzy) {
2318 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2319 (truth < claim) ? "warn" : "XXX",
2320 name, truth, claim);
2321 } else if ((claim > -truth) && !fuzzy) {
2322 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2323 name, -truth, claim);
2324 } else if (fuzzy && (claim < truth))
2325 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2326 name, truth, claim);
2327
2328 pandecode_log(".%s = %" PRId16, name, claim);
2329
2330 if (fuzzy)
2331 pandecode_log_cont(" /* %u used */", truth);
2332
2333 pandecode_log_cont(",\n");
2334 }
2335
2336 static void
2337 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2338 bool is_bifrost, unsigned gpu_id)
2339 {
2340 struct midgard_disasm_stats stats =
2341 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
2342
2343 bool has_texture = (stats.texture_count > 0);
2344 bool has_sampler = (stats.sampler_count > 0);
2345 bool has_attribute = (stats.attribute_count > 0);
2346 bool has_varying = (stats.varying_count > 0);
2347 bool has_uniform = (stats.uniform_count > 0);
2348 bool has_ubo = (stats.uniform_buffer_count > 0);
2349
2350 if (has_texture || has_sampler)
2351 pandecode_msg("XXX: blend shader accessing textures\n");
2352
2353 if (has_attribute || has_varying)
2354 pandecode_msg("XXX: blend shader accessing interstage\n");
2355
2356 if (has_uniform || has_ubo)
2357 pandecode_msg("XXX: blend shader accessing uniforms\n");
2358 }
2359
2360 static void
2361 pandecode_textures(mali_ptr textures, unsigned texture_count, int job_no, bool is_bifrost)
2362 {
2363 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(textures);
2364
2365 if (!mmem)
2366 return;
2367
2368 if (is_bifrost) {
2369 const struct bifrost_texture_descriptor *PANDECODE_PTR_VAR(t, mmem, textures);
2370
2371 pandecode_log("uint64_t textures_%"PRIx64"_%d[] = {\n", textures, job_no);
2372 pandecode_indent++;
2373
2374 for (unsigned tex = 0; tex < texture_count; ++tex)
2375 pandecode_bifrost_texture(&t[tex], job_no, tex);
2376
2377 pandecode_indent--;
2378 pandecode_log("};\n");
2379 } else {
2380 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures);
2381
2382 pandecode_log("uint64_t textures_%"PRIx64"_%d[] = {\n", textures, job_no);
2383 pandecode_indent++;
2384
2385 for (int tex = 0; tex < texture_count; ++tex) {
2386 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2387 char *a = pointer_as_memory_reference(*u);
2388 pandecode_log("%s,\n", a);
2389 free(a);
2390 }
2391
2392 pandecode_indent--;
2393 pandecode_log("};\n");
2394
2395 /* Now, finally, descend down into the texture descriptor */
2396 for (unsigned tex = 0; tex < texture_count; ++tex) {
2397 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2398 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2399 if (tmem)
2400 pandecode_texture(*u, tmem, job_no, tex);
2401 }
2402 }
2403 }
2404
2405 static void
2406 pandecode_samplers(mali_ptr samplers, unsigned sampler_count, int job_no, bool is_bifrost)
2407 {
2408 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(samplers);
2409
2410 if (!smem)
2411 return;
2412
2413 if (is_bifrost) {
2414 struct bifrost_sampler_descriptor *s;
2415
2416 for (int i = 0; i < sampler_count; ++i) {
2417 s = pandecode_fetch_gpu_mem(smem, samplers + sizeof(*s) * i, sizeof(*s));
2418
2419 pandecode_log("struct bifrost_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", samplers + sizeof(*s) * i, job_no, i);
2420 pandecode_indent++;
2421
2422 if (s->unk1 != 1) {
2423 pandecode_msg("XXX: unk1 tripped\n");
2424 pandecode_prop("unk1 = 0x%x", s->unk1);
2425 }
2426
2427 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2428 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2429 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2430
2431 if (s->unk8 != 0x8) {
2432 pandecode_msg("XXX: unk8 tripped\n");
2433 pandecode_prop("unk8 = 0x%x", s->unk8);
2434 }
2435
2436 pandecode_prop("unk2 = 0x%x", s->unk2);
2437 pandecode_prop("unk3 = 0x%x", s->unk3);
2438 pandecode_prop("min_filter = %s", s->min_filter ? "nearest" : "linear");
2439 pandecode_prop("norm_coords = 0x%x", s->norm_coords & 0x1);
2440 pandecode_prop("zero1 = 0x%x", s->zero1 & 0x1);
2441 pandecode_prop("mip_filter = %s", s->mip_filter ? "linear" : "nearest");
2442 pandecode_prop("mag_filter = %s", s->mag_filter ? "linear" : "nearest");
2443
2444 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2445 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2446
2447 if (s->zero1 || s->zero2 || s->zero3 || s->zero4) {
2448 pandecode_msg("XXX: sampler zero tripped\n");
2449 pandecode_prop("zero = 0x%" PRIx64 ", 0x%" PRIx64 ", 0x%" PRIx64 ", 0x%" PRIx64 "\n", s->zero1, s->zero2, s->zero3, s->zero4);
2450 }
2451
2452 pandecode_indent--;
2453 pandecode_log("};\n");
2454 }
2455 } else {
2456 struct mali_sampler_descriptor *s;
2457
2458 for (int i = 0; i < sampler_count; ++i) {
2459 s = pandecode_fetch_gpu_mem(smem, samplers + sizeof(*s) * i, sizeof(*s));
2460
2461 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", samplers + sizeof(*s) * i, job_no, i);
2462 pandecode_indent++;
2463
2464 pandecode_log(".filter_mode = ");
2465 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2466 pandecode_log_cont(",\n");
2467
2468 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2469 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2470
2471 if (s->lod_bias)
2472 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s->lod_bias));
2473
2474 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2475 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2476 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2477
2478 pandecode_prop("compare_func = %s", pandecode_func(s->compare_func));
2479
2480 if (s->zero || s->zero2) {
2481 pandecode_msg("XXX: sampler zero tripped\n");
2482 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2483 }
2484
2485 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2486
2487 pandecode_prop("border_color = { %f, %f, %f, %f }",
2488 s->border_color[0],
2489 s->border_color[1],
2490 s->border_color[2],
2491 s->border_color[3]);
2492
2493 pandecode_indent--;
2494 pandecode_log("};\n");
2495 }
2496 }
2497 }
2498
2499 static void
2500 pandecode_vertex_tiler_postfix_pre(
2501 const struct mali_vertex_tiler_postfix *p,
2502 int job_no, enum mali_job_type job_type,
2503 char *suffix, bool is_bifrost, unsigned gpu_id)
2504 {
2505 struct pandecode_mapped_memory *attr_mem;
2506
2507 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2508 * are the only things actually needed from the FBD, vertex/tiler jobs
2509 * no longer reference the FBD -- instead, this field points to some
2510 * info about the scratchpad.
2511 */
2512
2513 struct pandecode_fbd fbd_info = {
2514 /* Default for Bifrost */
2515 .rt_count = 1
2516 };
2517
2518 if (is_bifrost) {
2519 pandecode_log_cont("\t/* %X %/\n", p->shared_memory & 1);
2520 pandecode_compute_fbd(p->shared_memory & ~1, job_no);
2521 } else if (p->shared_memory & MALI_MFBD)
2522 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->shared_memory) & FBD_MASK, job_no, false, job_type == JOB_TYPE_COMPUTE, false);
2523 else if (job_type == JOB_TYPE_COMPUTE)
2524 pandecode_compute_fbd((u64) (uintptr_t) p->shared_memory, job_no);
2525 else
2526 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->shared_memory, job_no, false, gpu_id);
2527
2528 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2529 int texture_count = 0, sampler_count = 0;
2530
2531 if (p->shader) {
2532 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2533 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2534
2535 /* Disassemble ahead-of-time to get stats. Initialize with
2536 * stats for the missing-shader case so we get validation
2537 * there, too */
2538
2539 struct midgard_disasm_stats info = {
2540 .texture_count = 0,
2541 .sampler_count = 0,
2542 .attribute_count = 0,
2543 .varying_count = 0,
2544 .work_count = 1,
2545
2546 .uniform_count = -128,
2547 .uniform_buffer_count = 0
2548 };
2549
2550 if (s->shader & ~0xF)
2551 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
2552
2553 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2554 pandecode_indent++;
2555
2556 /* Save for dumps */
2557 attribute_count = s->attribute_count;
2558 varying_count = s->varying_count;
2559 texture_count = s->texture_count;
2560 sampler_count = s->sampler_count;
2561
2562 if (is_bifrost) {
2563 uniform_count = s->bifrost2.uniform_count;
2564 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2565 } else {
2566 uniform_count = s->midgard1.uniform_count;
2567 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2568 }
2569
2570 pandecode_shader_address("shader", s->shader);
2571
2572 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2573 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2574 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2575 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2576 pandecode_shader_prop("uniform_buffer_count",
2577 uniform_buffer_count,
2578 info.uniform_buffer_count, true);
2579
2580 if (!is_bifrost) {
2581 pandecode_shader_prop("uniform_count",
2582 uniform_count,
2583 info.uniform_count, false);
2584
2585 pandecode_shader_prop("work_count",
2586 s->midgard1.work_count, info.work_count, false);
2587 }
2588
2589 if (is_bifrost) {
2590 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32, s->bifrost1.unk1);
2591 } else {
2592 bool helpers = s->midgard1.flags_lo & MALI_HELPER_INVOCATIONS;
2593 s->midgard1.flags_lo &= ~MALI_HELPER_INVOCATIONS;
2594
2595 if (helpers != info.helper_invocations) {
2596 pandecode_msg("XXX: expected helpers %u but got %u\n",
2597 info.helper_invocations, helpers);
2598 }
2599
2600 pandecode_log(".midgard1.flags_lo = ");
2601 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info, s->midgard1.flags_lo);
2602 pandecode_log_cont(",\n");
2603
2604 pandecode_log(".midgard1.flags_hi = ");
2605 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info, s->midgard1.flags_hi);
2606 pandecode_log_cont(",\n");
2607 }
2608
2609 if (s->depth_units || s->depth_factor) {
2610 pandecode_prop("depth_factor = %f", s->depth_factor);
2611 pandecode_prop("depth_units = %f", s->depth_units);
2612 }
2613
2614 if (s->alpha_coverage) {
2615 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
2616 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
2617
2618 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2619 invert_alpha_coverage ? "~" : "",
2620 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
2621 }
2622
2623 if (s->unknown2_2)
2624 pandecode_prop(".unknown2_2 = %X", s->unknown2_2);
2625
2626 if (s->unknown2_3 || s->unknown2_4) {
2627 pandecode_log(".unknown2_3 = ");
2628
2629 int unknown2_3 = s->unknown2_3;
2630 int unknown2_4 = s->unknown2_4;
2631
2632 /* We're not quite sure what these flags mean without the depth test, if anything */
2633
2634 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2635 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2636 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2637
2638 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2639 }
2640
2641 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2642 pandecode_log_cont(",\n");
2643
2644 pandecode_log(".unknown2_4 = ");
2645 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2646 pandecode_log_cont(",\n");
2647 }
2648
2649 if (s->stencil_mask_front || s->stencil_mask_back) {
2650 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2651 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2652 }
2653
2654 pandecode_stencil("front", &s->stencil_front);
2655 pandecode_stencil("back", &s->stencil_back);
2656
2657 if (is_bifrost) {
2658 pandecode_log(".bifrost2 = {\n");
2659 pandecode_indent++;
2660
2661 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2662 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2663 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2664 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2665
2666 pandecode_indent--;
2667 pandecode_log("},\n");
2668 } else if (s->midgard2.unknown2_7) {
2669 pandecode_log(".midgard2 = {\n");
2670 pandecode_indent++;
2671
2672 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2673 pandecode_indent--;
2674 pandecode_log("},\n");
2675 }
2676
2677 if (s->padding) {
2678 pandecode_msg("XXX: shader padding tripped\n");
2679 pandecode_prop("padding = 0x%" PRIx32, s->padding);
2680 }
2681
2682 if (!is_bifrost) {
2683 /* TODO: Blend shaders routing/disasm */
2684 union midgard_blend blend = s->blend;
2685 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2686 if (shader & ~0xF)
2687 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2688 } else {
2689 pandecode_msg("mdg_blend = %" PRIx64 "\n", s->blend.shader);
2690 }
2691
2692 pandecode_indent--;
2693 pandecode_log("};\n");
2694
2695 /* MRT blend fields are used whenever MFBD is used, with
2696 * per-RT descriptors */
2697
2698 if (job_type == JOB_TYPE_TILER && (is_bifrost || p->shared_memory & MALI_MFBD)) {
2699 void* blend_base = (void *) (s + 1);
2700
2701 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2702 mali_ptr shader = 0;
2703
2704 if (is_bifrost)
2705 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2706 else
2707 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2708
2709 if (shader & ~0xF)
2710 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2711
2712 }
2713 }
2714 } else
2715 pandecode_msg("XXX: missing shader descriptor\n");
2716
2717 if (p->viewport) {
2718 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2719 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2720
2721 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2722 pandecode_indent++;
2723
2724 pandecode_prop("clip_minx = %f", f->clip_minx);
2725 pandecode_prop("clip_miny = %f", f->clip_miny);
2726 pandecode_prop("clip_minz = %f", f->clip_minz);
2727 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2728 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2729 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2730
2731 /* Only the higher coordinates are MALI_POSITIVE scaled */
2732
2733 pandecode_prop("viewport0 = { %d, %d }",
2734 f->viewport0[0], f->viewport0[1]);
2735
2736 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2737 f->viewport1[0] + 1, f->viewport1[1] + 1);
2738
2739 pandecode_indent--;
2740 pandecode_log("};\n");
2741 }
2742
2743 unsigned max_attr_index = 0;
2744
2745 if (p->attribute_meta)
2746 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2747
2748 if (p->attributes) {
2749 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2750 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2751 }
2752
2753 /* Varyings are encoded like attributes but not actually sent; we just
2754 * pass a zero buffer with the right stride/size set, (or whatever)
2755 * since the GPU will write to it itself */
2756
2757 if (p->varying_meta) {
2758 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2759 }
2760
2761 if (p->varyings) {
2762 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2763
2764 /* Number of descriptors depends on whether there are
2765 * non-internal varyings */
2766
2767 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2768 }
2769
2770 if (p->uniform_buffers) {
2771 if (uniform_buffer_count)
2772 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2773 else
2774 pandecode_msg("warn: UBOs specified but not referenced\n");
2775 } else if (uniform_buffer_count)
2776 pandecode_msg("XXX: UBOs referenced but not specified\n");
2777
2778 /* We don't want to actually dump uniforms, but we do need to validate
2779 * that the counts we were given are sane */
2780
2781 if (p->uniforms) {
2782 if (uniform_count)
2783 pandecode_uniforms(p->uniforms, uniform_count);
2784 else
2785 pandecode_msg("warn: Uniforms specified but not referenced\n");
2786 } else if (uniform_count)
2787 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2788
2789 if (p->textures)
2790 pandecode_textures(p->textures, texture_count, job_no, is_bifrost);
2791
2792 if (p->sampler_descriptor)
2793 pandecode_samplers(p->sampler_descriptor, sampler_count, job_no, is_bifrost);
2794 }
2795
2796 static void
2797 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2798 {
2799 pandecode_log(".gl_enables = ");
2800
2801 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2802
2803 pandecode_log_cont(",\n");
2804 }
2805
2806 static void
2807 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2808 {
2809 if (p->shader & 0xF)
2810 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2811
2812 pandecode_log(".postfix = {\n");
2813 pandecode_indent++;
2814
2815 pandecode_gl_enables(p->gl_enables, JOB_TYPE_TILER);
2816 pandecode_prop("instance_shift = 0x%x", p->instance_shift);
2817 pandecode_prop("instance_odd = 0x%x", p->instance_odd);
2818
2819 if (p->zero4) {
2820 pandecode_msg("XXX: vertex only zero tripped");
2821 pandecode_prop("zero4 = 0x%" PRIx32, p->zero4);
2822 }
2823
2824 pandecode_prop("offset_start = 0x%x", p->offset_start);
2825
2826 if (p->zero5) {
2827 pandecode_msg("XXX: vertex only zero tripped");
2828 pandecode_prop("zero5 = 0x%" PRIx64, p->zero5);
2829 }
2830
2831 MEMORY_PROP(p, position_varying);
2832 MEMORY_PROP(p, occlusion_counter);
2833
2834 pandecode_indent--;
2835 pandecode_log("},\n");
2836 }
2837
2838 static void
2839 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2840 {
2841 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2842 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2843
2844 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2845 pandecode_indent++;
2846
2847 if (h->zero) {
2848 pandecode_msg("XXX: tiler heap zero tripped\n");
2849 pandecode_prop("zero = 0x%x", h->zero);
2850 }
2851
2852 pandecode_prop("heap_size = 0x%x", h->heap_size);
2853 MEMORY_PROP(h, tiler_heap_start);
2854 MEMORY_PROP(h, tiler_heap_free);
2855
2856 /* this might point to the beginning of another buffer, when it's
2857 * really the end of the tiler heap buffer, so we have to be careful
2858 * here. but for zero length, we need the same pointer.
2859 */
2860
2861 if (h->tiler_heap_end == h->tiler_heap_start) {
2862 MEMORY_PROP(h, tiler_heap_start);
2863 } else {
2864 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2865 pandecode_prop("tiler_heap_end = %s + 1", a);
2866 free(a);
2867 }
2868
2869 for (int i = 0; i < 10; i++) {
2870 if (h->zeros[i] != 0) {
2871 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2872 i, h->zeros[i]);
2873 }
2874 }
2875
2876 if (h->unk1 != 0x1) {
2877 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2878 pandecode_prop("unk1 = 0x%x", h->unk1);
2879 }
2880
2881 if (h->unk7e007e != 0x7e007e) {
2882 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
2883 pandecode_prop("unk7e007e = 0x%x", h->unk7e007e);
2884 }
2885
2886 pandecode_indent--;
2887 pandecode_log("};\n");
2888 }
2889
2890 static void
2891 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2892 {
2893 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2894 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2895
2896 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2897
2898 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2899 pandecode_indent++;
2900
2901 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32, t->tiler_heap_next_start);
2902 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32, t->used_hierarchy_mask);
2903
2904 if (t->hierarchy_mask != 0xa &&
2905 t->hierarchy_mask != 0x14 &&
2906 t->hierarchy_mask != 0x28 &&
2907 t->hierarchy_mask != 0x50 &&
2908 t->hierarchy_mask != 0xa0)
2909 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
2910
2911 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2912
2913 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2914
2915 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2916 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2917
2918 if (t->zero0) {
2919 pandecode_msg("XXX: tiler meta zero tripped\n");
2920 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2921 }
2922
2923 for (int i = 0; i < 12; i++) {
2924 if (t->zeros[i] != 0) {
2925 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2926 i, t->zeros[i]);
2927 }
2928 }
2929
2930 pandecode_indent--;
2931 pandecode_log("};\n");
2932 }
2933
2934 static void
2935 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2936 {
2937 if (u.pointer == 0x0)
2938 return;
2939
2940 pandecode_log(".primitive_size = {\n");
2941 pandecode_indent++;
2942
2943 if (constant) {
2944 pandecode_prop("constant = %f", u.constant);
2945 } else {
2946 MEMORY_PROP((&u), pointer);
2947 }
2948
2949 pandecode_indent--;
2950 pandecode_log("},\n");
2951 }
2952
2953 static void
2954 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2955 {
2956 pandecode_log_cont("{\n");
2957 pandecode_indent++;
2958
2959 /* TODO: gl_PointSize on Bifrost */
2960 pandecode_primitive_size(t->primitive_size, true);
2961
2962 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2963 || t->zero6) {
2964 pandecode_msg("XXX: tiler only zero tripped\n");
2965 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2966 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2967 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2968 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2969 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2970 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2971 }
2972
2973 pandecode_indent--;
2974 pandecode_log("},\n");
2975 }
2976
2977 static int
2978 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2979 const struct pandecode_mapped_memory *mem,
2980 mali_ptr payload, int job_no, unsigned gpu_id)
2981 {
2982 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2983
2984 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
2985
2986 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64"_%d = {\n", payload, job_no);
2987 pandecode_indent++;
2988
2989 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2990 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2991
2992 pandecode_indent--;
2993 pandecode_log("};\n");
2994
2995 return sizeof(*v);
2996 }
2997
2998 static int
2999 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
3000 const struct pandecode_mapped_memory *mem,
3001 mali_ptr payload, int job_no, unsigned gpu_id)
3002 {
3003 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
3004
3005 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
3006 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
3007
3008 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64"_%d = {\n", payload, job_no);
3009 pandecode_indent++;
3010
3011 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
3012
3013 pandecode_log(".tiler = ");
3014 pandecode_tiler_only_bfr(&t->tiler, job_no);
3015
3016 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
3017
3018 pandecode_indent--;
3019 pandecode_log("};\n");
3020
3021 return sizeof(*t);
3022 }
3023
3024 static int
3025 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
3026 const struct pandecode_mapped_memory *mem,
3027 mali_ptr payload, int job_no, unsigned gpu_id)
3028 {
3029 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
3030 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
3031
3032 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
3033
3034 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
3035 pandecode_indent++;
3036
3037 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
3038 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
3039
3040 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
3041 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
3042
3043 pandecode_indent--;
3044 pandecode_log("};\n");
3045
3046 return sizeof(*v);
3047 }
3048
3049 static int
3050 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
3051 mali_ptr payload, int job_no,
3052 bool is_bifrost, unsigned gpu_id)
3053 {
3054 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
3055
3056 bool is_mfbd = s->framebuffer & MALI_MFBD;
3057
3058 if (!is_mfbd && is_bifrost)
3059 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
3060
3061 struct pandecode_fbd info;
3062
3063 if (is_mfbd)
3064 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true, false, is_bifrost);
3065 else
3066 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
3067
3068 /* Compute the tag for the tagged pointer. This contains the type of
3069 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
3070 * additional structures follow the MFBD header (an extra payload or
3071 * not, as well as a count of render targets) */
3072
3073 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
3074
3075 if (is_mfbd) {
3076 if (info.has_extra)
3077 expected_tag |= MALI_MFBD_TAG_EXTRA;
3078
3079 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
3080 }
3081
3082 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
3083 pandecode_msg("XXX: unexpected tile coordinate bits\n");
3084 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
3085 pandecode_prop("max_tile_coord = 0x%X\n", s->max_tile_coord);
3086 }
3087
3088 /* Extract tile coordinates */
3089
3090 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
3091 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
3092
3093 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
3094 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
3095
3096 /* For the max, we also want the floored (rather than ceiled) version for checking */
3097
3098 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
3099 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
3100
3101 /* Validate the coordinates are well-ordered */
3102
3103 if (min_x == max_x)
3104 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
3105 else if (min_x > max_x)
3106 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
3107
3108 if (min_y == max_y)
3109 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
3110 else if (min_y > max_y)
3111 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
3112
3113 /* Validate the coordinates fit inside the framebuffer. We use floor,
3114 * rather than ceil, for the max coordinates, since the tile
3115 * coordinates for something like an 800x600 framebuffer will actually
3116 * resolve to 800x608, which would otherwise trigger a Y-overflow */
3117
3118 if ((min_x > info.width) || (max_x_f > info.width))
3119 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
3120
3121 if ((min_y > info.height) || (max_y_f > info.height))
3122 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
3123
3124 /* After validation, we print */
3125
3126 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
3127
3128 /* The FBD is a tagged pointer */
3129
3130 unsigned tag = (s->framebuffer & ~FBD_MASK);
3131
3132 if (tag != expected_tag)
3133 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
3134
3135 return sizeof(*s);
3136 }
3137
3138 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
3139 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
3140 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
3141 * might wish to combine this with the bifrost parameter) because some details
3142 * are model-specific even within a particular architecture. Minimal traces
3143 * *only* examine the job descriptors, skipping printing entirely if there is
3144 * no faults, and only descends into the payload if there are faults. This is
3145 * useful for looking for faults without the overhead of invasive traces. */
3146
3147 void
3148 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id, bool minimal)
3149 {
3150 struct mali_job_descriptor_header *h;
3151 unsigned job_descriptor_number = 0;
3152
3153 do {
3154 struct pandecode_mapped_memory *mem =
3155 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
3156
3157 void *payload;
3158
3159 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
3160
3161 /* On Midgard, for 32-bit jobs except for fragment jobs, the
3162 * high 32-bits of the 64-bit pointer are reused to store
3163 * something else.
3164 */
3165 int offset = h->job_descriptor_size == MALI_JOB_32 &&
3166 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
3167 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
3168
3169 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
3170
3171 int job_no = job_descriptor_number++;
3172
3173 /* If the job is good to go, skip it in minimal mode */
3174 if (minimal && (h->exception_status == 0x0 || h->exception_status == 0x1))
3175 continue;
3176
3177 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
3178 pandecode_indent++;
3179
3180 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
3181
3182 if (h->job_descriptor_size)
3183 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
3184
3185 if (h->exception_status && h->exception_status != 0x1)
3186 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
3187 h->exception_status,
3188 (h->exception_status >> 16) & 0xFFFF,
3189 pandecode_exception_access((h->exception_status >> 8) & 0x3),
3190 h->exception_status & 0xFF);
3191
3192 if (h->first_incomplete_task)
3193 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
3194
3195 if (h->fault_pointer)
3196 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
3197
3198 if (h->job_barrier)
3199 pandecode_prop("job_barrier = %d", h->job_barrier);
3200
3201 pandecode_prop("job_index = %d", h->job_index);
3202
3203 if (h->unknown_flags)
3204 pandecode_prop("unknown_flags = %d", h->unknown_flags);
3205
3206 if (h->job_dependency_index_1)
3207 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
3208
3209 if (h->job_dependency_index_2)
3210 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
3211
3212 pandecode_indent--;
3213 pandecode_log("};\n");
3214
3215 switch (h->job_type) {
3216 case JOB_TYPE_WRITE_VALUE: {
3217 struct mali_payload_write_value *s = payload;
3218 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
3219 pandecode_indent++;
3220 MEMORY_PROP(s, address);
3221
3222 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
3223 pandecode_msg("XXX: unknown value descriptor\n");
3224 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
3225 }
3226
3227 if (s->reserved) {
3228 pandecode_msg("XXX: set value tripped\n");
3229 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
3230 }
3231
3232 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
3233 pandecode_indent--;
3234 pandecode_log("};\n");
3235
3236 break;
3237 }
3238
3239 case JOB_TYPE_TILER:
3240 case JOB_TYPE_VERTEX:
3241 case JOB_TYPE_COMPUTE:
3242 if (bifrost) {
3243 if (h->job_type == JOB_TYPE_TILER)
3244 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
3245 else
3246 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
3247 } else
3248 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
3249
3250 break;
3251
3252 case JOB_TYPE_FRAGMENT:
3253 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
3254 break;
3255
3256 default:
3257 break;
3258 }
3259 } while ((jc_gpu_va = h->next_job));
3260 }