projects / mesa.git / blob
? search:


24ec19ee3240b30bcd36caa88de76766df8af794
[mesa.git] / src / panfrost / pandecode / decode.c
1 /*
2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 */
25
26 #include <panfrost-job.h>
27 #include <stdio.h>
28 #include <stdlib.h>
29 #include <memory.h>
30 #include <stdbool.h>
31 #include <stdarg.h>
32 #include <ctype.h>
33 #include "decode.h"
34 #include "util/macros.h"
35 #include "util/u_math.h"
36
37 #include "pan_pretty_print.h"
38 #include "midgard/disassemble.h"
39 #include "bifrost/disassemble.h"
40
41 #include "pan_encoder.h"
42
43 static void pandecode_swizzle(unsigned swizzle, enum mali_format format);
44
45 #define MEMORY_PROP(obj, p) {\
46 if (obj->p) { \
47 char *a = pointer_as_memory_reference(obj->p); \
48 pandecode_prop("%s = %s", #p, a); \
49 free(a); \
50 } \
51 }
52
53 #define MEMORY_PROP_DIR(obj, p) {\
54 if (obj.p) { \
55 char *a = pointer_as_memory_reference(obj.p); \
56 pandecode_prop("%s = %s", #p, a); \
57 free(a); \
58 } \
59 }
60
61 FILE *pandecode_dump_stream;
62
63 /* Semantic logging type.
64 *
65 * Raw: for raw messages to be printed as is.
66 * Message: for helpful information to be commented out in replays.
67 * Property: for properties of a struct
68 *
69 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
70 */
71
72 enum pandecode_log_type {
73 PANDECODE_RAW,
74 PANDECODE_MESSAGE,
75 PANDECODE_PROPERTY
76 };
77
78 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
79 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
80 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
81
82 unsigned pandecode_indent = 0;
83
84 static void
85 pandecode_make_indent(void)
86 {
87 for (unsigned i = 0; i < pandecode_indent; ++i)
88 fprintf(pandecode_dump_stream, " ");
89 }
90
91 static void
92 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
93 {
94 va_list ap;
95
96 pandecode_make_indent();
97
98 if (type == PANDECODE_MESSAGE)
99 fprintf(pandecode_dump_stream, "// ");
100 else if (type == PANDECODE_PROPERTY)
101 fprintf(pandecode_dump_stream, ".");
102
103 va_start(ap, format);
104 vfprintf(pandecode_dump_stream, format, ap);
105 va_end(ap);
106
107 if (type == PANDECODE_PROPERTY)
108 fprintf(pandecode_dump_stream, ",\n");
109 }
110
111 static void
112 pandecode_log_cont(const char *format, ...)
113 {
114 va_list ap;
115
116 va_start(ap, format);
117 vfprintf(pandecode_dump_stream, format, ap);
118 va_end(ap);
119 }
120
121 /* To check for memory safety issues, validates that the given pointer in GPU
122 * memory is valid, containing at least sz bytes. The goal is to eliminate
123 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
124 * overruns) by statically validating pointers.
125 */
126
127 static void
128 pandecode_validate_buffer(mali_ptr addr, size_t sz)
129 {
130 if (!addr) {
131 pandecode_msg("XXX: null pointer deref");
132 return;
133 }
134
135 /* Find a BO */
136
137 struct pandecode_mapped_memory *bo =
138 pandecode_find_mapped_gpu_mem_containing(addr);
139
140 if (!bo) {
141 pandecode_msg("XXX: invalid memory dereference\n");
142 return;
143 }
144
145 /* Bounds check */
146
147 unsigned offset = addr - bo->gpu_va;
148 unsigned total = offset + sz;
149
150 if (total > bo->length) {
151 pandecode_msg("XXX: buffer overrun. "
152 "Chunk of size %zu at offset %d in buffer of size %zu. "
153 "Overrun by %zu bytes. \n",
154 sz, offset, bo->length, total - bo->length);
155 return;
156 }
157 }
158
159 struct pandecode_flag_info {
160 u64 flag;
161 const char *name;
162 };
163
164 static void
165 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
166 u64 flags)
167 {
168 bool decodable_flags_found = false;
169
170 for (int i = 0; flag_info[i].name; i++) {
171 if ((flags & flag_info[i].flag) != flag_info[i].flag)
172 continue;
173
174 if (!decodable_flags_found) {
175 decodable_flags_found = true;
176 } else {
177 pandecode_log_cont(" | ");
178 }
179
180 pandecode_log_cont("%s", flag_info[i].name);
181
182 flags &= ~flag_info[i].flag;
183 }
184
185 if (decodable_flags_found) {
186 if (flags)
187 pandecode_log_cont(" | 0x%" PRIx64, flags);
188 } else {
189 pandecode_log_cont("0x%" PRIx64, flags);
190 }
191 }
192
193 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
194 static const struct pandecode_flag_info gl_enable_flag_info[] = {
195 FLAG_INFO(OCCLUSION_QUERY),
196 FLAG_INFO(OCCLUSION_PRECISE),
197 FLAG_INFO(FRONT_CCW_TOP),
198 FLAG_INFO(CULL_FACE_FRONT),
199 FLAG_INFO(CULL_FACE_BACK),
200 {}
201 };
202 #undef FLAG_INFO
203
204 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
205 static const struct pandecode_flag_info clear_flag_info[] = {
206 FLAG_INFO(FAST),
207 FLAG_INFO(SLOW),
208 FLAG_INFO(SLOW_STENCIL),
209 {}
210 };
211 #undef FLAG_INFO
212
213 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
214 static const struct pandecode_flag_info mask_flag_info[] = {
215 FLAG_INFO(R),
216 FLAG_INFO(G),
217 FLAG_INFO(B),
218 FLAG_INFO(A),
219 {}
220 };
221 #undef FLAG_INFO
222
223 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
224 static const struct pandecode_flag_info u3_flag_info[] = {
225 FLAG_INFO(HAS_MSAA),
226 FLAG_INFO(CAN_DISCARD),
227 FLAG_INFO(HAS_BLEND_SHADER),
228 FLAG_INFO(DEPTH_WRITEMASK),
229 {}
230 };
231
232 static const struct pandecode_flag_info u4_flag_info[] = {
233 FLAG_INFO(NO_MSAA),
234 FLAG_INFO(NO_DITHER),
235 FLAG_INFO(DEPTH_RANGE_A),
236 FLAG_INFO(DEPTH_RANGE_B),
237 FLAG_INFO(STENCIL_TEST),
238 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER),
239 {}
240 };
241 #undef FLAG_INFO
242
243 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
244 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
245 FLAG_INFO(MSAA),
246 FLAG_INFO(SRGB),
247 {}
248 };
249 #undef FLAG_INFO
250
251 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
252 static const struct pandecode_flag_info mfbd_extra_flag_hi_info[] = {
253 FLAG_INFO(PRESENT),
254 {}
255 };
256 #undef FLAG_INFO
257
258 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
259 static const struct pandecode_flag_info mfbd_extra_flag_lo_info[] = {
260 FLAG_INFO(ZS),
261 {}
262 };
263 #undef FLAG_INFO
264
265 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
266 static const struct pandecode_flag_info shader_midgard1_flag_lo_info [] = {
267 FLAG_INFO(WRITES_Z),
268 FLAG_INFO(EARLY_Z),
269 FLAG_INFO(READS_TILEBUFFER),
270 FLAG_INFO(READS_ZS),
271 {}
272 };
273
274 static const struct pandecode_flag_info shader_midgard1_flag_hi_info [] = {
275 FLAG_INFO(WRITES_S),
276 {}
277 };
278 #undef FLAG_INFO
279
280 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
281 static const struct pandecode_flag_info mfbd_flag_info [] = {
282 FLAG_INFO(DEPTH_WRITE),
283 FLAG_INFO(EXTRA),
284 {}
285 };
286 #undef FLAG_INFO
287
288 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
289 static const struct pandecode_flag_info sampler_flag_info [] = {
290 FLAG_INFO(MAG_NEAREST),
291 FLAG_INFO(MIN_NEAREST),
292 FLAG_INFO(MIP_LINEAR_1),
293 FLAG_INFO(MIP_LINEAR_2),
294 FLAG_INFO(NORM_COORDS),
295 {}
296 };
297 #undef FLAG_INFO
298
299 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
300 static const struct pandecode_flag_info sfbd_unk1_info [] = {
301 FLAG_INFO(MSAA_8),
302 FLAG_INFO(MSAA_A),
303 {}
304 };
305 #undef FLAG_INFO
306
307 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
308 static const struct pandecode_flag_info sfbd_unk2_info [] = {
309 FLAG_INFO(MSAA_B),
310 FLAG_INFO(SRGB),
311 {}
312 };
313 #undef FLAG_INFO
314
315 extern char *replace_fragment;
316 extern char *replace_vertex;
317
318 static char *
319 pandecode_job_type(enum mali_job_type type)
320 {
321 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
322
323 switch (type) {
324 DEFINE_CASE(NULL);
325 DEFINE_CASE(WRITE_VALUE);
326 DEFINE_CASE(CACHE_FLUSH);
327 DEFINE_CASE(COMPUTE);
328 DEFINE_CASE(VERTEX);
329 DEFINE_CASE(TILER);
330 DEFINE_CASE(FUSED);
331 DEFINE_CASE(FRAGMENT);
332
333 case JOB_NOT_STARTED:
334 return "NOT_STARTED";
335
336 default:
337 pandecode_log("Warning! Unknown job type %x\n", type);
338 return "!?!?!?";
339 }
340
341 #undef DEFINE_CASE
342 }
343
344 static char *
345 pandecode_draw_mode(enum mali_draw_mode mode)
346 {
347 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
348
349 switch (mode) {
350 DEFINE_CASE(DRAW_NONE);
351 DEFINE_CASE(POINTS);
352 DEFINE_CASE(LINES);
353 DEFINE_CASE(TRIANGLES);
354 DEFINE_CASE(TRIANGLE_STRIP);
355 DEFINE_CASE(TRIANGLE_FAN);
356 DEFINE_CASE(LINE_STRIP);
357 DEFINE_CASE(LINE_LOOP);
358 DEFINE_CASE(POLYGON);
359 DEFINE_CASE(QUADS);
360 DEFINE_CASE(QUAD_STRIP);
361
362 default:
363 pandecode_msg("XXX: invalid draw mode %X\n", mode);
364 return "";
365 }
366
367 #undef DEFINE_CASE
368 }
369
370 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
371 static char *
372 pandecode_func(enum mali_func mode)
373 {
374 switch (mode) {
375 DEFINE_CASE(NEVER);
376 DEFINE_CASE(LESS);
377 DEFINE_CASE(EQUAL);
378 DEFINE_CASE(LEQUAL);
379 DEFINE_CASE(GREATER);
380 DEFINE_CASE(NOTEQUAL);
381 DEFINE_CASE(GEQUAL);
382 DEFINE_CASE(ALWAYS);
383
384 default:
385 pandecode_msg("XXX: invalid func %X\n", mode);
386 return "";
387 }
388 }
389 #undef DEFINE_CASE
390
391 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
392 static char *
393 pandecode_stencil_op(enum mali_stencil_op op)
394 {
395 switch (op) {
396 DEFINE_CASE(KEEP);
397 DEFINE_CASE(REPLACE);
398 DEFINE_CASE(ZERO);
399 DEFINE_CASE(INVERT);
400 DEFINE_CASE(INCR_WRAP);
401 DEFINE_CASE(DECR_WRAP);
402 DEFINE_CASE(INCR);
403 DEFINE_CASE(DECR);
404
405 default:
406 pandecode_msg("XXX: invalid stencil op %X\n", op);
407 return "";
408 }
409 }
410
411 #undef DEFINE_CASE
412
413 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
414 {
415 switch(mode) {
416 /* TODO: Combine to just "instanced" once this can be done
417 * unambiguously in all known cases */
418 case MALI_ATTR_POT_DIVIDE:
419 return "instanced_pot";
420 case MALI_ATTR_MODULO:
421 return "instanced_mod";
422 case MALI_ATTR_NPOT_DIVIDE:
423 return "instanced_npot";
424 case MALI_ATTR_IMAGE:
425 return "image";
426 default:
427 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
428 return "";
429 }
430 }
431
432 static const char *
433 pandecode_special_record(uint64_t v, bool* attribute)
434 {
435 switch(v) {
436 case MALI_ATTR_VERTEXID:
437 *attribute = true;
438 return "gl_VertexID";
439 case MALI_ATTR_INSTANCEID:
440 *attribute = true;
441 return "gl_InstanceID";
442 case MALI_VARYING_FRAG_COORD:
443 return "gl_FragCoord";
444 case MALI_VARYING_FRONT_FACING:
445 return "gl_FrontFacing";
446 case MALI_VARYING_POINT_COORD:
447 return "gl_PointCoord";
448 default:
449 pandecode_msg("XXX: invalid special record %" PRIx64 "\n", v);
450 return "";
451 }
452 }
453
454 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
455 static char *
456 pandecode_wrap_mode(enum mali_wrap_mode op)
457 {
458 switch (op) {
459 DEFINE_CASE(REPEAT);
460 DEFINE_CASE(CLAMP_TO_EDGE);
461 DEFINE_CASE(CLAMP_TO_BORDER);
462 DEFINE_CASE(MIRRORED_REPEAT);
463
464 default:
465 pandecode_msg("XXX: invalid wrap mode %X\n", op);
466 return "";
467 }
468 }
469 #undef DEFINE_CASE
470
471 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
472 static char *
473 pandecode_block_format(enum mali_block_format fmt)
474 {
475 switch (fmt) {
476 DEFINE_CASE(TILED);
477 DEFINE_CASE(UNKNOWN);
478 DEFINE_CASE(LINEAR);
479 DEFINE_CASE(AFBC);
480
481 default:
482 unreachable("Invalid case");
483 }
484 }
485 #undef DEFINE_CASE
486
487 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
488 char *
489 pandecode_exception_access(unsigned access)
490 {
491 switch (access) {
492 DEFINE_CASE(NONE);
493 DEFINE_CASE(EXECUTE);
494 DEFINE_CASE(READ);
495 DEFINE_CASE(WRITE);
496
497 default:
498 unreachable("Invalid case");
499 }
500 }
501 #undef DEFINE_CASE
502
503 /* Midgard's tiler descriptor is embedded within the
504 * larger FBD */
505
506 static void
507 pandecode_midgard_tiler_descriptor(
508 const struct midgard_tiler_descriptor *t,
509 unsigned width,
510 unsigned height,
511 bool is_fragment,
512 bool has_hierarchy)
513 {
514 pandecode_log(".tiler = {\n");
515 pandecode_indent++;
516
517 if (t->hierarchy_mask == MALI_TILER_DISABLED)
518 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
519 else
520 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
521
522 /* We know this name from the kernel, but we never see it nonzero */
523
524 if (t->flags)
525 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
526
527 MEMORY_PROP(t, polygon_list);
528
529 /* The body is offset from the base of the polygon list */
530 //assert(t->polygon_list_body > t->polygon_list);
531 unsigned body_offset = t->polygon_list_body - t->polygon_list;
532
533 /* It needs to fit inside the reported size */
534 //assert(t->polygon_list_size >= body_offset);
535
536 /* Now that we've sanity checked, we'll try to calculate the sizes
537 * ourselves for comparison */
538
539 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
540 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
541
542 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
543 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
544 ref_header, ref_size);
545 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
546 pandecode_msg("body offset %d\n", body_offset);
547 }
548
549 /* The tiler heap has a start and end specified -- it should be
550 * identical to what we have in the BO. The exception is if tiling is
551 * disabled. */
552
553 MEMORY_PROP(t, heap_start);
554 assert(t->heap_end >= t->heap_start);
555
556 struct pandecode_mapped_memory *heap =
557 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
558
559 unsigned heap_size = t->heap_end - t->heap_start;
560
561 /* Tiling is enabled with a special flag */
562 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
563 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
564
565 bool tiling_enabled = hierarchy_mask;
566
567 if (tiling_enabled) {
568 /* When tiling is enabled, the heap should be a tight fit */
569 unsigned heap_offset = t->heap_start - heap->gpu_va;
570 if ((heap_offset + heap_size) != heap->length) {
571 pandecode_msg("XXX: heap size %u (expected %zu)\n",
572 heap_size, heap->length - heap_offset);
573 }
574
575 /* We should also have no other flags */
576 if (tiler_flags)
577 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
578 } else {
579 /* When tiling is disabled, we should have that flag and no others */
580
581 if (tiler_flags != MALI_TILER_DISABLED) {
582 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
583 tiler_flags);
584 }
585
586 /* We should also have an empty heap */
587 if (heap_size) {
588 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
589 heap_size);
590 }
591
592 /* Disabled tiling is used only for clear-only jobs, which are
593 * purely FRAGMENT, so we should never see this for
594 * non-FRAGMENT descriptors. */
595
596 if (!is_fragment)
597 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
598 }
599
600 /* We've never seen weights used in practice, but we know from the
601 * kernel these fields is there */
602
603 bool nonzero_weights = false;
604
605 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
606 nonzero_weights |= t->weights[w] != 0x0;
607 }
608
609 if (nonzero_weights) {
610 pandecode_log(".weights = { ");
611
612 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
613 pandecode_log_cont("%d, ", t->weights[w]);
614 }
615
616 pandecode_log("},");
617 }
618
619 pandecode_indent--;
620 pandecode_log("}\n");
621 }
622
623 /* TODO: The Bifrost tiler is not understood at all yet */
624
625 static void
626 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer *fb)
627 {
628 pandecode_log(".tiler = {\n");
629 pandecode_indent++;
630
631 MEMORY_PROP(fb, tiler_meta);
632
633 for (int i = 0; i < 16; i++) {
634 if (fb->zeros[i] != 0) {
635 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
636 i, fb->zeros[i]);
637 }
638 }
639
640 pandecode_log("},\n");
641
642 pandecode_indent--;
643 pandecode_log("}\n");
644
645 }
646
647 /* Information about the framebuffer passed back for
648 * additional analysis */
649
650 struct pandecode_fbd {
651 unsigned width;
652 unsigned height;
653 unsigned rt_count;
654 bool has_extra;
655 };
656
657 static void
658 pandecode_sfbd_format(struct mali_sfbd_format format)
659 {
660 pandecode_log(".format = {\n");
661 pandecode_indent++;
662
663 pandecode_log(".unk1 = ");
664 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
665 pandecode_log_cont(",\n");
666
667 /* TODO: Map formats so we can check swizzles and print nicely */
668 pandecode_log("swizzle");
669 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
670 pandecode_log_cont(",\n");
671
672 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
673 (format.nr_channels + 1));
674
675 pandecode_log(".unk2 = ");
676 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
677 pandecode_log_cont(",\n");
678
679 pandecode_prop("block = %s", pandecode_block_format(format.block));
680
681 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
682
683 pandecode_indent--;
684 pandecode_log("},\n");
685 }
686
687 static void
688 pandecode_shared_memory(const struct mali_shared_memory *desc, bool is_compute)
689 {
690 pandecode_prop("stack_shift = 0x%x", desc->stack_shift);
691
692 if (desc->unk0)
693 pandecode_prop("unk0 = 0x%x", desc->unk0);
694
695 if (desc->shared_workgroup_count != 0x1F) {
696 pandecode_prop("shared_workgroup_count = %d", desc->shared_workgroup_count);
697 if (!is_compute)
698 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
699 }
700
701 if (desc->shared_unk1 || desc->shared_shift) {
702 pandecode_prop("shared_unk1 = %X", desc->shared_unk1);
703 pandecode_prop("shared_shift = %X", desc->shared_shift);
704
705 if (!is_compute)
706 pandecode_msg("XXX: shared memory configured in noncompute shader");
707 }
708
709 if (desc->shared_zero) {
710 pandecode_msg("XXX: shared memory zero tripped\n");
711 pandecode_prop("shared_zero = 0x%" PRIx32, desc->shared_zero);
712 }
713
714 if (desc->shared_memory && !is_compute)
715 pandecode_msg("XXX: shared memory used in noncompute shader\n");
716
717 MEMORY_PROP(desc, scratchpad);
718 MEMORY_PROP(desc, shared_memory);
719 MEMORY_PROP(desc, unknown1);
720
721 if (desc->scratchpad) {
722 struct pandecode_mapped_memory *smem =
723 pandecode_find_mapped_gpu_mem_containing(desc->scratchpad);
724
725 pandecode_msg("scratchpad size %u\n", smem->length);
726 }
727
728 }
729
730 static struct pandecode_fbd
731 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
732 {
733 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
734 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
735
736 struct pandecode_fbd info = {
737 .has_extra = false,
738 .rt_count = 1
739 };
740
741 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
742 pandecode_indent++;
743
744 pandecode_log(".shared_memory = {\n");
745 pandecode_indent++;
746 pandecode_shared_memory(&s->shared_memory, false);
747 pandecode_indent--;
748 pandecode_log("},\n");
749
750 pandecode_sfbd_format(s->format);
751
752 info.width = s->width + 1;
753 info.height = s->height + 1;
754
755 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
756 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
757
758 MEMORY_PROP(s, checksum);
759
760 if (s->checksum_stride)
761 pandecode_prop("checksum_stride = %d", s->checksum_stride);
762
763 MEMORY_PROP(s, framebuffer);
764 pandecode_prop("stride = %d", s->stride);
765
766 /* Earlier in the actual commandstream -- right before width -- but we
767 * delay to flow nicer */
768
769 pandecode_log(".clear_flags = ");
770 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
771 pandecode_log_cont(",\n");
772
773 if (s->depth_buffer) {
774 MEMORY_PROP(s, depth_buffer);
775 pandecode_prop("depth_stride = %d", s->depth_stride);
776 }
777
778 if (s->stencil_buffer) {
779 MEMORY_PROP(s, stencil_buffer);
780 pandecode_prop("stencil_stride = %d", s->stencil_stride);
781 }
782
783 if (s->depth_stride_zero ||
784 s->stencil_stride_zero ||
785 s->zero7 || s->zero8) {
786 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
787 pandecode_prop("depth_stride_zero = 0x%x",
788 s->depth_stride_zero);
789 pandecode_prop("stencil_stride_zero = 0x%x",
790 s->stencil_stride_zero);
791 pandecode_prop("zero7 = 0x%" PRIx32,
792 s->zero7);
793 pandecode_prop("zero8 = 0x%" PRIx32,
794 s->zero8);
795 }
796
797 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
798 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
799 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
800 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
801 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
802 }
803
804 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
805 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
806 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
807 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
808 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
809 }
810
811 if (s->clear_stencil) {
812 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
813 }
814
815 const struct midgard_tiler_descriptor t = s->tiler;
816
817 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
818 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
819
820 pandecode_indent--;
821 pandecode_log("};\n");
822
823 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
824 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
825 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
826
827 pandecode_log_cont(".zero3 = {");
828
829 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
830 pandecode_log_cont("%X, ", s->zero3[i]);
831
832 pandecode_log_cont("},\n");
833
834 pandecode_log_cont(".zero6 = {");
835
836 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
837 pandecode_log_cont("%X, ", s->zero6[i]);
838
839 pandecode_log_cont("},\n");
840
841 return info;
842 }
843
844 static void
845 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
846 {
847 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
848 const struct mali_shared_memory *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
849
850 pandecode_log("struct mali_shared_memory shared_%"PRIx64"_%d = {\n", gpu_va, job_no);
851 pandecode_indent++;
852 pandecode_shared_memory(s, true);
853 pandecode_indent--;
854 pandecode_log("},\n");
855 }
856
857 /* Extracts the number of components associated with a Mali format */
858
859 static unsigned
860 pandecode_format_component_count(enum mali_format fmt)
861 {
862 /* Mask out the format class */
863 unsigned top = fmt & 0b11100000;
864
865 switch (top) {
866 case MALI_FORMAT_SNORM:
867 case MALI_FORMAT_UINT:
868 case MALI_FORMAT_UNORM:
869 case MALI_FORMAT_SINT:
870 return ((fmt >> 3) & 3) + 1;
871 default:
872 /* TODO: Validate */
873 return 4;
874 }
875 }
876
877 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
878
879 static unsigned
880 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
881 {
882 unsigned mask = 0;
883 assert(MALI_CHANNEL_RED == 0);
884
885 for (unsigned c = 0; c < 4; ++c) {
886 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
887
888 if (chan <= MALI_CHANNEL_ALPHA)
889 mask |= (1 << chan);
890 }
891
892 return mask;
893 }
894
895 /* Validates that a (format, swizzle) pair is valid, in the sense that the
896 * swizzle doesn't access any components that are undefined in the format.
897 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
898 * omitted */
899
900 static bool
901 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
902 {
903 unsigned nr_comp = pandecode_format_component_count(fmt);
904 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
905 unsigned valid_mask = (1 << nr_comp) - 1;
906 unsigned invalid_mask = ~valid_mask;
907
908 if (access_mask & invalid_mask) {
909 pandecode_msg("XXX: invalid components accessed\n");
910 return false;
911 }
912
913 /* Check for the default non-swizzling swizzle so we can suppress
914 * useless printing for the defaults */
915
916 unsigned default_swizzles[4] = {
917 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
918 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
919 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
920 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
921 };
922
923 return (swizzle == default_swizzles[nr_comp - 1]);
924 }
925
926 /* Maps MALI_RGBA32F to rgba32f, etc */
927
928 static void
929 pandecode_format_short(enum mali_format fmt, bool srgb)
930 {
931 /* We want a type-like format, so cut off the initial MALI_ */
932 char *format = pandecode_format(fmt);
933 format += strlen("MALI_");
934
935 unsigned len = strlen(format);
936 char *lower_format = calloc(1, len + 1);
937
938 for (unsigned i = 0; i < len; ++i)
939 lower_format[i] = tolower(format[i]);
940
941 /* Sanity check sRGB flag is applied to RGB, per the name */
942 if (srgb && lower_format[0] != 'r')
943 pandecode_msg("XXX: sRGB applied to non-colour format\n");
944
945 /* Just prefix with an s, so you get formats like srgba8_unorm */
946 if (srgb)
947 pandecode_log_cont("s");
948
949 pandecode_log_cont("%s", lower_format);
950 free(lower_format);
951 }
952
953 static void
954 pandecode_swizzle(unsigned swizzle, enum mali_format format)
955 {
956 /* First, do some validation */
957 bool trivial_swizzle = pandecode_validate_format_swizzle(
958 format, swizzle);
959
960 if (trivial_swizzle)
961 return;
962
963 /* Next, print the swizzle */
964 pandecode_log_cont(".");
965
966 static const char components[] = "rgba01";
967
968 for (unsigned c = 0; c < 4; ++c) {
969 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
970
971 if (chan >= MALI_CHANNEL_RESERVED_0) {
972 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
973 continue;
974 }
975 pandecode_log_cont("%c", components[chan]);
976 }
977 }
978
979 static void
980 pandecode_rt_format(struct mali_rt_format format)
981 {
982 pandecode_log(".format = {\n");
983 pandecode_indent++;
984
985 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
986 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
987 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
988 pandecode_prop("unk4 = 0x%" PRIx32, format.unk4);
989
990 pandecode_prop("block = %s", pandecode_block_format(format.block));
991
992 /* TODO: Map formats so we can check swizzles and print nicely */
993 pandecode_log("swizzle");
994 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
995 pandecode_log_cont(",\n");
996
997 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
998 (format.nr_channels + 1));
999
1000 pandecode_log(".flags = ");
1001 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
1002 pandecode_log_cont(",\n");
1003
1004 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
1005 * which is a faster hardware-based alternative to the wallpaper method
1006 * to preserve framebuffer contents across frames. In practice, MFBD
1007 * preload is buggy on Midgard, and so this is a chicken bit. If this
1008 * bit isn't set, most likely something broke unrelated to preload */
1009
1010 if (!format.no_preload) {
1011 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
1012 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
1013 }
1014
1015 if (format.zero)
1016 pandecode_prop("zero = 0x%" PRIx32, format.zero);
1017
1018 pandecode_indent--;
1019 pandecode_log("},\n");
1020 }
1021
1022 static void
1023 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct mali_framebuffer *fb)
1024 {
1025 pandecode_log("struct mali_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
1026 pandecode_indent++;
1027
1028 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
1029 mali_ptr rt_va = gpu_va + i * sizeof(struct mali_render_target);
1030 struct pandecode_mapped_memory *mem =
1031 pandecode_find_mapped_gpu_mem_containing(rt_va);
1032 const struct mali_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
1033
1034 pandecode_log("{\n");
1035 pandecode_indent++;
1036
1037 pandecode_rt_format(rt->format);
1038
1039 if (rt->format.block == MALI_BLOCK_AFBC) {
1040 pandecode_log(".afbc = {\n");
1041 pandecode_indent++;
1042
1043 char *a = pointer_as_memory_reference(rt->afbc.metadata);
1044 pandecode_prop("metadata = %s", a);
1045 free(a);
1046
1047 pandecode_prop("stride = %d", rt->afbc.stride);
1048 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
1049
1050 pandecode_indent--;
1051 pandecode_log("},\n");
1052 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
1053 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1054 rt->afbc.metadata,
1055 rt->afbc.stride,
1056 rt->afbc.unk);
1057 }
1058
1059 MEMORY_PROP(rt, framebuffer);
1060 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1061
1062 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1063 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1064 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1065 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1066 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1067 }
1068
1069 if (rt->zero1 || rt->zero2 || rt->zero3) {
1070 pandecode_msg("XXX: render target zeros tripped\n");
1071 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1072 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1073 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
1074 }
1075
1076 pandecode_indent--;
1077 pandecode_log("},\n");
1078 }
1079
1080 pandecode_indent--;
1081 pandecode_log("};\n");
1082 }
1083
1084 static struct pandecode_fbd
1085 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment, bool is_compute, bool is_bifrost)
1086 {
1087 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1088 const struct mali_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1089
1090 struct pandecode_fbd info;
1091
1092 if (is_bifrost && fb->msaa.sample_locations) {
1093 /* The blob stores all possible sample locations in a single buffer
1094 * allocated on startup, and just switches the pointer when switching
1095 * MSAA state. For now, we just put the data into the cmdstream, but we
1096 * should do something like what the blob does with a real driver.
1097 *
1098 * There seem to be 32 slots for sample locations, followed by another
1099 * 16. The second 16 is just the center location followed by 15 zeros
1100 * in all the cases I've identified (maybe shader vs. depth/color
1101 * samples?).
1102 */
1103
1104 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->msaa.sample_locations);
1105
1106 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->msaa.sample_locations);
1107
1108 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
1109 pandecode_indent++;
1110
1111 for (int i = 0; i < 32 + 16; i++) {
1112 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1113 }
1114
1115 pandecode_indent--;
1116 pandecode_log("};\n");
1117 }
1118
1119 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1120 pandecode_indent++;
1121
1122 if (is_bifrost) {
1123 pandecode_log(".msaa = {\n");
1124 pandecode_indent++;
1125
1126 if (fb->msaa.sample_locations)
1127 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1128 else
1129 pandecode_msg("XXX: sample_locations missing\n");
1130
1131 if (fb->msaa.zero1 || fb->msaa.zero2 || fb->msaa.zero4) {
1132 pandecode_msg("XXX: multisampling zero tripped\n");
1133 pandecode_prop("zero1 = %" PRIx64, fb->msaa.zero1);
1134 pandecode_prop("zero2 = %" PRIx64, fb->msaa.zero2);
1135 pandecode_prop("zero4 = %" PRIx64, fb->msaa.zero4);
1136 }
1137
1138 pandecode_indent--;
1139 pandecode_log("},\n");
1140 } else {
1141 pandecode_log(".shared_memory = {\n");
1142 pandecode_indent++;
1143 pandecode_shared_memory(&fb->shared_memory, is_compute);
1144 pandecode_indent--;
1145 pandecode_log("},\n");
1146 }
1147
1148 info.width = fb->width1 + 1;
1149 info.height = fb->height1 + 1;
1150 info.rt_count = fb->rt_count_1 + 1;
1151
1152 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1153 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1154 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1155 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1156
1157 pandecode_prop("unk1 = 0x%x", fb->unk1);
1158 pandecode_prop("unk2 = 0x%x", fb->unk2);
1159 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1160 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1161
1162 pandecode_log(".mfbd_flags = ");
1163 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1164 pandecode_log_cont(",\n");
1165
1166 if (fb->clear_stencil)
1167 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1168
1169 if (fb->clear_depth)
1170 pandecode_prop("clear_depth = %f", fb->clear_depth);
1171
1172 if (!is_compute)
1173 if (is_bifrost)
1174 pandecode_bifrost_tiler_descriptor(fb);
1175 else
1176 pandecode_midgard_tiler_descriptor(&fb->tiler, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
1177 else
1178 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1179
1180 if (fb->zero3 || fb->zero4) {
1181 pandecode_msg("XXX: framebuffer zeros tripped\n");
1182 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1183 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1184 }
1185
1186 pandecode_indent--;
1187 pandecode_log("};\n");
1188
1189 gpu_va += sizeof(struct mali_framebuffer);
1190
1191 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1192
1193 if (info.has_extra) {
1194 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1195 const struct mali_framebuffer_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1196
1197 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1198 pandecode_indent++;
1199
1200 MEMORY_PROP(fbx, checksum);
1201
1202 if (fbx->checksum_stride)
1203 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1204
1205 pandecode_log(".flags_hi = ");
1206 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_hi);
1207 pandecode_log_cont(",\n");
1208
1209 pandecode_log(".flags_lo = ");
1210 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
1211 pandecode_log_cont(",\n");
1212
1213 pandecode_prop("zs_block = %s", pandecode_block_format(fbx->zs_block));
1214
1215 if (fbx->zs_block == MALI_BLOCK_AFBC) {
1216 pandecode_log(".ds_afbc = {\n");
1217 pandecode_indent++;
1218
1219 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1220 pandecode_prop("depth_stencil_afbc_stride = %d",
1221 fbx->ds_afbc.depth_stencil_afbc_stride);
1222 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1223
1224 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
1225 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1226 pandecode_prop("zero1 = 0x%" PRIx32,
1227 fbx->ds_afbc.zero1);
1228 pandecode_prop("padding = 0x%" PRIx64,
1229 fbx->ds_afbc.padding);
1230 }
1231
1232 pandecode_indent--;
1233 pandecode_log("},\n");
1234 } else {
1235 pandecode_log(".ds_linear = {\n");
1236 pandecode_indent++;
1237
1238 if (fbx->ds_linear.depth) {
1239 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1240 pandecode_prop("depth_stride = %d",
1241 fbx->ds_linear.depth_stride);
1242 } else if (fbx->ds_linear.depth_stride) {
1243 pandecode_msg("XXX: depth stride zero tripped %d\n", fbx->ds_linear.depth_stride);
1244 }
1245
1246 if (fbx->ds_linear.stencil) {
1247 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1248 pandecode_prop("stencil_stride = %d",
1249 fbx->ds_linear.stencil_stride);
1250 } else if (fbx->ds_linear.stencil_stride) {
1251 pandecode_msg("XXX: stencil stride zero tripped %d\n", fbx->ds_linear.stencil_stride);
1252 }
1253
1254 if (fbx->ds_linear.depth_stride_zero ||
1255 fbx->ds_linear.stencil_stride_zero ||
1256 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1257 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1258 pandecode_prop("depth_stride_zero = 0x%x",
1259 fbx->ds_linear.depth_stride_zero);
1260 pandecode_prop("stencil_stride_zero = 0x%x",
1261 fbx->ds_linear.stencil_stride_zero);
1262 pandecode_prop("zero1 = 0x%" PRIx32,
1263 fbx->ds_linear.zero1);
1264 pandecode_prop("zero2 = 0x%" PRIx32,
1265 fbx->ds_linear.zero2);
1266 }
1267
1268 pandecode_indent--;
1269 pandecode_log("},\n");
1270 }
1271
1272 if (fbx->clear_color_1 | fbx->clear_color_2) {
1273 pandecode_prop("clear_color_1 = 0x%" PRIx32, fbx->clear_color_1);
1274 pandecode_prop("clear_color_2 = 0x%" PRIx32, fbx->clear_color_2);
1275 }
1276
1277 if (fbx->zero3) {
1278 pandecode_msg("XXX: fb_extra zeros tripped\n");
1279 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1280 }
1281
1282 pandecode_indent--;
1283 pandecode_log("};\n");
1284
1285 gpu_va += sizeof(struct mali_framebuffer_extra);
1286 }
1287
1288 if (is_fragment)
1289 pandecode_render_target(gpu_va, job_no, fb);
1290
1291 return info;
1292 }
1293
1294 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1295 * count */
1296
1297 static void
1298 pandecode_padded_vertices(unsigned shift, unsigned k)
1299 {
1300 unsigned odd = 2*k + 1;
1301 unsigned pot = 1 << shift;
1302 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1303 }
1304
1305 /* Given a magic divisor, recover what we were trying to divide by.
1306 *
1307 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1308 * 0 <= m < 2^N, for N bits in m.
1309 *
1310 * Let q represent the number we would like to divide by.
1311 *
1312 * By definition of a magic divisor for N-bit unsigned integers (a number you
1313 * multiply by to magically get division), m is a number such that:
1314 *
1315 * (m * x) & (2^N - 1) = floor(x/q).
1316 * for all x on Z where 0 <= x < 2^N
1317 *
1318 * Ignore the case where any of the above values equals zero; it is irrelevant
1319 * for our purposes (instanced arrays).
1320 *
1321 * Choose x = q. Then:
1322 *
1323 * (m * x) & (2^N - 1) = floor(x/q).
1324 * (m * q) & (2^N - 1) = floor(q/q).
1325 *
1326 * floor(q/q) = floor(1) = 1, therefore:
1327 *
1328 * (m * q) & (2^N - 1) = 1
1329 *
1330 * Recall the identity that the bitwise AND of one less than a power-of-two
1331 * equals the modulo with that power of two, i.e. for all x:
1332 *
1333 * x & (2^N - 1) = x % N
1334 *
1335 * Therefore:
1336 *
1337 * mq % (2^N) = 1
1338 *
1339 * By definition, a modular multiplicative inverse of a number m is the number
1340 * q such that with respect to a modulos M:
1341 *
1342 * mq % M = 1
1343 *
1344 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1345 *
1346 */
1347
1348 static void
1349 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1350 {
1351 #if 0
1352 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1353 * shift) the most lame way possible... just repeatedly add.
1354 * Asymptoptically slow but nobody cares in practice, unless you have
1355 * massive numbers of vertices or high divisors. */
1356
1357 unsigned inverse = 0;
1358
1359 /* Magic implicitly has the highest bit set */
1360 magic |= (1 << 31);
1361
1362 /* Depending on rounding direction */
1363 if (extra)
1364 magic++;
1365
1366 for (;;) {
1367 uint32_t product = magic * inverse;
1368
1369 if (shift) {
1370 product >>= shift;
1371 }
1372
1373 if (product == 1)
1374 break;
1375
1376 ++inverse;
1377 }
1378
1379 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1380
1381 /* Recall we're supposed to divide by (gl_level_divisor *
1382 * padded_num_vertices) */
1383
1384 unsigned padded_num_vertices = inverse / orig_divisor;
1385
1386 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1387 #endif
1388 }
1389
1390 static void
1391 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1392 mali_ptr addr, int job_no, char *suffix,
1393 int count, bool varying, enum mali_job_type job_type)
1394 {
1395 char *prefix = varying ? "varying" : "attribute";
1396 assert(addr);
1397
1398 if (!count) {
1399 pandecode_msg("warn: No %s records\n", prefix);
1400 return;
1401 }
1402
1403 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1404
1405 for (int i = 0; i < count; ++i) {
1406 /* First, check for special records */
1407 if (attr[i].elements < MALI_RECORD_SPECIAL) {
1408 if (attr[i].size)
1409 pandecode_msg("XXX: tripped size=%d\n", attr[i].size);
1410
1411 if (attr[i].stride) {
1412 /* gl_InstanceID passes a magic divisor in the
1413 * stride field to divide by the padded vertex
1414 * count. No other records should do so, so
1415 * stride should otherwise be zero. Note that
1416 * stride in the usual attribute sense doesn't
1417 * apply to special records. */
1418
1419 bool has_divisor = attr[i].elements == MALI_ATTR_INSTANCEID;
1420
1421 pandecode_log_cont("/* %smagic divisor = %X */ ",
1422 has_divisor ? "" : "XXX: ", attr[i].stride);
1423 }
1424
1425 if (attr[i].shift || attr[i].extra_flags) {
1426 /* Attributes use these fields for
1427 * instancing/padding/etc type issues, but
1428 * varyings don't */
1429
1430 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1431 varying ? "XXX: " : "",
1432 attr[i].shift, attr[i].extra_flags);
1433 }
1434
1435 /* Print the special record name */
1436 bool attribute = false;
1437 pandecode_log("%s_%d = %s;\n", prefix, i, pandecode_special_record(attr[i].elements, &attribute));
1438
1439 /* Sanity check */
1440 if (attribute == varying)
1441 pandecode_msg("XXX: mismatched special record\n");
1442
1443 continue;
1444 }
1445
1446 enum mali_attr_mode mode = attr[i].elements & 7;
1447
1448 if (mode == MALI_ATTR_UNUSED)
1449 pandecode_msg("XXX: unused attribute record\n");
1450
1451 /* For non-linear records, we need to print the type of record */
1452 if (mode != MALI_ATTR_LINEAR)
1453 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1454
1455 /* Print the name to link with attr_meta */
1456 pandecode_log_cont("%s_%d", prefix, i);
1457
1458 /* Print the stride and size */
1459 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1460
1461 /* TODO: Sanity check the quotient itself. It must be equal to
1462 * (or be greater than, if the driver added padding) the padded
1463 * vertex count. */
1464
1465 /* Finally, print the pointer */
1466 mali_ptr raw_elements = attr[i].elements & ~7;
1467 char *a = pointer_as_memory_reference(raw_elements);
1468 pandecode_log_cont(" = (%s);\n", a);
1469 free(a);
1470
1471 /* Check the pointer */
1472 pandecode_validate_buffer(raw_elements, attr[i].size);
1473
1474 /* shift/extra_flags exist only for instanced */
1475 if (attr[i].shift | attr[i].extra_flags) {
1476 /* These are set to random values by the blob for
1477 * varyings, most likely a symptom of uninitialized
1478 * memory where the hardware masked the bug. As such we
1479 * put this at a warning, not an error. */
1480
1481 if (mode == MALI_ATTR_LINEAR)
1482 pandecode_msg("warn: instancing fields set for linear\n");
1483
1484 pandecode_prop("shift = %d", attr[i].shift);
1485 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1486 }
1487
1488 /* Decode further where possible */
1489
1490 if (mode == MALI_ATTR_MODULO) {
1491 pandecode_padded_vertices(
1492 attr[i].shift,
1493 attr[i].extra_flags);
1494 }
1495
1496 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1497 i++;
1498 pandecode_log("{\n");
1499 pandecode_indent++;
1500 pandecode_prop("unk = 0x%x", attr[i].unk);
1501 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1502 if (attr[i].zero != 0)
1503 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1504 pandecode_prop("divisor = %d", attr[i].divisor);
1505 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1506 pandecode_indent--;
1507 pandecode_log("}, \n");
1508 }
1509
1510 }
1511
1512 pandecode_log("\n");
1513 }
1514
1515 static mali_ptr
1516 pandecode_shader_address(const char *name, mali_ptr ptr)
1517 {
1518 /* TODO: Decode flags */
1519 mali_ptr shader_ptr = ptr & ~15;
1520
1521 char *a = pointer_as_memory_reference(shader_ptr);
1522 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1523 free(a);
1524
1525 return shader_ptr;
1526 }
1527
1528 static void
1529 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1530 {
1531 unsigned any_nonzero =
1532 stencil->ref | stencil->mask | stencil->func |
1533 stencil->sfail | stencil->dpfail | stencil->dppass;
1534
1535 if (any_nonzero == 0)
1536 return;
1537
1538 const char *func = pandecode_func(stencil->func);
1539 const char *sfail = pandecode_stencil_op(stencil->sfail);
1540 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1541 const char *dppass = pandecode_stencil_op(stencil->dppass);
1542
1543 if (stencil->zero)
1544 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1545
1546 pandecode_log(".stencil_%s = {\n", name);
1547 pandecode_indent++;
1548 pandecode_prop("ref = %d", stencil->ref);
1549 pandecode_prop("mask = 0x%02X", stencil->mask);
1550 pandecode_prop("func = %s", func);
1551 pandecode_prop("sfail = %s", sfail);
1552 pandecode_prop("dpfail = %s", dpfail);
1553 pandecode_prop("dppass = %s", dppass);
1554 pandecode_indent--;
1555 pandecode_log("},\n");
1556 }
1557
1558 static void
1559 pandecode_blend_equation(const struct mali_blend_equation *blend)
1560 {
1561 if (blend->zero1)
1562 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1563
1564 pandecode_log(".equation = {\n");
1565 pandecode_indent++;
1566
1567 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1568 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1569
1570 pandecode_log(".color_mask = ");
1571 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1572 pandecode_log_cont(",\n");
1573
1574 pandecode_indent--;
1575 pandecode_log("},\n");
1576 }
1577
1578 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1579
1580 static unsigned
1581 decode_bifrost_constant(u16 constant)
1582 {
1583 float lo = (float) (constant & 0xFF);
1584 float hi = (float) (constant >> 8);
1585
1586 return (hi / 255.0) + (lo / 65535.0);
1587 }
1588
1589 static mali_ptr
1590 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1591 {
1592 struct bifrost_blend_rt *b =
1593 ((struct bifrost_blend_rt *) descs) + rt_no;
1594
1595 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1596 pandecode_indent++;
1597
1598 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1599 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1600 b->constant, decode_bifrost_constant(b->constant));
1601
1602 /* TODO figure out blend shader enable bit */
1603 pandecode_blend_equation(&b->equation);
1604
1605 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1606 pandecode_prop("index = 0x%" PRIx16, b->index);
1607
1608 pandecode_log(".format = ");
1609 pandecode_format_short(b->format, false);
1610 pandecode_swizzle(b->swizzle, b->format);
1611 pandecode_log_cont(",\n");
1612
1613 pandecode_prop("swizzle = 0x%" PRIx32, b->swizzle);
1614 pandecode_prop("format = 0x%" PRIx32, b->format);
1615
1616 if (b->zero1) {
1617 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1618 pandecode_prop("zero1 = 0x%" PRIx32, b->zero1);
1619 }
1620
1621 pandecode_log(".shader_type = ");
1622 switch(b->shader_type) {
1623 case BIFROST_BLEND_F16:
1624 pandecode_log_cont("BIFROST_BLEND_F16");
1625 break;
1626 case BIFROST_BLEND_F32:
1627 pandecode_log_cont("BIFROST_BLEND_F32");
1628 break;
1629 case BIFROST_BLEND_I32:
1630 pandecode_log_cont("BIFROST_BLEND_I32");
1631 break;
1632 case BIFROST_BLEND_U32:
1633 pandecode_log_cont("BIFROST_BLEND_U32");
1634 break;
1635 case BIFROST_BLEND_I16:
1636 pandecode_log_cont("BIFROST_BLEND_I16");
1637 break;
1638 case BIFROST_BLEND_U16:
1639 pandecode_log_cont("BIFROST_BLEND_U16");
1640 break;
1641 }
1642 pandecode_log_cont(",\n");
1643
1644 if (b->zero2) {
1645 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1646 pandecode_prop("zero2 = 0x%" PRIx32, b->zero2);
1647 }
1648
1649 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1650
1651 pandecode_indent--;
1652 pandecode_log("},\n");
1653
1654 return 0;
1655 }
1656
1657 static mali_ptr
1658 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1659 {
1660 /* constant/equation is in a union */
1661 if (!blend->shader)
1662 return 0;
1663
1664 pandecode_log(".blend = {\n");
1665 pandecode_indent++;
1666
1667 if (is_shader) {
1668 pandecode_shader_address("shader", blend->shader);
1669 } else {
1670 pandecode_blend_equation(&blend->equation);
1671 pandecode_prop("constant = %f", blend->constant);
1672 }
1673
1674 pandecode_indent--;
1675 pandecode_log("},\n");
1676
1677 /* Return blend shader to disassemble if present */
1678 return is_shader ? (blend->shader & ~0xF) : 0;
1679 }
1680
1681 static mali_ptr
1682 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1683 {
1684 struct midgard_blend_rt *b =
1685 ((struct midgard_blend_rt *) descs) + rt_no;
1686
1687 /* Flags determine presence of blend shader */
1688 bool is_shader = (b->flags & 0xF) >= 0x2;
1689
1690 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1691 pandecode_indent++;
1692
1693 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1694
1695 union midgard_blend blend = b->blend;
1696 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1697
1698 pandecode_indent--;
1699 pandecode_log("};\n");
1700
1701 return shader;
1702 }
1703
1704 /* Attributes and varyings have descriptor records, which contain information
1705 * about their format and ordering with the attribute/varying buffers. We'll
1706 * want to validate that the combinations specified are self-consistent.
1707 */
1708
1709 static int
1710 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1711 {
1712 char base[128];
1713 char *prefix = varying ? "varying" : "attribute";
1714 unsigned max_index = 0;
1715 snprintf(base, sizeof(base), "%s_meta", prefix);
1716
1717 struct mali_attr_meta *attr_meta;
1718 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1719
1720 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1721
1722 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1723 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1724 sizeof(*attr_mem));
1725
1726 /* If the record is discard, it should be zero for everything else */
1727
1728 if (attr_meta->format == MALI_VARYING_DISCARD) {
1729 uint64_t zero =
1730 attr_meta->index |
1731 attr_meta->unknown1 |
1732 attr_meta->unknown3 |
1733 attr_meta->src_offset;
1734
1735 if (zero)
1736 pandecode_msg("XXX: expected empty record for varying discard\n");
1737
1738 /* We want to look for a literal 0000 swizzle -- this
1739 * is not encoded with all zeroes, however */
1740
1741 enum mali_channel z = MALI_CHANNEL_ZERO;
1742 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1743 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1744
1745 if (!good_swizzle)
1746 pandecode_msg("XXX: expected zero swizzle for discard\n");
1747
1748 if (!varying)
1749 pandecode_msg("XXX: cannot discard attribute\n");
1750
1751 /* If we're all good, omit the record */
1752 if (!zero && varying && good_swizzle) {
1753 pandecode_log("/* discarded varying */\n");
1754 continue;
1755 }
1756 }
1757
1758 if (attr_meta->index > max_index)
1759 max_index = attr_meta->index;
1760
1761 if (attr_meta->unknown1 != 0x2) {
1762 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1763 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1764 }
1765
1766 if (attr_meta->unknown3) {
1767 pandecode_msg("XXX: unexpected unknown3 set\n");
1768 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1769 }
1770
1771 pandecode_format_short(attr_meta->format, false);
1772 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1773
1774 if (attr_meta->src_offset)
1775 pandecode_log_cont("[%u]", attr_meta->src_offset);
1776
1777 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1778
1779 pandecode_log_cont(";\n");
1780 }
1781
1782 pandecode_log("\n");
1783
1784 return count ? (max_index + 1) : 0;
1785 }
1786
1787 /* return bits [lo, hi) of word */
1788 static u32
1789 bits(u32 word, u32 lo, u32 hi)
1790 {
1791 if (hi - lo >= 32)
1792 return word; // avoid undefined behavior with the shift
1793
1794 return (word >> lo) & ((1 << (hi - lo)) - 1);
1795 }
1796
1797 static void
1798 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1799 {
1800 pandecode_log(".prefix = {\n");
1801 pandecode_indent++;
1802
1803 /* Decode invocation_count. See the comment before the definition of
1804 * invocation_count for an explanation.
1805 */
1806
1807 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1808 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1809 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1810 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1811 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1812 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1813
1814 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1815 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1816 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1817
1818 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1819 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1820 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1821
1822 /* Even though we have this decoded, we want to ensure that the
1823 * representation is "unique" so we don't lose anything by printing only
1824 * the final result. More specifically, we need to check that we were
1825 * passed something in canonical form, since the definition per the
1826 * hardware is inherently not unique. How? Well, take the resulting
1827 * decode and pack it ourselves! If it is bit exact with what we
1828 * decoded, we're good to go. */
1829
1830 struct mali_vertex_tiler_prefix ref;
1831 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1832
1833 bool canonical =
1834 (p->invocation_count == ref.invocation_count) &&
1835 (p->invocation_shifts == ref.invocation_shifts);
1836
1837 if (!canonical) {
1838 pandecode_msg("XXX: non-canonical workgroups packing\n");
1839 pandecode_msg("expected: %X, %X",
1840 ref.invocation_count,
1841 ref.invocation_shifts);
1842
1843 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1844 pandecode_prop("size_y_shift = %d", size_y_shift);
1845 pandecode_prop("size_z_shift = %d", size_z_shift);
1846 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
1847 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
1848 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
1849 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
1850 }
1851
1852 /* Regardless, print the decode */
1853 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1854 size_x, size_y, size_z,
1855 groups_x, groups_y, groups_z);
1856
1857 /* TODO: Decode */
1858 if (p->unknown_draw)
1859 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1860
1861 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1862
1863 if (p->draw_mode != MALI_DRAW_NONE)
1864 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1865
1866 /* Index count only exists for tiler jobs anyway */
1867
1868 if (p->index_count)
1869 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1870
1871
1872 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1873 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1874
1875 /* Validate an index buffer is present if we need one. TODO: verify
1876 * relationship between invocation_count and index_count */
1877
1878 if (p->indices) {
1879 unsigned count = p->index_count;
1880
1881 /* Grab the size */
1882 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1883
1884 /* Ensure we got a size, and if so, validate the index buffer
1885 * is large enough to hold a full set of indices of the given
1886 * size */
1887
1888 if (!index_raw_size)
1889 pandecode_msg("XXX: index size missing\n");
1890 else
1891 pandecode_validate_buffer(p->indices, count * size);
1892 } else if (index_raw_size)
1893 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1894
1895 if (p->offset_bias_correction)
1896 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1897
1898 /* TODO: Figure out what this is. It's not zero */
1899 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1900
1901 pandecode_indent--;
1902 pandecode_log("},\n");
1903 }
1904
1905 static void
1906 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1907 {
1908 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1909 uint64_t *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1910
1911 for (int i = 0; i < ubufs_count; i++) {
1912 unsigned size = (ubufs[i] & ((1 << 10) - 1)) * 16;
1913 mali_ptr addr = (ubufs[i] >> 10) << 2;
1914
1915 pandecode_validate_buffer(addr, size);
1916
1917 char *ptr = pointer_as_memory_reference(addr);
1918 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1919 free(ptr);
1920 }
1921
1922 pandecode_log("\n");
1923 }
1924
1925 static void
1926 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1927 {
1928 pandecode_validate_buffer(uniforms, uniform_count * 16);
1929
1930 char *ptr = pointer_as_memory_reference(uniforms);
1931 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1932 free(ptr);
1933 }
1934
1935 static const char *
1936 shader_type_for_job(unsigned type)
1937 {
1938 switch (type) {
1939 case JOB_TYPE_VERTEX: return "VERTEX";
1940 case JOB_TYPE_TILER: return "FRAGMENT";
1941 case JOB_TYPE_COMPUTE: return "COMPUTE";
1942 default:
1943 return "UNKNOWN";
1944 }
1945 }
1946
1947 static unsigned shader_id = 0;
1948
1949 static struct midgard_disasm_stats
1950 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1951 bool is_bifrost, unsigned gpu_id)
1952 {
1953 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1954 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1955
1956 /* Compute maximum possible size */
1957 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1958
1959 /* Print some boilerplate to clearly denote the assembly (which doesn't
1960 * obey indentation rules), and actually do the disassembly! */
1961
1962 pandecode_log_cont("\n\n");
1963
1964 struct midgard_disasm_stats stats;
1965
1966 if (is_bifrost) {
1967 disassemble_bifrost(pandecode_dump_stream, code, sz, true);
1968
1969 /* TODO: Extend stats to Bifrost */
1970 stats.texture_count = -128;
1971 stats.sampler_count = -128;
1972 stats.attribute_count = -128;
1973 stats.varying_count = -128;
1974 stats.uniform_count = -128;
1975 stats.uniform_buffer_count = -128;
1976 stats.work_count = -128;
1977
1978 stats.instruction_count = 0;
1979 stats.bundle_count = 0;
1980 stats.quadword_count = 0;
1981 stats.helper_invocations = false;
1982 } else {
1983 stats = disassemble_midgard(pandecode_dump_stream,
1984 code, sz, gpu_id,
1985 type == JOB_TYPE_TILER ?
1986 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
1987 }
1988
1989 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1990 * driver-internal purposes with the blob and interfere */
1991
1992 bool should_shaderdb = type != JOB_TYPE_COMPUTE;
1993
1994 if (should_shaderdb) {
1995 unsigned nr_threads =
1996 (stats.work_count <= 4) ? 4 :
1997 (stats.work_count <= 8) ? 2 :
1998 1;
1999
2000 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
2001 "%u inst, %u bundles, %u quadwords, "
2002 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
2003 shader_id++,
2004 shader_type_for_job(type),
2005 stats.instruction_count, stats.bundle_count, stats.quadword_count,
2006 stats.work_count, nr_threads);
2007 }
2008
2009
2010 return stats;
2011 }
2012
2013 static void
2014 pandecode_texture_payload(mali_ptr payload,
2015 enum mali_texture_type type,
2016 enum mali_texture_layout layout,
2017 bool manual_stride,
2018 uint8_t levels,
2019 uint16_t depth,
2020 uint16_t array_size,
2021 struct pandecode_mapped_memory *tmem)
2022 {
2023 pandecode_log(".payload = {\n");
2024 pandecode_indent++;
2025
2026 /* A bunch of bitmap pointers follow.
2027 * We work out the correct number,
2028 * based on the mipmap/cubemap
2029 * properties, but dump extra
2030 * possibilities to futureproof */
2031
2032 int bitmap_count = levels + 1;
2033
2034 /* Miptree for each face */
2035 if (type == MALI_TEX_CUBE)
2036 bitmap_count *= 6;
2037 else if (type == MALI_TEX_3D && layout == MALI_TEXTURE_LINEAR)
2038 bitmap_count *= (depth + 1);
2039
2040 /* Array of textures */
2041 bitmap_count *= (array_size + 1);
2042
2043 /* Stride for each element */
2044 if (manual_stride)
2045 bitmap_count *= 2;
2046
2047 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
2048 payload, sizeof(mali_ptr) * bitmap_count);
2049 for (int i = 0; i < bitmap_count; ++i) {
2050 /* How we dump depends if this is a stride or a pointer */
2051
2052 if (manual_stride && (i & 1)) {
2053 /* signed 32-bit snuck in as a 64-bit pointer */
2054 uint64_t stride_set = pointers_and_strides[i];
2055 uint32_t clamped_stride = stride_set;
2056 int32_t stride = clamped_stride;
2057 assert(stride_set == clamped_stride);
2058 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2059 } else {
2060 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
2061 pandecode_log("%s, \n", a);
2062 free(a);
2063 }
2064 }
2065
2066 pandecode_indent--;
2067 pandecode_log("},\n");
2068 }
2069
2070 static void
2071 pandecode_texture(mali_ptr u,
2072 struct pandecode_mapped_memory *tmem,
2073 unsigned job_no, unsigned tex)
2074 {
2075 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
2076
2077 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
2078 pandecode_indent++;
2079
2080 pandecode_prop("width = %" PRId32, t->width);
2081 pandecode_prop("height = %" PRId32, t->height);
2082 pandecode_prop("depth = %" PRId32, t->depth);
2083 pandecode_prop("array_size = %" PRId32, t->array_size);
2084
2085 pandecode_log("\n");
2086 pandecode_prop("f.swizzle = 0x%" PRIx32, t->format.swizzle);
2087 pandecode_prop("f.format = 0x%" PRIx32, t->format.format);
2088 pandecode_prop("f.srgb = 0x%" PRIx32, t->format.srgb);
2089 pandecode_prop("f.unknown1 = 0x%" PRIx32, t->format.unknown1);
2090 pandecode_prop("f.type = %" PRId32, t->format.type);
2091 pandecode_prop("f.layout = %" PRId32, t->format.layout);
2092 pandecode_prop("f.unknown2 = 0x%" PRIx32, t->format.unknown2);
2093 pandecode_prop("f.manual_stride = %" PRId32, t->format.manual_stride);
2094 pandecode_prop("f.zero = 0x%" PRIx32, t->format.zero);
2095 pandecode_log("\n");
2096
2097 pandecode_prop("unknown3 = 0x%" PRIx32, t->unknown3);
2098 pandecode_prop("unknown3A = 0x%" PRIx32, t->unknown3A);
2099 pandecode_prop("levels = %" PRId32, t->levels);
2100 pandecode_prop("swizzle = 0x%" PRIx32, t->swizzle);
2101 pandecode_prop("swizzle_zero = 0x%" PRIx32, t->swizzle_zero);
2102
2103 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2104 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2105 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2106 pandecode_log("\n");
2107
2108 struct mali_texture_format f = t->format;
2109
2110 /* See the definiton of enum mali_texture_type */
2111
2112 bool is_cube = f.type == MALI_TEX_CUBE;
2113 unsigned dimension = is_cube ? 2 : f.type;
2114
2115 pandecode_make_indent();
2116
2117 /* Print the layout. Default is linear; a modifier can denote AFBC or
2118 * u-interleaved/tiled modes */
2119
2120 if (f.layout == MALI_TEXTURE_AFBC)
2121 pandecode_log_cont("afbc");
2122 else if (f.layout == MALI_TEXTURE_TILED)
2123 pandecode_log_cont("tiled");
2124 else if (f.layout == MALI_TEXTURE_LINEAR)
2125 pandecode_log_cont("linear");
2126 else
2127 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
2128
2129 pandecode_swizzle(t->swizzle, f.format);
2130 pandecode_log_cont(" ");
2131
2132 /* Distinguish cube/2D with modifier */
2133
2134 if (is_cube)
2135 pandecode_log_cont("cube ");
2136
2137 pandecode_format_short(f.format, f.srgb);
2138 pandecode_swizzle(f.swizzle, f.format);
2139
2140 /* All four width/height/depth/array_size dimensions are present
2141 * regardless of the type of texture, but it is an error to have
2142 * non-zero dimensions for unused dimensions. Verify this. array_size
2143 * can always be set, as can width. */
2144
2145 if (t->height && dimension < 2)
2146 pandecode_msg("XXX: nonzero height for <2D texture\n");
2147
2148 if (t->depth && dimension < 3)
2149 pandecode_msg("XXX: nonzero depth for <2D texture\n");
2150
2151 /* Print only the dimensions that are actually there */
2152
2153 pandecode_log_cont(": %d", t->width + 1);
2154
2155 if (dimension >= 2)
2156 pandecode_log_cont("x%u", t->height + 1);
2157
2158 if (dimension >= 3)
2159 pandecode_log_cont("x%u", t->depth + 1);
2160
2161 if (t->array_size)
2162 pandecode_log_cont("[%u]", t->array_size + 1);
2163
2164 if (t->levels)
2165 pandecode_log_cont(" mip %u", t->levels);
2166
2167 pandecode_log_cont("\n");
2168
2169 if (f.unknown1 | f.zero) {
2170 pandecode_msg("XXX: texture format zero tripped\n");
2171 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
2172 pandecode_prop("zero = %" PRId32, f.zero);
2173 }
2174
2175 if (!f.unknown2) {
2176 pandecode_msg("XXX: expected unknown texture bit set\n");
2177 pandecode_prop("unknown2 = %" PRId32, f.unknown2);
2178 }
2179
2180 if (t->swizzle_zero) {
2181 pandecode_msg("XXX: swizzle zero tripped\n");
2182 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
2183 }
2184
2185 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
2186 pandecode_msg("XXX: texture zero tripped\n");
2187 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
2188 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
2189 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2190 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2191 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2192 }
2193
2194 pandecode_texture_payload(u + sizeof(*t), f.type, f.layout, f.manual_stride, t->levels, t->depth, t->array_size, tmem);
2195
2196 pandecode_indent--;
2197 pandecode_log("};\n");
2198 }
2199
2200 static void
2201 pandecode_bifrost_texture(
2202 const struct bifrost_texture_descriptor *t,
2203 unsigned job_no,
2204 unsigned tex)
2205 {
2206 pandecode_log("struct bifrost_texture_descriptor texture_descriptor_%d_%d = {\n", job_no, tex);
2207 pandecode_indent++;
2208
2209 pandecode_prop("format_unk = 0x%" PRIx32, t->format_unk);
2210 pandecode_prop("type = %" PRId32, t->type);
2211 pandecode_prop("format_unk2 = 0x%" PRIx32, t->format_unk2);
2212 pandecode_prop("format = 0x%" PRIx32, t->format);
2213 pandecode_prop("srgb = 0x%" PRIx32, t->srgb);
2214 pandecode_prop("format_unk3 = 0x%" PRIx32, t->format_unk3);
2215 pandecode_prop("width = %" PRId32, t->width);
2216 pandecode_prop("height = %" PRId32, t->height);
2217 pandecode_prop("swizzle = 0x%" PRIx32, t->swizzle);
2218 pandecode_prop("levels = %" PRId32, t->levels);
2219 pandecode_prop("unk1 = 0x%" PRIx32, t->unk1);
2220 pandecode_prop("levels_unk = %" PRId32, t->levels_unk);
2221 pandecode_prop("level_2 = %" PRId32, t->level_2);
2222 pandecode_prop("payload = 0x%" PRIx64, t->payload);
2223 pandecode_prop("array_size = %" PRId32, t->array_size);
2224 pandecode_prop("unk4 = 0x%" PRIx32, t->unk4);
2225 pandecode_prop("depth = %" PRId32, t->depth);
2226 pandecode_prop("unk5 = 0x%" PRIx32, t->unk5);
2227 pandecode_log("\n");
2228
2229 /* See the definiton of enum mali_texture_type */
2230
2231 bool is_cube = t->type == MALI_TEX_CUBE;
2232 unsigned dimension = is_cube ? 2 : t->type;
2233
2234 /* Print the layout. Default is linear; a modifier can denote AFBC or
2235 * u-interleaved/tiled modes */
2236
2237 if (t->layout == MALI_TEXTURE_AFBC)
2238 pandecode_log_cont("afbc");
2239 else if (t->layout == MALI_TEXTURE_TILED)
2240 pandecode_log_cont("tiled");
2241 else if (t->layout == MALI_TEXTURE_LINEAR)
2242 pandecode_log_cont("linear");
2243 else
2244 pandecode_msg("XXX: invalid texture layout 0x%X\n", t->layout);
2245
2246 pandecode_swizzle(t->swizzle, t->format);
2247 pandecode_log_cont(" ");
2248
2249 /* Distinguish cube/2D with modifier */
2250
2251 if (is_cube)
2252 pandecode_log_cont("cube ");
2253
2254 pandecode_format_short(t->format, t->srgb);
2255
2256 /* All four width/height/depth/array_size dimensions are present
2257 * regardless of the type of texture, but it is an error to have
2258 * non-zero dimensions for unused dimensions. Verify this. array_size
2259 * can always be set, as can width. */
2260
2261 if (t->height && dimension < 2)
2262 pandecode_msg("XXX: nonzero height for <2D texture\n");
2263
2264 if (t->depth && dimension < 3)
2265 pandecode_msg("XXX: nonzero depth for <2D texture\n");
2266
2267 /* Print only the dimensions that are actually there */
2268
2269 pandecode_log_cont(": %d", t->width + 1);
2270
2271 if (dimension >= 2)
2272 pandecode_log_cont("x%u", t->height + 1);
2273
2274 if (dimension >= 3)
2275 pandecode_log_cont("x%u", t->depth + 1);
2276
2277 if (t->array_size)
2278 pandecode_log_cont("[%u]", t->array_size + 1);
2279
2280 if (t->levels)
2281 pandecode_log_cont(" mip %u", t->levels);
2282
2283 pandecode_log_cont("\n");
2284
2285 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(t->payload);
2286 if (t->payload) {
2287 pandecode_texture_payload(t->payload, t->type, t->layout,
2288 true, t->levels, t->depth,
2289 t->array_size, tmem);
2290 }
2291
2292 pandecode_indent--;
2293 pandecode_log("};\n");
2294 }
2295
2296 /* 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 */
2297
2298 static void
2299 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2300 {
2301 /* Nothing to do */
2302 if (claim == truth)
2303 return;
2304
2305 if (fuzzy && (truth < 0))
2306 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name, claim, truth);
2307
2308 if ((truth >= 0) && !fuzzy) {
2309 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2310 (truth < claim) ? "warn" : "XXX",
2311 name, truth, claim);
2312 } else if ((claim > -truth) && !fuzzy) {
2313 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2314 name, -truth, claim);
2315 } else if (fuzzy && (claim < truth))
2316 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2317 name, truth, claim);
2318
2319 pandecode_log(".%s = %" PRId16, name, claim);
2320
2321 if (fuzzy)
2322 pandecode_log_cont(" /* %u used */", truth);
2323
2324 pandecode_log_cont(",\n");
2325 }
2326
2327 static void
2328 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2329 bool is_bifrost, unsigned gpu_id)
2330 {
2331 struct midgard_disasm_stats stats =
2332 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
2333
2334 bool has_texture = (stats.texture_count > 0);
2335 bool has_sampler = (stats.sampler_count > 0);
2336 bool has_attribute = (stats.attribute_count > 0);
2337 bool has_varying = (stats.varying_count > 0);
2338 bool has_uniform = (stats.uniform_count > 0);
2339 bool has_ubo = (stats.uniform_buffer_count > 0);
2340
2341 if (has_texture || has_sampler)
2342 pandecode_msg("XXX: blend shader accessing textures\n");
2343
2344 if (has_attribute || has_varying)
2345 pandecode_msg("XXX: blend shader accessing interstage\n");
2346
2347 if (has_uniform || has_ubo)
2348 pandecode_msg("XXX: blend shader accessing uniforms\n");
2349 }
2350
2351 static void
2352 pandecode_textures(mali_ptr textures, unsigned texture_count, int job_no, bool is_bifrost)
2353 {
2354 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(textures);
2355
2356 if (!mmem)
2357 return;
2358
2359 if (is_bifrost) {
2360 const struct bifrost_texture_descriptor *PANDECODE_PTR_VAR(t, mmem, textures);
2361
2362 pandecode_log("uint64_t textures_%"PRIx64"_%d[] = {\n", textures, job_no);
2363 pandecode_indent++;
2364
2365 for (unsigned tex = 0; tex < texture_count; ++tex)
2366 pandecode_bifrost_texture(&t[tex], job_no, tex);
2367
2368 pandecode_indent--;
2369 pandecode_log("};\n");
2370 } else {
2371 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures);
2372
2373 pandecode_log("uint64_t textures_%"PRIx64"_%d[] = {\n", textures, job_no);
2374 pandecode_indent++;
2375
2376 for (int tex = 0; tex < texture_count; ++tex) {
2377 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2378 char *a = pointer_as_memory_reference(*u);
2379 pandecode_log("%s,\n", a);
2380 free(a);
2381 }
2382
2383 pandecode_indent--;
2384 pandecode_log("};\n");
2385
2386 /* Now, finally, descend down into the texture descriptor */
2387 for (unsigned tex = 0; tex < texture_count; ++tex) {
2388 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2389 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2390 if (tmem)
2391 pandecode_texture(*u, tmem, job_no, tex);
2392 }
2393 }
2394 }
2395
2396 static void
2397 pandecode_samplers(mali_ptr samplers, unsigned sampler_count, int job_no, bool is_bifrost)
2398 {
2399 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(samplers);
2400
2401 if (!smem)
2402 return;
2403
2404 if (is_bifrost) {
2405 struct bifrost_sampler_descriptor *s;
2406
2407 for (int i = 0; i < sampler_count; ++i) {
2408 s = pandecode_fetch_gpu_mem(smem, samplers + sizeof(*s) * i, sizeof(*s));
2409
2410 pandecode_log("struct bifrost_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", samplers + sizeof(*s) * i, job_no, i);
2411 pandecode_indent++;
2412
2413 if (s->unk1 != 1) {
2414 pandecode_msg("XXX: unk1 tripped\n");
2415 pandecode_prop("unk1 = 0x%x", s->unk1);
2416 }
2417
2418 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2419 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2420 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2421
2422 if (s->unk8 != 0x8) {
2423 pandecode_msg("XXX: unk8 tripped\n");
2424 pandecode_prop("unk8 = 0x%x", s->unk8);
2425 }
2426
2427 pandecode_prop("unk2 = 0x%x", s->unk2);
2428 pandecode_prop("unk3 = 0x%x", s->unk3);
2429 pandecode_prop("min_filter = %s", s->min_filter ? "nearest" : "linear");
2430 pandecode_prop("norm_coords = 0x%x", s->norm_coords & 0x1);
2431 pandecode_prop("zero1 = 0x%x", s->zero1 & 0x1);
2432 pandecode_prop("mip_filter = %s", s->mip_filter ? "linear" : "nearest");
2433 pandecode_prop("mag_filter = %s", s->mag_filter ? "linear" : "nearest");
2434
2435 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2436 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2437
2438 if (s->zero1 || s->zero2 || s->zero3 || s->zero4) {
2439 pandecode_msg("XXX: sampler zero tripped\n");
2440 pandecode_prop("zero = 0x%" PRIx64 ", 0x%" PRIx64 ", 0x%" PRIx64 ", 0x%" PRIx64 "\n", s->zero1, s->zero2, s->zero3, s->zero4);
2441 }
2442
2443 pandecode_indent--;
2444 pandecode_log("};\n");
2445 }
2446 } else {
2447 struct mali_sampler_descriptor *s;
2448
2449 for (int i = 0; i < sampler_count; ++i) {
2450 s = pandecode_fetch_gpu_mem(smem, samplers + sizeof(*s) * i, sizeof(*s));
2451
2452 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", samplers + sizeof(*s) * i, job_no, i);
2453 pandecode_indent++;
2454
2455 pandecode_log(".filter_mode = ");
2456 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2457 pandecode_log_cont(",\n");
2458
2459 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2460 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2461
2462 if (s->lod_bias)
2463 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s->lod_bias));
2464
2465 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2466 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2467 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2468
2469 pandecode_prop("compare_func = %s", pandecode_func(s->compare_func));
2470
2471 if (s->zero || s->zero2) {
2472 pandecode_msg("XXX: sampler zero tripped\n");
2473 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2474 }
2475
2476 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2477
2478 pandecode_prop("border_color = { %f, %f, %f, %f }",
2479 s->border_color[0],
2480 s->border_color[1],
2481 s->border_color[2],
2482 s->border_color[3]);
2483
2484 pandecode_indent--;
2485 pandecode_log("};\n");
2486 }
2487 }
2488 }
2489
2490 static void
2491 pandecode_vertex_tiler_postfix_pre(
2492 const struct mali_vertex_tiler_postfix *p,
2493 int job_no, enum mali_job_type job_type,
2494 char *suffix, bool is_bifrost, unsigned gpu_id)
2495 {
2496 struct pandecode_mapped_memory *attr_mem;
2497
2498 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2499 * are the only things actually needed from the FBD, vertex/tiler jobs
2500 * no longer reference the FBD -- instead, this field points to some
2501 * info about the scratchpad.
2502 */
2503
2504 struct pandecode_fbd fbd_info = {
2505 /* Default for Bifrost */
2506 .rt_count = 1
2507 };
2508
2509 if (is_bifrost) {
2510 pandecode_log_cont("\t/* %X %/\n", p->shared_memory & 1);
2511 pandecode_compute_fbd(p->shared_memory & ~1, job_no);
2512 } else if (p->shared_memory & MALI_MFBD)
2513 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->shared_memory) & FBD_MASK, job_no, false, job_type == JOB_TYPE_COMPUTE, false);
2514 else if (job_type == JOB_TYPE_COMPUTE)
2515 pandecode_compute_fbd((u64) (uintptr_t) p->shared_memory, job_no);
2516 else
2517 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->shared_memory, job_no, false, gpu_id);
2518
2519 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2520 int texture_count = 0, sampler_count = 0;
2521
2522 if (p->shader) {
2523 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2524 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2525
2526 /* Disassemble ahead-of-time to get stats. Initialize with
2527 * stats for the missing-shader case so we get validation
2528 * there, too */
2529
2530 struct midgard_disasm_stats info = {
2531 .texture_count = 0,
2532 .sampler_count = 0,
2533 .attribute_count = 0,
2534 .varying_count = 0,
2535 .work_count = 1,
2536
2537 .uniform_count = -128,
2538 .uniform_buffer_count = 0
2539 };
2540
2541 if (s->shader & ~0xF)
2542 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
2543
2544 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2545 pandecode_indent++;
2546
2547 /* Save for dumps */
2548 attribute_count = s->attribute_count;
2549 varying_count = s->varying_count;
2550 texture_count = s->texture_count;
2551 sampler_count = s->sampler_count;
2552
2553 if (is_bifrost) {
2554 uniform_count = s->bifrost2.uniform_count;
2555 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2556 } else {
2557 uniform_count = s->midgard1.uniform_count;
2558 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2559 }
2560
2561 pandecode_shader_address("shader", s->shader);
2562
2563 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2564 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2565 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2566 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2567 pandecode_shader_prop("uniform_buffer_count",
2568 uniform_buffer_count,
2569 info.uniform_buffer_count, true);
2570
2571 if (!is_bifrost) {
2572 pandecode_shader_prop("uniform_count",
2573 uniform_count,
2574 info.uniform_count, false);
2575
2576 pandecode_shader_prop("work_count",
2577 s->midgard1.work_count, info.work_count, false);
2578 }
2579
2580 if (is_bifrost) {
2581 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32, s->bifrost1.unk1);
2582 } else {
2583 bool helpers = s->midgard1.flags_lo & MALI_HELPER_INVOCATIONS;
2584 s->midgard1.flags_lo &= ~MALI_HELPER_INVOCATIONS;
2585
2586 if (helpers != info.helper_invocations) {
2587 pandecode_msg("XXX: expected helpers %u but got %u\n",
2588 info.helper_invocations, helpers);
2589 }
2590
2591 pandecode_log(".midgard1.flags_lo = ");
2592 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info, s->midgard1.flags_lo);
2593 pandecode_log_cont(",\n");
2594
2595 pandecode_log(".midgard1.flags_hi = ");
2596 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info, s->midgard1.flags_hi);
2597 pandecode_log_cont(",\n");
2598 }
2599
2600 if (s->depth_units || s->depth_factor) {
2601 pandecode_prop("depth_factor = %f", s->depth_factor);
2602 pandecode_prop("depth_units = %f", s->depth_units);
2603 }
2604
2605 if (s->alpha_coverage) {
2606 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
2607 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
2608
2609 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2610 invert_alpha_coverage ? "~" : "",
2611 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
2612 }
2613
2614 if (s->unknown2_3 || s->unknown2_4) {
2615 pandecode_log(".unknown2_3 = ");
2616
2617 int unknown2_3 = s->unknown2_3;
2618 int unknown2_4 = s->unknown2_4;
2619
2620 /* We're not quite sure what these flags mean without the depth test, if anything */
2621
2622 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2623 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2624 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2625
2626 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2627 }
2628
2629 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2630 pandecode_log_cont(",\n");
2631
2632 pandecode_log(".unknown2_4 = ");
2633 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2634 pandecode_log_cont(",\n");
2635 }
2636
2637 if (s->stencil_mask_front || s->stencil_mask_back) {
2638 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2639 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2640 }
2641
2642 pandecode_stencil("front", &s->stencil_front);
2643 pandecode_stencil("back", &s->stencil_back);
2644
2645 if (is_bifrost) {
2646 pandecode_log(".bifrost2 = {\n");
2647 pandecode_indent++;
2648
2649 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2650 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2651 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2652 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2653
2654 pandecode_indent--;
2655 pandecode_log("},\n");
2656 } else if (s->midgard2.unknown2_7) {
2657 pandecode_log(".midgard2 = {\n");
2658 pandecode_indent++;
2659
2660 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2661 pandecode_indent--;
2662 pandecode_log("},\n");
2663 }
2664
2665 if (s->padding) {
2666 pandecode_msg("XXX: shader padding tripped\n");
2667 pandecode_prop("padding = 0x%" PRIx32, s->padding);
2668 }
2669
2670 if (!is_bifrost) {
2671 /* TODO: Blend shaders routing/disasm */
2672 union midgard_blend blend = s->blend;
2673 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2674 if (shader & ~0xF)
2675 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2676 }
2677
2678 pandecode_indent--;
2679 pandecode_log("};\n");
2680
2681 /* MRT blend fields are used whenever MFBD is used, with
2682 * per-RT descriptors */
2683
2684 if (job_type == JOB_TYPE_TILER && (is_bifrost || p->shared_memory & MALI_MFBD)) {
2685 void* blend_base = (void *) (s + 1);
2686
2687 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2688 mali_ptr shader = 0;
2689
2690 if (is_bifrost)
2691 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2692 else
2693 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2694
2695 if (shader & ~0xF)
2696 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2697
2698 }
2699 }
2700 } else
2701 pandecode_msg("XXX: missing shader descriptor\n");
2702
2703 if (p->viewport) {
2704 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2705 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2706
2707 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2708 pandecode_indent++;
2709
2710 pandecode_prop("clip_minx = %f", f->clip_minx);
2711 pandecode_prop("clip_miny = %f", f->clip_miny);
2712 pandecode_prop("clip_minz = %f", f->clip_minz);
2713 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2714 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2715 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2716
2717 /* Only the higher coordinates are MALI_POSITIVE scaled */
2718
2719 pandecode_prop("viewport0 = { %d, %d }",
2720 f->viewport0[0], f->viewport0[1]);
2721
2722 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2723 f->viewport1[0] + 1, f->viewport1[1] + 1);
2724
2725 pandecode_indent--;
2726 pandecode_log("};\n");
2727 }
2728
2729 unsigned max_attr_index = 0;
2730
2731 if (p->attribute_meta)
2732 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2733
2734 if (p->attributes) {
2735 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2736 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2737 }
2738
2739 /* Varyings are encoded like attributes but not actually sent; we just
2740 * pass a zero buffer with the right stride/size set, (or whatever)
2741 * since the GPU will write to it itself */
2742
2743 if (p->varying_meta) {
2744 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2745 }
2746
2747 if (p->varyings) {
2748 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2749
2750 /* Number of descriptors depends on whether there are
2751 * non-internal varyings */
2752
2753 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2754 }
2755
2756 if (p->uniform_buffers) {
2757 if (uniform_buffer_count)
2758 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2759 else
2760 pandecode_msg("warn: UBOs specified but not referenced\n");
2761 } else if (uniform_buffer_count)
2762 pandecode_msg("XXX: UBOs referenced but not specified\n");
2763
2764 /* We don't want to actually dump uniforms, but we do need to validate
2765 * that the counts we were given are sane */
2766
2767 if (p->uniforms) {
2768 if (uniform_count)
2769 pandecode_uniforms(p->uniforms, uniform_count);
2770 else
2771 pandecode_msg("warn: Uniforms specified but not referenced\n");
2772 } else if (uniform_count)
2773 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2774
2775 if (p->textures)
2776 pandecode_textures(p->textures, texture_count, job_no, is_bifrost);
2777
2778 if (p->sampler_descriptor)
2779 pandecode_samplers(p->sampler_descriptor, sampler_count, job_no, is_bifrost);
2780 }
2781
2782 static void
2783 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2784 {
2785 pandecode_log(".gl_enables = ");
2786
2787 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2788
2789 pandecode_log_cont(",\n");
2790 }
2791
2792 static void
2793 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2794 {
2795 if (p->shader & 0xF)
2796 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2797
2798 pandecode_log(".postfix = {\n");
2799 pandecode_indent++;
2800
2801 pandecode_gl_enables(p->gl_enables, JOB_TYPE_TILER);
2802 pandecode_prop("instance_shift = 0x%x", p->instance_shift);
2803 pandecode_prop("instance_odd = 0x%x", p->instance_odd);
2804
2805 if (p->zero4) {
2806 pandecode_msg("XXX: vertex only zero tripped");
2807 pandecode_prop("zero4 = 0x%" PRIx32, p->zero4);
2808 }
2809
2810 pandecode_prop("offset_start = 0x%x", p->offset_start);
2811
2812 if (p->zero5) {
2813 pandecode_msg("XXX: vertex only zero tripped");
2814 pandecode_prop("zero5 = 0x%" PRIx64, p->zero5);
2815 }
2816
2817 MEMORY_PROP(p, position_varying);
2818 MEMORY_PROP(p, occlusion_counter);
2819
2820 pandecode_indent--;
2821 pandecode_log("},\n");
2822 }
2823
2824 static void
2825 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2826 {
2827 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2828 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2829
2830 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2831 pandecode_indent++;
2832
2833 if (h->zero) {
2834 pandecode_msg("XXX: tiler heap zero tripped\n");
2835 pandecode_prop("zero = 0x%x", h->zero);
2836 }
2837
2838 pandecode_prop("heap_size = 0x%x", h->heap_size);
2839 MEMORY_PROP(h, tiler_heap_start);
2840 MEMORY_PROP(h, tiler_heap_free);
2841
2842 /* this might point to the beginning of another buffer, when it's
2843 * really the end of the tiler heap buffer, so we have to be careful
2844 * here. but for zero length, we need the same pointer.
2845 */
2846
2847 if (h->tiler_heap_end == h->tiler_heap_start) {
2848 MEMORY_PROP(h, tiler_heap_start);
2849 } else {
2850 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2851 pandecode_prop("tiler_heap_end = %s + 1", a);
2852 free(a);
2853 }
2854
2855 for (int i = 0; i < 10; i++) {
2856 if (h->zeros[i] != 0) {
2857 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2858 i, h->zeros[i]);
2859 }
2860 }
2861
2862 if (h->unk1 != 0x1) {
2863 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2864 pandecode_prop("unk1 = 0x%x", h->unk1);
2865 }
2866
2867 if (h->unk7e007e != 0x7e007e) {
2868 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
2869 pandecode_prop("unk7e007e = 0x%x", h->unk7e007e);
2870 }
2871
2872 pandecode_indent--;
2873 pandecode_log("};\n");
2874 }
2875
2876 static void
2877 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2878 {
2879 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2880 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2881
2882 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2883
2884 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2885 pandecode_indent++;
2886
2887 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32, t->tiler_heap_next_start);
2888 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32, t->used_hierarchy_mask);
2889
2890 if (t->hierarchy_mask != 0xa &&
2891 t->hierarchy_mask != 0x14 &&
2892 t->hierarchy_mask != 0x28 &&
2893 t->hierarchy_mask != 0x50 &&
2894 t->hierarchy_mask != 0xa0)
2895 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
2896
2897 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2898
2899 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2900
2901 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2902 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2903
2904 if (t->zero0) {
2905 pandecode_msg("XXX: tiler meta zero tripped\n");
2906 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2907 }
2908
2909 for (int i = 0; i < 12; i++) {
2910 if (t->zeros[i] != 0) {
2911 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2912 i, t->zeros[i]);
2913 }
2914 }
2915
2916 pandecode_indent--;
2917 pandecode_log("};\n");
2918 }
2919
2920 static void
2921 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2922 {
2923 if (u.pointer == 0x0)
2924 return;
2925
2926 pandecode_log(".primitive_size = {\n");
2927 pandecode_indent++;
2928
2929 if (constant) {
2930 pandecode_prop("constant = %f", u.constant);
2931 } else {
2932 MEMORY_PROP((&u), pointer);
2933 }
2934
2935 pandecode_indent--;
2936 pandecode_log("},\n");
2937 }
2938
2939 static void
2940 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2941 {
2942 pandecode_log_cont("{\n");
2943 pandecode_indent++;
2944
2945 /* TODO: gl_PointSize on Bifrost */
2946 pandecode_primitive_size(t->primitive_size, true);
2947
2948 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2949 || t->zero6) {
2950 pandecode_msg("XXX: tiler only zero tripped\n");
2951 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2952 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2953 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2954 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2955 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2956 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2957 }
2958
2959 pandecode_indent--;
2960 pandecode_log("},\n");
2961 }
2962
2963 static int
2964 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2965 const struct pandecode_mapped_memory *mem,
2966 mali_ptr payload, int job_no, unsigned gpu_id)
2967 {
2968 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2969
2970 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
2971
2972 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64"_%d = {\n", payload, job_no);
2973 pandecode_indent++;
2974
2975 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2976 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2977
2978 pandecode_indent--;
2979 pandecode_log("};\n");
2980
2981 return sizeof(*v);
2982 }
2983
2984 static int
2985 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2986 const struct pandecode_mapped_memory *mem,
2987 mali_ptr payload, int job_no, unsigned gpu_id)
2988 {
2989 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2990
2991 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
2992 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2993
2994 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64"_%d = {\n", payload, job_no);
2995 pandecode_indent++;
2996
2997 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2998
2999 pandecode_log(".tiler = ");
3000 pandecode_tiler_only_bfr(&t->tiler, job_no);
3001
3002 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
3003
3004 pandecode_indent--;
3005 pandecode_log("};\n");
3006
3007 return sizeof(*t);
3008 }
3009
3010 static int
3011 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
3012 const struct pandecode_mapped_memory *mem,
3013 mali_ptr payload, int job_no, unsigned gpu_id)
3014 {
3015 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
3016 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
3017
3018 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
3019
3020 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
3021 pandecode_indent++;
3022
3023 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
3024 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
3025
3026 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
3027 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
3028
3029 pandecode_indent--;
3030 pandecode_log("};\n");
3031
3032 return sizeof(*v);
3033 }
3034
3035 static int
3036 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
3037 mali_ptr payload, int job_no,
3038 bool is_bifrost, unsigned gpu_id)
3039 {
3040 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
3041
3042 bool is_mfbd = s->framebuffer & MALI_MFBD;
3043
3044 if (!is_mfbd && is_bifrost)
3045 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
3046
3047 struct pandecode_fbd info;
3048
3049 if (is_mfbd)
3050 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true, false, is_bifrost);
3051 else
3052 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
3053
3054 /* Compute the tag for the tagged pointer. This contains the type of
3055 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
3056 * additional structures follow the MFBD header (an extra payload or
3057 * not, as well as a count of render targets) */
3058
3059 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
3060
3061 if (is_mfbd) {
3062 if (info.has_extra)
3063 expected_tag |= MALI_MFBD_TAG_EXTRA;
3064
3065 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
3066 }
3067
3068 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
3069 pandecode_msg("XXX: unexpected tile coordinate bits\n");
3070 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
3071 pandecode_prop("max_tile_coord = 0x%X\n", s->max_tile_coord);
3072 }
3073
3074 /* Extract tile coordinates */
3075
3076 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
3077 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
3078
3079 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
3080 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
3081
3082 /* For the max, we also want the floored (rather than ceiled) version for checking */
3083
3084 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
3085 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
3086
3087 /* Validate the coordinates are well-ordered */
3088
3089 if (min_x == max_x)
3090 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
3091 else if (min_x > max_x)
3092 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
3093
3094 if (min_y == max_y)
3095 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
3096 else if (min_y > max_y)
3097 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
3098
3099 /* Validate the coordinates fit inside the framebuffer. We use floor,
3100 * rather than ceil, for the max coordinates, since the tile
3101 * coordinates for something like an 800x600 framebuffer will actually
3102 * resolve to 800x608, which would otherwise trigger a Y-overflow */
3103
3104 if ((min_x > info.width) || (max_x_f > info.width))
3105 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
3106
3107 if ((min_y > info.height) || (max_y_f > info.height))
3108 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
3109
3110 /* After validation, we print */
3111
3112 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
3113
3114 /* The FBD is a tagged pointer */
3115
3116 unsigned tag = (s->framebuffer & ~FBD_MASK);
3117
3118 if (tag != expected_tag)
3119 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
3120
3121 return sizeof(*s);
3122 }
3123
3124 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
3125 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
3126 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
3127 * might wish to combine this with the bifrost parameter) because some details
3128 * are model-specific even within a particular architecture. Minimal traces
3129 * *only* examine the job descriptors, skipping printing entirely if there is
3130 * no faults, and only descends into the payload if there are faults. This is
3131 * useful for looking for faults without the overhead of invasive traces. */
3132
3133 void
3134 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id, bool minimal)
3135 {
3136 struct mali_job_descriptor_header *h;
3137 unsigned job_descriptor_number = 0;
3138
3139 do {
3140 struct pandecode_mapped_memory *mem =
3141 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
3142
3143 void *payload;
3144
3145 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
3146
3147 /* On Midgard, for 32-bit jobs except for fragment jobs, the
3148 * high 32-bits of the 64-bit pointer are reused to store
3149 * something else.
3150 */
3151 int offset = h->job_descriptor_size == MALI_JOB_32 &&
3152 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
3153 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
3154
3155 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
3156
3157 int job_no = job_descriptor_number++;
3158
3159 /* If the job is good to go, skip it in minimal mode */
3160 if (minimal && (h->exception_status == 0x0 || h->exception_status == 0x1))
3161 continue;
3162
3163 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
3164 pandecode_indent++;
3165
3166 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
3167
3168 if (h->job_descriptor_size)
3169 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
3170
3171 if (h->exception_status && h->exception_status != 0x1)
3172 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
3173 h->exception_status,
3174 (h->exception_status >> 16) & 0xFFFF,
3175 pandecode_exception_access((h->exception_status >> 8) & 0x3),
3176 h->exception_status & 0xFF);
3177
3178 if (h->first_incomplete_task)
3179 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
3180
3181 if (h->fault_pointer)
3182 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
3183
3184 if (h->job_barrier)
3185 pandecode_prop("job_barrier = %d", h->job_barrier);
3186
3187 pandecode_prop("job_index = %d", h->job_index);
3188
3189 if (h->unknown_flags)
3190 pandecode_prop("unknown_flags = %d", h->unknown_flags);
3191
3192 if (h->job_dependency_index_1)
3193 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
3194
3195 if (h->job_dependency_index_2)
3196 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
3197
3198 pandecode_indent--;
3199 pandecode_log("};\n");
3200
3201 switch (h->job_type) {
3202 case JOB_TYPE_WRITE_VALUE: {
3203 struct mali_payload_write_value *s = payload;
3204 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
3205 pandecode_indent++;
3206 MEMORY_PROP(s, address);
3207
3208 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
3209 pandecode_msg("XXX: unknown value descriptor\n");
3210 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
3211 }
3212
3213 if (s->reserved) {
3214 pandecode_msg("XXX: set value tripped\n");
3215 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
3216 }
3217
3218 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
3219 pandecode_indent--;
3220 pandecode_log("};\n");
3221
3222 break;
3223 }
3224
3225 case JOB_TYPE_TILER:
3226 case JOB_TYPE_VERTEX:
3227 case JOB_TYPE_COMPUTE:
3228 if (bifrost) {
3229 if (h->job_type == JOB_TYPE_TILER)
3230 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
3231 else
3232 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
3233 } else
3234 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
3235
3236 break;
3237
3238 case JOB_TYPE_FRAGMENT:
3239 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
3240 break;
3241
3242 default:
3243 break;
3244 }
3245 } while ((jc_gpu_va = h->next_job));
3246 }