projects / mesa.git / blob
? search:


ea50020c0218a3829c9a44a18e182d0e3b7eb973
[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 FLAG_INFO(DEPTH_CLIP_NEAR),
230 FLAG_INFO(DEPTH_CLIP_FAR),
231 {}
232 };
233
234 static const struct pandecode_flag_info u4_flag_info[] = {
235 FLAG_INFO(NO_MSAA),
236 FLAG_INFO(NO_DITHER),
237 FLAG_INFO(DEPTH_RANGE_A),
238 FLAG_INFO(DEPTH_RANGE_B),
239 FLAG_INFO(STENCIL_TEST),
240 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER),
241 {}
242 };
243 #undef FLAG_INFO
244
245 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
246 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
247 FLAG_INFO(LAYERED),
248 FLAG_INFO(MSAA),
249 FLAG_INFO(SRGB),
250 {}
251 };
252 #undef FLAG_INFO
253
254 #define FLAG_INFO(flag) { MALI_AFBC_##flag, "MALI_AFBC_" #flag }
255 static const struct pandecode_flag_info afbc_fmt_flag_info[] = {
256 FLAG_INFO(YTR),
257 {}
258 };
259 #undef FLAG_INFO
260
261 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
262 static const struct pandecode_flag_info mfbd_extra_flag_hi_info[] = {
263 FLAG_INFO(PRESENT),
264 {}
265 };
266 #undef FLAG_INFO
267
268 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
269 static const struct pandecode_flag_info mfbd_extra_flag_lo_info[] = {
270 FLAG_INFO(ZS),
271 {}
272 };
273 #undef FLAG_INFO
274
275 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
276 static const struct pandecode_flag_info shader_midgard1_flag_lo_info [] = {
277 FLAG_INFO(WRITES_Z),
278 FLAG_INFO(EARLY_Z),
279 FLAG_INFO(READS_TILEBUFFER),
280 FLAG_INFO(WRITES_GLOBAL),
281 FLAG_INFO(READS_ZS),
282 {}
283 };
284
285 static const struct pandecode_flag_info shader_midgard1_flag_hi_info [] = {
286 FLAG_INFO(WRITES_S),
287 FLAG_INFO(SUPPRESS_INF_NAN),
288 {}
289 };
290 #undef FLAG_INFO
291
292 #define FLAG_INFO(flag) { MALI_BIFROST_##flag, "MALI_BIFROST_" #flag }
293 static const struct pandecode_flag_info shader_bifrost_info [] = {
294 FLAG_INFO(FULL_THREAD),
295 FLAG_INFO(EARLY_Z),
296 FLAG_INFO(FIRST_ATEST),
297 {}
298 };
299
300 #undef FLAG_INFO
301
302 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
303 static const struct pandecode_flag_info mfbd_flag_info [] = {
304 FLAG_INFO(DEPTH_WRITE),
305 FLAG_INFO(EXTRA),
306 {}
307 };
308 #undef FLAG_INFO
309
310 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
311 static const struct pandecode_flag_info sampler_flag_info [] = {
312 FLAG_INFO(MAG_NEAREST),
313 FLAG_INFO(MIN_NEAREST),
314 FLAG_INFO(MIP_LINEAR_1),
315 FLAG_INFO(MIP_LINEAR_2),
316 FLAG_INFO(NORM_COORDS),
317 {}
318 };
319 #undef FLAG_INFO
320
321 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
322 static const struct pandecode_flag_info sfbd_unk1_info [] = {
323 FLAG_INFO(MSAA_8),
324 FLAG_INFO(MSAA_A),
325 {}
326 };
327 #undef FLAG_INFO
328
329 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
330 static const struct pandecode_flag_info sfbd_unk2_info [] = {
331 FLAG_INFO(MSAA_B),
332 FLAG_INFO(SRGB),
333 {}
334 };
335 #undef FLAG_INFO
336
337 extern char *replace_fragment;
338 extern char *replace_vertex;
339
340 static char *
341 pandecode_job_type(enum mali_job_type type)
342 {
343 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
344
345 switch (type) {
346 DEFINE_CASE(NULL);
347 DEFINE_CASE(WRITE_VALUE);
348 DEFINE_CASE(CACHE_FLUSH);
349 DEFINE_CASE(COMPUTE);
350 DEFINE_CASE(VERTEX);
351 DEFINE_CASE(TILER);
352 DEFINE_CASE(FUSED);
353 DEFINE_CASE(FRAGMENT);
354
355 case JOB_NOT_STARTED:
356 return "NOT_STARTED";
357
358 default:
359 pandecode_log("Warning! Unknown job type %x\n", type);
360 return "!?!?!?";
361 }
362
363 #undef DEFINE_CASE
364 }
365
366 static char *
367 pandecode_draw_mode(enum mali_draw_mode mode)
368 {
369 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
370
371 switch (mode) {
372 DEFINE_CASE(DRAW_NONE);
373 DEFINE_CASE(POINTS);
374 DEFINE_CASE(LINES);
375 DEFINE_CASE(TRIANGLES);
376 DEFINE_CASE(TRIANGLE_STRIP);
377 DEFINE_CASE(TRIANGLE_FAN);
378 DEFINE_CASE(LINE_STRIP);
379 DEFINE_CASE(LINE_LOOP);
380 DEFINE_CASE(POLYGON);
381 DEFINE_CASE(QUADS);
382 DEFINE_CASE(QUAD_STRIP);
383
384 default:
385 pandecode_msg("XXX: invalid draw mode %X\n", mode);
386 return "";
387 }
388
389 #undef DEFINE_CASE
390 }
391
392 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
393 static char *
394 pandecode_func(enum mali_func mode)
395 {
396 switch (mode) {
397 DEFINE_CASE(NEVER);
398 DEFINE_CASE(LESS);
399 DEFINE_CASE(EQUAL);
400 DEFINE_CASE(LEQUAL);
401 DEFINE_CASE(GREATER);
402 DEFINE_CASE(NOTEQUAL);
403 DEFINE_CASE(GEQUAL);
404 DEFINE_CASE(ALWAYS);
405
406 default:
407 pandecode_msg("XXX: invalid func %X\n", mode);
408 return "";
409 }
410 }
411 #undef DEFINE_CASE
412
413 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
414 static char *
415 pandecode_stencil_op(enum mali_stencil_op op)
416 {
417 switch (op) {
418 DEFINE_CASE(KEEP);
419 DEFINE_CASE(REPLACE);
420 DEFINE_CASE(ZERO);
421 DEFINE_CASE(INVERT);
422 DEFINE_CASE(INCR_WRAP);
423 DEFINE_CASE(DECR_WRAP);
424 DEFINE_CASE(INCR);
425 DEFINE_CASE(DECR);
426
427 default:
428 pandecode_msg("XXX: invalid stencil op %X\n", op);
429 return "";
430 }
431 }
432
433 #undef DEFINE_CASE
434
435 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
436 {
437 switch(mode) {
438 /* TODO: Combine to just "instanced" once this can be done
439 * unambiguously in all known cases */
440 case MALI_ATTR_POT_DIVIDE:
441 return "instanced_pot";
442 case MALI_ATTR_MODULO:
443 return "instanced_mod";
444 case MALI_ATTR_NPOT_DIVIDE:
445 return "instanced_npot";
446 case MALI_ATTR_IMAGE:
447 return "image";
448 default:
449 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
450 return "";
451 }
452 }
453
454 static const char *
455 pandecode_special_record(uint64_t v, bool* attribute)
456 {
457 switch(v) {
458 case MALI_ATTR_VERTEXID:
459 *attribute = true;
460 return "gl_VertexID";
461 case MALI_ATTR_INSTANCEID:
462 *attribute = true;
463 return "gl_InstanceID";
464 case MALI_VARYING_FRAG_COORD:
465 return "gl_FragCoord";
466 case MALI_VARYING_FRONT_FACING:
467 return "gl_FrontFacing";
468 case MALI_VARYING_POINT_COORD:
469 return "gl_PointCoord";
470 default:
471 pandecode_msg("XXX: invalid special record %" PRIx64 "\n", v);
472 return "";
473 }
474 }
475
476 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
477 static char *
478 pandecode_wrap_mode(enum mali_wrap_mode op)
479 {
480 switch (op) {
481 DEFINE_CASE(REPEAT);
482 DEFINE_CASE(CLAMP_TO_EDGE);
483 DEFINE_CASE(CLAMP);
484 DEFINE_CASE(CLAMP_TO_BORDER);
485 DEFINE_CASE(MIRRORED_REPEAT);
486 DEFINE_CASE(MIRRORED_CLAMP_TO_EDGE);
487 DEFINE_CASE(MIRRORED_CLAMP);
488 DEFINE_CASE(MIRRORED_CLAMP_TO_BORDER);
489
490 default:
491 pandecode_msg("XXX: invalid wrap mode %X\n", op);
492 return "";
493 }
494 }
495 #undef DEFINE_CASE
496
497 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
498 static char *
499 pandecode_block_format(enum mali_block_format fmt)
500 {
501 switch (fmt) {
502 DEFINE_CASE(TILED);
503 DEFINE_CASE(UNKNOWN);
504 DEFINE_CASE(LINEAR);
505 DEFINE_CASE(AFBC);
506
507 default:
508 unreachable("Invalid case");
509 }
510 }
511 #undef DEFINE_CASE
512
513 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
514 char *
515 pandecode_exception_access(unsigned access)
516 {
517 switch (access) {
518 DEFINE_CASE(NONE);
519 DEFINE_CASE(EXECUTE);
520 DEFINE_CASE(READ);
521 DEFINE_CASE(WRITE);
522
523 default:
524 unreachable("Invalid case");
525 }
526 }
527 #undef DEFINE_CASE
528
529 /* Midgard's tiler descriptor is embedded within the
530 * larger FBD */
531
532 static void
533 pandecode_midgard_tiler_descriptor(
534 const struct midgard_tiler_descriptor *t,
535 unsigned width,
536 unsigned height,
537 bool is_fragment,
538 bool has_hierarchy)
539 {
540 pandecode_log(".tiler = {\n");
541 pandecode_indent++;
542
543 if (t->hierarchy_mask == MALI_TILER_DISABLED)
544 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
545 else
546 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
547
548 /* We know this name from the kernel, but we never see it nonzero */
549
550 if (t->flags)
551 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
552
553 MEMORY_PROP(t, polygon_list);
554
555 /* The body is offset from the base of the polygon list */
556 //assert(t->polygon_list_body > t->polygon_list);
557 unsigned body_offset = t->polygon_list_body - t->polygon_list;
558
559 /* It needs to fit inside the reported size */
560 //assert(t->polygon_list_size >= body_offset);
561
562 /* Now that we've sanity checked, we'll try to calculate the sizes
563 * ourselves for comparison */
564
565 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
566 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
567
568 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
569 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
570 ref_header, ref_size);
571 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
572 pandecode_msg("body offset %d\n", body_offset);
573 }
574
575 /* The tiler heap has a start and end specified -- it should be
576 * identical to what we have in the BO. The exception is if tiling is
577 * disabled. */
578
579 MEMORY_PROP(t, heap_start);
580 assert(t->heap_end >= t->heap_start);
581
582 struct pandecode_mapped_memory *heap =
583 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
584
585 unsigned heap_size = t->heap_end - t->heap_start;
586
587 /* Tiling is enabled with a special flag */
588 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
589 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
590
591 bool tiling_enabled = hierarchy_mask;
592
593 if (tiling_enabled) {
594 /* When tiling is enabled, the heap should be a tight fit */
595 unsigned heap_offset = t->heap_start - heap->gpu_va;
596 if ((heap_offset + heap_size) != heap->length) {
597 pandecode_msg("XXX: heap size %u (expected %zu)\n",
598 heap_size, heap->length - heap_offset);
599 }
600
601 /* We should also have no other flags */
602 if (tiler_flags)
603 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
604 } else {
605 /* When tiling is disabled, we should have that flag and no others */
606
607 if (tiler_flags != MALI_TILER_DISABLED) {
608 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
609 tiler_flags);
610 }
611
612 /* We should also have an empty heap */
613 if (heap_size) {
614 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
615 heap_size);
616 }
617
618 /* Disabled tiling is used only for clear-only jobs, which are
619 * purely FRAGMENT, so we should never see this for
620 * non-FRAGMENT descriptors. */
621
622 if (!is_fragment)
623 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
624 }
625
626 /* We've never seen weights used in practice, but we know from the
627 * kernel these fields is there */
628
629 bool nonzero_weights = false;
630
631 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
632 nonzero_weights |= t->weights[w] != 0x0;
633 }
634
635 if (nonzero_weights) {
636 pandecode_log(".weights = { ");
637
638 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
639 pandecode_log_cont("%d, ", t->weights[w]);
640 }
641
642 pandecode_log("},");
643 }
644
645 pandecode_indent--;
646 pandecode_log("}\n");
647 }
648
649 /* TODO: The Bifrost tiler is not understood at all yet */
650
651 static void
652 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer *fb)
653 {
654 pandecode_log(".tiler = {\n");
655 pandecode_indent++;
656
657 MEMORY_PROP(fb, tiler_meta);
658
659 for (int i = 0; i < 16; i++) {
660 if (fb->zeros[i] != 0) {
661 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
662 i, fb->zeros[i]);
663 }
664 }
665
666 pandecode_log("},\n");
667
668 pandecode_indent--;
669 pandecode_log("}\n");
670
671 }
672
673 /* Information about the framebuffer passed back for
674 * additional analysis */
675
676 struct pandecode_fbd {
677 unsigned width;
678 unsigned height;
679 unsigned rt_count;
680 bool has_extra;
681 };
682
683 static void
684 pandecode_sfbd_format(struct mali_sfbd_format format)
685 {
686 pandecode_log(".format = {\n");
687 pandecode_indent++;
688
689 pandecode_log(".unk1 = ");
690 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
691 pandecode_log_cont(",\n");
692
693 /* TODO: Map formats so we can check swizzles and print nicely */
694 pandecode_log("swizzle");
695 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
696 pandecode_log_cont(",\n");
697
698 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
699 (format.nr_channels + 1));
700
701 pandecode_log(".unk2 = ");
702 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
703 pandecode_log_cont(",\n");
704
705 pandecode_prop("block = %s", pandecode_block_format(format.block));
706
707 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
708
709 pandecode_indent--;
710 pandecode_log("},\n");
711 }
712
713 static void
714 pandecode_shared_memory(const struct mali_shared_memory *desc, bool is_compute)
715 {
716 pandecode_prop("stack_shift = 0x%x", desc->stack_shift);
717
718 if (desc->unk0)
719 pandecode_prop("unk0 = 0x%x", desc->unk0);
720
721 if (desc->shared_workgroup_count != 0x1F) {
722 pandecode_prop("shared_workgroup_count = %d", desc->shared_workgroup_count);
723 if (!is_compute)
724 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
725 }
726
727 if (desc->shared_unk1 || desc->shared_shift) {
728 pandecode_prop("shared_unk1 = %X", desc->shared_unk1);
729 pandecode_prop("shared_shift = %X", desc->shared_shift);
730
731 if (!is_compute)
732 pandecode_msg("XXX: shared memory configured in noncompute shader");
733 }
734
735 if (desc->shared_zero) {
736 pandecode_msg("XXX: shared memory zero tripped\n");
737 pandecode_prop("shared_zero = 0x%" PRIx32, desc->shared_zero);
738 }
739
740 if (desc->shared_memory && !is_compute)
741 pandecode_msg("XXX: shared memory used in noncompute shader\n");
742
743 MEMORY_PROP(desc, scratchpad);
744 MEMORY_PROP(desc, shared_memory);
745 MEMORY_PROP(desc, unknown1);
746
747 if (desc->scratchpad) {
748 struct pandecode_mapped_memory *smem =
749 pandecode_find_mapped_gpu_mem_containing(desc->scratchpad);
750
751 pandecode_msg("scratchpad size %u\n", smem->length);
752 }
753
754 }
755
756 static struct pandecode_fbd
757 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
758 {
759 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
760 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
761
762 struct pandecode_fbd info = {
763 .has_extra = false,
764 .rt_count = 1
765 };
766
767 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
768 pandecode_indent++;
769
770 pandecode_log(".shared_memory = {\n");
771 pandecode_indent++;
772 pandecode_shared_memory(&s->shared_memory, false);
773 pandecode_indent--;
774 pandecode_log("},\n");
775
776 pandecode_sfbd_format(s->format);
777
778 info.width = s->width + 1;
779 info.height = s->height + 1;
780
781 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
782 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
783
784 MEMORY_PROP(s, checksum);
785
786 if (s->checksum_stride)
787 pandecode_prop("checksum_stride = %d", s->checksum_stride);
788
789 MEMORY_PROP(s, framebuffer);
790 pandecode_prop("stride = %d", s->stride);
791
792 /* Earlier in the actual commandstream -- right before width -- but we
793 * delay to flow nicer */
794
795 pandecode_log(".clear_flags = ");
796 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
797 pandecode_log_cont(",\n");
798
799 if (s->depth_buffer) {
800 MEMORY_PROP(s, depth_buffer);
801 pandecode_prop("depth_stride = %d", s->depth_stride);
802 }
803
804 if (s->stencil_buffer) {
805 MEMORY_PROP(s, stencil_buffer);
806 pandecode_prop("stencil_stride = %d", s->stencil_stride);
807 }
808
809 if (s->depth_stride_zero ||
810 s->stencil_stride_zero ||
811 s->zero7 || s->zero8) {
812 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
813 pandecode_prop("depth_stride_zero = 0x%x",
814 s->depth_stride_zero);
815 pandecode_prop("stencil_stride_zero = 0x%x",
816 s->stencil_stride_zero);
817 pandecode_prop("zero7 = 0x%" PRIx32,
818 s->zero7);
819 pandecode_prop("zero8 = 0x%" PRIx32,
820 s->zero8);
821 }
822
823 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
824 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
825 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
826 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
827 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
828 }
829
830 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
831 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
832 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
833 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
834 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
835 }
836
837 if (s->clear_stencil) {
838 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
839 }
840
841 const struct midgard_tiler_descriptor t = s->tiler;
842
843 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
844 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
845
846 pandecode_indent--;
847 pandecode_log("};\n");
848
849 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
850 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
851 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
852
853 pandecode_log_cont(".zero3 = {");
854
855 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
856 pandecode_log_cont("%X, ", s->zero3[i]);
857
858 pandecode_log_cont("},\n");
859
860 pandecode_log_cont(".zero6 = {");
861
862 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
863 pandecode_log_cont("%X, ", s->zero6[i]);
864
865 pandecode_log_cont("},\n");
866
867 return info;
868 }
869
870 static void
871 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
872 {
873 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
874 const struct mali_shared_memory *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
875
876 pandecode_log("struct mali_shared_memory shared_%"PRIx64"_%d = {\n", gpu_va, job_no);
877 pandecode_indent++;
878 pandecode_shared_memory(s, true);
879 pandecode_indent--;
880 pandecode_log("},\n");
881 }
882
883 /* Extracts the number of components associated with a Mali format */
884
885 static unsigned
886 pandecode_format_component_count(enum mali_format fmt)
887 {
888 /* Mask out the format class */
889 unsigned top = fmt & 0b11100000;
890
891 switch (top) {
892 case MALI_FORMAT_SNORM:
893 case MALI_FORMAT_UINT:
894 case MALI_FORMAT_UNORM:
895 case MALI_FORMAT_SINT:
896 return ((fmt >> 3) & 3) + 1;
897 default:
898 /* TODO: Validate */
899 return 4;
900 }
901 }
902
903 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
904
905 static unsigned
906 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
907 {
908 unsigned mask = 0;
909 assert(MALI_CHANNEL_RED == 0);
910
911 for (unsigned c = 0; c < 4; ++c) {
912 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
913
914 if (chan <= MALI_CHANNEL_ALPHA)
915 mask |= (1 << chan);
916 }
917
918 return mask;
919 }
920
921 /* Validates that a (format, swizzle) pair is valid, in the sense that the
922 * swizzle doesn't access any components that are undefined in the format.
923 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
924 * omitted */
925
926 static bool
927 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
928 {
929 unsigned nr_comp = pandecode_format_component_count(fmt);
930 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
931 unsigned valid_mask = (1 << nr_comp) - 1;
932 unsigned invalid_mask = ~valid_mask;
933
934 if (access_mask & invalid_mask) {
935 pandecode_msg("XXX: invalid components accessed\n");
936 return false;
937 }
938
939 /* Check for the default non-swizzling swizzle so we can suppress
940 * useless printing for the defaults */
941
942 unsigned default_swizzles[4] = {
943 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
944 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
945 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
946 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
947 };
948
949 return (swizzle == default_swizzles[nr_comp - 1]);
950 }
951
952 /* Maps MALI_RGBA32F to rgba32f, etc */
953
954 static void
955 pandecode_format_short(enum mali_format fmt, bool srgb)
956 {
957 /* We want a type-like format, so cut off the initial MALI_ */
958 char *format = pandecode_format(fmt);
959 format += strlen("MALI_");
960
961 unsigned len = strlen(format);
962 char *lower_format = calloc(1, len + 1);
963
964 for (unsigned i = 0; i < len; ++i)
965 lower_format[i] = tolower(format[i]);
966
967 /* Sanity check sRGB flag is applied to RGB, per the name */
968 if (srgb && lower_format[0] != 'r')
969 pandecode_msg("XXX: sRGB applied to non-colour format\n");
970
971 /* Just prefix with an s, so you get formats like srgba8_unorm */
972 if (srgb)
973 pandecode_log_cont("s");
974
975 pandecode_log_cont("%s", lower_format);
976 free(lower_format);
977 }
978
979 static void
980 pandecode_swizzle(unsigned swizzle, enum mali_format format)
981 {
982 /* First, do some validation */
983 bool trivial_swizzle = pandecode_validate_format_swizzle(
984 format, swizzle);
985
986 if (trivial_swizzle)
987 return;
988
989 /* Next, print the swizzle */
990 pandecode_log_cont(".");
991
992 static const char components[] = "rgba01";
993
994 for (unsigned c = 0; c < 4; ++c) {
995 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
996
997 if (chan >= MALI_CHANNEL_RESERVED_0) {
998 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
999 continue;
1000 }
1001 pandecode_log_cont("%c", components[chan]);
1002 }
1003 }
1004
1005 static void
1006 pandecode_rt_format(struct mali_rt_format format)
1007 {
1008 pandecode_log(".format = {\n");
1009 pandecode_indent++;
1010
1011 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
1012 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
1013 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
1014 pandecode_prop("unk4 = 0x%" PRIx32, format.unk4);
1015
1016 pandecode_prop("block = %s", pandecode_block_format(format.block));
1017
1018 /* TODO: Map formats so we can check swizzles and print nicely */
1019 pandecode_log("swizzle");
1020 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
1021 pandecode_log_cont(",\n");
1022
1023 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
1024 (format.nr_channels + 1));
1025
1026 pandecode_log(".flags = ");
1027 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
1028 pandecode_log_cont(",\n");
1029
1030 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
1031 * which is a faster hardware-based alternative to the wallpaper method
1032 * to preserve framebuffer contents across frames. In practice, MFBD
1033 * preload is buggy on Midgard, and so this is a chicken bit. If this
1034 * bit isn't set, most likely something broke unrelated to preload */
1035
1036 if (!format.no_preload) {
1037 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
1038 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
1039 }
1040
1041 if (format.zero)
1042 pandecode_prop("zero = 0x%" PRIx32, format.zero);
1043
1044 pandecode_indent--;
1045 pandecode_log("},\n");
1046 }
1047
1048 static void
1049 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct mali_framebuffer *fb)
1050 {
1051 pandecode_log("struct mali_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
1052 pandecode_indent++;
1053
1054 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
1055 mali_ptr rt_va = gpu_va + i * sizeof(struct mali_render_target);
1056 struct pandecode_mapped_memory *mem =
1057 pandecode_find_mapped_gpu_mem_containing(rt_va);
1058 const struct mali_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
1059
1060 pandecode_log("{\n");
1061 pandecode_indent++;
1062
1063 pandecode_rt_format(rt->format);
1064
1065 if (rt->format.block == MALI_BLOCK_AFBC) {
1066 pandecode_log(".afbc = {\n");
1067 pandecode_indent++;
1068
1069 char *a = pointer_as_memory_reference(rt->afbc.metadata);
1070 pandecode_prop("metadata = %s", a);
1071 free(a);
1072
1073 pandecode_prop("stride = %d", rt->afbc.stride);
1074
1075 pandecode_log(".flags = ");
1076 pandecode_log_decoded_flags(afbc_fmt_flag_info, rt->afbc.flags);
1077 pandecode_log_cont(",\n");
1078
1079 pandecode_indent--;
1080 pandecode_log("},\n");
1081 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.flags) {
1082 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1083 rt->afbc.metadata,
1084 rt->afbc.stride,
1085 rt->afbc.flags);
1086 }
1087
1088 MEMORY_PROP(rt, framebuffer);
1089 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1090
1091 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1092 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1093 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1094 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1095 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1096 }
1097
1098 if (rt->zero1 || rt->zero2 || rt->zero3) {
1099 pandecode_msg("XXX: render target zeros tripped\n");
1100 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1101 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1102 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
1103 }
1104
1105 pandecode_indent--;
1106 pandecode_log("},\n");
1107 }
1108
1109 pandecode_indent--;
1110 pandecode_log("};\n");
1111 }
1112
1113 static struct pandecode_fbd
1114 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment, bool is_compute, bool is_bifrost)
1115 {
1116 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1117 const struct mali_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1118
1119 struct pandecode_fbd info;
1120
1121 if (is_bifrost && fb->msaa.sample_locations) {
1122 /* The blob stores all possible sample locations in a single buffer
1123 * allocated on startup, and just switches the pointer when switching
1124 * MSAA state. For now, we just put the data into the cmdstream, but we
1125 * should do something like what the blob does with a real driver.
1126 *
1127 * There seem to be 32 slots for sample locations, followed by another
1128 * 16. The second 16 is just the center location followed by 15 zeros
1129 * in all the cases I've identified (maybe shader vs. depth/color
1130 * samples?).
1131 */
1132
1133 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->msaa.sample_locations);
1134
1135 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->msaa.sample_locations);
1136
1137 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
1138 pandecode_indent++;
1139
1140 for (int i = 0; i < 32 + 16; i++) {
1141 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1142 }
1143
1144 pandecode_indent--;
1145 pandecode_log("};\n");
1146 }
1147
1148 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1149 pandecode_indent++;
1150
1151 if (is_bifrost) {
1152 pandecode_log(".msaa = {\n");
1153 pandecode_indent++;
1154
1155 if (fb->msaa.sample_locations)
1156 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1157 else
1158 pandecode_msg("XXX: sample_locations missing\n");
1159
1160 if (fb->msaa.zero1 || fb->msaa.zero2 || fb->msaa.zero4) {
1161 pandecode_msg("XXX: multisampling zero tripped\n");
1162 pandecode_prop("zero1 = %" PRIx64, fb->msaa.zero1);
1163 pandecode_prop("zero2 = %" PRIx64, fb->msaa.zero2);
1164 pandecode_prop("zero4 = %" PRIx64, fb->msaa.zero4);
1165 }
1166
1167 pandecode_indent--;
1168 pandecode_log("},\n");
1169 } else {
1170 pandecode_log(".shared_memory = {\n");
1171 pandecode_indent++;
1172 pandecode_shared_memory(&fb->shared_memory, is_compute);
1173 pandecode_indent--;
1174 pandecode_log("},\n");
1175 }
1176
1177 info.width = fb->width1 + 1;
1178 info.height = fb->height1 + 1;
1179 info.rt_count = fb->rt_count_1 + 1;
1180
1181 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1182 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1183 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1184 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1185
1186 pandecode_prop("unk1 = 0x%x", fb->unk1);
1187 pandecode_prop("unk2 = 0x%x", fb->unk2);
1188 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1189 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1190
1191 pandecode_log(".mfbd_flags = ");
1192 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1193 pandecode_log_cont(",\n");
1194
1195 if (fb->clear_stencil)
1196 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1197
1198 if (fb->clear_depth)
1199 pandecode_prop("clear_depth = %f", fb->clear_depth);
1200
1201 if (!is_compute)
1202 if (is_bifrost)
1203 pandecode_bifrost_tiler_descriptor(fb);
1204 else {
1205 const struct midgard_tiler_descriptor t = fb->tiler;
1206 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
1207 }
1208 else
1209 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1210
1211 if (fb->zero3 || fb->zero4) {
1212 pandecode_msg("XXX: framebuffer zeros tripped\n");
1213 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1214 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1215 }
1216
1217 pandecode_indent--;
1218 pandecode_log("};\n");
1219
1220 gpu_va += sizeof(struct mali_framebuffer);
1221
1222 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1223
1224 if (info.has_extra) {
1225 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1226 const struct mali_framebuffer_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1227
1228 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1229 pandecode_indent++;
1230
1231 MEMORY_PROP(fbx, checksum);
1232
1233 if (fbx->checksum_stride)
1234 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1235
1236 pandecode_log(".flags_hi = ");
1237 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_hi);
1238 pandecode_log_cont(",\n");
1239
1240 pandecode_log(".flags_lo = ");
1241 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
1242 pandecode_log_cont(",\n");
1243
1244 pandecode_prop("zs_block = %s", pandecode_block_format(fbx->zs_block));
1245
1246 if (fbx->zs_block == MALI_BLOCK_AFBC) {
1247 pandecode_log(".ds_afbc = {\n");
1248 pandecode_indent++;
1249
1250 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1251 pandecode_prop("depth_stencil_afbc_stride = %d",
1252 fbx->ds_afbc.depth_stencil_afbc_stride);
1253 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1254
1255 pandecode_log(".flags = ");
1256 pandecode_log_decoded_flags(afbc_fmt_flag_info, fbx->ds_afbc.flags);
1257 pandecode_log_cont(",\n");
1258
1259 if (fbx->ds_afbc.padding) {
1260 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1261 pandecode_prop("padding = 0x%" PRIx64, fbx->ds_afbc.padding);
1262 }
1263
1264 pandecode_indent--;
1265 pandecode_log("},\n");
1266 } else {
1267 pandecode_log(".ds_linear = {\n");
1268 pandecode_indent++;
1269
1270 if (fbx->ds_linear.depth) {
1271 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1272 pandecode_prop("depth_stride = %d",
1273 fbx->ds_linear.depth_stride);
1274 } else if (fbx->ds_linear.depth_stride) {
1275 pandecode_msg("XXX: depth stride zero tripped %d\n", fbx->ds_linear.depth_stride);
1276 }
1277
1278 if (fbx->ds_linear.stencil) {
1279 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1280 pandecode_prop("stencil_stride = %d",
1281 fbx->ds_linear.stencil_stride);
1282 } else if (fbx->ds_linear.stencil_stride) {
1283 pandecode_msg("XXX: stencil stride zero tripped %d\n", fbx->ds_linear.stencil_stride);
1284 }
1285
1286 if (fbx->ds_linear.depth_stride_zero ||
1287 fbx->ds_linear.stencil_stride_zero ||
1288 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1289 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1290 pandecode_prop("depth_stride_zero = 0x%x",
1291 fbx->ds_linear.depth_stride_zero);
1292 pandecode_prop("stencil_stride_zero = 0x%x",
1293 fbx->ds_linear.stencil_stride_zero);
1294 pandecode_prop("zero1 = 0x%" PRIx32,
1295 fbx->ds_linear.zero1);
1296 pandecode_prop("zero2 = 0x%" PRIx32,
1297 fbx->ds_linear.zero2);
1298 }
1299
1300 pandecode_indent--;
1301 pandecode_log("},\n");
1302 }
1303
1304 if (fbx->clear_color_1 | fbx->clear_color_2) {
1305 pandecode_prop("clear_color_1 = 0x%" PRIx32, fbx->clear_color_1);
1306 pandecode_prop("clear_color_2 = 0x%" PRIx32, fbx->clear_color_2);
1307 }
1308
1309 if (fbx->zero3) {
1310 pandecode_msg("XXX: fb_extra zeros tripped\n");
1311 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1312 }
1313
1314 pandecode_indent--;
1315 pandecode_log("};\n");
1316
1317 gpu_va += sizeof(struct mali_framebuffer_extra);
1318 }
1319
1320 if (is_fragment)
1321 pandecode_render_target(gpu_va, job_no, fb);
1322
1323 return info;
1324 }
1325
1326 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1327 * count */
1328
1329 static void
1330 pandecode_padded_vertices(unsigned shift, unsigned k)
1331 {
1332 unsigned odd = 2*k + 1;
1333 unsigned pot = 1 << shift;
1334 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1335 }
1336
1337 /* Given a magic divisor, recover what we were trying to divide by.
1338 *
1339 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1340 * 0 <= m < 2^N, for N bits in m.
1341 *
1342 * Let q represent the number we would like to divide by.
1343 *
1344 * By definition of a magic divisor for N-bit unsigned integers (a number you
1345 * multiply by to magically get division), m is a number such that:
1346 *
1347 * (m * x) & (2^N - 1) = floor(x/q).
1348 * for all x on Z where 0 <= x < 2^N
1349 *
1350 * Ignore the case where any of the above values equals zero; it is irrelevant
1351 * for our purposes (instanced arrays).
1352 *
1353 * Choose x = q. Then:
1354 *
1355 * (m * x) & (2^N - 1) = floor(x/q).
1356 * (m * q) & (2^N - 1) = floor(q/q).
1357 *
1358 * floor(q/q) = floor(1) = 1, therefore:
1359 *
1360 * (m * q) & (2^N - 1) = 1
1361 *
1362 * Recall the identity that the bitwise AND of one less than a power-of-two
1363 * equals the modulo with that power of two, i.e. for all x:
1364 *
1365 * x & (2^N - 1) = x % N
1366 *
1367 * Therefore:
1368 *
1369 * mq % (2^N) = 1
1370 *
1371 * By definition, a modular multiplicative inverse of a number m is the number
1372 * q such that with respect to a modulos M:
1373 *
1374 * mq % M = 1
1375 *
1376 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1377 *
1378 */
1379
1380 static void
1381 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1382 {
1383 #if 0
1384 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1385 * shift) the most lame way possible... just repeatedly add.
1386 * Asymptoptically slow but nobody cares in practice, unless you have
1387 * massive numbers of vertices or high divisors. */
1388
1389 unsigned inverse = 0;
1390
1391 /* Magic implicitly has the highest bit set */
1392 magic |= (1 << 31);
1393
1394 /* Depending on rounding direction */
1395 if (extra)
1396 magic++;
1397
1398 for (;;) {
1399 uint32_t product = magic * inverse;
1400
1401 if (shift) {
1402 product >>= shift;
1403 }
1404
1405 if (product == 1)
1406 break;
1407
1408 ++inverse;
1409 }
1410
1411 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1412
1413 /* Recall we're supposed to divide by (gl_level_divisor *
1414 * padded_num_vertices) */
1415
1416 unsigned padded_num_vertices = inverse / orig_divisor;
1417
1418 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1419 #endif
1420 }
1421
1422 static void
1423 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1424 mali_ptr addr, int job_no, char *suffix,
1425 int count, bool varying, enum mali_job_type job_type)
1426 {
1427 char *prefix = varying ? "varying" : "attribute";
1428 assert(addr);
1429
1430 if (!count) {
1431 pandecode_msg("warn: No %s records\n", prefix);
1432 return;
1433 }
1434
1435 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1436
1437 for (int i = 0; i < count; ++i) {
1438 /* First, check for special records */
1439 if (attr[i].elements < MALI_RECORD_SPECIAL) {
1440 if (attr[i].size)
1441 pandecode_msg("XXX: tripped size=%d\n", attr[i].size);
1442
1443 if (attr[i].stride) {
1444 /* gl_InstanceID passes a magic divisor in the
1445 * stride field to divide by the padded vertex
1446 * count. No other records should do so, so
1447 * stride should otherwise be zero. Note that
1448 * stride in the usual attribute sense doesn't
1449 * apply to special records. */
1450
1451 bool has_divisor = attr[i].elements == MALI_ATTR_INSTANCEID;
1452
1453 pandecode_log_cont("/* %smagic divisor = %X */ ",
1454 has_divisor ? "" : "XXX: ", attr[i].stride);
1455 }
1456
1457 if (attr[i].shift || attr[i].extra_flags) {
1458 /* Attributes use these fields for
1459 * instancing/padding/etc type issues, but
1460 * varyings don't */
1461
1462 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1463 varying ? "XXX: " : "",
1464 attr[i].shift, attr[i].extra_flags);
1465 }
1466
1467 /* Print the special record name */
1468 bool attribute = false;
1469 pandecode_log("%s_%d = %s;\n", prefix, i, pandecode_special_record(attr[i].elements, &attribute));
1470
1471 /* Sanity check */
1472 if (attribute == varying)
1473 pandecode_msg("XXX: mismatched special record\n");
1474
1475 continue;
1476 }
1477
1478 enum mali_attr_mode mode = attr[i].elements & 7;
1479
1480 if (mode == MALI_ATTR_UNUSED)
1481 pandecode_msg("XXX: unused attribute record\n");
1482
1483 /* For non-linear records, we need to print the type of record */
1484 if (mode != MALI_ATTR_LINEAR)
1485 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1486
1487 /* Print the name to link with attr_meta */
1488 pandecode_log_cont("%s_%d", prefix, i);
1489
1490 /* Print the stride and size */
1491 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1492
1493 /* TODO: Sanity check the quotient itself. It must be equal to
1494 * (or be greater than, if the driver added padding) the padded
1495 * vertex count. */
1496
1497 /* Finally, print the pointer */
1498 mali_ptr raw_elements = attr[i].elements & ~7;
1499 char *a = pointer_as_memory_reference(raw_elements);
1500 pandecode_log_cont(" = (%s);\n", a);
1501 free(a);
1502
1503 /* Check the pointer */
1504 pandecode_validate_buffer(raw_elements, attr[i].size);
1505
1506 /* shift/extra_flags exist only for instanced */
1507 if (attr[i].shift | attr[i].extra_flags) {
1508 /* These are set to random values by the blob for
1509 * varyings, most likely a symptom of uninitialized
1510 * memory where the hardware masked the bug. As such we
1511 * put this at a warning, not an error. */
1512
1513 if (mode == MALI_ATTR_LINEAR)
1514 pandecode_msg("warn: instancing fields set for linear\n");
1515
1516 pandecode_prop("shift = %d", attr[i].shift);
1517 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1518 }
1519
1520 /* Decode further where possible */
1521
1522 if (mode == MALI_ATTR_MODULO) {
1523 pandecode_padded_vertices(
1524 attr[i].shift,
1525 attr[i].extra_flags);
1526 }
1527
1528 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1529 i++;
1530 pandecode_log("{\n");
1531 pandecode_indent++;
1532 pandecode_prop("unk = 0x%x", attr[i].unk);
1533 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1534 if (attr[i].zero != 0)
1535 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1536 pandecode_prop("divisor = %d", attr[i].divisor);
1537 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1538 pandecode_indent--;
1539 pandecode_log("}, \n");
1540 }
1541
1542 }
1543
1544 pandecode_log("\n");
1545 }
1546
1547 static mali_ptr
1548 pandecode_shader_address(const char *name, mali_ptr ptr)
1549 {
1550 /* TODO: Decode flags */
1551 mali_ptr shader_ptr = ptr & ~15;
1552
1553 char *a = pointer_as_memory_reference(shader_ptr);
1554 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1555 free(a);
1556
1557 return shader_ptr;
1558 }
1559
1560 static void
1561 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1562 {
1563 unsigned any_nonzero =
1564 stencil->ref | stencil->mask | stencil->func |
1565 stencil->sfail | stencil->dpfail | stencil->dppass;
1566
1567 if (any_nonzero == 0)
1568 return;
1569
1570 const char *func = pandecode_func(stencil->func);
1571 const char *sfail = pandecode_stencil_op(stencil->sfail);
1572 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1573 const char *dppass = pandecode_stencil_op(stencil->dppass);
1574
1575 if (stencil->zero)
1576 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1577
1578 pandecode_log(".stencil_%s = {\n", name);
1579 pandecode_indent++;
1580 pandecode_prop("ref = %d", stencil->ref);
1581 pandecode_prop("mask = 0x%02X", stencil->mask);
1582 pandecode_prop("func = %s", func);
1583 pandecode_prop("sfail = %s", sfail);
1584 pandecode_prop("dpfail = %s", dpfail);
1585 pandecode_prop("dppass = %s", dppass);
1586 pandecode_indent--;
1587 pandecode_log("},\n");
1588 }
1589
1590 static void
1591 pandecode_blend_equation(const struct mali_blend_equation *blend)
1592 {
1593 if (blend->zero1)
1594 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1595
1596 pandecode_log(".equation = {\n");
1597 pandecode_indent++;
1598
1599 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1600 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1601
1602 pandecode_log(".color_mask = ");
1603 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1604 pandecode_log_cont(",\n");
1605
1606 pandecode_indent--;
1607 pandecode_log("},\n");
1608 }
1609
1610 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1611
1612 static unsigned
1613 decode_bifrost_constant(u16 constant)
1614 {
1615 float lo = (float) (constant & 0xFF);
1616 float hi = (float) (constant >> 8);
1617
1618 return (hi / 255.0) + (lo / 65535.0);
1619 }
1620
1621 static mali_ptr
1622 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1623 {
1624 struct bifrost_blend_rt *b =
1625 ((struct bifrost_blend_rt *) descs) + rt_no;
1626
1627 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1628 pandecode_indent++;
1629
1630 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1631 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1632 b->constant, decode_bifrost_constant(b->constant));
1633
1634 /* TODO figure out blend shader enable bit */
1635 pandecode_blend_equation(&b->equation);
1636
1637 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1638 pandecode_prop("index = 0x%" PRIx16, b->index);
1639
1640 pandecode_log(".format = ");
1641 pandecode_format_short(b->format, false);
1642 pandecode_swizzle(b->swizzle, b->format);
1643 pandecode_log_cont(",\n");
1644
1645 pandecode_prop("swizzle = 0x%" PRIx32, b->swizzle);
1646 pandecode_prop("format = 0x%" PRIx32, b->format);
1647
1648 if (b->zero1) {
1649 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1650 pandecode_prop("zero1 = 0x%" PRIx32, b->zero1);
1651 }
1652
1653 pandecode_log(".shader_type = ");
1654 switch(b->shader_type) {
1655 case BIFROST_BLEND_F16:
1656 pandecode_log_cont("BIFROST_BLEND_F16");
1657 break;
1658 case BIFROST_BLEND_F32:
1659 pandecode_log_cont("BIFROST_BLEND_F32");
1660 break;
1661 case BIFROST_BLEND_I32:
1662 pandecode_log_cont("BIFROST_BLEND_I32");
1663 break;
1664 case BIFROST_BLEND_U32:
1665 pandecode_log_cont("BIFROST_BLEND_U32");
1666 break;
1667 case BIFROST_BLEND_I16:
1668 pandecode_log_cont("BIFROST_BLEND_I16");
1669 break;
1670 case BIFROST_BLEND_U16:
1671 pandecode_log_cont("BIFROST_BLEND_U16");
1672 break;
1673 }
1674 pandecode_log_cont(",\n");
1675
1676 if (b->zero2) {
1677 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1678 pandecode_prop("zero2 = 0x%" PRIx32, b->zero2);
1679 }
1680
1681 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1682
1683 pandecode_indent--;
1684 pandecode_log("},\n");
1685
1686 return 0;
1687 }
1688
1689 static mali_ptr
1690 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1691 {
1692 /* constant/equation is in a union */
1693 if (!blend->shader)
1694 return 0;
1695
1696 pandecode_log(".blend = {\n");
1697 pandecode_indent++;
1698
1699 if (is_shader) {
1700 pandecode_shader_address("shader", blend->shader);
1701 } else {
1702 pandecode_blend_equation(&blend->equation);
1703 pandecode_prop("constant = %f", blend->constant);
1704 }
1705
1706 pandecode_indent--;
1707 pandecode_log("},\n");
1708
1709 /* Return blend shader to disassemble if present */
1710 return is_shader ? (blend->shader & ~0xF) : 0;
1711 }
1712
1713 static mali_ptr
1714 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1715 {
1716 struct midgard_blend_rt *b =
1717 ((struct midgard_blend_rt *) descs) + rt_no;
1718
1719 /* Flags determine presence of blend shader */
1720 bool is_shader = (b->flags & 0xF) >= 0x2;
1721
1722 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1723 pandecode_indent++;
1724
1725 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1726
1727 union midgard_blend blend = b->blend;
1728 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1729
1730 pandecode_indent--;
1731 pandecode_log("};\n");
1732
1733 return shader;
1734 }
1735
1736 /* Attributes and varyings have descriptor records, which contain information
1737 * about their format and ordering with the attribute/varying buffers. We'll
1738 * want to validate that the combinations specified are self-consistent.
1739 */
1740
1741 static int
1742 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1743 {
1744 char base[128];
1745 char *prefix = varying ? "varying" : "attribute";
1746 unsigned max_index = 0;
1747 snprintf(base, sizeof(base), "%s_meta", prefix);
1748
1749 struct mali_attr_meta *attr_meta;
1750 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1751
1752 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1753
1754 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1755 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1756 sizeof(*attr_mem));
1757
1758 /* If the record is discard, it should be zero for everything else */
1759
1760 if (attr_meta->format == MALI_VARYING_DISCARD) {
1761 uint64_t zero =
1762 attr_meta->index |
1763 attr_meta->unknown1 |
1764 attr_meta->unknown3 |
1765 attr_meta->src_offset;
1766
1767 if (zero)
1768 pandecode_msg("XXX: expected empty record for varying discard\n");
1769
1770 /* We want to look for a literal 0000 swizzle -- this
1771 * is not encoded with all zeroes, however */
1772
1773 enum mali_channel z = MALI_CHANNEL_ZERO;
1774 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1775 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1776
1777 if (!good_swizzle)
1778 pandecode_msg("XXX: expected zero swizzle for discard\n");
1779
1780 if (!varying)
1781 pandecode_msg("XXX: cannot discard attribute\n");
1782
1783 /* If we're all good, omit the record */
1784 if (!zero && varying && good_swizzle) {
1785 pandecode_log("/* discarded varying */\n");
1786 continue;
1787 }
1788 }
1789
1790 if (attr_meta->index > max_index)
1791 max_index = attr_meta->index;
1792
1793 if (attr_meta->unknown1 != 0x2) {
1794 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1795 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1796 }
1797
1798 if (attr_meta->unknown3) {
1799 pandecode_msg("XXX: unexpected unknown3 set\n");
1800 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1801 }
1802
1803 pandecode_format_short(attr_meta->format, false);
1804 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1805
1806 if (attr_meta->src_offset)
1807 pandecode_log_cont("[%u]", attr_meta->src_offset);
1808
1809 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1810
1811 pandecode_log_cont(";\n");
1812 }
1813
1814 pandecode_log("\n");
1815
1816 return count ? (max_index + 1) : 0;
1817 }
1818
1819 /* return bits [lo, hi) of word */
1820 static u32
1821 bits(u32 word, u32 lo, u32 hi)
1822 {
1823 if (hi - lo >= 32)
1824 return word; // avoid undefined behavior with the shift
1825
1826 return (word >> lo) & ((1 << (hi - lo)) - 1);
1827 }
1828
1829 static void
1830 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1831 {
1832 pandecode_log(".prefix = {\n");
1833 pandecode_indent++;
1834
1835 /* Decode invocation_count. See the comment before the definition of
1836 * invocation_count for an explanation.
1837 */
1838
1839 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1840 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1841 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1842 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1843 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1844 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1845
1846 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1847 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1848 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1849
1850 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1851 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1852 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1853
1854 /* Even though we have this decoded, we want to ensure that the
1855 * representation is "unique" so we don't lose anything by printing only
1856 * the final result. More specifically, we need to check that we were
1857 * passed something in canonical form, since the definition per the
1858 * hardware is inherently not unique. How? Well, take the resulting
1859 * decode and pack it ourselves! If it is bit exact with what we
1860 * decoded, we're good to go. */
1861
1862 struct mali_vertex_tiler_prefix ref;
1863 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1864
1865 bool canonical =
1866 (p->invocation_count == ref.invocation_count) &&
1867 (p->invocation_shifts == ref.invocation_shifts);
1868
1869 if (!canonical) {
1870 pandecode_msg("XXX: non-canonical workgroups packing\n");
1871 pandecode_msg("expected: %X, %X",
1872 ref.invocation_count,
1873 ref.invocation_shifts);
1874
1875 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1876 pandecode_prop("size_y_shift = %d", size_y_shift);
1877 pandecode_prop("size_z_shift = %d", size_z_shift);
1878 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
1879 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
1880 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
1881 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
1882 }
1883
1884 /* Regardless, print the decode */
1885 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1886 size_x, size_y, size_z,
1887 groups_x, groups_y, groups_z);
1888
1889 /* TODO: Decode */
1890 if (p->unknown_draw)
1891 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1892
1893 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1894
1895 if (p->draw_mode != MALI_DRAW_NONE)
1896 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1897
1898 /* Index count only exists for tiler jobs anyway */
1899
1900 if (p->index_count)
1901 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1902
1903
1904 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1905 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1906
1907 /* Validate an index buffer is present if we need one. TODO: verify
1908 * relationship between invocation_count and index_count */
1909
1910 if (p->indices) {
1911 unsigned count = p->index_count;
1912
1913 /* Grab the size */
1914 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1915
1916 /* Ensure we got a size, and if so, validate the index buffer
1917 * is large enough to hold a full set of indices of the given
1918 * size */
1919
1920 if (!index_raw_size)
1921 pandecode_msg("XXX: index size missing\n");
1922 else
1923 pandecode_validate_buffer(p->indices, count * size);
1924 } else if (index_raw_size)
1925 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1926
1927 if (p->offset_bias_correction)
1928 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1929
1930 /* TODO: Figure out what this is. It's not zero */
1931 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1932
1933 pandecode_indent--;
1934 pandecode_log("},\n");
1935 }
1936
1937 static void
1938 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1939 {
1940 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1941 uint64_t *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1942
1943 for (int i = 0; i < ubufs_count; i++) {
1944 unsigned size = (ubufs[i] & ((1 << 10) - 1)) * 16;
1945 mali_ptr addr = (ubufs[i] >> 10) << 2;
1946
1947 pandecode_validate_buffer(addr, size);
1948
1949 char *ptr = pointer_as_memory_reference(addr);
1950 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1951 free(ptr);
1952 }
1953
1954 pandecode_log("\n");
1955 }
1956
1957 static void
1958 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1959 {
1960 pandecode_validate_buffer(uniforms, uniform_count * 16);
1961
1962 char *ptr = pointer_as_memory_reference(uniforms);
1963 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1964 free(ptr);
1965 }
1966
1967 static const char *
1968 shader_type_for_job(unsigned type)
1969 {
1970 switch (type) {
1971 case JOB_TYPE_VERTEX: return "VERTEX";
1972 case JOB_TYPE_TILER: return "FRAGMENT";
1973 case JOB_TYPE_COMPUTE: return "COMPUTE";
1974 default:
1975 return "UNKNOWN";
1976 }
1977 }
1978
1979 static unsigned shader_id = 0;
1980
1981 static struct midgard_disasm_stats
1982 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1983 bool is_bifrost, unsigned gpu_id)
1984 {
1985 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1986 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1987
1988 /* Compute maximum possible size */
1989 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1990
1991 /* Print some boilerplate to clearly denote the assembly (which doesn't
1992 * obey indentation rules), and actually do the disassembly! */
1993
1994 pandecode_log_cont("\n\n");
1995
1996 struct midgard_disasm_stats stats;
1997
1998 if (is_bifrost) {
1999 disassemble_bifrost(pandecode_dump_stream, code, sz, true);
2000
2001 /* TODO: Extend stats to Bifrost */
2002 stats.texture_count = -128;
2003 stats.sampler_count = -128;
2004 stats.attribute_count = -128;
2005 stats.varying_count = -128;
2006 stats.uniform_count = -128;
2007 stats.uniform_buffer_count = -128;
2008 stats.work_count = -128;
2009
2010 stats.instruction_count = 0;
2011 stats.bundle_count = 0;
2012 stats.quadword_count = 0;
2013 stats.helper_invocations = false;
2014 } else {
2015 stats = disassemble_midgard(pandecode_dump_stream,
2016 code, sz, gpu_id,
2017 type == JOB_TYPE_TILER ?
2018 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
2019 }
2020
2021 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
2022 * driver-internal purposes with the blob and interfere */
2023
2024 bool should_shaderdb = type != JOB_TYPE_COMPUTE;
2025
2026 if (should_shaderdb) {
2027 unsigned nr_threads =
2028 (stats.work_count <= 4) ? 4 :
2029 (stats.work_count <= 8) ? 2 :
2030 1;
2031
2032 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
2033 "%u inst, %u bundles, %u quadwords, "
2034 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
2035 shader_id++,
2036 shader_type_for_job(type),
2037 stats.instruction_count, stats.bundle_count, stats.quadword_count,
2038 stats.work_count, nr_threads);
2039 }
2040
2041
2042 return stats;
2043 }
2044
2045 static void
2046 pandecode_texture_payload(mali_ptr payload,
2047 enum mali_texture_type type,
2048 enum mali_texture_layout layout,
2049 bool manual_stride,
2050 uint8_t levels,
2051 uint16_t depth,
2052 uint16_t array_size,
2053 struct pandecode_mapped_memory *tmem)
2054 {
2055 pandecode_log(".payload = {\n");
2056 pandecode_indent++;
2057
2058 /* A bunch of bitmap pointers follow.
2059 * We work out the correct number,
2060 * based on the mipmap/cubemap
2061 * properties, but dump extra
2062 * possibilities to futureproof */
2063
2064 int bitmap_count = levels + 1;
2065
2066 /* Miptree for each face */
2067 if (type == MALI_TEX_CUBE)
2068 bitmap_count *= 6;
2069
2070 /* Array of layers */
2071 bitmap_count *= (depth + 1);
2072
2073 /* Array of textures */
2074 bitmap_count *= (array_size + 1);
2075
2076 /* Stride for each element */
2077 if (manual_stride)
2078 bitmap_count *= 2;
2079
2080 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
2081 payload, sizeof(mali_ptr) * bitmap_count);
2082 for (int i = 0; i < bitmap_count; ++i) {
2083 /* How we dump depends if this is a stride or a pointer */
2084
2085 if (manual_stride && (i & 1)) {
2086 /* signed 32-bit snuck in as a 64-bit pointer */
2087 uint64_t stride_set = pointers_and_strides[i];
2088 uint32_t clamped_stride = stride_set;
2089 int32_t stride = clamped_stride;
2090 assert(stride_set == clamped_stride);
2091 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2092 } else {
2093 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
2094 pandecode_log("%s, \n", a);
2095 free(a);
2096 }
2097 }
2098
2099 pandecode_indent--;
2100 pandecode_log("},\n");
2101 }
2102
2103 static void
2104 pandecode_texture(mali_ptr u,
2105 struct pandecode_mapped_memory *tmem,
2106 unsigned job_no, unsigned tex)
2107 {
2108 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
2109
2110 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
2111 pandecode_indent++;
2112
2113 pandecode_prop("width = %" PRId32, t->width);
2114 pandecode_prop("height = %" PRId32, t->height);
2115 pandecode_prop("depth = %" PRId32, t->depth);
2116 pandecode_prop("array_size = %" PRId32, t->array_size);
2117
2118 pandecode_log("\n");
2119 pandecode_prop("f.swizzle = 0x%" PRIx32, t->format.swizzle);
2120 pandecode_prop("f.format = 0x%" PRIx32, t->format.format);
2121 pandecode_prop("f.srgb = 0x%" PRIx32, t->format.srgb);
2122 pandecode_prop("f.unknown1 = 0x%" PRIx32, t->format.unknown1);
2123 pandecode_prop("f.type = %" PRId32, t->format.type);
2124 pandecode_prop("f.layout = %" PRId32, t->format.layout);
2125 pandecode_prop("f.unknown2 = 0x%" PRIx32, t->format.unknown2);
2126 pandecode_prop("f.manual_stride = %" PRId32, t->format.manual_stride);
2127 pandecode_prop("f.zero = 0x%" PRIx32, t->format.zero);
2128 pandecode_log("\n");
2129
2130 pandecode_prop("unknown3 = 0x%" PRIx32, t->unknown3);
2131 pandecode_prop("unknown3A = 0x%" PRIx32, t->unknown3A);
2132 pandecode_prop("levels = %" PRId32, t->levels);
2133 pandecode_prop("swizzle = 0x%" PRIx32, t->swizzle);
2134 pandecode_prop("swizzle_zero = 0x%" PRIx32, t->swizzle_zero);
2135
2136 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2137 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2138 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2139 pandecode_log("\n");
2140
2141 struct mali_texture_format f = t->format;
2142
2143 /* See the definiton of enum mali_texture_type */
2144
2145 bool is_cube = f.type == MALI_TEX_CUBE;
2146 unsigned dimension = is_cube ? 2 : f.type;
2147
2148 pandecode_make_indent();
2149
2150 /* Print the layout. Default is linear; a modifier can denote AFBC or
2151 * u-interleaved/tiled modes */
2152
2153 if (f.layout == MALI_TEXTURE_AFBC)
2154 pandecode_log_cont("afbc");
2155 else if (f.layout == MALI_TEXTURE_TILED)
2156 pandecode_log_cont("tiled");
2157 else if (f.layout == MALI_TEXTURE_LINEAR)
2158 pandecode_log_cont("linear");
2159 else
2160 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
2161
2162 pandecode_swizzle(t->swizzle, f.format);
2163 pandecode_log_cont(" ");
2164
2165 /* Distinguish cube/2D with modifier */
2166
2167 if (is_cube)
2168 pandecode_log_cont("cube ");
2169
2170 pandecode_format_short(f.format, f.srgb);
2171 pandecode_swizzle(f.swizzle, f.format);
2172
2173 /* All four width/height/depth/array_size dimensions are present
2174 * regardless of the type of texture, but it is an error to have
2175 * non-zero dimensions for unused dimensions. Verify this. array_size
2176 * can always be set, as can width. Depth used for MSAA. */
2177
2178 if (t->height && dimension < 2)
2179 pandecode_msg("XXX: nonzero height for <2D texture\n");
2180
2181 /* Print only the dimensions that are actually there */
2182
2183 pandecode_log_cont(": %d", t->width + 1);
2184
2185 if (t->height || t->depth)
2186 pandecode_log_cont("x%u", t->height + 1);
2187
2188 if (t->depth)
2189 pandecode_log_cont("x%u", t->depth + 1);
2190
2191 if (t->array_size)
2192 pandecode_log_cont("[%u]", t->array_size + 1);
2193
2194 if (t->levels)
2195 pandecode_log_cont(" mip %u", t->levels);
2196
2197 pandecode_log_cont("\n");
2198
2199 if (f.unknown1 | f.zero) {
2200 pandecode_msg("XXX: texture format zero tripped\n");
2201 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
2202 pandecode_prop("zero = %" PRId32, f.zero);
2203 }
2204
2205 if (!f.unknown2) {
2206 pandecode_msg("XXX: expected unknown texture bit set\n");
2207 pandecode_prop("unknown2 = %" PRId32, f.unknown2);
2208 }
2209
2210 if (t->swizzle_zero) {
2211 pandecode_msg("XXX: swizzle zero tripped\n");
2212 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
2213 }
2214
2215 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
2216 pandecode_msg("XXX: texture zero tripped\n");
2217 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
2218 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
2219 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2220 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2221 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2222 }
2223
2224 pandecode_texture_payload(u + sizeof(*t), f.type, f.layout, f.manual_stride, t->levels, t->depth, t->array_size, tmem);
2225
2226 pandecode_indent--;
2227 pandecode_log("};\n");
2228 }
2229
2230 static void
2231 pandecode_bifrost_texture(
2232 const struct bifrost_texture_descriptor *t,
2233 unsigned job_no,
2234 unsigned tex)
2235 {
2236 pandecode_log("struct bifrost_texture_descriptor texture_descriptor_%d_%d = {\n", job_no, tex);
2237 pandecode_indent++;
2238
2239 pandecode_prop("format_unk = 0x%" PRIx32, t->format_unk);
2240 pandecode_prop("type = %" PRId32, t->type);
2241
2242 if (t->zero) {
2243 pandecode_msg("XXX: zero tripped\n");
2244 pandecode_prop("zero = 0x%" PRIx32, t->zero);
2245 }
2246
2247 pandecode_prop("format_swizzle = 0x%" PRIx32, t->format_swizzle);
2248 pandecode_prop("format = 0x%" PRIx32, t->format);
2249 pandecode_prop("srgb = 0x%" PRIx32, t->srgb);
2250 pandecode_prop("format_unk3 = 0x%" PRIx32, t->format_unk3);
2251 pandecode_prop("width = %" PRId32, t->width);
2252 pandecode_prop("height = %" PRId32, t->height);
2253 pandecode_prop("swizzle = 0x%" PRIx32, t->swizzle);
2254 pandecode_prop("levels = %" PRId32, t->levels);
2255 pandecode_prop("unk1 = 0x%" PRIx32, t->unk1);
2256 pandecode_prop("levels_unk = %" PRId32, t->levels_unk);
2257 pandecode_prop("level_2 = %" PRId32, t->level_2);
2258 pandecode_prop("payload = 0x%" PRIx64, t->payload);
2259 pandecode_prop("array_size = %" PRId32, t->array_size);
2260 pandecode_prop("unk4 = 0x%" PRIx32, t->unk4);
2261 pandecode_prop("depth = %" PRId32, t->depth);
2262 pandecode_prop("unk5 = 0x%" PRIx32, t->unk5);
2263 pandecode_log("\n");
2264
2265 /* See the definiton of enum mali_texture_type */
2266
2267 bool is_cube = t->type == MALI_TEX_CUBE;
2268 unsigned dimension = is_cube ? 2 : t->type;
2269
2270 /* Print the layout. Default is linear; a modifier can denote AFBC or
2271 * u-interleaved/tiled modes */
2272
2273 if (t->layout == MALI_TEXTURE_AFBC)
2274 pandecode_log_cont("afbc");
2275 else if (t->layout == MALI_TEXTURE_TILED)
2276 pandecode_log_cont("tiled");
2277 else if (t->layout == MALI_TEXTURE_LINEAR)
2278 pandecode_log_cont("linear");
2279 else
2280 pandecode_msg("XXX: invalid texture layout 0x%X\n", t->layout);
2281
2282 pandecode_swizzle(t->swizzle, t->format);
2283 pandecode_log_cont(" ");
2284
2285 /* Distinguish cube/2D with modifier */
2286
2287 if (is_cube)
2288 pandecode_log_cont("cube ");
2289
2290 pandecode_format_short(t->format, t->srgb);
2291
2292 /* All four width/height/depth/array_size dimensions are present
2293 * regardless of the type of texture, but it is an error to have
2294 * non-zero dimensions for unused dimensions. Verify this. array_size
2295 * can always be set, as can width. */
2296
2297 if (t->height && dimension < 2)
2298 pandecode_msg("XXX: nonzero height for <2D texture\n");
2299
2300 if (t->depth && dimension < 3)
2301 pandecode_msg("XXX: nonzero depth for <2D texture\n");
2302
2303 /* Print only the dimensions that are actually there */
2304
2305 pandecode_log_cont(": %d", t->width + 1);
2306
2307 if (dimension >= 2)
2308 pandecode_log_cont("x%u", t->height + 1);
2309
2310 if (dimension >= 3)
2311 pandecode_log_cont("x%u", t->depth + 1);
2312
2313 if (t->array_size)
2314 pandecode_log_cont("[%u]", t->array_size + 1);
2315
2316 if (t->levels)
2317 pandecode_log_cont(" mip %u", t->levels);
2318
2319 pandecode_log_cont("\n");
2320
2321 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(t->payload);
2322 if (t->payload) {
2323 pandecode_texture_payload(t->payload, t->type, t->layout,
2324 true, t->levels, t->depth,
2325 t->array_size, tmem);
2326 }
2327
2328 pandecode_indent--;
2329 pandecode_log("};\n");
2330 }
2331
2332 /* 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 */
2333
2334 static void
2335 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2336 {
2337 /* Nothing to do */
2338 if (claim == truth)
2339 return;
2340
2341 if (fuzzy && (truth < 0))
2342 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name, claim, truth);
2343
2344 if ((truth >= 0) && !fuzzy) {
2345 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2346 (truth < claim) ? "warn" : "XXX",
2347 name, truth, claim);
2348 } else if ((claim > -truth) && !fuzzy) {
2349 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2350 name, -truth, claim);
2351 } else if (fuzzy && (claim < truth))
2352 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2353 name, truth, claim);
2354
2355 pandecode_log(".%s = %" PRId16, name, claim);
2356
2357 if (fuzzy)
2358 pandecode_log_cont(" /* %u used */", truth);
2359
2360 pandecode_log_cont(",\n");
2361 }
2362
2363 static void
2364 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2365 bool is_bifrost, unsigned gpu_id)
2366 {
2367 struct midgard_disasm_stats stats =
2368 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
2369
2370 bool has_texture = (stats.texture_count > 0);
2371 bool has_sampler = (stats.sampler_count > 0);
2372 bool has_attribute = (stats.attribute_count > 0);
2373 bool has_varying = (stats.varying_count > 0);
2374 bool has_uniform = (stats.uniform_count > 0);
2375 bool has_ubo = (stats.uniform_buffer_count > 0);
2376
2377 if (has_texture || has_sampler)
2378 pandecode_msg("XXX: blend shader accessing textures\n");
2379
2380 if (has_attribute || has_varying)
2381 pandecode_msg("XXX: blend shader accessing interstage\n");
2382
2383 if (has_uniform || has_ubo)
2384 pandecode_msg("XXX: blend shader accessing uniforms\n");
2385 }
2386
2387 static void
2388 pandecode_textures(mali_ptr textures, unsigned texture_count, int job_no, bool is_bifrost)
2389 {
2390 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(textures);
2391
2392 if (!mmem)
2393 return;
2394
2395 if (is_bifrost) {
2396 const struct bifrost_texture_descriptor *PANDECODE_PTR_VAR(t, mmem, textures);
2397
2398 pandecode_log("uint64_t textures_%"PRIx64"_%d[] = {\n", textures, job_no);
2399 pandecode_indent++;
2400
2401 for (unsigned tex = 0; tex < texture_count; ++tex)
2402 pandecode_bifrost_texture(&t[tex], job_no, tex);
2403
2404 pandecode_indent--;
2405 pandecode_log("};\n");
2406 } else {
2407 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures);
2408
2409 pandecode_log("uint64_t textures_%"PRIx64"_%d[] = {\n", textures, job_no);
2410 pandecode_indent++;
2411
2412 for (int tex = 0; tex < texture_count; ++tex) {
2413 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2414 char *a = pointer_as_memory_reference(*u);
2415 pandecode_log("%s,\n", a);
2416 free(a);
2417 }
2418
2419 pandecode_indent--;
2420 pandecode_log("};\n");
2421
2422 /* Now, finally, descend down into the texture descriptor */
2423 for (unsigned tex = 0; tex < texture_count; ++tex) {
2424 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2425 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2426 if (tmem)
2427 pandecode_texture(*u, tmem, job_no, tex);
2428 }
2429 }
2430 }
2431
2432 static void
2433 pandecode_samplers(mali_ptr samplers, unsigned sampler_count, int job_no, bool is_bifrost)
2434 {
2435 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(samplers);
2436
2437 if (!smem)
2438 return;
2439
2440 if (is_bifrost) {
2441 struct bifrost_sampler_descriptor *s;
2442
2443 for (int i = 0; i < sampler_count; ++i) {
2444 s = pandecode_fetch_gpu_mem(smem, samplers + sizeof(*s) * i, sizeof(*s));
2445
2446 pandecode_log("struct bifrost_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", samplers + sizeof(*s) * i, job_no, i);
2447 pandecode_indent++;
2448
2449 if (s->unk1 != 1) {
2450 pandecode_msg("XXX: unk1 tripped\n");
2451 pandecode_prop("unk1 = 0x%x", s->unk1);
2452 }
2453
2454 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2455 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2456 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2457
2458 if (s->unk8 != 0x8) {
2459 pandecode_msg("XXX: unk8 tripped\n");
2460 pandecode_prop("unk8 = 0x%x", s->unk8);
2461 }
2462
2463 pandecode_prop("unk2 = 0x%x", s->unk2);
2464 pandecode_prop("unk3 = 0x%x", s->unk3);
2465 pandecode_prop("min_filter = %s", s->min_filter ? "nearest" : "linear");
2466 pandecode_prop("norm_coords = 0x%x", s->norm_coords & 0x1);
2467 pandecode_prop("zero1 = 0x%x", s->zero1 & 0x1);
2468 pandecode_prop("mip_filter = %s", s->mip_filter ? "linear" : "nearest");
2469 pandecode_prop("mag_filter = %s", s->mag_filter ? "linear" : "nearest");
2470
2471 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2472 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2473
2474 if (s->zero1 || s->zero2 || s->zero3 || s->zero4) {
2475 pandecode_msg("XXX: sampler zero tripped\n");
2476 pandecode_prop("zero = 0x%" PRIx8 ", 0x%" PRIx64 ", 0x%" PRIx64 ", 0x%" PRIx64 "\n", s->zero1, s->zero2, s->zero3, s->zero4);
2477 }
2478
2479 pandecode_indent--;
2480 pandecode_log("};\n");
2481 }
2482 } else {
2483 struct mali_sampler_descriptor *s;
2484
2485 for (int i = 0; i < sampler_count; ++i) {
2486 s = pandecode_fetch_gpu_mem(smem, samplers + sizeof(*s) * i, sizeof(*s));
2487
2488 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", samplers + sizeof(*s) * i, job_no, i);
2489 pandecode_indent++;
2490
2491 pandecode_log(".filter_mode = ");
2492 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2493 pandecode_log_cont(",\n");
2494
2495 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2496 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2497
2498 if (s->lod_bias)
2499 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s->lod_bias));
2500
2501 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2502 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2503 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2504
2505 pandecode_prop("compare_func = %s", pandecode_func(s->compare_func));
2506
2507 if (s->zero || s->zero2) {
2508 pandecode_msg("XXX: sampler zero tripped\n");
2509 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2510 }
2511
2512 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2513
2514 pandecode_prop("border_color = { %f, %f, %f, %f }",
2515 s->border_color[0],
2516 s->border_color[1],
2517 s->border_color[2],
2518 s->border_color[3]);
2519
2520 pandecode_indent--;
2521 pandecode_log("};\n");
2522 }
2523 }
2524 }
2525
2526 static void
2527 pandecode_vertex_tiler_postfix_pre(
2528 const struct mali_vertex_tiler_postfix *p,
2529 int job_no, enum mali_job_type job_type,
2530 char *suffix, bool is_bifrost, unsigned gpu_id)
2531 {
2532 struct pandecode_mapped_memory *attr_mem;
2533
2534 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2535 * are the only things actually needed from the FBD, vertex/tiler jobs
2536 * no longer reference the FBD -- instead, this field points to some
2537 * info about the scratchpad.
2538 */
2539
2540 struct pandecode_fbd fbd_info = {
2541 /* Default for Bifrost */
2542 .rt_count = 1
2543 };
2544
2545 if (is_bifrost) {
2546 pandecode_log_cont("\t/* %X %/\n", p->shared_memory & 1);
2547 pandecode_compute_fbd(p->shared_memory & ~1, job_no);
2548 } else if (p->shared_memory & MALI_MFBD)
2549 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->shared_memory) & FBD_MASK, job_no, false, job_type == JOB_TYPE_COMPUTE, false);
2550 else if (job_type == JOB_TYPE_COMPUTE)
2551 pandecode_compute_fbd((u64) (uintptr_t) p->shared_memory, job_no);
2552 else
2553 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->shared_memory, job_no, false, gpu_id);
2554
2555 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2556 int texture_count = 0, sampler_count = 0;
2557
2558 if (p->shader) {
2559 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2560 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2561
2562 /* Disassemble ahead-of-time to get stats. Initialize with
2563 * stats for the missing-shader case so we get validation
2564 * there, too */
2565
2566 struct midgard_disasm_stats info = {
2567 .texture_count = 0,
2568 .sampler_count = 0,
2569 .attribute_count = 0,
2570 .varying_count = 0,
2571 .work_count = 1,
2572
2573 .uniform_count = -128,
2574 .uniform_buffer_count = 0
2575 };
2576
2577 if (s->shader & ~0xF)
2578 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
2579
2580 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2581 pandecode_indent++;
2582
2583 /* Save for dumps */
2584 attribute_count = s->attribute_count;
2585 varying_count = s->varying_count;
2586 texture_count = s->texture_count;
2587 sampler_count = s->sampler_count;
2588
2589 if (is_bifrost) {
2590 uniform_count = s->bifrost2.uniform_count;
2591 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2592 } else {
2593 uniform_count = s->midgard1.uniform_count;
2594 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2595 }
2596
2597 pandecode_shader_address("shader", s->shader);
2598
2599 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2600 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2601 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2602 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2603 pandecode_shader_prop("uniform_buffer_count",
2604 uniform_buffer_count,
2605 info.uniform_buffer_count, true);
2606
2607 if (!is_bifrost) {
2608 pandecode_shader_prop("uniform_count",
2609 uniform_count,
2610 info.uniform_count, false);
2611
2612 pandecode_shader_prop("work_count",
2613 s->midgard1.work_count, info.work_count, false);
2614 }
2615
2616 if (is_bifrost) {
2617 pandecode_log("bifrost1.unk1 = ");
2618 pandecode_log_decoded_flags(shader_bifrost_info, s->bifrost1.unk1);
2619 pandecode_log_cont(",\n");
2620 } else {
2621 bool helpers = s->midgard1.flags_lo & MALI_HELPER_INVOCATIONS;
2622
2623 if (helpers != info.helper_invocations) {
2624 pandecode_msg("XXX: expected helpers %u but got %u\n",
2625 info.helper_invocations, helpers);
2626 }
2627
2628 pandecode_log(".midgard1.flags_lo = ");
2629 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info,
2630 s->midgard1.flags_lo & ~MALI_HELPER_INVOCATIONS);
2631 pandecode_log_cont(",\n");
2632
2633 pandecode_log(".midgard1.flags_hi = ");
2634 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info, s->midgard1.flags_hi);
2635 pandecode_log_cont(",\n");
2636 }
2637
2638 if (s->depth_units || s->depth_factor) {
2639 pandecode_prop("depth_factor = %f", s->depth_factor);
2640 pandecode_prop("depth_units = %f", s->depth_units);
2641 }
2642
2643 if (s->alpha_coverage) {
2644 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
2645 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
2646
2647 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2648 invert_alpha_coverage ? "~" : "",
2649 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
2650 }
2651
2652 if (s->unknown2_2)
2653 pandecode_prop(".unknown2_2 = %X", s->unknown2_2);
2654
2655 if (s->unknown2_3 || s->unknown2_4) {
2656 pandecode_log(".unknown2_3 = ");
2657
2658 int unknown2_3 = s->unknown2_3;
2659 int unknown2_4 = s->unknown2_4;
2660
2661 /* We're not quite sure what these flags mean without the depth test, if anything */
2662
2663 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2664 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2665 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2666
2667 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2668 }
2669
2670 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2671 pandecode_log_cont(",\n");
2672
2673 pandecode_log(".unknown2_4 = ");
2674 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2675 pandecode_log_cont(",\n");
2676 }
2677
2678 if (s->stencil_mask_front || s->stencil_mask_back) {
2679 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2680 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2681 }
2682
2683 pandecode_stencil("front", &s->stencil_front);
2684 pandecode_stencil("back", &s->stencil_back);
2685
2686 if (is_bifrost) {
2687 pandecode_log(".bifrost2 = {\n");
2688 pandecode_indent++;
2689
2690 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2691 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2692 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2693 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2694
2695 pandecode_indent--;
2696 pandecode_log("},\n");
2697 } else if (s->midgard2.unknown2_7) {
2698 pandecode_log(".midgard2 = {\n");
2699 pandecode_indent++;
2700
2701 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2702 pandecode_indent--;
2703 pandecode_log("},\n");
2704 }
2705
2706 if (s->padding) {
2707 pandecode_msg("XXX: shader padding tripped\n");
2708 pandecode_prop("padding = 0x%" PRIx32, s->padding);
2709 }
2710
2711 if (!is_bifrost) {
2712 /* TODO: Blend shaders routing/disasm */
2713 union midgard_blend blend = s->blend;
2714 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2715 if (shader & ~0xF)
2716 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2717 } else {
2718 pandecode_msg("mdg_blend = %" PRIx64 "\n", s->blend.shader);
2719 }
2720
2721 pandecode_indent--;
2722 pandecode_log("};\n");
2723
2724 /* MRT blend fields are used whenever MFBD is used, with
2725 * per-RT descriptors */
2726
2727 if (job_type == JOB_TYPE_TILER && (is_bifrost || p->shared_memory & MALI_MFBD)) {
2728 void* blend_base = (void *) (s + 1);
2729
2730 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2731 mali_ptr shader = 0;
2732
2733 if (is_bifrost)
2734 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2735 else
2736 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2737
2738 if (shader & ~0xF)
2739 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2740
2741 }
2742 }
2743 } else
2744 pandecode_msg("XXX: missing shader descriptor\n");
2745
2746 if (p->viewport) {
2747 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2748 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2749
2750 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2751 pandecode_indent++;
2752
2753 pandecode_prop("clip_minx = %f", f->clip_minx);
2754 pandecode_prop("clip_miny = %f", f->clip_miny);
2755 pandecode_prop("clip_minz = %f", f->clip_minz);
2756 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2757 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2758 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2759
2760 /* Only the higher coordinates are MALI_POSITIVE scaled */
2761
2762 pandecode_prop("viewport0 = { %d, %d }",
2763 f->viewport0[0], f->viewport0[1]);
2764
2765 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2766 f->viewport1[0] + 1, f->viewport1[1] + 1);
2767
2768 pandecode_indent--;
2769 pandecode_log("};\n");
2770 }
2771
2772 unsigned max_attr_index = 0;
2773
2774 if (p->attribute_meta)
2775 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2776
2777 if (p->attributes) {
2778 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2779 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2780 }
2781
2782 /* Varyings are encoded like attributes but not actually sent; we just
2783 * pass a zero buffer with the right stride/size set, (or whatever)
2784 * since the GPU will write to it itself */
2785
2786 if (p->varying_meta) {
2787 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2788 }
2789
2790 if (p->varyings) {
2791 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2792
2793 /* Number of descriptors depends on whether there are
2794 * non-internal varyings */
2795
2796 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2797 }
2798
2799 if (p->uniform_buffers) {
2800 if (uniform_buffer_count)
2801 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2802 else
2803 pandecode_msg("warn: UBOs specified but not referenced\n");
2804 } else if (uniform_buffer_count)
2805 pandecode_msg("XXX: UBOs referenced but not specified\n");
2806
2807 /* We don't want to actually dump uniforms, but we do need to validate
2808 * that the counts we were given are sane */
2809
2810 if (p->uniforms) {
2811 if (uniform_count)
2812 pandecode_uniforms(p->uniforms, uniform_count);
2813 else
2814 pandecode_msg("warn: Uniforms specified but not referenced\n");
2815 } else if (uniform_count)
2816 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2817
2818 if (p->textures)
2819 pandecode_textures(p->textures, texture_count, job_no, is_bifrost);
2820
2821 if (p->sampler_descriptor)
2822 pandecode_samplers(p->sampler_descriptor, sampler_count, job_no, is_bifrost);
2823 }
2824
2825 static void
2826 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2827 {
2828 pandecode_log(".gl_enables = ");
2829
2830 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2831
2832 pandecode_log_cont(",\n");
2833 }
2834
2835 static void
2836 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2837 {
2838 if (p->shader & 0xF)
2839 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2840
2841 pandecode_log(".postfix = {\n");
2842 pandecode_indent++;
2843
2844 pandecode_gl_enables(p->gl_enables, JOB_TYPE_TILER);
2845 pandecode_prop("instance_shift = 0x%x", p->instance_shift);
2846 pandecode_prop("instance_odd = 0x%x", p->instance_odd);
2847
2848 if (p->zero4) {
2849 pandecode_msg("XXX: vertex only zero tripped");
2850 pandecode_prop("zero4 = 0x%" PRIx32, p->zero4);
2851 }
2852
2853 pandecode_prop("offset_start = 0x%x", p->offset_start);
2854
2855 if (p->zero5) {
2856 pandecode_msg("XXX: vertex only zero tripped");
2857 pandecode_prop("zero5 = 0x%" PRIx64, p->zero5);
2858 }
2859
2860 MEMORY_PROP(p, position_varying);
2861 MEMORY_PROP(p, occlusion_counter);
2862
2863 pandecode_indent--;
2864 pandecode_log("},\n");
2865 }
2866
2867 static void
2868 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2869 {
2870 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2871 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2872
2873 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2874 pandecode_indent++;
2875
2876 if (h->zero) {
2877 pandecode_msg("XXX: tiler heap zero tripped\n");
2878 pandecode_prop("zero = 0x%x", h->zero);
2879 }
2880
2881 pandecode_prop("heap_size = 0x%x", h->heap_size);
2882 MEMORY_PROP(h, tiler_heap_start);
2883 MEMORY_PROP(h, tiler_heap_free);
2884
2885 /* this might point to the beginning of another buffer, when it's
2886 * really the end of the tiler heap buffer, so we have to be careful
2887 * here. but for zero length, we need the same pointer.
2888 */
2889
2890 if (h->tiler_heap_end == h->tiler_heap_start) {
2891 MEMORY_PROP(h, tiler_heap_start);
2892 } else {
2893 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2894 pandecode_prop("tiler_heap_end = %s + 1", a);
2895 free(a);
2896 }
2897
2898 for (int i = 0; i < 10; i++) {
2899 if (h->zeros[i] != 0) {
2900 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2901 i, h->zeros[i]);
2902 }
2903 }
2904
2905 if (h->unk1 != 0x1) {
2906 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2907 pandecode_prop("unk1 = 0x%x", h->unk1);
2908 }
2909
2910 if (h->unk7e007e != 0x7e007e) {
2911 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
2912 pandecode_prop("unk7e007e = 0x%x", h->unk7e007e);
2913 }
2914
2915 pandecode_indent--;
2916 pandecode_log("};\n");
2917 }
2918
2919 static void
2920 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2921 {
2922 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2923 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2924
2925 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2926
2927 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2928 pandecode_indent++;
2929
2930 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32, t->tiler_heap_next_start);
2931 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32, t->used_hierarchy_mask);
2932
2933 if (t->hierarchy_mask != 0xa &&
2934 t->hierarchy_mask != 0x14 &&
2935 t->hierarchy_mask != 0x28 &&
2936 t->hierarchy_mask != 0x50 &&
2937 t->hierarchy_mask != 0xa0)
2938 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
2939
2940 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2941
2942 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2943
2944 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2945 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2946
2947 if (t->zero0) {
2948 pandecode_msg("XXX: tiler meta zero tripped\n");
2949 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2950 }
2951
2952 for (int i = 0; i < 12; i++) {
2953 if (t->zeros[i] != 0) {
2954 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2955 i, t->zeros[i]);
2956 }
2957 }
2958
2959 pandecode_indent--;
2960 pandecode_log("};\n");
2961 }
2962
2963 static void
2964 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2965 {
2966 if (u.pointer == 0x0)
2967 return;
2968
2969 pandecode_log(".primitive_size = {\n");
2970 pandecode_indent++;
2971
2972 if (constant) {
2973 pandecode_prop("constant = %f", u.constant);
2974 } else {
2975 MEMORY_PROP((&u), pointer);
2976 }
2977
2978 pandecode_indent--;
2979 pandecode_log("},\n");
2980 }
2981
2982 static void
2983 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2984 {
2985 pandecode_log_cont("{\n");
2986 pandecode_indent++;
2987
2988 /* TODO: gl_PointSize on Bifrost */
2989 pandecode_primitive_size(t->primitive_size, true);
2990
2991 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2992 || t->zero6) {
2993 pandecode_msg("XXX: tiler only zero tripped\n");
2994 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2995 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2996 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2997 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2998 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2999 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
3000 }
3001
3002 pandecode_indent--;
3003 pandecode_log("},\n");
3004 }
3005
3006 static int
3007 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
3008 const struct pandecode_mapped_memory *mem,
3009 mali_ptr payload, int job_no, unsigned gpu_id)
3010 {
3011 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
3012
3013 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
3014
3015 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64"_%d = {\n", payload, job_no);
3016 pandecode_indent++;
3017
3018 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
3019 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
3020
3021 pandecode_indent--;
3022 pandecode_log("};\n");
3023
3024 return sizeof(*v);
3025 }
3026
3027 static int
3028 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
3029 const struct pandecode_mapped_memory *mem,
3030 mali_ptr payload, int job_no, unsigned gpu_id)
3031 {
3032 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
3033
3034 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
3035 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
3036
3037 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64"_%d = {\n", payload, job_no);
3038 pandecode_indent++;
3039
3040 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
3041
3042 pandecode_log(".tiler = ");
3043 pandecode_tiler_only_bfr(&t->tiler, job_no);
3044
3045 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
3046
3047 pandecode_indent--;
3048 pandecode_log("};\n");
3049
3050 return sizeof(*t);
3051 }
3052
3053 static int
3054 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
3055 const struct pandecode_mapped_memory *mem,
3056 mali_ptr payload, int job_no, unsigned gpu_id)
3057 {
3058 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
3059 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
3060
3061 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
3062
3063 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
3064 pandecode_indent++;
3065
3066 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
3067 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
3068
3069 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
3070 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
3071
3072 pandecode_indent--;
3073 pandecode_log("};\n");
3074
3075 return sizeof(*v);
3076 }
3077
3078 static int
3079 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
3080 mali_ptr payload, int job_no,
3081 bool is_bifrost, unsigned gpu_id)
3082 {
3083 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
3084
3085 bool is_mfbd = s->framebuffer & MALI_MFBD;
3086
3087 if (!is_mfbd && is_bifrost)
3088 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
3089
3090 struct pandecode_fbd info;
3091
3092 if (is_mfbd)
3093 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true, false, is_bifrost);
3094 else
3095 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
3096
3097 /* Compute the tag for the tagged pointer. This contains the type of
3098 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
3099 * additional structures follow the MFBD header (an extra payload or
3100 * not, as well as a count of render targets) */
3101
3102 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
3103
3104 if (is_mfbd) {
3105 if (info.has_extra)
3106 expected_tag |= MALI_MFBD_TAG_EXTRA;
3107
3108 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
3109 }
3110
3111 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
3112 pandecode_msg("XXX: unexpected tile coordinate bits\n");
3113 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
3114 pandecode_prop("max_tile_coord = 0x%X\n", s->max_tile_coord);
3115 }
3116
3117 /* Extract tile coordinates */
3118
3119 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
3120 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
3121
3122 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
3123 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
3124
3125 /* For the max, we also want the floored (rather than ceiled) version for checking */
3126
3127 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
3128 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
3129
3130 /* Validate the coordinates are well-ordered */
3131
3132 if (min_x == max_x)
3133 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
3134 else if (min_x > max_x)
3135 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
3136
3137 if (min_y == max_y)
3138 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
3139 else if (min_y > max_y)
3140 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
3141
3142 /* Validate the coordinates fit inside the framebuffer. We use floor,
3143 * rather than ceil, for the max coordinates, since the tile
3144 * coordinates for something like an 800x600 framebuffer will actually
3145 * resolve to 800x608, which would otherwise trigger a Y-overflow */
3146
3147 if ((min_x > info.width) || (max_x_f > info.width))
3148 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
3149
3150 if ((min_y > info.height) || (max_y_f > info.height))
3151 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
3152
3153 /* After validation, we print */
3154
3155 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
3156
3157 /* The FBD is a tagged pointer */
3158
3159 unsigned tag = (s->framebuffer & ~FBD_MASK);
3160
3161 if (tag != expected_tag)
3162 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
3163
3164 return sizeof(*s);
3165 }
3166
3167 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
3168 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
3169 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
3170 * might wish to combine this with the bifrost parameter) because some details
3171 * are model-specific even within a particular architecture. Minimal traces
3172 * *only* examine the job descriptors, skipping printing entirely if there is
3173 * no faults, and only descends into the payload if there are faults. This is
3174 * useful for looking for faults without the overhead of invasive traces. */
3175
3176 void
3177 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id, bool minimal)
3178 {
3179 struct mali_job_descriptor_header *h;
3180 unsigned job_descriptor_number = 0;
3181
3182 do {
3183 struct pandecode_mapped_memory *mem =
3184 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
3185
3186 void *payload;
3187
3188 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
3189
3190 /* On Midgard, for 32-bit jobs except for fragment jobs, the
3191 * high 32-bits of the 64-bit pointer are reused to store
3192 * something else.
3193 */
3194 int offset = h->job_descriptor_size == MALI_JOB_32 &&
3195 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
3196 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
3197
3198 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
3199
3200 int job_no = job_descriptor_number++;
3201
3202 /* If the job is good to go, skip it in minimal mode */
3203 if (minimal && (h->exception_status == 0x0 || h->exception_status == 0x1))
3204 continue;
3205
3206 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
3207 pandecode_indent++;
3208
3209 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
3210
3211 if (h->job_descriptor_size)
3212 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
3213
3214 if (h->exception_status && h->exception_status != 0x1)
3215 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
3216 h->exception_status,
3217 (h->exception_status >> 16) & 0xFFFF,
3218 pandecode_exception_access((h->exception_status >> 8) & 0x3),
3219 h->exception_status & 0xFF);
3220
3221 if (h->first_incomplete_task)
3222 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
3223
3224 if (h->fault_pointer)
3225 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
3226
3227 if (h->job_barrier)
3228 pandecode_prop("job_barrier = %d", h->job_barrier);
3229
3230 pandecode_prop("job_index = %d", h->job_index);
3231
3232 if (h->unknown_flags)
3233 pandecode_prop("unknown_flags = %d", h->unknown_flags);
3234
3235 if (h->job_dependency_index_1)
3236 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
3237
3238 if (h->job_dependency_index_2)
3239 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
3240
3241 pandecode_indent--;
3242 pandecode_log("};\n");
3243
3244 switch (h->job_type) {
3245 case JOB_TYPE_WRITE_VALUE: {
3246 struct mali_payload_write_value *s = payload;
3247 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
3248 pandecode_indent++;
3249 MEMORY_PROP(s, address);
3250
3251 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
3252 pandecode_msg("XXX: unknown value descriptor\n");
3253 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
3254 }
3255
3256 if (s->reserved) {
3257 pandecode_msg("XXX: set value tripped\n");
3258 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
3259 }
3260
3261 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
3262 pandecode_indent--;
3263 pandecode_log("};\n");
3264
3265 break;
3266 }
3267
3268 case JOB_TYPE_TILER:
3269 case JOB_TYPE_VERTEX:
3270 case JOB_TYPE_COMPUTE:
3271 if (bifrost) {
3272 if (h->job_type == JOB_TYPE_TILER)
3273 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
3274 else
3275 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
3276 } else
3277 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
3278
3279 break;
3280
3281 case JOB_TYPE_FRAGMENT:
3282 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
3283 break;
3284
3285 default:
3286 break;
3287 }
3288 } while ((jc_gpu_va = h->next_job));
3289 }