projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
pan/decode: Dump scratchpad size if present
[mesa.git] / src / panfrost / pandecode / decode.c
1 /*
2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 */
25
26 #include <panfrost-job.h>
27 #include <stdio.h>
28 #include <stdlib.h>
29 #include <memory.h>
30 #include <stdbool.h>
31 #include <stdarg.h>
32 #include <ctype.h>
33 #include "decode.h"
34 #include "util/macros.h"
35 #include "util/u_math.h"
36
37 #include "pan_pretty_print.h"
38 #include "midgard/disassemble.h"
39 #include "bifrost/disassemble.h"
40
41 #include "pan_encoder.h"
42
43 static void pandecode_swizzle(unsigned swizzle, enum mali_format format);
44
45 #define MEMORY_PROP(obj, p) {\
46 if (obj->p) { \
47 char *a = pointer_as_memory_reference(obj->p); \
48 pandecode_prop("%s = %s", #p, a); \
49 free(a); \
50 } \
51 }
52
53 #define MEMORY_PROP_DIR(obj, p) {\
54 if (obj.p) { \
55 char *a = pointer_as_memory_reference(obj.p); \
56 pandecode_prop("%s = %s", #p, a); \
57 free(a); \
58 } \
59 }
60
61 FILE *pandecode_dump_stream;
62
63 /* Semantic logging type.
64 *
65 * Raw: for raw messages to be printed as is.
66 * Message: for helpful information to be commented out in replays.
67 * Property: for properties of a struct
68 *
69 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
70 */
71
72 enum pandecode_log_type {
73 PANDECODE_RAW,
74 PANDECODE_MESSAGE,
75 PANDECODE_PROPERTY
76 };
77
78 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
79 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
80 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
81
82 unsigned pandecode_indent = 0;
83
84 static void
85 pandecode_make_indent(void)
86 {
87 for (unsigned i = 0; i < pandecode_indent; ++i)
88 fprintf(pandecode_dump_stream, " ");
89 }
90
91 static void
92 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
93 {
94 va_list ap;
95
96 pandecode_make_indent();
97
98 if (type == PANDECODE_MESSAGE)
99 fprintf(pandecode_dump_stream, "// ");
100 else if (type == PANDECODE_PROPERTY)
101 fprintf(pandecode_dump_stream, ".");
102
103 va_start(ap, format);
104 vfprintf(pandecode_dump_stream, format, ap);
105 va_end(ap);
106
107 if (type == PANDECODE_PROPERTY)
108 fprintf(pandecode_dump_stream, ",\n");
109 }
110
111 static void
112 pandecode_log_cont(const char *format, ...)
113 {
114 va_list ap;
115
116 va_start(ap, format);
117 vfprintf(pandecode_dump_stream, format, ap);
118 va_end(ap);
119 }
120
121 /* To check for memory safety issues, validates that the given pointer in GPU
122 * memory is valid, containing at least sz bytes. The goal is to eliminate
123 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
124 * overruns) by statically validating pointers.
125 */
126
127 static void
128 pandecode_validate_buffer(mali_ptr addr, size_t sz)
129 {
130 if (!addr) {
131 pandecode_msg("XXX: null pointer deref");
132 return;
133 }
134
135 /* Find a BO */
136
137 struct pandecode_mapped_memory *bo =
138 pandecode_find_mapped_gpu_mem_containing(addr);
139
140 if (!bo) {
141 pandecode_msg("XXX: invalid memory dereference\n");
142 return;
143 }
144
145 /* Bounds check */
146
147 unsigned offset = addr - bo->gpu_va;
148 unsigned total = offset + sz;
149
150 if (total > bo->length) {
151 pandecode_msg("XXX: buffer overrun. "
152 "Chunk of size %zu at offset %d in buffer of size %zu. "
153 "Overrun by %zu bytes. \n",
154 sz, offset, bo->length, total - bo->length);
155 return;
156 }
157 }
158
159 struct pandecode_flag_info {
160 u64 flag;
161 const char *name;
162 };
163
164 static void
165 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
166 u64 flags)
167 {
168 bool decodable_flags_found = false;
169
170 for (int i = 0; flag_info[i].name; i++) {
171 if ((flags & flag_info[i].flag) != flag_info[i].flag)
172 continue;
173
174 if (!decodable_flags_found) {
175 decodable_flags_found = true;
176 } else {
177 pandecode_log_cont(" | ");
178 }
179
180 pandecode_log_cont("%s", flag_info[i].name);
181
182 flags &= ~flag_info[i].flag;
183 }
184
185 if (decodable_flags_found) {
186 if (flags)
187 pandecode_log_cont(" | 0x%" PRIx64, flags);
188 } else {
189 pandecode_log_cont("0x%" PRIx64, flags);
190 }
191 }
192
193 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
194 static const struct pandecode_flag_info gl_enable_flag_info[] = {
195 FLAG_INFO(OCCLUSION_QUERY),
196 FLAG_INFO(OCCLUSION_PRECISE),
197 FLAG_INFO(FRONT_CCW_TOP),
198 FLAG_INFO(CULL_FACE_FRONT),
199 FLAG_INFO(CULL_FACE_BACK),
200 {}
201 };
202 #undef FLAG_INFO
203
204 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
205 static const struct pandecode_flag_info clear_flag_info[] = {
206 FLAG_INFO(FAST),
207 FLAG_INFO(SLOW),
208 FLAG_INFO(SLOW_STENCIL),
209 {}
210 };
211 #undef FLAG_INFO
212
213 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
214 static const struct pandecode_flag_info mask_flag_info[] = {
215 FLAG_INFO(R),
216 FLAG_INFO(G),
217 FLAG_INFO(B),
218 FLAG_INFO(A),
219 {}
220 };
221 #undef FLAG_INFO
222
223 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
224 static const struct pandecode_flag_info u3_flag_info[] = {
225 FLAG_INFO(HAS_MSAA),
226 FLAG_INFO(CAN_DISCARD),
227 FLAG_INFO(HAS_BLEND_SHADER),
228 FLAG_INFO(DEPTH_WRITEMASK),
229 {}
230 };
231
232 static const struct pandecode_flag_info u4_flag_info[] = {
233 FLAG_INFO(NO_MSAA),
234 FLAG_INFO(NO_DITHER),
235 FLAG_INFO(DEPTH_RANGE_A),
236 FLAG_INFO(DEPTH_RANGE_B),
237 FLAG_INFO(STENCIL_TEST),
238 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER),
239 {}
240 };
241 #undef FLAG_INFO
242
243 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
244 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
245 FLAG_INFO(MSAA),
246 FLAG_INFO(SRGB),
247 {}
248 };
249 #undef FLAG_INFO
250
251 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
252 static const struct pandecode_flag_info mfbd_extra_flag_hi_info[] = {
253 FLAG_INFO(PRESENT),
254 {}
255 };
256 #undef FLAG_INFO
257
258 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
259 static const struct pandecode_flag_info mfbd_extra_flag_lo_info[] = {
260 FLAG_INFO(ZS),
261 {}
262 };
263 #undef FLAG_INFO
264
265 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
266 static const struct pandecode_flag_info shader_midgard1_flag_lo_info [] = {
267 FLAG_INFO(WRITES_Z),
268 FLAG_INFO(EARLY_Z),
269 FLAG_INFO(READS_TILEBUFFER),
270 FLAG_INFO(READS_ZS),
271 {}
272 };
273
274 static const struct pandecode_flag_info shader_midgard1_flag_hi_info [] = {
275 FLAG_INFO(WRITES_S),
276 {}
277 };
278 #undef FLAG_INFO
279
280 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
281 static const struct pandecode_flag_info mfbd_flag_info [] = {
282 FLAG_INFO(DEPTH_WRITE),
283 FLAG_INFO(EXTRA),
284 {}
285 };
286 #undef FLAG_INFO
287
288 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
289 static const struct pandecode_flag_info sampler_flag_info [] = {
290 FLAG_INFO(MAG_NEAREST),
291 FLAG_INFO(MIN_NEAREST),
292 FLAG_INFO(MIP_LINEAR_1),
293 FLAG_INFO(MIP_LINEAR_2),
294 FLAG_INFO(NORM_COORDS),
295 {}
296 };
297 #undef FLAG_INFO
298
299 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
300 static const struct pandecode_flag_info sfbd_unk1_info [] = {
301 FLAG_INFO(MSAA_8),
302 FLAG_INFO(MSAA_A),
303 {}
304 };
305 #undef FLAG_INFO
306
307 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
308 static const struct pandecode_flag_info sfbd_unk2_info [] = {
309 FLAG_INFO(MSAA_B),
310 FLAG_INFO(SRGB),
311 {}
312 };
313 #undef FLAG_INFO
314
315 extern char *replace_fragment;
316 extern char *replace_vertex;
317
318 static char *
319 pandecode_job_type(enum mali_job_type type)
320 {
321 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
322
323 switch (type) {
324 DEFINE_CASE(NULL);
325 DEFINE_CASE(WRITE_VALUE);
326 DEFINE_CASE(CACHE_FLUSH);
327 DEFINE_CASE(COMPUTE);
328 DEFINE_CASE(VERTEX);
329 DEFINE_CASE(TILER);
330 DEFINE_CASE(FUSED);
331 DEFINE_CASE(FRAGMENT);
332
333 case JOB_NOT_STARTED:
334 return "NOT_STARTED";
335
336 default:
337 pandecode_log("Warning! Unknown job type %x\n", type);
338 return "!?!?!?";
339 }
340
341 #undef DEFINE_CASE
342 }
343
344 static char *
345 pandecode_draw_mode(enum mali_draw_mode mode)
346 {
347 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
348
349 switch (mode) {
350 DEFINE_CASE(DRAW_NONE);
351 DEFINE_CASE(POINTS);
352 DEFINE_CASE(LINES);
353 DEFINE_CASE(TRIANGLES);
354 DEFINE_CASE(TRIANGLE_STRIP);
355 DEFINE_CASE(TRIANGLE_FAN);
356 DEFINE_CASE(LINE_STRIP);
357 DEFINE_CASE(LINE_LOOP);
358 DEFINE_CASE(POLYGON);
359 DEFINE_CASE(QUADS);
360 DEFINE_CASE(QUAD_STRIP);
361
362 default:
363 pandecode_msg("XXX: invalid draw mode %X\n", mode);
364 return "";
365 }
366
367 #undef DEFINE_CASE
368 }
369
370 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
371 static char *
372 pandecode_func(enum mali_func mode)
373 {
374 switch (mode) {
375 DEFINE_CASE(NEVER);
376 DEFINE_CASE(LESS);
377 DEFINE_CASE(EQUAL);
378 DEFINE_CASE(LEQUAL);
379 DEFINE_CASE(GREATER);
380 DEFINE_CASE(NOTEQUAL);
381 DEFINE_CASE(GEQUAL);
382 DEFINE_CASE(ALWAYS);
383
384 default:
385 pandecode_msg("XXX: invalid func %X\n", mode);
386 return "";
387 }
388 }
389 #undef DEFINE_CASE
390
391 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
392 static char *
393 pandecode_stencil_op(enum mali_stencil_op op)
394 {
395 switch (op) {
396 DEFINE_CASE(KEEP);
397 DEFINE_CASE(REPLACE);
398 DEFINE_CASE(ZERO);
399 DEFINE_CASE(INVERT);
400 DEFINE_CASE(INCR_WRAP);
401 DEFINE_CASE(DECR_WRAP);
402 DEFINE_CASE(INCR);
403 DEFINE_CASE(DECR);
404
405 default:
406 pandecode_msg("XXX: invalid stencil op %X\n", op);
407 return "";
408 }
409 }
410
411 #undef DEFINE_CASE
412
413 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
414 {
415 switch(mode) {
416 /* TODO: Combine to just "instanced" once this can be done
417 * unambiguously in all known cases */
418 case MALI_ATTR_POT_DIVIDE:
419 return "instanced_pot";
420 case MALI_ATTR_MODULO:
421 return "instanced_mod";
422 case MALI_ATTR_NPOT_DIVIDE:
423 return "instanced_npot";
424 case MALI_ATTR_IMAGE:
425 return "image";
426 default:
427 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
428 return "";
429 }
430 }
431
432 static const char *
433 pandecode_special_record(uint64_t v, bool* attribute)
434 {
435 switch(v) {
436 case MALI_ATTR_VERTEXID:
437 *attribute = true;
438 return "gl_VertexID";
439 case MALI_ATTR_INSTANCEID:
440 *attribute = true;
441 return "gl_InstanceID";
442 case MALI_VARYING_FRAG_COORD:
443 return "gl_FragCoord";
444 case MALI_VARYING_FRONT_FACING:
445 return "gl_FrontFacing";
446 case MALI_VARYING_POINT_COORD:
447 return "gl_PointCoord";
448 default:
449 pandecode_msg("XXX: invalid special record %" PRIx64 "\n", v);
450 return "";
451 }
452 }
453
454 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
455 static char *
456 pandecode_wrap_mode(enum mali_wrap_mode op)
457 {
458 switch (op) {
459 DEFINE_CASE(REPEAT);
460 DEFINE_CASE(CLAMP_TO_EDGE);
461 DEFINE_CASE(CLAMP_TO_BORDER);
462 DEFINE_CASE(MIRRORED_REPEAT);
463
464 default:
465 pandecode_msg("XXX: invalid wrap mode %X\n", op);
466 return "";
467 }
468 }
469 #undef DEFINE_CASE
470
471 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
472 static char *
473 pandecode_block_format(enum mali_block_format fmt)
474 {
475 switch (fmt) {
476 DEFINE_CASE(TILED);
477 DEFINE_CASE(UNKNOWN);
478 DEFINE_CASE(LINEAR);
479 DEFINE_CASE(AFBC);
480
481 default:
482 unreachable("Invalid case");
483 }
484 }
485 #undef DEFINE_CASE
486
487 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
488 char *
489 pandecode_exception_access(unsigned access)
490 {
491 switch (access) {
492 DEFINE_CASE(NONE);
493 DEFINE_CASE(EXECUTE);
494 DEFINE_CASE(READ);
495 DEFINE_CASE(WRITE);
496
497 default:
498 unreachable("Invalid case");
499 }
500 }
501 #undef DEFINE_CASE
502
503 /* Midgard's tiler descriptor is embedded within the
504 * larger FBD */
505
506 static void
507 pandecode_midgard_tiler_descriptor(
508 const struct midgard_tiler_descriptor *t,
509 unsigned width,
510 unsigned height,
511 bool is_fragment,
512 bool has_hierarchy)
513 {
514 pandecode_log(".tiler = {\n");
515 pandecode_indent++;
516
517 if (t->hierarchy_mask == MALI_TILER_DISABLED)
518 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
519 else
520 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
521
522 /* We know this name from the kernel, but we never see it nonzero */
523
524 if (t->flags)
525 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
526
527 MEMORY_PROP(t, polygon_list);
528
529 /* The body is offset from the base of the polygon list */
530 assert(t->polygon_list_body > t->polygon_list);
531 unsigned body_offset = t->polygon_list_body - t->polygon_list;
532
533 /* It needs to fit inside the reported size */
534 assert(t->polygon_list_size >= body_offset);
535
536 /* Check that we fit */
537 struct pandecode_mapped_memory *plist =
538 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
539
540 assert(t->polygon_list_size <= plist->length);
541
542 /* Now that we've sanity checked, we'll try to calculate the sizes
543 * ourselves for comparison */
544
545 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
546 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
547
548 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
549 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
550 ref_header, ref_size);
551 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
552 pandecode_msg("body offset %d\n", body_offset);
553 }
554
555 /* The tiler heap has a start and end specified -- it should be
556 * identical to what we have in the BO. The exception is if tiling is
557 * disabled. */
558
559 MEMORY_PROP(t, heap_start);
560 assert(t->heap_end >= t->heap_start);
561
562 struct pandecode_mapped_memory *heap =
563 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
564
565 unsigned heap_size = t->heap_end - t->heap_start;
566
567 /* Tiling is enabled with a special flag */
568 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
569 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
570
571 bool tiling_enabled = hierarchy_mask;
572
573 if (tiling_enabled) {
574 /* When tiling is enabled, the heap should be a tight fit */
575 unsigned heap_offset = t->heap_start - heap->gpu_va;
576 if ((heap_offset + heap_size) != heap->length) {
577 pandecode_msg("XXX: heap size %u (expected %zu)\n",
578 heap_size, heap->length - heap_offset);
579 }
580
581 /* We should also have no other flags */
582 if (tiler_flags)
583 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
584 } else {
585 /* When tiling is disabled, we should have that flag and no others */
586
587 if (tiler_flags != MALI_TILER_DISABLED) {
588 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
589 tiler_flags);
590 }
591
592 /* We should also have an empty heap */
593 if (heap_size) {
594 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
595 heap_size);
596 }
597
598 /* Disabled tiling is used only for clear-only jobs, which are
599 * purely FRAGMENT, so we should never see this for
600 * non-FRAGMENT descriptors. */
601
602 if (!is_fragment)
603 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
604 }
605
606 /* We've never seen weights used in practice, but we know from the
607 * kernel these fields is there */
608
609 bool nonzero_weights = false;
610
611 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
612 nonzero_weights |= t->weights[w] != 0x0;
613 }
614
615 if (nonzero_weights) {
616 pandecode_log(".weights = {");
617
618 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
619 pandecode_log("%d, ", t->weights[w]);
620 }
621
622 pandecode_log("},");
623 }
624
625 pandecode_indent--;
626 pandecode_log("}\n");
627 }
628
629 /* Information about the framebuffer passed back for
630 * additional analysis */
631
632 struct pandecode_fbd {
633 unsigned width;
634 unsigned height;
635 unsigned rt_count;
636 bool has_extra;
637 };
638
639 static void
640 pandecode_sfbd_format(struct mali_sfbd_format format)
641 {
642 pandecode_log(".format = {\n");
643 pandecode_indent++;
644
645 pandecode_log(".unk1 = ");
646 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
647 pandecode_log_cont(",\n");
648
649 /* TODO: Map formats so we can check swizzles and print nicely */
650 pandecode_log("swizzle");
651 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
652 pandecode_log_cont(",\n");
653
654 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
655 (format.nr_channels + 1));
656
657 pandecode_log(".unk2 = ");
658 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
659 pandecode_log_cont(",\n");
660
661 pandecode_prop("block = %s", pandecode_block_format(format.block));
662
663 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
664
665 pandecode_indent--;
666 pandecode_log("},\n");
667 }
668
669 static void
670 pandecode_shared_memory(const struct mali_shared_memory *desc, bool is_compute)
671 {
672 pandecode_prop("stack_shift = 0x%x", desc->stack_shift);
673
674 if (desc->unk0)
675 pandecode_prop("unk0 = 0x%x", desc->unk0);
676
677 if (desc->shared_workgroup_count != 0x1F) {
678 pandecode_prop("shared_workgroup_count = %d", desc->shared_workgroup_count);
679 if (!is_compute)
680 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
681 }
682
683 if (desc->shared_unk1 || desc->shared_shift) {
684 pandecode_prop("shared_unk1 = %X", desc->shared_unk1);
685 pandecode_prop("shared_shift = %X", desc->shared_shift);
686
687 if (!is_compute)
688 pandecode_msg("XXX: shared memory configured in noncompute shader");
689 }
690
691 if (desc->shared_zero) {
692 pandecode_msg("XXX: shared memory zero tripped\n");
693 pandecode_prop("shared_zero = 0x%" PRIx32, desc->shared_zero);
694 }
695
696 if (desc->shared_memory && !is_compute)
697 pandecode_msg("XXX: shared memory used in noncompute shader\n");
698
699 MEMORY_PROP(desc, scratchpad);
700 MEMORY_PROP(desc, shared_memory);
701 MEMORY_PROP(desc, unknown1);
702
703 if (desc->scratchpad) {
704 struct pandecode_mapped_memory *smem =
705 pandecode_find_mapped_gpu_mem_containing(desc->scratchpad);
706
707 pandecode_msg("scratchpad size %u\n", smem->length);
708 }
709
710 }
711
712 static struct pandecode_fbd
713 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
714 {
715 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
716 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
717
718 struct pandecode_fbd info = {
719 .has_extra = false,
720 .rt_count = 1
721 };
722
723 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
724 pandecode_indent++;
725
726 pandecode_log(".shared_memory = {\n");
727 pandecode_indent++;
728 pandecode_shared_memory(&s->shared_memory, false);
729 pandecode_indent--;
730 pandecode_log("},\n");
731
732 pandecode_sfbd_format(s->format);
733
734 info.width = s->width + 1;
735 info.height = s->height + 1;
736
737 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
738 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
739
740 MEMORY_PROP(s, checksum);
741
742 if (s->checksum_stride)
743 pandecode_prop("checksum_stride = %d", s->checksum_stride);
744
745 MEMORY_PROP(s, framebuffer);
746 pandecode_prop("stride = %d", s->stride);
747
748 /* Earlier in the actual commandstream -- right before width -- but we
749 * delay to flow nicer */
750
751 pandecode_log(".clear_flags = ");
752 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
753 pandecode_log_cont(",\n");
754
755 if (s->depth_buffer) {
756 MEMORY_PROP(s, depth_buffer);
757 pandecode_prop("depth_stride = %d", s->depth_stride);
758 }
759
760 if (s->stencil_buffer) {
761 MEMORY_PROP(s, stencil_buffer);
762 pandecode_prop("stencil_stride = %d", s->stencil_stride);
763 }
764
765 if (s->depth_stride_zero ||
766 s->stencil_stride_zero ||
767 s->zero7 || s->zero8) {
768 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
769 pandecode_prop("depth_stride_zero = 0x%x",
770 s->depth_stride_zero);
771 pandecode_prop("stencil_stride_zero = 0x%x",
772 s->stencil_stride_zero);
773 pandecode_prop("zero7 = 0x%" PRIx32,
774 s->zero7);
775 pandecode_prop("zero8 = 0x%" PRIx32,
776 s->zero8);
777 }
778
779 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
780 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
781 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
782 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
783 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
784 }
785
786 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
787 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
788 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
789 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
790 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
791 }
792
793 if (s->clear_stencil) {
794 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
795 }
796
797 const struct midgard_tiler_descriptor t = s->tiler;
798
799 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
800 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
801
802 pandecode_indent--;
803 pandecode_log("};\n");
804
805 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
806 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
807 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
808
809 pandecode_log_cont(".zero3 = {");
810
811 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
812 pandecode_log_cont("%X, ", s->zero3[i]);
813
814 pandecode_log_cont("},\n");
815
816 pandecode_log_cont(".zero6 = {");
817
818 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
819 pandecode_log_cont("%X, ", s->zero6[i]);
820
821 pandecode_log_cont("},\n");
822
823 return info;
824 }
825
826 static void
827 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
828 {
829 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
830 const struct mali_shared_memory *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
831
832 pandecode_log("struct mali_shared_memory shared_%"PRIx64"_%d = {\n", gpu_va, job_no);
833 pandecode_indent++;
834 pandecode_shared_memory(s, true);
835 pandecode_indent--;
836 pandecode_log("},\n");
837 }
838
839 /* Extracts the number of components associated with a Mali format */
840
841 static unsigned
842 pandecode_format_component_count(enum mali_format fmt)
843 {
844 /* Mask out the format class */
845 unsigned top = fmt & 0b11100000;
846
847 switch (top) {
848 case MALI_FORMAT_SNORM:
849 case MALI_FORMAT_UINT:
850 case MALI_FORMAT_UNORM:
851 case MALI_FORMAT_SINT:
852 return ((fmt >> 3) & 3) + 1;
853 default:
854 /* TODO: Validate */
855 return 4;
856 }
857 }
858
859 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
860
861 static unsigned
862 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
863 {
864 unsigned mask = 0;
865 assert(MALI_CHANNEL_RED == 0);
866
867 for (unsigned c = 0; c < 4; ++c) {
868 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
869
870 if (chan <= MALI_CHANNEL_ALPHA)
871 mask |= (1 << chan);
872 }
873
874 return mask;
875 }
876
877 /* Validates that a (format, swizzle) pair is valid, in the sense that the
878 * swizzle doesn't access any components that are undefined in the format.
879 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
880 * omitted */
881
882 static bool
883 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
884 {
885 unsigned nr_comp = pandecode_format_component_count(fmt);
886 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
887 unsigned valid_mask = (1 << nr_comp) - 1;
888 unsigned invalid_mask = ~valid_mask;
889
890 if (access_mask & invalid_mask) {
891 pandecode_msg("XXX: invalid components accessed\n");
892 return false;
893 }
894
895 /* Check for the default non-swizzling swizzle so we can suppress
896 * useless printing for the defaults */
897
898 unsigned default_swizzles[4] = {
899 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
900 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
901 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
902 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
903 };
904
905 return (swizzle == default_swizzles[nr_comp - 1]);
906 }
907
908 /* Maps MALI_RGBA32F to rgba32f, etc */
909
910 static void
911 pandecode_format_short(enum mali_format fmt, bool srgb)
912 {
913 /* We want a type-like format, so cut off the initial MALI_ */
914 char *format = pandecode_format(fmt);
915 format += strlen("MALI_");
916
917 unsigned len = strlen(format);
918 char *lower_format = calloc(1, len + 1);
919
920 for (unsigned i = 0; i < len; ++i)
921 lower_format[i] = tolower(format[i]);
922
923 /* Sanity check sRGB flag is applied to RGB, per the name */
924 if (srgb && lower_format[0] != 'r')
925 pandecode_msg("XXX: sRGB applied to non-colour format\n");
926
927 /* Just prefix with an s, so you get formats like srgba8_unorm */
928 if (srgb)
929 pandecode_log_cont("s");
930
931 pandecode_log_cont("%s", lower_format);
932 free(lower_format);
933 }
934
935 static void
936 pandecode_swizzle(unsigned swizzle, enum mali_format format)
937 {
938 /* First, do some validation */
939 bool trivial_swizzle = pandecode_validate_format_swizzle(
940 format, swizzle);
941
942 if (trivial_swizzle)
943 return;
944
945 /* Next, print the swizzle */
946 pandecode_log_cont(".");
947
948 static const char components[] = "rgba01";
949
950 for (unsigned c = 0; c < 4; ++c) {
951 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
952
953 if (chan >= MALI_CHANNEL_RESERVED_0) {
954 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
955 continue;
956 }
957 pandecode_log_cont("%c", components[chan]);
958 }
959 }
960
961 static void
962 pandecode_rt_format(struct mali_rt_format format)
963 {
964 pandecode_log(".format = {\n");
965 pandecode_indent++;
966
967 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
968 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
969 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
970
971 pandecode_prop("block = %s", pandecode_block_format(format.block));
972
973 /* TODO: Map formats so we can check swizzles and print nicely */
974 pandecode_log("swizzle");
975 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
976 pandecode_log_cont(",\n");
977
978 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
979 (format.nr_channels + 1));
980
981 pandecode_log(".flags = ");
982 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
983 pandecode_log_cont(",\n");
984
985 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
986 * which is a faster hardware-based alternative to the wallpaper method
987 * to preserve framebuffer contents across frames. In practice, MFBD
988 * preload is buggy on Midgard, and so this is a chicken bit. If this
989 * bit isn't set, most likely something broke unrelated to preload */
990
991 if (!format.no_preload) {
992 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
993 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
994 }
995
996 if (format.zero)
997 pandecode_prop("zero = 0x%" PRIx32, format.zero);
998
999 pandecode_indent--;
1000 pandecode_log("},\n");
1001 }
1002
1003 static void
1004 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct mali_framebuffer *fb)
1005 {
1006 pandecode_log("struct mali_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
1007 pandecode_indent++;
1008
1009 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
1010 mali_ptr rt_va = gpu_va + i * sizeof(struct mali_render_target);
1011 struct pandecode_mapped_memory *mem =
1012 pandecode_find_mapped_gpu_mem_containing(rt_va);
1013 const struct mali_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
1014
1015 pandecode_log("{\n");
1016 pandecode_indent++;
1017
1018 pandecode_rt_format(rt->format);
1019
1020 if (rt->format.block == MALI_BLOCK_AFBC) {
1021 pandecode_log(".afbc = {\n");
1022 pandecode_indent++;
1023
1024 char *a = pointer_as_memory_reference(rt->afbc.metadata);
1025 pandecode_prop("metadata = %s", a);
1026 free(a);
1027
1028 pandecode_prop("stride = %d", rt->afbc.stride);
1029 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
1030
1031 pandecode_indent--;
1032 pandecode_log("},\n");
1033 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
1034 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1035 rt->afbc.metadata,
1036 rt->afbc.stride,
1037 rt->afbc.unk);
1038 }
1039
1040 MEMORY_PROP(rt, framebuffer);
1041 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1042
1043 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1044 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1045 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1046 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1047 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1048 }
1049
1050 if (rt->zero1 || rt->zero2 || rt->zero3) {
1051 pandecode_msg("XXX: render target zeros tripped\n");
1052 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1053 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1054 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
1055 }
1056
1057 pandecode_indent--;
1058 pandecode_log("},\n");
1059 }
1060
1061 pandecode_indent--;
1062 pandecode_log("};\n");
1063 }
1064
1065 static struct pandecode_fbd
1066 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment, bool is_compute)
1067 {
1068 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1069 const struct mali_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1070
1071 struct pandecode_fbd info;
1072
1073 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1074 pandecode_indent++;
1075
1076 pandecode_log(".shared_memory = {\n");
1077 pandecode_indent++;
1078 pandecode_shared_memory(&fb->shared_memory, is_compute);
1079 pandecode_indent--;
1080 pandecode_log("},\n");
1081
1082 info.width = fb->width1 + 1;
1083 info.height = fb->height1 + 1;
1084 info.rt_count = fb->rt_count_1 + 1;
1085
1086 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1087 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1088 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1089 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1090
1091 pandecode_prop("unk1 = 0x%x", fb->unk1);
1092 pandecode_prop("unk2 = 0x%x", fb->unk2);
1093 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1094 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1095
1096 pandecode_log(".mfbd_flags = ");
1097 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1098 pandecode_log_cont(",\n");
1099
1100 if (fb->clear_stencil)
1101 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1102
1103 if (fb->clear_depth)
1104 pandecode_prop("clear_depth = %f", fb->clear_depth);
1105
1106 const struct midgard_tiler_descriptor t = fb->tiler;
1107 if (!is_compute)
1108 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
1109 else
1110 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1111
1112 if (fb->zero3 || fb->zero4) {
1113 pandecode_msg("XXX: framebuffer zeros tripped\n");
1114 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1115 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1116 }
1117
1118 pandecode_indent--;
1119 pandecode_log("};\n");
1120
1121 gpu_va += sizeof(struct mali_framebuffer);
1122
1123 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1124
1125 if (info.has_extra) {
1126 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1127 const struct mali_framebuffer_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1128
1129 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1130 pandecode_indent++;
1131
1132 MEMORY_PROP(fbx, checksum);
1133
1134 if (fbx->checksum_stride)
1135 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1136
1137 pandecode_log(".flags_hi = ");
1138 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_lo);
1139 pandecode_log_cont(",\n");
1140
1141 pandecode_log(".flags_lo = ");
1142 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
1143 pandecode_log_cont(",\n");
1144
1145 pandecode_prop("zs_block = %s", pandecode_block_format(fbx->zs_block));
1146
1147 if (fbx->zs_block == MALI_BLOCK_AFBC) {
1148 pandecode_log(".ds_afbc = {\n");
1149 pandecode_indent++;
1150
1151 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1152 pandecode_prop("depth_stencil_afbc_stride = %d",
1153 fbx->ds_afbc.depth_stencil_afbc_stride);
1154 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1155
1156 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
1157 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1158 pandecode_prop("zero1 = 0x%" PRIx32,
1159 fbx->ds_afbc.zero1);
1160 pandecode_prop("padding = 0x%" PRIx64,
1161 fbx->ds_afbc.padding);
1162 }
1163
1164 pandecode_indent--;
1165 pandecode_log("},\n");
1166 } else {
1167 pandecode_log(".ds_linear = {\n");
1168 pandecode_indent++;
1169
1170 if (fbx->ds_linear.depth) {
1171 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1172 pandecode_prop("depth_stride = %d",
1173 fbx->ds_linear.depth_stride);
1174 } else if (fbx->ds_linear.depth_stride) {
1175 pandecode_msg("XXX: depth stride zero tripped %d\n", fbx->ds_linear.depth_stride);
1176 }
1177
1178 if (fbx->ds_linear.stencil) {
1179 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1180 pandecode_prop("stencil_stride = %d",
1181 fbx->ds_linear.stencil_stride);
1182 } else if (fbx->ds_linear.stencil_stride) {
1183 pandecode_msg("XXX: stencil stride zero tripped %d\n", fbx->ds_linear.stencil_stride);
1184 }
1185
1186 if (fbx->ds_linear.depth_stride_zero ||
1187 fbx->ds_linear.stencil_stride_zero ||
1188 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1189 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1190 pandecode_prop("depth_stride_zero = 0x%x",
1191 fbx->ds_linear.depth_stride_zero);
1192 pandecode_prop("stencil_stride_zero = 0x%x",
1193 fbx->ds_linear.stencil_stride_zero);
1194 pandecode_prop("zero1 = 0x%" PRIx32,
1195 fbx->ds_linear.zero1);
1196 pandecode_prop("zero2 = 0x%" PRIx32,
1197 fbx->ds_linear.zero2);
1198 }
1199
1200 pandecode_indent--;
1201 pandecode_log("},\n");
1202 }
1203
1204 if (fbx->zero3 || fbx->zero4) {
1205 pandecode_msg("XXX: fb_extra zeros tripped\n");
1206 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1207 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1208 }
1209
1210 pandecode_indent--;
1211 pandecode_log("};\n");
1212
1213 gpu_va += sizeof(struct mali_framebuffer_extra);
1214 }
1215
1216 if (is_fragment)
1217 pandecode_render_target(gpu_va, job_no, fb);
1218
1219 return info;
1220 }
1221
1222 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1223 * count */
1224
1225 static void
1226 pandecode_padded_vertices(unsigned shift, unsigned k)
1227 {
1228 unsigned odd = 2*k + 1;
1229 unsigned pot = 1 << shift;
1230 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1231 }
1232
1233 /* Given a magic divisor, recover what we were trying to divide by.
1234 *
1235 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1236 * 0 <= m < 2^N, for N bits in m.
1237 *
1238 * Let q represent the number we would like to divide by.
1239 *
1240 * By definition of a magic divisor for N-bit unsigned integers (a number you
1241 * multiply by to magically get division), m is a number such that:
1242 *
1243 * (m * x) & (2^N - 1) = floor(x/q).
1244 * for all x on Z where 0 <= x < 2^N
1245 *
1246 * Ignore the case where any of the above values equals zero; it is irrelevant
1247 * for our purposes (instanced arrays).
1248 *
1249 * Choose x = q. Then:
1250 *
1251 * (m * x) & (2^N - 1) = floor(x/q).
1252 * (m * q) & (2^N - 1) = floor(q/q).
1253 *
1254 * floor(q/q) = floor(1) = 1, therefore:
1255 *
1256 * (m * q) & (2^N - 1) = 1
1257 *
1258 * Recall the identity that the bitwise AND of one less than a power-of-two
1259 * equals the modulo with that power of two, i.e. for all x:
1260 *
1261 * x & (2^N - 1) = x % N
1262 *
1263 * Therefore:
1264 *
1265 * mq % (2^N) = 1
1266 *
1267 * By definition, a modular multiplicative inverse of a number m is the number
1268 * q such that with respect to a modulos M:
1269 *
1270 * mq % M = 1
1271 *
1272 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1273 *
1274 */
1275
1276 static void
1277 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1278 {
1279 #if 0
1280 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1281 * shift) the most lame way possible... just repeatedly add.
1282 * Asymptoptically slow but nobody cares in practice, unless you have
1283 * massive numbers of vertices or high divisors. */
1284
1285 unsigned inverse = 0;
1286
1287 /* Magic implicitly has the highest bit set */
1288 magic |= (1 << 31);
1289
1290 /* Depending on rounding direction */
1291 if (extra)
1292 magic++;
1293
1294 for (;;) {
1295 uint32_t product = magic * inverse;
1296
1297 if (shift) {
1298 product >>= shift;
1299 }
1300
1301 if (product == 1)
1302 break;
1303
1304 ++inverse;
1305 }
1306
1307 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1308
1309 /* Recall we're supposed to divide by (gl_level_divisor *
1310 * padded_num_vertices) */
1311
1312 unsigned padded_num_vertices = inverse / orig_divisor;
1313
1314 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1315 #endif
1316 }
1317
1318 static void
1319 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1320 mali_ptr addr, int job_no, char *suffix,
1321 int count, bool varying, enum mali_job_type job_type)
1322 {
1323 char *prefix = varying ? "varying" : "attribute";
1324 assert(addr);
1325
1326 if (!count) {
1327 pandecode_msg("warn: No %s records\n", prefix);
1328 return;
1329 }
1330
1331 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1332
1333 for (int i = 0; i < count; ++i) {
1334 /* First, check for special records */
1335 if (attr[i].elements < MALI_RECORD_SPECIAL) {
1336 if (attr[i].size)
1337 pandecode_msg("XXX: tripped size=%d\n", attr[i].size);
1338
1339 if (attr[i].stride) {
1340 /* gl_InstanceID passes a magic divisor in the
1341 * stride field to divide by the padded vertex
1342 * count. No other records should do so, so
1343 * stride should otherwise be zero. Note that
1344 * stride in the usual attribute sense doesn't
1345 * apply to special records. */
1346
1347 bool has_divisor = attr[i].elements == MALI_ATTR_INSTANCEID;
1348
1349 pandecode_log_cont("/* %smagic divisor = %X */ ",
1350 has_divisor ? "" : "XXX: ", attr[i].stride);
1351 }
1352
1353 if (attr[i].shift || attr[i].extra_flags) {
1354 /* Attributes use these fields for
1355 * instancing/padding/etc type issues, but
1356 * varyings don't */
1357
1358 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1359 varying ? "XXX: " : "",
1360 attr[i].shift, attr[i].extra_flags);
1361 }
1362
1363 /* Print the special record name */
1364 bool attribute = false;
1365 pandecode_log("%s_%d = %s;\n", prefix, i, pandecode_special_record(attr[i].elements, &attribute));
1366
1367 /* Sanity check */
1368 if (attribute == varying)
1369 pandecode_msg("XXX: mismatched special record\n");
1370
1371 continue;
1372 }
1373
1374 enum mali_attr_mode mode = attr[i].elements & 7;
1375
1376 if (mode == MALI_ATTR_UNUSED)
1377 pandecode_msg("XXX: unused attribute record\n");
1378
1379 /* For non-linear records, we need to print the type of record */
1380 if (mode != MALI_ATTR_LINEAR)
1381 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1382
1383 /* Print the name to link with attr_meta */
1384 pandecode_log_cont("%s_%d", prefix, i);
1385
1386 /* Print the stride and size */
1387 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1388
1389 /* TODO: Sanity check the quotient itself. It must be equal to
1390 * (or be greater than, if the driver added padding) the padded
1391 * vertex count. */
1392
1393 /* Finally, print the pointer */
1394 mali_ptr raw_elements = attr[i].elements & ~7;
1395 char *a = pointer_as_memory_reference(raw_elements);
1396 pandecode_log_cont(" = (%s);\n", a);
1397 free(a);
1398
1399 /* Check the pointer */
1400 pandecode_validate_buffer(raw_elements, attr[i].size);
1401
1402 /* shift/extra_flags exist only for instanced */
1403 if (attr[i].shift | attr[i].extra_flags) {
1404 /* These are set to random values by the blob for
1405 * varyings, most likely a symptom of uninitialized
1406 * memory where the hardware masked the bug. As such we
1407 * put this at a warning, not an error. */
1408
1409 if (mode == MALI_ATTR_LINEAR)
1410 pandecode_msg("warn: instancing fields set for linear\n");
1411
1412 pandecode_prop("shift = %d", attr[i].shift);
1413 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1414 }
1415
1416 /* Decode further where possible */
1417
1418 if (mode == MALI_ATTR_MODULO) {
1419 pandecode_padded_vertices(
1420 attr[i].shift,
1421 attr[i].extra_flags);
1422 }
1423
1424 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1425 i++;
1426 pandecode_log("{\n");
1427 pandecode_indent++;
1428 pandecode_prop("unk = 0x%x", attr[i].unk);
1429 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1430 if (attr[i].zero != 0)
1431 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1432 pandecode_prop("divisor = %d", attr[i].divisor);
1433 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1434 pandecode_indent--;
1435 pandecode_log("}, \n");
1436 }
1437
1438 }
1439
1440 pandecode_log("\n");
1441 }
1442
1443 static mali_ptr
1444 pandecode_shader_address(const char *name, mali_ptr ptr)
1445 {
1446 /* TODO: Decode flags */
1447 mali_ptr shader_ptr = ptr & ~15;
1448
1449 char *a = pointer_as_memory_reference(shader_ptr);
1450 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1451 free(a);
1452
1453 return shader_ptr;
1454 }
1455
1456 static void
1457 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1458 {
1459 unsigned any_nonzero =
1460 stencil->ref | stencil->mask | stencil->func |
1461 stencil->sfail | stencil->dpfail | stencil->dppass;
1462
1463 if (any_nonzero == 0)
1464 return;
1465
1466 const char *func = pandecode_func(stencil->func);
1467 const char *sfail = pandecode_stencil_op(stencil->sfail);
1468 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1469 const char *dppass = pandecode_stencil_op(stencil->dppass);
1470
1471 if (stencil->zero)
1472 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1473
1474 pandecode_log(".stencil_%s = {\n", name);
1475 pandecode_indent++;
1476 pandecode_prop("ref = %d", stencil->ref);
1477 pandecode_prop("mask = 0x%02X", stencil->mask);
1478 pandecode_prop("func = %s", func);
1479 pandecode_prop("sfail = %s", sfail);
1480 pandecode_prop("dpfail = %s", dpfail);
1481 pandecode_prop("dppass = %s", dppass);
1482 pandecode_indent--;
1483 pandecode_log("},\n");
1484 }
1485
1486 static void
1487 pandecode_blend_equation(const struct mali_blend_equation *blend)
1488 {
1489 if (blend->zero1)
1490 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1491
1492 pandecode_log(".equation = {\n");
1493 pandecode_indent++;
1494
1495 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1496 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1497
1498 pandecode_log(".color_mask = ");
1499 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1500 pandecode_log_cont(",\n");
1501
1502 pandecode_indent--;
1503 pandecode_log("},\n");
1504 }
1505
1506 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1507
1508 static unsigned
1509 decode_bifrost_constant(u16 constant)
1510 {
1511 float lo = (float) (constant & 0xFF);
1512 float hi = (float) (constant >> 8);
1513
1514 return (hi / 255.0) + (lo / 65535.0);
1515 }
1516
1517 static mali_ptr
1518 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1519 {
1520 struct bifrost_blend_rt *b =
1521 ((struct bifrost_blend_rt *) descs) + rt_no;
1522
1523 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1524 pandecode_indent++;
1525
1526 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1527 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1528 b->constant, decode_bifrost_constant(b->constant));
1529
1530 /* TODO figure out blend shader enable bit */
1531 pandecode_blend_equation(&b->equation);
1532 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1533 pandecode_prop("index = 0x%" PRIx16, b->index);
1534 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1535
1536 pandecode_indent--;
1537 pandecode_log("},\n");
1538
1539 return 0;
1540 }
1541
1542 static mali_ptr
1543 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1544 {
1545 /* constant/equation is in a union */
1546 if (!blend->shader)
1547 return 0;
1548
1549 pandecode_log(".blend = {\n");
1550 pandecode_indent++;
1551
1552 if (is_shader) {
1553 pandecode_shader_address("shader", blend->shader);
1554 } else {
1555 pandecode_blend_equation(&blend->equation);
1556 pandecode_prop("constant = %f", blend->constant);
1557 }
1558
1559 pandecode_indent--;
1560 pandecode_log("},\n");
1561
1562 /* Return blend shader to disassemble if present */
1563 return is_shader ? (blend->shader & ~0xF) : 0;
1564 }
1565
1566 static mali_ptr
1567 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1568 {
1569 struct midgard_blend_rt *b =
1570 ((struct midgard_blend_rt *) descs) + rt_no;
1571
1572 /* Flags determine presence of blend shader */
1573 bool is_shader = (b->flags & 0xF) >= 0x2;
1574
1575 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1576 pandecode_indent++;
1577
1578 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1579
1580 union midgard_blend blend = b->blend;
1581 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1582
1583 pandecode_indent--;
1584 pandecode_log("};\n");
1585
1586 return shader;
1587 }
1588
1589 /* Attributes and varyings have descriptor records, which contain information
1590 * about their format and ordering with the attribute/varying buffers. We'll
1591 * want to validate that the combinations specified are self-consistent.
1592 */
1593
1594 static int
1595 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1596 {
1597 char base[128];
1598 char *prefix = varying ? "varying" : "attribute";
1599 unsigned max_index = 0;
1600 snprintf(base, sizeof(base), "%s_meta", prefix);
1601
1602 struct mali_attr_meta *attr_meta;
1603 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1604
1605 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1606
1607 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1608 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1609 sizeof(*attr_mem));
1610
1611 /* If the record is discard, it should be zero for everything else */
1612
1613 if (attr_meta->format == MALI_VARYING_DISCARD) {
1614 uint64_t zero =
1615 attr_meta->index |
1616 attr_meta->unknown1 |
1617 attr_meta->unknown3 |
1618 attr_meta->src_offset;
1619
1620 if (zero)
1621 pandecode_msg("XXX: expected empty record for varying discard\n");
1622
1623 /* We want to look for a literal 0000 swizzle -- this
1624 * is not encoded with all zeroes, however */
1625
1626 enum mali_channel z = MALI_CHANNEL_ZERO;
1627 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1628 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1629
1630 if (!good_swizzle)
1631 pandecode_msg("XXX: expected zero swizzle for discard\n");
1632
1633 if (!varying)
1634 pandecode_msg("XXX: cannot discard attribute\n");
1635
1636 /* If we're all good, omit the record */
1637 if (!zero && varying && good_swizzle) {
1638 pandecode_log("/* discarded varying */\n");
1639 continue;
1640 }
1641 }
1642
1643 if (attr_meta->index > max_index)
1644 max_index = attr_meta->index;
1645
1646 if (attr_meta->unknown1 != 0x2) {
1647 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1648 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1649 }
1650
1651 if (attr_meta->unknown3) {
1652 pandecode_msg("XXX: unexpected unknown3 set\n");
1653 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1654 }
1655
1656 pandecode_format_short(attr_meta->format, false);
1657 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1658
1659 if (attr_meta->src_offset)
1660 pandecode_log_cont("[%u]", attr_meta->src_offset);
1661
1662 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1663
1664 pandecode_log_cont(";\n");
1665 }
1666
1667 pandecode_log("\n");
1668
1669 return count ? (max_index + 1) : 0;
1670 }
1671
1672 /* return bits [lo, hi) of word */
1673 static u32
1674 bits(u32 word, u32 lo, u32 hi)
1675 {
1676 if (hi - lo >= 32)
1677 return word; // avoid undefined behavior with the shift
1678
1679 return (word >> lo) & ((1 << (hi - lo)) - 1);
1680 }
1681
1682 static void
1683 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1684 {
1685 pandecode_log_cont("{\n");
1686 pandecode_indent++;
1687
1688 /* Decode invocation_count. See the comment before the definition of
1689 * invocation_count for an explanation.
1690 */
1691
1692 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1693 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1694 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1695 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1696 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1697 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1698
1699 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1700 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1701 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1702
1703 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1704 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1705 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1706
1707 /* Even though we have this decoded, we want to ensure that the
1708 * representation is "unique" so we don't lose anything by printing only
1709 * the final result. More specifically, we need to check that we were
1710 * passed something in canonical form, since the definition per the
1711 * hardware is inherently not unique. How? Well, take the resulting
1712 * decode and pack it ourselves! If it is bit exact with what we
1713 * decoded, we're good to go. */
1714
1715 struct mali_vertex_tiler_prefix ref;
1716 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1717
1718 bool canonical =
1719 (p->invocation_count == ref.invocation_count) &&
1720 (p->invocation_shifts == ref.invocation_shifts);
1721
1722 if (!canonical) {
1723 pandecode_msg("XXX: non-canonical workgroups packing\n");
1724 pandecode_msg("expected: %X, %X",
1725 ref.invocation_count,
1726 ref.invocation_shifts);
1727
1728 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1729 pandecode_prop("size_y_shift = %d", size_y_shift);
1730 pandecode_prop("size_z_shift = %d", size_z_shift);
1731 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
1732 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
1733 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
1734 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
1735 }
1736
1737 /* Regardless, print the decode */
1738 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1739 size_x, size_y, size_z,
1740 groups_x, groups_y, groups_z);
1741
1742 /* TODO: Decode */
1743 if (p->unknown_draw)
1744 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1745
1746 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1747
1748 if (p->draw_mode != MALI_DRAW_NONE)
1749 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1750
1751 /* Index count only exists for tiler jobs anyway */
1752
1753 if (p->index_count)
1754 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1755
1756
1757 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1758 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1759
1760 /* Validate an index buffer is present if we need one. TODO: verify
1761 * relationship between invocation_count and index_count */
1762
1763 if (p->indices) {
1764 unsigned count = p->index_count;
1765
1766 /* Grab the size */
1767 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1768
1769 /* Ensure we got a size, and if so, validate the index buffer
1770 * is large enough to hold a full set of indices of the given
1771 * size */
1772
1773 if (!index_raw_size)
1774 pandecode_msg("XXX: index size missing\n");
1775 else
1776 pandecode_validate_buffer(p->indices, count * size);
1777 } else if (index_raw_size)
1778 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1779
1780 if (p->offset_bias_correction)
1781 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1782
1783 /* TODO: Figure out what this is. It's not zero */
1784 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1785
1786 pandecode_indent--;
1787 pandecode_log("},\n");
1788 }
1789
1790 static void
1791 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1792 {
1793 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1794 uint64_t *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1795
1796 for (int i = 0; i < ubufs_count; i++) {
1797 unsigned size = (ubufs[i] & ((1 << 10) - 1)) * 16;
1798 mali_ptr addr = (ubufs[i] >> 10) << 2;
1799
1800 pandecode_validate_buffer(addr, size);
1801
1802 char *ptr = pointer_as_memory_reference(addr);
1803 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1804 free(ptr);
1805 }
1806
1807 pandecode_log("\n");
1808 }
1809
1810 static void
1811 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1812 {
1813 pandecode_validate_buffer(uniforms, uniform_count * 16);
1814
1815 char *ptr = pointer_as_memory_reference(uniforms);
1816 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1817 free(ptr);
1818 }
1819
1820 static const char *
1821 shader_type_for_job(unsigned type)
1822 {
1823 switch (type) {
1824 case JOB_TYPE_VERTEX: return "VERTEX";
1825 case JOB_TYPE_TILER: return "FRAGMENT";
1826 case JOB_TYPE_COMPUTE: return "COMPUTE";
1827 default:
1828 return "UNKNOWN";
1829 }
1830 }
1831
1832 static unsigned shader_id = 0;
1833
1834 static struct midgard_disasm_stats
1835 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1836 bool is_bifrost, unsigned gpu_id)
1837 {
1838 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1839 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1840
1841 /* Compute maximum possible size */
1842 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1843
1844 /* Print some boilerplate to clearly denote the assembly (which doesn't
1845 * obey indentation rules), and actually do the disassembly! */
1846
1847 pandecode_log_cont("\n\n");
1848
1849 struct midgard_disasm_stats stats;
1850
1851 if (is_bifrost) {
1852 disassemble_bifrost(pandecode_dump_stream, code, sz, false);
1853
1854 /* TODO: Extend stats to Bifrost */
1855 stats.texture_count = -128;
1856 stats.sampler_count = -128;
1857 stats.attribute_count = -128;
1858 stats.varying_count = -128;
1859 stats.uniform_count = -128;
1860 stats.uniform_buffer_count = -128;
1861 stats.work_count = -128;
1862
1863 stats.instruction_count = 0;
1864 stats.bundle_count = 0;
1865 stats.quadword_count = 0;
1866 stats.helper_invocations = false;
1867 } else {
1868 stats = disassemble_midgard(pandecode_dump_stream,
1869 code, sz, gpu_id,
1870 type == JOB_TYPE_TILER ?
1871 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
1872 }
1873
1874 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1875 * driver-internal purposes with the blob and interfere */
1876
1877 bool should_shaderdb = type != JOB_TYPE_COMPUTE;
1878
1879 if (should_shaderdb) {
1880 unsigned nr_threads =
1881 (stats.work_count <= 4) ? 4 :
1882 (stats.work_count <= 8) ? 2 :
1883 1;
1884
1885 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
1886 "%u inst, %u bundles, %u quadwords, "
1887 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1888 shader_id++,
1889 shader_type_for_job(type),
1890 stats.instruction_count, stats.bundle_count, stats.quadword_count,
1891 stats.work_count, nr_threads);
1892 }
1893
1894
1895 return stats;
1896 }
1897
1898 static void
1899 pandecode_texture(mali_ptr u,
1900 struct pandecode_mapped_memory *tmem,
1901 unsigned job_no, unsigned tex)
1902 {
1903 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
1904
1905 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
1906 pandecode_indent++;
1907
1908 struct mali_texture_format f = t->format;
1909
1910 /* See the definiton of enum mali_texture_type */
1911
1912 bool is_cube = f.type == MALI_TEX_CUBE;
1913 unsigned dimension = is_cube ? 2 : f.type;
1914
1915 pandecode_make_indent();
1916
1917 /* TODO: Are there others? */
1918 bool is_zs = f.format == MALI_Z32_UNORM;
1919
1920 /* Recall Z/S switched the meaning of linear/tiled .. */
1921 if (is_zs && f.layout == MALI_TEXTURE_LINEAR)
1922 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1923
1924 /* Print the layout. Default is linear; a modifier can denote AFBC or
1925 * u-interleaved/tiled modes */
1926
1927 if (f.layout == MALI_TEXTURE_AFBC)
1928 pandecode_log_cont("afbc");
1929 else if (f.layout == MALI_TEXTURE_TILED)
1930 pandecode_log_cont("tiled");
1931 else if (f.layout == MALI_TEXTURE_LINEAR)
1932 pandecode_log_cont("linear");
1933 else
1934 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
1935
1936 pandecode_swizzle(t->swizzle, f.format);
1937 pandecode_log_cont(" ");
1938
1939 /* Distinguish cube/2D with modifier */
1940
1941 if (is_cube)
1942 pandecode_log_cont("cube ");
1943
1944 pandecode_format_short(f.format, f.srgb);
1945 pandecode_swizzle(f.swizzle, f.format);
1946
1947 /* All four width/height/depth/array_size dimensions are present
1948 * regardless of the type of texture, but it is an error to have
1949 * non-zero dimensions for unused dimensions. Verify this. array_size
1950 * can always be set, as can width. */
1951
1952 if (t->height && dimension < 2)
1953 pandecode_msg("XXX: nonzero height for <2D texture\n");
1954
1955 if (t->depth && dimension < 3)
1956 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1957
1958 /* Print only the dimensions that are actually there */
1959
1960 pandecode_log_cont(": %d", t->width + 1);
1961
1962 if (dimension >= 2)
1963 pandecode_log_cont("x%u", t->height + 1);
1964
1965 if (dimension >= 3)
1966 pandecode_log_cont("x%u", t->depth + 1);
1967
1968 if (t->array_size)
1969 pandecode_log_cont("[%u]", t->array_size + 1);
1970
1971 if (t->levels)
1972 pandecode_log_cont(" mip %u", t->levels);
1973
1974 pandecode_log_cont("\n");
1975
1976 if (f.unknown1 | f.zero) {
1977 pandecode_msg("XXX: texture format zero tripped\n");
1978 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
1979 pandecode_prop("zero = %" PRId32, f.zero);
1980 }
1981
1982 if (!f.unknown2) {
1983 pandecode_msg("XXX: expected unknown texture bit set\n");
1984 pandecode_prop("unknown2 = %" PRId32, f.unknown2);
1985 }
1986
1987 if (t->swizzle_zero) {
1988 pandecode_msg("XXX: swizzle zero tripped\n");
1989 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
1990 }
1991
1992 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
1993 pandecode_msg("XXX: texture zero tripped\n");
1994 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
1995 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
1996 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
1997 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
1998 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
1999 }
2000
2001 pandecode_log(".payload = {\n");
2002 pandecode_indent++;
2003
2004 /* A bunch of bitmap pointers follow.
2005 * We work out the correct number,
2006 * based on the mipmap/cubemap
2007 * properties, but dump extra
2008 * possibilities to futureproof */
2009
2010 int bitmap_count = t->levels + 1;
2011
2012 /* Miptree for each face */
2013 if (f.type == MALI_TEX_CUBE)
2014 bitmap_count *= 6;
2015 else if (f.type == MALI_TEX_3D && f.layout == MALI_TEXTURE_LINEAR)
2016 bitmap_count *= (t->depth + 1);
2017
2018 /* Array of textures */
2019 bitmap_count *= (t->array_size + 1);
2020
2021 /* Stride for each element */
2022 if (f.manual_stride)
2023 bitmap_count *= 2;
2024
2025 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
2026 u + sizeof(*t), sizeof(mali_ptr) * bitmap_count);
2027 for (int i = 0; i < bitmap_count; ++i) {
2028 /* How we dump depends if this is a stride or a pointer */
2029
2030 if (f.manual_stride && (i & 1)) {
2031 /* signed 32-bit snuck in as a 64-bit pointer */
2032 uint64_t stride_set = pointers_and_strides[i];
2033 uint32_t clamped_stride = stride_set;
2034 int32_t stride = clamped_stride;
2035 assert(stride_set == clamped_stride);
2036 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2037 } else {
2038 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
2039 pandecode_log("%s, \n", a);
2040 free(a);
2041 }
2042 }
2043
2044 pandecode_indent--;
2045 pandecode_log("},\n");
2046
2047 pandecode_indent--;
2048 pandecode_log("};\n");
2049 }
2050
2051 /* 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 */
2052
2053 static void
2054 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2055 {
2056 /* Nothing to do */
2057 if (claim == truth)
2058 return;
2059
2060 if (fuzzy)
2061 assert(truth >= 0);
2062
2063 if ((truth >= 0) && !fuzzy) {
2064 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2065 (truth < claim) ? "warn" : "XXX",
2066 name, truth, claim);
2067 } else if ((claim > -truth) && !fuzzy) {
2068 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2069 name, -truth, claim);
2070 } else if (fuzzy && (claim < truth))
2071 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2072 name, truth, claim);
2073
2074 pandecode_log(".%s = %" PRId16, name, claim);
2075
2076 if (fuzzy)
2077 pandecode_log_cont(" /* %u used */", truth);
2078
2079 pandecode_log_cont(",\n");
2080 }
2081
2082 static void
2083 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2084 bool is_bifrost, unsigned gpu_id)
2085 {
2086 struct midgard_disasm_stats stats =
2087 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
2088
2089 bool has_texture = (stats.texture_count > 0);
2090 bool has_sampler = (stats.sampler_count > 0);
2091 bool has_attribute = (stats.attribute_count > 0);
2092 bool has_varying = (stats.varying_count > 0);
2093 bool has_uniform = (stats.uniform_count > 0);
2094 bool has_ubo = (stats.uniform_buffer_count > 0);
2095
2096 if (has_texture || has_sampler)
2097 pandecode_msg("XXX: blend shader accessing textures\n");
2098
2099 if (has_attribute || has_varying)
2100 pandecode_msg("XXX: blend shader accessing interstage\n");
2101
2102 if (has_uniform || has_ubo)
2103 pandecode_msg("XXX: blend shader accessing uniforms\n");
2104 }
2105
2106 static void
2107 pandecode_vertex_tiler_postfix_pre(
2108 const struct mali_vertex_tiler_postfix *p,
2109 int job_no, enum mali_job_type job_type,
2110 char *suffix, bool is_bifrost, unsigned gpu_id)
2111 {
2112 struct pandecode_mapped_memory *attr_mem;
2113
2114 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2115 * are the only things actually needed from the FBD, vertex/tiler jobs
2116 * no longer reference the FBD -- instead, this field points to some
2117 * info about the scratchpad.
2118 */
2119
2120 struct pandecode_fbd fbd_info = {
2121 /* Default for Bifrost */
2122 .rt_count = 1
2123 };
2124
2125 if (is_bifrost) {
2126 pandecode_log_cont("\t/* %X %/\n", p->shared_memory & 1);
2127 pandecode_compute_fbd(p->shared_memory & ~1, job_no);
2128 } else if (p->shared_memory & MALI_MFBD)
2129 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->shared_memory) & FBD_MASK, job_no, false, job_type == JOB_TYPE_COMPUTE);
2130 else if (job_type == JOB_TYPE_COMPUTE)
2131 pandecode_compute_fbd((u64) (uintptr_t) p->shared_memory, job_no);
2132 else
2133 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->shared_memory, job_no, false, gpu_id);
2134
2135 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2136 int texture_count = 0, sampler_count = 0;
2137
2138 if (p->shader) {
2139 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2140 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2141
2142 /* Disassemble ahead-of-time to get stats. Initialize with
2143 * stats for the missing-shader case so we get validation
2144 * there, too */
2145
2146 struct midgard_disasm_stats info = {
2147 .texture_count = 0,
2148 .sampler_count = 0,
2149 .attribute_count = 0,
2150 .varying_count = 0,
2151 .work_count = 1,
2152
2153 .uniform_count = -128,
2154 .uniform_buffer_count = 0
2155 };
2156
2157 if (s->shader & ~0xF)
2158 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
2159
2160 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2161 pandecode_indent++;
2162
2163 /* Save for dumps */
2164 attribute_count = s->attribute_count;
2165 varying_count = s->varying_count;
2166 texture_count = s->texture_count;
2167 sampler_count = s->sampler_count;
2168
2169 if (is_bifrost) {
2170 uniform_count = s->bifrost2.uniform_count;
2171 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2172 } else {
2173 uniform_count = s->midgard1.uniform_count;
2174 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2175 }
2176
2177 pandecode_shader_address("shader", s->shader);
2178
2179 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2180 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2181 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2182 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2183 pandecode_shader_prop("uniform_buffer_count",
2184 uniform_buffer_count,
2185 info.uniform_buffer_count, true);
2186
2187 if (!is_bifrost) {
2188 pandecode_shader_prop("uniform_count",
2189 uniform_count,
2190 info.uniform_count, false);
2191
2192 pandecode_shader_prop("work_count",
2193 s->midgard1.work_count, info.work_count, false);
2194 }
2195
2196 if (is_bifrost) {
2197 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32, s->bifrost1.unk1);
2198 } else {
2199 bool helpers = s->midgard1.flags_lo & MALI_HELPER_INVOCATIONS;
2200 s->midgard1.flags_lo &= ~MALI_HELPER_INVOCATIONS;
2201
2202 if (helpers != info.helper_invocations) {
2203 pandecode_msg("XXX: expected helpers %u but got %u\n",
2204 info.helper_invocations, helpers);
2205 }
2206
2207 pandecode_log(".midgard1.flags_lo = ");
2208 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info, s->midgard1.flags_lo);
2209 pandecode_log_cont(",\n");
2210
2211 pandecode_log(".midgard1.flags_hi = ");
2212 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info, s->midgard1.flags_hi);
2213 pandecode_log_cont(",\n");
2214 }
2215
2216 if (s->depth_units || s->depth_factor) {
2217 pandecode_prop("depth_factor = %f", s->depth_factor);
2218 pandecode_prop("depth_units = %f", s->depth_units);
2219 }
2220
2221 if (s->alpha_coverage) {
2222 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
2223 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
2224
2225 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2226 invert_alpha_coverage ? "~" : "",
2227 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
2228 }
2229
2230 if (s->unknown2_3 || s->unknown2_4) {
2231 pandecode_log(".unknown2_3 = ");
2232
2233 int unknown2_3 = s->unknown2_3;
2234 int unknown2_4 = s->unknown2_4;
2235
2236 /* We're not quite sure what these flags mean without the depth test, if anything */
2237
2238 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2239 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2240 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2241
2242 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2243 }
2244
2245 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2246 pandecode_log_cont(",\n");
2247
2248 pandecode_log(".unknown2_4 = ");
2249 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2250 pandecode_log_cont(",\n");
2251 }
2252
2253 if (s->stencil_mask_front || s->stencil_mask_back) {
2254 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2255 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2256 }
2257
2258 pandecode_stencil("front", &s->stencil_front);
2259 pandecode_stencil("back", &s->stencil_back);
2260
2261 if (is_bifrost) {
2262 pandecode_log(".bifrost2 = {\n");
2263 pandecode_indent++;
2264
2265 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2266 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2267 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2268 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2269
2270 pandecode_indent--;
2271 pandecode_log("},\n");
2272 } else if (s->midgard2.unknown2_7) {
2273 pandecode_log(".midgard2 = {\n");
2274 pandecode_indent++;
2275
2276 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2277 pandecode_indent--;
2278 pandecode_log("},\n");
2279 }
2280
2281 if (s->padding) {
2282 pandecode_msg("XXX: shader padding tripped\n");
2283 pandecode_prop("padding = 0x%" PRIx32, s->padding);
2284 }
2285
2286 if (!is_bifrost) {
2287 /* TODO: Blend shaders routing/disasm */
2288 union midgard_blend blend = s->blend;
2289 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2290 if (shader & ~0xF)
2291 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2292 }
2293
2294 pandecode_indent--;
2295 pandecode_log("};\n");
2296
2297 /* MRT blend fields are used whenever MFBD is used, with
2298 * per-RT descriptors */
2299
2300 if (job_type == JOB_TYPE_TILER && p->shared_memory & MALI_MFBD) {
2301 void* blend_base = (void *) (s + 1);
2302
2303 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2304 mali_ptr shader = 0;
2305
2306 if (is_bifrost)
2307 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2308 else
2309 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2310
2311 if (shader & ~0xF)
2312 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2313
2314 }
2315 }
2316 } else
2317 pandecode_msg("XXX: missing shader descriptor\n");
2318
2319 if (p->viewport) {
2320 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2321 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2322
2323 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2324 pandecode_indent++;
2325
2326 pandecode_prop("clip_minx = %f", f->clip_minx);
2327 pandecode_prop("clip_miny = %f", f->clip_miny);
2328 pandecode_prop("clip_minz = %f", f->clip_minz);
2329 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2330 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2331 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2332
2333 /* Only the higher coordinates are MALI_POSITIVE scaled */
2334
2335 pandecode_prop("viewport0 = { %d, %d }",
2336 f->viewport0[0], f->viewport0[1]);
2337
2338 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2339 f->viewport1[0] + 1, f->viewport1[1] + 1);
2340
2341 pandecode_indent--;
2342 pandecode_log("};\n");
2343 }
2344
2345 unsigned max_attr_index = 0;
2346
2347 if (p->attribute_meta)
2348 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2349
2350 if (p->attributes) {
2351 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2352 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2353 }
2354
2355 /* Varyings are encoded like attributes but not actually sent; we just
2356 * pass a zero buffer with the right stride/size set, (or whatever)
2357 * since the GPU will write to it itself */
2358
2359 if (p->varying_meta) {
2360 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2361 }
2362
2363 if (p->varyings) {
2364 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2365
2366 /* Number of descriptors depends on whether there are
2367 * non-internal varyings */
2368
2369 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2370 }
2371
2372 if (p->uniform_buffers) {
2373 if (uniform_buffer_count)
2374 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2375 else
2376 pandecode_msg("warn: UBOs specified but not referenced\n");
2377 } else if (uniform_buffer_count)
2378 pandecode_msg("XXX: UBOs referenced but not specified\n");
2379
2380 /* We don't want to actually dump uniforms, but we do need to validate
2381 * that the counts we were given are sane */
2382
2383 if (p->uniforms) {
2384 if (uniform_count)
2385 pandecode_uniforms(p->uniforms, uniform_count);
2386 else
2387 pandecode_msg("warn: Uniforms specified but not referenced\n");
2388 } else if (uniform_count)
2389 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2390
2391 if (p->texture_trampoline) {
2392 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
2393
2394 if (mmem) {
2395 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
2396
2397 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
2398 pandecode_indent++;
2399
2400 for (int tex = 0; tex < texture_count; ++tex) {
2401 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2402 char *a = pointer_as_memory_reference(*u);
2403 pandecode_log("%s,\n", a);
2404 free(a);
2405 }
2406
2407 pandecode_indent--;
2408 pandecode_log("};\n");
2409
2410 /* Now, finally, descend down into the texture descriptor */
2411 for (unsigned tex = 0; tex < texture_count; ++tex) {
2412 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2413 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2414 if (tmem)
2415 pandecode_texture(*u, tmem, job_no, tex);
2416 }
2417 }
2418 }
2419
2420 if (p->sampler_descriptor) {
2421 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2422
2423 if (smem) {
2424 struct mali_sampler_descriptor *s;
2425
2426 mali_ptr d = p->sampler_descriptor;
2427
2428 for (int i = 0; i < sampler_count; ++i) {
2429 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2430
2431 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2432 pandecode_indent++;
2433
2434 pandecode_log(".filter_mode = ");
2435 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2436 pandecode_log_cont(",\n");
2437
2438 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2439 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2440
2441 if (s->lod_bias)
2442 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s->lod_bias));
2443
2444 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2445 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2446 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2447
2448 pandecode_prop("compare_func = %s", pandecode_func(s->compare_func));
2449
2450 if (s->zero || s->zero2) {
2451 pandecode_msg("XXX: sampler zero tripped\n");
2452 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2453 }
2454
2455 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2456
2457 pandecode_prop("border_color = { %f, %f, %f, %f }",
2458 s->border_color[0],
2459 s->border_color[1],
2460 s->border_color[2],
2461 s->border_color[3]);
2462
2463 pandecode_indent--;
2464 pandecode_log("};\n");
2465 }
2466 }
2467 }
2468 }
2469
2470 static void
2471 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2472 {
2473 if (p->shader & 0xF)
2474 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2475
2476 if (!(p->position_varying || p->occlusion_counter))
2477 return;
2478
2479 pandecode_log(".postfix = {\n");
2480 pandecode_indent++;
2481
2482 MEMORY_PROP(p, position_varying);
2483 MEMORY_PROP(p, occlusion_counter);
2484
2485 pandecode_indent--;
2486 pandecode_log("},\n");
2487 }
2488
2489 static void
2490 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2491 {
2492 pandecode_log_cont("{\n");
2493 pandecode_indent++;
2494
2495 pandecode_prop("unk2 = 0x%x", v->unk2);
2496
2497 if (v->zero0 || v->zero1) {
2498 pandecode_msg("XXX: vertex only zero tripped");
2499 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2500 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2501 }
2502
2503 pandecode_indent--;
2504 pandecode_log("}\n");
2505 }
2506
2507 static void
2508 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2509 {
2510
2511 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2512 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2513
2514 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2515 pandecode_indent++;
2516
2517 if (h->zero) {
2518 pandecode_msg("XXX: tiler heap zero tripped\n");
2519 pandecode_prop("zero = 0x%x", h->zero);
2520 }
2521
2522 for (int i = 0; i < 12; i++) {
2523 if (h->zeros[i] != 0) {
2524 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2525 i, h->zeros[i]);
2526 }
2527 }
2528
2529 pandecode_prop("heap_size = 0x%x", h->heap_size);
2530 MEMORY_PROP(h, tiler_heap_start);
2531 MEMORY_PROP(h, tiler_heap_free);
2532
2533 /* this might point to the beginning of another buffer, when it's
2534 * really the end of the tiler heap buffer, so we have to be careful
2535 * here. but for zero length, we need the same pointer.
2536 */
2537
2538 if (h->tiler_heap_end == h->tiler_heap_start) {
2539 MEMORY_PROP(h, tiler_heap_start);
2540 } else {
2541 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2542 pandecode_prop("tiler_heap_end = %s + 1", a);
2543 free(a);
2544 }
2545
2546 pandecode_indent--;
2547 pandecode_log("};\n");
2548 }
2549
2550 static void
2551 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2552 {
2553 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2554 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2555
2556 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2557
2558 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2559 pandecode_indent++;
2560
2561 if (t->zero0 || t->zero1) {
2562 pandecode_msg("XXX: tiler meta zero tripped\n");
2563 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2564 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2565 }
2566
2567 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2568 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2569
2570 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2571 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2572
2573 for (int i = 0; i < 12; i++) {
2574 if (t->zeros[i] != 0) {
2575 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2576 i, t->zeros[i]);
2577 }
2578 }
2579
2580 pandecode_indent--;
2581 pandecode_log("};\n");
2582 }
2583
2584 static void
2585 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2586 {
2587 pandecode_log(".gl_enables = ");
2588
2589 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2590
2591 pandecode_log_cont(",\n");
2592 }
2593
2594 static void
2595 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2596 {
2597 if (u.pointer == 0x0)
2598 return;
2599
2600 pandecode_log(".primitive_size = {\n");
2601 pandecode_indent++;
2602
2603 if (constant) {
2604 pandecode_prop("constant = %f", u.constant);
2605 } else {
2606 MEMORY_PROP((&u), pointer);
2607 }
2608
2609 pandecode_indent--;
2610 pandecode_log("},\n");
2611 }
2612
2613 static void
2614 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2615 {
2616 pandecode_log_cont("{\n");
2617 pandecode_indent++;
2618
2619 /* TODO: gl_PointSize on Bifrost */
2620 pandecode_primitive_size(t->primitive_size, true);
2621
2622 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2623
2624 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2625 || t->zero6 || t->zero7 || t->zero8) {
2626 pandecode_msg("XXX: tiler only zero tripped\n");
2627 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2628 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2629 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2630 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2631 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2632 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2633 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2634 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2635 }
2636
2637 pandecode_indent--;
2638 pandecode_log("},\n");
2639 }
2640
2641 static int
2642 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2643 const struct pandecode_mapped_memory *mem,
2644 mali_ptr payload, int job_no, unsigned gpu_id)
2645 {
2646 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2647
2648 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
2649
2650 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2651 pandecode_indent++;
2652
2653 pandecode_log(".prefix = ");
2654 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2655
2656 pandecode_log(".vertex = ");
2657 pandecode_vertex_only_bfr(&v->vertex);
2658
2659 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2660
2661 pandecode_indent--;
2662 pandecode_log("};\n");
2663
2664 return sizeof(*v);
2665 }
2666
2667 static int
2668 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2669 const struct pandecode_mapped_memory *mem,
2670 mali_ptr payload, int job_no, unsigned gpu_id)
2671 {
2672 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2673
2674 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
2675 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2676
2677 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2678 pandecode_indent++;
2679
2680 pandecode_log(".prefix = ");
2681 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2682
2683 pandecode_log(".tiler = ");
2684 pandecode_tiler_only_bfr(&t->tiler, job_no);
2685
2686 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2687
2688 pandecode_indent--;
2689 pandecode_log("};\n");
2690
2691 return sizeof(*t);
2692 }
2693
2694 static int
2695 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2696 const struct pandecode_mapped_memory *mem,
2697 mali_ptr payload, int job_no, unsigned gpu_id)
2698 {
2699 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2700
2701 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
2702
2703 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2704 pandecode_indent++;
2705
2706 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2707 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2708
2709 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2710
2711 pandecode_log(".prefix = ");
2712 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
2713
2714 pandecode_gl_enables(v->gl_enables, h->job_type);
2715
2716 if (v->instance_shift || v->instance_odd) {
2717 pandecode_prop("instance_shift = 0x%d /* %d */",
2718 v->instance_shift, 1 << v->instance_shift);
2719 pandecode_prop("instance_odd = 0x%X /* %d */",
2720 v->instance_odd, (2 * v->instance_odd) + 1);
2721
2722 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2723 }
2724
2725 if (v->offset_start)
2726 pandecode_prop("offset_start = %d", v->offset_start);
2727
2728 if (v->zero5) {
2729 pandecode_msg("XXX: midgard payload zero tripped\n");
2730 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2731 }
2732
2733 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2734
2735 pandecode_indent--;
2736 pandecode_log("};\n");
2737
2738 return sizeof(*v);
2739 }
2740
2741 static int
2742 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2743 mali_ptr payload, int job_no,
2744 bool is_bifrost, unsigned gpu_id)
2745 {
2746 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2747
2748 bool is_mfbd = s->framebuffer & MALI_MFBD;
2749
2750 if (!is_mfbd && is_bifrost)
2751 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2752
2753 struct pandecode_fbd info;
2754
2755 if (is_mfbd)
2756 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true, false);
2757 else
2758 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
2759
2760 /* Compute the tag for the tagged pointer. This contains the type of
2761 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2762 * additional structures follow the MFBD header (an extra payload or
2763 * not, as well as a count of render targets) */
2764
2765 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
2766
2767 if (is_mfbd) {
2768 if (info.has_extra)
2769 expected_tag |= MALI_MFBD_TAG_EXTRA;
2770
2771 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
2772 }
2773
2774 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
2775 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2776 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
2777 pandecode_prop("max_tile_coord = 0x%X\n", s->min_tile_coord);
2778 }
2779
2780 /* Extract tile coordinates */
2781
2782 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
2783 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
2784
2785 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2786 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2787
2788 /* For the max, we also want the floored (rather than ceiled) version for checking */
2789
2790 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
2791 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
2792
2793 /* Validate the coordinates are well-ordered */
2794
2795 if (min_x == max_x)
2796 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2797 else if (min_x > max_x)
2798 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2799
2800 if (min_y == max_y)
2801 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2802 else if (min_y > max_y)
2803 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2804
2805 /* Validate the coordinates fit inside the framebuffer. We use floor,
2806 * rather than ceil, for the max coordinates, since the tile
2807 * coordinates for something like an 800x600 framebuffer will actually
2808 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2809
2810 if ((min_x > info.width) || (max_x_f > info.width))
2811 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2812
2813 if ((min_y > info.height) || (max_y_f > info.height))
2814 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2815
2816 /* After validation, we print */
2817
2818 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
2819
2820 /* The FBD is a tagged pointer */
2821
2822 unsigned tag = (s->framebuffer & ~FBD_MASK);
2823
2824 if (tag != expected_tag)
2825 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
2826
2827 return sizeof(*s);
2828 }
2829
2830 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
2831 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
2832 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
2833 * might wish to combine this with the bifrost parameter) because some details
2834 * are model-specific even within a particular architecture. Minimal traces
2835 * *only* examine the job descriptors, skipping printing entirely if there is
2836 * no faults, and only descends into the payload if there are faults. This is
2837 * useful for looking for faults without the overhead of invasive traces. */
2838
2839 void
2840 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id, bool minimal)
2841 {
2842 struct mali_job_descriptor_header *h;
2843 unsigned job_descriptor_number = 0;
2844
2845 do {
2846 struct pandecode_mapped_memory *mem =
2847 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2848
2849 void *payload;
2850
2851 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2852
2853 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2854 * high 32-bits of the 64-bit pointer are reused to store
2855 * something else.
2856 */
2857 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2858 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2859 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2860
2861 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
2862
2863 int job_no = job_descriptor_number++;
2864
2865 /* If the job is good to go, skip it in minimal mode */
2866 if (minimal && (h->exception_status == 0x0 || h->exception_status == 0x1))
2867 continue;
2868
2869 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2870 pandecode_indent++;
2871
2872 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2873
2874 if (h->job_descriptor_size)
2875 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2876
2877 if (h->exception_status && h->exception_status != 0x1)
2878 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2879 h->exception_status,
2880 (h->exception_status >> 16) & 0xFFFF,
2881 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2882 h->exception_status & 0xFF);
2883
2884 if (h->first_incomplete_task)
2885 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2886
2887 if (h->fault_pointer)
2888 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2889
2890 if (h->job_barrier)
2891 pandecode_prop("job_barrier = %d", h->job_barrier);
2892
2893 pandecode_prop("job_index = %d", h->job_index);
2894
2895 if (h->unknown_flags)
2896 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2897
2898 if (h->job_dependency_index_1)
2899 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2900
2901 if (h->job_dependency_index_2)
2902 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2903
2904 pandecode_indent--;
2905 pandecode_log("};\n");
2906
2907 switch (h->job_type) {
2908 case JOB_TYPE_WRITE_VALUE: {
2909 struct mali_payload_write_value *s = payload;
2910 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2911 pandecode_indent++;
2912 MEMORY_PROP(s, address);
2913
2914 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
2915 pandecode_msg("XXX: unknown value descriptor\n");
2916 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
2917 }
2918
2919 if (s->reserved) {
2920 pandecode_msg("XXX: set value tripped\n");
2921 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
2922 }
2923
2924 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
2925 pandecode_indent--;
2926 pandecode_log("};\n");
2927
2928 break;
2929 }
2930
2931 case JOB_TYPE_TILER:
2932 case JOB_TYPE_VERTEX:
2933 case JOB_TYPE_COMPUTE:
2934 if (bifrost) {
2935 if (h->job_type == JOB_TYPE_TILER)
2936 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2937 else
2938 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2939 } else
2940 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
2941
2942 break;
2943
2944 case JOB_TYPE_FRAGMENT:
2945 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
2946 break;
2947
2948 default:
2949 break;
2950 }
2951 } while ((jc_gpu_va = h->next_job));
2952 }