projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
panfrost: Don't use implicit mali_exception_status enum
[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_info [] = {
267 FLAG_INFO(EARLY_Z),
268 FLAG_INFO(READS_TILEBUFFER),
269 FLAG_INFO(READS_ZS),
270 {}
271 };
272 #undef FLAG_INFO
273
274 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
275 static const struct pandecode_flag_info mfbd_flag_info [] = {
276 FLAG_INFO(DEPTH_WRITE),
277 FLAG_INFO(EXTRA),
278 {}
279 };
280 #undef FLAG_INFO
281
282 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
283 static const struct pandecode_flag_info sampler_flag_info [] = {
284 FLAG_INFO(MAG_NEAREST),
285 FLAG_INFO(MIN_NEAREST),
286 FLAG_INFO(MIP_LINEAR_1),
287 FLAG_INFO(MIP_LINEAR_2),
288 FLAG_INFO(NORM_COORDS),
289 {}
290 };
291 #undef FLAG_INFO
292
293 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
294 static const struct pandecode_flag_info sfbd_unk1_info [] = {
295 FLAG_INFO(MSAA_8),
296 FLAG_INFO(MSAA_A),
297 {}
298 };
299 #undef FLAG_INFO
300
301 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
302 static const struct pandecode_flag_info sfbd_unk2_info [] = {
303 FLAG_INFO(MSAA_B),
304 FLAG_INFO(SRGB),
305 {}
306 };
307 #undef FLAG_INFO
308
309 extern char *replace_fragment;
310 extern char *replace_vertex;
311
312 static char *
313 pandecode_job_type(enum mali_job_type type)
314 {
315 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
316
317 switch (type) {
318 DEFINE_CASE(NULL);
319 DEFINE_CASE(WRITE_VALUE);
320 DEFINE_CASE(CACHE_FLUSH);
321 DEFINE_CASE(COMPUTE);
322 DEFINE_CASE(VERTEX);
323 DEFINE_CASE(TILER);
324 DEFINE_CASE(FUSED);
325 DEFINE_CASE(FRAGMENT);
326
327 case JOB_NOT_STARTED:
328 return "NOT_STARTED";
329
330 default:
331 pandecode_log("Warning! Unknown job type %x\n", type);
332 return "!?!?!?";
333 }
334
335 #undef DEFINE_CASE
336 }
337
338 static char *
339 pandecode_draw_mode(enum mali_draw_mode mode)
340 {
341 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
342
343 switch (mode) {
344 DEFINE_CASE(DRAW_NONE);
345 DEFINE_CASE(POINTS);
346 DEFINE_CASE(LINES);
347 DEFINE_CASE(TRIANGLES);
348 DEFINE_CASE(TRIANGLE_STRIP);
349 DEFINE_CASE(TRIANGLE_FAN);
350 DEFINE_CASE(LINE_STRIP);
351 DEFINE_CASE(LINE_LOOP);
352 DEFINE_CASE(POLYGON);
353 DEFINE_CASE(QUADS);
354 DEFINE_CASE(QUAD_STRIP);
355
356 default:
357 pandecode_msg("XXX: invalid draw mode %X\n", mode);
358 return "";
359 }
360
361 #undef DEFINE_CASE
362 }
363
364 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
365 static char *
366 pandecode_func(enum mali_func mode)
367 {
368 switch (mode) {
369 DEFINE_CASE(NEVER);
370 DEFINE_CASE(LESS);
371 DEFINE_CASE(EQUAL);
372 DEFINE_CASE(LEQUAL);
373 DEFINE_CASE(GREATER);
374 DEFINE_CASE(NOTEQUAL);
375 DEFINE_CASE(GEQUAL);
376 DEFINE_CASE(ALWAYS);
377
378 default:
379 pandecode_msg("XXX: invalid func %X\n", mode);
380 return "";
381 }
382 }
383 #undef DEFINE_CASE
384
385 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
386 static char *
387 pandecode_stencil_op(enum mali_stencil_op op)
388 {
389 switch (op) {
390 DEFINE_CASE(KEEP);
391 DEFINE_CASE(REPLACE);
392 DEFINE_CASE(ZERO);
393 DEFINE_CASE(INVERT);
394 DEFINE_CASE(INCR_WRAP);
395 DEFINE_CASE(DECR_WRAP);
396 DEFINE_CASE(INCR);
397 DEFINE_CASE(DECR);
398
399 default:
400 pandecode_msg("XXX: invalid stencil op %X\n", op);
401 return "";
402 }
403 }
404
405 #undef DEFINE_CASE
406
407 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
408 {
409 switch(mode) {
410 /* TODO: Combine to just "instanced" once this can be done
411 * unambiguously in all known cases */
412 case MALI_ATTR_POT_DIVIDE:
413 return "instanced_pot";
414 case MALI_ATTR_MODULO:
415 return "instanced_mod";
416 case MALI_ATTR_NPOT_DIVIDE:
417 return "instanced_npot";
418 case MALI_ATTR_IMAGE:
419 return "image";
420 default:
421 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
422 return "";
423 }
424 }
425
426 static const char *
427 pandecode_special_record(uint64_t v, bool* attribute)
428 {
429 switch(v) {
430 case MALI_ATTR_VERTEXID:
431 *attribute = true;
432 return "gl_VertexID";
433 case MALI_ATTR_INSTANCEID:
434 *attribute = true;
435 return "gl_InstanceID";
436 case MALI_VARYING_FRAG_COORD:
437 return "gl_FragCoord";
438 case MALI_VARYING_FRONT_FACING:
439 return "gl_FrontFacing";
440 case MALI_VARYING_POINT_COORD:
441 return "gl_PointCoord";
442 default:
443 pandecode_msg("XXX: invalid special record %" PRIx64 "\n", v);
444 return "";
445 }
446 }
447
448 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
449 static char *
450 pandecode_wrap_mode(enum mali_wrap_mode op)
451 {
452 switch (op) {
453 DEFINE_CASE(REPEAT);
454 DEFINE_CASE(CLAMP_TO_EDGE);
455 DEFINE_CASE(CLAMP_TO_BORDER);
456 DEFINE_CASE(MIRRORED_REPEAT);
457
458 default:
459 pandecode_msg("XXX: invalid wrap mode %X\n", op);
460 return "";
461 }
462 }
463 #undef DEFINE_CASE
464
465 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
466 static char *
467 pandecode_block_format(enum mali_block_format fmt)
468 {
469 switch (fmt) {
470 DEFINE_CASE(TILED);
471 DEFINE_CASE(UNKNOWN);
472 DEFINE_CASE(LINEAR);
473 DEFINE_CASE(AFBC);
474
475 default:
476 unreachable("Invalid case");
477 }
478 }
479 #undef DEFINE_CASE
480
481 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
482 char *
483 pandecode_exception_access(unsigned access)
484 {
485 switch (access) {
486 DEFINE_CASE(NONE);
487 DEFINE_CASE(EXECUTE);
488 DEFINE_CASE(READ);
489 DEFINE_CASE(WRITE);
490
491 default:
492 unreachable("Invalid case");
493 }
494 }
495 #undef DEFINE_CASE
496
497 /* Midgard's tiler descriptor is embedded within the
498 * larger FBD */
499
500 static void
501 pandecode_midgard_tiler_descriptor(
502 const struct midgard_tiler_descriptor *t,
503 unsigned width,
504 unsigned height,
505 bool is_fragment,
506 bool has_hierarchy)
507 {
508 pandecode_log(".tiler = {\n");
509 pandecode_indent++;
510
511 if (t->hierarchy_mask == MALI_TILER_DISABLED)
512 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
513 else
514 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
515
516 /* We know this name from the kernel, but we never see it nonzero */
517
518 if (t->flags)
519 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
520
521 MEMORY_PROP(t, polygon_list);
522
523 /* The body is offset from the base of the polygon list */
524 assert(t->polygon_list_body > t->polygon_list);
525 unsigned body_offset = t->polygon_list_body - t->polygon_list;
526
527 /* It needs to fit inside the reported size */
528 assert(t->polygon_list_size >= body_offset);
529
530 /* Check that we fit */
531 struct pandecode_mapped_memory *plist =
532 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
533
534 assert(t->polygon_list_size <= plist->length);
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("%d, ", t->weights[w]);
614 }
615
616 pandecode_log("},");
617 }
618
619 pandecode_indent--;
620 pandecode_log("}\n");
621 }
622
623 /* Information about the framebuffer passed back for
624 * additional analysis */
625
626 struct pandecode_fbd {
627 unsigned width;
628 unsigned height;
629 unsigned rt_count;
630 bool has_extra;
631 };
632
633 static void
634 pandecode_sfbd_format(struct mali_sfbd_format format)
635 {
636 pandecode_log(".format = {\n");
637 pandecode_indent++;
638
639 pandecode_log(".unk1 = ");
640 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
641 pandecode_log_cont(",\n");
642
643 /* TODO: Map formats so we can check swizzles and print nicely */
644 pandecode_log("swizzle");
645 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
646 pandecode_log_cont(",\n");
647
648 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
649 (format.nr_channels + 1));
650
651 pandecode_log(".unk2 = ");
652 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
653 pandecode_log_cont(",\n");
654
655 pandecode_prop("block = %s", pandecode_block_format(format.block));
656
657 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
658
659 pandecode_indent--;
660 pandecode_log("},\n");
661 }
662
663 static struct pandecode_fbd
664 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
665 {
666 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
667 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
668
669 struct pandecode_fbd info = {
670 .has_extra = false,
671 .rt_count = 1
672 };
673
674 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
675 pandecode_indent++;
676
677 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
678 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
679
680 pandecode_sfbd_format(s->format);
681
682 info.width = s->width + 1;
683 info.height = s->height + 1;
684
685 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
686 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
687
688 MEMORY_PROP(s, checksum);
689
690 if (s->checksum_stride)
691 pandecode_prop("checksum_stride = %d", s->checksum_stride);
692
693 MEMORY_PROP(s, framebuffer);
694 pandecode_prop("stride = %d", s->stride);
695
696 /* Earlier in the actual commandstream -- right before width -- but we
697 * delay to flow nicer */
698
699 pandecode_log(".clear_flags = ");
700 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
701 pandecode_log_cont(",\n");
702
703 if (s->depth_buffer) {
704 MEMORY_PROP(s, depth_buffer);
705 pandecode_prop("depth_stride = %d", s->depth_stride);
706 }
707
708 if (s->stencil_buffer) {
709 MEMORY_PROP(s, stencil_buffer);
710 pandecode_prop("stencil_stride = %d", s->stencil_stride);
711 }
712
713 if (s->depth_stride_zero ||
714 s->stencil_stride_zero ||
715 s->zero7 || s->zero8) {
716 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
717 pandecode_prop("depth_stride_zero = 0x%x",
718 s->depth_stride_zero);
719 pandecode_prop("stencil_stride_zero = 0x%x",
720 s->stencil_stride_zero);
721 pandecode_prop("zero7 = 0x%" PRIx32,
722 s->zero7);
723 pandecode_prop("zero8 = 0x%" PRIx32,
724 s->zero8);
725 }
726
727 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
728 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
729 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
730 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
731 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
732 }
733
734 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
735 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
736 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
737 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
738 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
739 }
740
741 if (s->clear_stencil) {
742 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
743 }
744
745 MEMORY_PROP(s, scratchpad);
746 const struct midgard_tiler_descriptor t = s->tiler;
747
748 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
749 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
750
751 pandecode_indent--;
752 pandecode_log("};\n");
753
754 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
755 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
756 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
757 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
758 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
759
760 pandecode_log_cont(".zero3 = {");
761
762 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
763 pandecode_log_cont("%X, ", s->zero3[i]);
764
765 pandecode_log_cont("},\n");
766
767 pandecode_log_cont(".zero6 = {");
768
769 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
770 pandecode_log_cont("%X, ", s->zero6[i]);
771
772 pandecode_log_cont("},\n");
773
774 return info;
775 }
776
777 static void
778 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
779 {
780 pandecode_log(".unknown%d = {", name);
781
782 for (int i = 0; i < count; ++i)
783 pandecode_log_cont("%X, ", slide[i]);
784
785 pandecode_log("},\n");
786 }
787
788 #define SHORT_SLIDE(num) \
789 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
790
791 static void
792 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
793 {
794 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
795 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
796
797 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
798 pandecode_indent++;
799
800 SHORT_SLIDE(1);
801
802 pandecode_indent--;
803 pandecode_log_cont("},\n");
804 }
805
806 /* Extracts the number of components associated with a Mali format */
807
808 static unsigned
809 pandecode_format_component_count(enum mali_format fmt)
810 {
811 /* Mask out the format class */
812 unsigned top = fmt & 0b11100000;
813
814 switch (top) {
815 case MALI_FORMAT_SNORM:
816 case MALI_FORMAT_UINT:
817 case MALI_FORMAT_UNORM:
818 case MALI_FORMAT_SINT:
819 return ((fmt >> 3) & 3) + 1;
820 default:
821 /* TODO: Validate */
822 return 4;
823 }
824 }
825
826 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
827
828 static unsigned
829 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
830 {
831 unsigned mask = 0;
832 assert(MALI_CHANNEL_RED == 0);
833
834 for (unsigned c = 0; c < 4; ++c) {
835 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
836
837 if (chan <= MALI_CHANNEL_ALPHA)
838 mask |= (1 << chan);
839 }
840
841 return mask;
842 }
843
844 /* Validates that a (format, swizzle) pair is valid, in the sense that the
845 * swizzle doesn't access any components that are undefined in the format.
846 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
847 * omitted */
848
849 static bool
850 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
851 {
852 unsigned nr_comp = pandecode_format_component_count(fmt);
853 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
854 unsigned valid_mask = (1 << nr_comp) - 1;
855 unsigned invalid_mask = ~valid_mask;
856
857 if (access_mask & invalid_mask) {
858 pandecode_msg("XXX: invalid components accessed\n");
859 return false;
860 }
861
862 /* Check for the default non-swizzling swizzle so we can suppress
863 * useless printing for the defaults */
864
865 unsigned default_swizzles[4] = {
866 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
867 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
868 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
869 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
870 };
871
872 return (swizzle == default_swizzles[nr_comp - 1]);
873 }
874
875 /* Maps MALI_RGBA32F to rgba32f, etc */
876
877 static void
878 pandecode_format_short(enum mali_format fmt, bool srgb)
879 {
880 /* We want a type-like format, so cut off the initial MALI_ */
881 char *format = pandecode_format(fmt);
882 format += strlen("MALI_");
883
884 unsigned len = strlen(format);
885 char *lower_format = calloc(1, len + 1);
886
887 for (unsigned i = 0; i < len; ++i)
888 lower_format[i] = tolower(format[i]);
889
890 /* Sanity check sRGB flag is applied to RGB, per the name */
891 if (srgb && lower_format[0] != 'r')
892 pandecode_msg("XXX: sRGB applied to non-colour format\n");
893
894 /* Just prefix with an s, so you get formats like srgba8_unorm */
895 if (srgb)
896 pandecode_log_cont("s");
897
898 pandecode_log_cont("%s", lower_format);
899 free(lower_format);
900 }
901
902 static void
903 pandecode_swizzle(unsigned swizzle, enum mali_format format)
904 {
905 /* First, do some validation */
906 bool trivial_swizzle = pandecode_validate_format_swizzle(
907 format, swizzle);
908
909 if (trivial_swizzle)
910 return;
911
912 /* Next, print the swizzle */
913 pandecode_log_cont(".");
914
915 static const char components[] = "rgba01";
916
917 for (unsigned c = 0; c < 4; ++c) {
918 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
919
920 if (chan >= MALI_CHANNEL_RESERVED_0) {
921 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
922 continue;
923 }
924 pandecode_log_cont("%c", components[chan]);
925 }
926 }
927
928 static void
929 pandecode_rt_format(struct mali_rt_format format)
930 {
931 pandecode_log(".format = {\n");
932 pandecode_indent++;
933
934 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
935 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
936 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
937
938 pandecode_prop("block = %s", pandecode_block_format(format.block));
939
940 /* TODO: Map formats so we can check swizzles and print nicely */
941 pandecode_log("swizzle");
942 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
943 pandecode_log_cont(",\n");
944
945 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
946 (format.nr_channels + 1));
947
948 pandecode_log(".flags = ");
949 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
950 pandecode_log_cont(",\n");
951
952 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
953 * which is a faster hardware-based alternative to the wallpaper method
954 * to preserve framebuffer contents across frames. In practice, MFBD
955 * preload is buggy on Midgard, and so this is a chicken bit. If this
956 * bit isn't set, most likely something broke unrelated to preload */
957
958 if (!format.no_preload) {
959 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
960 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
961 }
962
963 if (format.zero)
964 pandecode_prop("zero = 0x%" PRIx32, format.zero);
965
966 pandecode_indent--;
967 pandecode_log("},\n");
968 }
969
970 static void
971 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
972 {
973 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
974 pandecode_indent++;
975
976 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
977 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
978 struct pandecode_mapped_memory *mem =
979 pandecode_find_mapped_gpu_mem_containing(rt_va);
980 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
981
982 pandecode_log("{\n");
983 pandecode_indent++;
984
985 pandecode_rt_format(rt->format);
986
987 if (rt->format.block == MALI_BLOCK_AFBC) {
988 pandecode_log(".afbc = {\n");
989 pandecode_indent++;
990
991 char *a = pointer_as_memory_reference(rt->afbc.metadata);
992 pandecode_prop("metadata = %s", a);
993 free(a);
994
995 pandecode_prop("stride = %d", rt->afbc.stride);
996 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
997
998 pandecode_indent--;
999 pandecode_log("},\n");
1000 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
1001 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1002 rt->afbc.metadata,
1003 rt->afbc.stride,
1004 rt->afbc.unk);
1005 }
1006
1007 MEMORY_PROP(rt, framebuffer);
1008 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1009
1010 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1011 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1012 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1013 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1014 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1015 }
1016
1017 if (rt->zero1 || rt->zero2 || rt->zero3) {
1018 pandecode_msg("XXX: render target zeros tripped\n");
1019 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1020 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1021 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
1022 }
1023
1024 pandecode_indent--;
1025 pandecode_log("},\n");
1026 }
1027
1028 pandecode_indent--;
1029 pandecode_log("};\n");
1030 }
1031
1032 static struct pandecode_fbd
1033 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment)
1034 {
1035 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1036 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1037
1038 struct pandecode_fbd info;
1039
1040 if (fb->sample_locations) {
1041 /* The blob stores all possible sample locations in a single buffer
1042 * allocated on startup, and just switches the pointer when switching
1043 * MSAA state. For now, we just put the data into the cmdstream, but we
1044 * should do something like what the blob does with a real driver.
1045 *
1046 * There seem to be 32 slots for sample locations, followed by another
1047 * 16. The second 16 is just the center location followed by 15 zeros
1048 * in all the cases I've identified (maybe shader vs. depth/color
1049 * samples?).
1050 */
1051
1052 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
1053
1054 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
1055
1056 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
1057 pandecode_indent++;
1058
1059 for (int i = 0; i < 32 + 16; i++) {
1060 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1061 }
1062
1063 pandecode_indent--;
1064 pandecode_log("};\n");
1065 }
1066
1067 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1068 pandecode_indent++;
1069
1070 pandecode_prop("stack_shift = 0x%x", fb->stack_shift);
1071 pandecode_prop("unk0 = 0x%x", fb->unk0);
1072
1073 if (fb->sample_locations)
1074 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1075
1076 /* Assume that unknown1 was emitted in the last job for
1077 * now */
1078 MEMORY_PROP(fb, unknown1);
1079
1080 info.width = fb->width1 + 1;
1081 info.height = fb->height1 + 1;
1082 info.rt_count = fb->rt_count_1 + 1;
1083
1084 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1085 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1086 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1087 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1088
1089 pandecode_prop("unk1 = 0x%x", fb->unk1);
1090 pandecode_prop("unk2 = 0x%x", fb->unk2);
1091 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1092 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1093
1094 pandecode_log(".mfbd_flags = ");
1095 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1096 pandecode_log_cont(",\n");
1097
1098 if (fb->clear_stencil)
1099 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1100
1101 if (fb->clear_depth)
1102 pandecode_prop("clear_depth = %f", fb->clear_depth);
1103
1104 /* TODO: What is this? Let's not blow up.. */
1105 if (fb->unknown2 != 0x1F)
1106 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1107
1108 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1109 MEMORY_PROP(fb, scratchpad);
1110 const struct midgard_tiler_descriptor t = fb->tiler;
1111 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
1112
1113 if (fb->zero3 || fb->zero4) {
1114 pandecode_msg("XXX: framebuffer zeros tripped\n");
1115 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1116 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1117 }
1118
1119 pandecode_indent--;
1120 pandecode_log("};\n");
1121
1122 gpu_va += sizeof(struct bifrost_framebuffer);
1123
1124 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1125
1126 if (info.has_extra) {
1127 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1128 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1129
1130 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1131 pandecode_indent++;
1132
1133 MEMORY_PROP(fbx, checksum);
1134
1135 if (fbx->checksum_stride)
1136 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1137
1138 pandecode_log(".flags_hi = ");
1139 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_lo);
1140 pandecode_log_cont(",\n");
1141
1142 pandecode_log(".flags_lo = ");
1143 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
1144 pandecode_log_cont(",\n");
1145
1146 pandecode_prop("zs_block = %s\n", pandecode_block_format(fbx->zs_block));
1147
1148 if (fbx->zs_block == MALI_BLOCK_AFBC) {
1149 pandecode_log(".ds_afbc = {\n");
1150 pandecode_indent++;
1151
1152 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1153 pandecode_prop("depth_stencil_afbc_stride = %d",
1154 fbx->ds_afbc.depth_stencil_afbc_stride);
1155 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1156
1157 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
1158 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1159 pandecode_prop("zero1 = 0x%" PRIx32,
1160 fbx->ds_afbc.zero1);
1161 pandecode_prop("padding = 0x%" PRIx64,
1162 fbx->ds_afbc.padding);
1163 }
1164
1165 pandecode_indent--;
1166 pandecode_log("},\n");
1167 } else {
1168 pandecode_log(".ds_linear = {\n");
1169 pandecode_indent++;
1170
1171 if (fbx->ds_linear.depth) {
1172 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1173 pandecode_prop("depth_stride = %d",
1174 fbx->ds_linear.depth_stride);
1175 } else if (fbx->ds_linear.depth_stride) {
1176 pandecode_msg("XXX: depth stride zero tripped %d\n", fbx->ds_linear.depth_stride);
1177 }
1178
1179 if (fbx->ds_linear.stencil) {
1180 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1181 pandecode_prop("stencil_stride = %d",
1182 fbx->ds_linear.stencil_stride);
1183 } else if (fbx->ds_linear.stencil_stride) {
1184 pandecode_msg("XXX: stencil stride zero tripped %d\n", fbx->ds_linear.stencil_stride);
1185 }
1186
1187 if (fbx->ds_linear.depth_stride_zero ||
1188 fbx->ds_linear.stencil_stride_zero ||
1189 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1190 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1191 pandecode_prop("depth_stride_zero = 0x%x",
1192 fbx->ds_linear.depth_stride_zero);
1193 pandecode_prop("stencil_stride_zero = 0x%x",
1194 fbx->ds_linear.stencil_stride_zero);
1195 pandecode_prop("zero1 = 0x%" PRIx32,
1196 fbx->ds_linear.zero1);
1197 pandecode_prop("zero2 = 0x%" PRIx32,
1198 fbx->ds_linear.zero2);
1199 }
1200
1201 pandecode_indent--;
1202 pandecode_log("},\n");
1203 }
1204
1205 if (fbx->zero3 || fbx->zero4) {
1206 pandecode_msg("XXX: fb_extra zeros tripped\n");
1207 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1208 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1209 }
1210
1211 pandecode_indent--;
1212 pandecode_log("};\n");
1213
1214 gpu_va += sizeof(struct bifrost_fb_extra);
1215 }
1216
1217 if (is_fragment)
1218 pandecode_render_target(gpu_va, job_no, fb);
1219
1220 return info;
1221 }
1222
1223 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1224 * count */
1225
1226 static void
1227 pandecode_padded_vertices(unsigned shift, unsigned k)
1228 {
1229 unsigned odd = 2*k + 1;
1230 unsigned pot = 1 << shift;
1231 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1232 }
1233
1234 /* Given a magic divisor, recover what we were trying to divide by.
1235 *
1236 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1237 * 0 <= m < 2^N, for N bits in m.
1238 *
1239 * Let q represent the number we would like to divide by.
1240 *
1241 * By definition of a magic divisor for N-bit unsigned integers (a number you
1242 * multiply by to magically get division), m is a number such that:
1243 *
1244 * (m * x) & (2^N - 1) = floor(x/q).
1245 * for all x on Z where 0 <= x < 2^N
1246 *
1247 * Ignore the case where any of the above values equals zero; it is irrelevant
1248 * for our purposes (instanced arrays).
1249 *
1250 * Choose x = q. Then:
1251 *
1252 * (m * x) & (2^N - 1) = floor(x/q).
1253 * (m * q) & (2^N - 1) = floor(q/q).
1254 *
1255 * floor(q/q) = floor(1) = 1, therefore:
1256 *
1257 * (m * q) & (2^N - 1) = 1
1258 *
1259 * Recall the identity that the bitwise AND of one less than a power-of-two
1260 * equals the modulo with that power of two, i.e. for all x:
1261 *
1262 * x & (2^N - 1) = x % N
1263 *
1264 * Therefore:
1265 *
1266 * mq % (2^N) = 1
1267 *
1268 * By definition, a modular multiplicative inverse of a number m is the number
1269 * q such that with respect to a modulos M:
1270 *
1271 * mq % M = 1
1272 *
1273 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1274 *
1275 */
1276
1277 static void
1278 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1279 {
1280 #if 0
1281 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1282 * shift) the most lame way possible... just repeatedly add.
1283 * Asymptoptically slow but nobody cares in practice, unless you have
1284 * massive numbers of vertices or high divisors. */
1285
1286 unsigned inverse = 0;
1287
1288 /* Magic implicitly has the highest bit set */
1289 magic |= (1 << 31);
1290
1291 /* Depending on rounding direction */
1292 if (extra)
1293 magic++;
1294
1295 for (;;) {
1296 uint32_t product = magic * inverse;
1297
1298 if (shift) {
1299 product >>= shift;
1300 }
1301
1302 if (product == 1)
1303 break;
1304
1305 ++inverse;
1306 }
1307
1308 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1309
1310 /* Recall we're supposed to divide by (gl_level_divisor *
1311 * padded_num_vertices) */
1312
1313 unsigned padded_num_vertices = inverse / orig_divisor;
1314
1315 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1316 #endif
1317 }
1318
1319 static void
1320 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1321 mali_ptr addr, int job_no, char *suffix,
1322 int count, bool varying, enum mali_job_type job_type)
1323 {
1324 char *prefix = varying ? "varying" : "attribute";
1325 assert(addr);
1326
1327 if (!count) {
1328 pandecode_msg("warn: No %s records\n", prefix);
1329 return;
1330 }
1331
1332 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1333
1334 for (int i = 0; i < count; ++i) {
1335 /* First, check for special records */
1336 if (attr[i].elements < MALI_RECORD_SPECIAL) {
1337 if (attr[i].size)
1338 pandecode_msg("XXX: tripped size=%d\n", attr[i].size);
1339
1340 if (attr[i].stride) {
1341 /* gl_InstanceID passes a magic divisor in the
1342 * stride field to divide by the padded vertex
1343 * count. No other records should do so, so
1344 * stride should otherwise be zero. Note that
1345 * stride in the usual attribute sense doesn't
1346 * apply to special records. */
1347
1348 bool has_divisor = attr[i].elements == MALI_ATTR_INSTANCEID;
1349
1350 pandecode_log_cont("/* %smagic divisor = %X */ ",
1351 has_divisor ? "" : "XXX: ", attr[i].stride);
1352 }
1353
1354 if (attr[i].shift || attr[i].extra_flags) {
1355 /* Attributes use these fields for
1356 * instancing/padding/etc type issues, but
1357 * varyings don't */
1358
1359 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1360 varying ? "XXX: " : "",
1361 attr[i].shift, attr[i].extra_flags);
1362 }
1363
1364 /* Print the special record name */
1365 bool attribute = false;
1366 pandecode_log("%s_%d = %s;\n", prefix, i, pandecode_special_record(attr[i].elements, &attribute));
1367
1368 /* Sanity check */
1369 if (attribute == varying)
1370 pandecode_msg("XXX: mismatched special record\n");
1371
1372 continue;
1373 }
1374
1375 enum mali_attr_mode mode = attr[i].elements & 7;
1376
1377 if (mode == MALI_ATTR_UNUSED)
1378 pandecode_msg("XXX: unused attribute record\n");
1379
1380 /* For non-linear records, we need to print the type of record */
1381 if (mode != MALI_ATTR_LINEAR)
1382 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1383
1384 /* Print the name to link with attr_meta */
1385 pandecode_log_cont("%s_%d", prefix, i);
1386
1387 /* Print the stride and size */
1388 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1389
1390 /* TODO: Sanity check the quotient itself. It must be equal to
1391 * (or be greater than, if the driver added padding) the padded
1392 * vertex count. */
1393
1394 /* Finally, print the pointer */
1395 mali_ptr raw_elements = attr[i].elements & ~7;
1396 char *a = pointer_as_memory_reference(raw_elements);
1397 pandecode_log_cont(" = (%s);\n", a);
1398 free(a);
1399
1400 /* Check the pointer */
1401 pandecode_validate_buffer(raw_elements, attr[i].size);
1402
1403 /* shift/extra_flags exist only for instanced */
1404 if (attr[i].shift | attr[i].extra_flags) {
1405 /* These are set to random values by the blob for
1406 * varyings, most likely a symptom of uninitialized
1407 * memory where the hardware masked the bug. As such we
1408 * put this at a warning, not an error. */
1409
1410 if (mode == MALI_ATTR_LINEAR)
1411 pandecode_msg("warn: instancing fields set for linear\n");
1412
1413 pandecode_prop("shift = %d", attr[i].shift);
1414 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1415 }
1416
1417 /* Decode further where possible */
1418
1419 if (mode == MALI_ATTR_MODULO) {
1420 pandecode_padded_vertices(
1421 attr[i].shift,
1422 attr[i].extra_flags);
1423 }
1424
1425 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1426 i++;
1427 pandecode_log("{\n");
1428 pandecode_indent++;
1429 pandecode_prop("unk = 0x%x", attr[i].unk);
1430 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1431 if (attr[i].zero != 0)
1432 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1433 pandecode_prop("divisor = %d", attr[i].divisor);
1434 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1435 pandecode_indent--;
1436 pandecode_log("}, \n");
1437 }
1438
1439 }
1440
1441 pandecode_log("\n");
1442 }
1443
1444 static mali_ptr
1445 pandecode_shader_address(const char *name, mali_ptr ptr)
1446 {
1447 /* TODO: Decode flags */
1448 mali_ptr shader_ptr = ptr & ~15;
1449
1450 char *a = pointer_as_memory_reference(shader_ptr);
1451 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1452 free(a);
1453
1454 return shader_ptr;
1455 }
1456
1457 static void
1458 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1459 {
1460 unsigned any_nonzero =
1461 stencil->ref | stencil->mask | stencil->func |
1462 stencil->sfail | stencil->dpfail | stencil->dppass;
1463
1464 if (any_nonzero == 0)
1465 return;
1466
1467 const char *func = pandecode_func(stencil->func);
1468 const char *sfail = pandecode_stencil_op(stencil->sfail);
1469 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1470 const char *dppass = pandecode_stencil_op(stencil->dppass);
1471
1472 if (stencil->zero)
1473 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1474
1475 pandecode_log(".stencil_%s = {\n", name);
1476 pandecode_indent++;
1477 pandecode_prop("ref = %d", stencil->ref);
1478 pandecode_prop("mask = 0x%02X", stencil->mask);
1479 pandecode_prop("func = %s", func);
1480 pandecode_prop("sfail = %s", sfail);
1481 pandecode_prop("dpfail = %s", dpfail);
1482 pandecode_prop("dppass = %s", dppass);
1483 pandecode_indent--;
1484 pandecode_log("},\n");
1485 }
1486
1487 static void
1488 pandecode_blend_equation(const struct mali_blend_equation *blend)
1489 {
1490 if (blend->zero1)
1491 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1492
1493 pandecode_log(".equation = {\n");
1494 pandecode_indent++;
1495
1496 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1497 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1498
1499 pandecode_log(".color_mask = ");
1500 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1501 pandecode_log_cont(",\n");
1502
1503 pandecode_indent--;
1504 pandecode_log("},\n");
1505 }
1506
1507 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1508
1509 static unsigned
1510 decode_bifrost_constant(u16 constant)
1511 {
1512 float lo = (float) (constant & 0xFF);
1513 float hi = (float) (constant >> 8);
1514
1515 return (hi / 255.0) + (lo / 65535.0);
1516 }
1517
1518 static mali_ptr
1519 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1520 {
1521 struct bifrost_blend_rt *b =
1522 ((struct bifrost_blend_rt *) descs) + rt_no;
1523
1524 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1525 pandecode_indent++;
1526
1527 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1528 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1529 b->constant, decode_bifrost_constant(b->constant));
1530
1531 /* TODO figure out blend shader enable bit */
1532 pandecode_blend_equation(&b->equation);
1533 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1534 pandecode_prop("index = 0x%" PRIx16, b->index);
1535 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1536
1537 pandecode_indent--;
1538 pandecode_log("},\n");
1539
1540 return 0;
1541 }
1542
1543 static mali_ptr
1544 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1545 {
1546 /* constant/equation is in a union */
1547 if (!blend->shader)
1548 return 0;
1549
1550 pandecode_log(".blend = {\n");
1551 pandecode_indent++;
1552
1553 if (is_shader) {
1554 pandecode_shader_address("shader", blend->shader);
1555 } else {
1556 pandecode_blend_equation(&blend->equation);
1557 pandecode_prop("constant = %f", blend->constant);
1558 }
1559
1560 pandecode_indent--;
1561 pandecode_log("},\n");
1562
1563 /* Return blend shader to disassemble if present */
1564 return is_shader ? (blend->shader & ~0xF) : 0;
1565 }
1566
1567 static mali_ptr
1568 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1569 {
1570 struct midgard_blend_rt *b =
1571 ((struct midgard_blend_rt *) descs) + rt_no;
1572
1573 /* Flags determine presence of blend shader */
1574 bool is_shader = (b->flags & 0xF) >= 0x2;
1575
1576 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1577 pandecode_indent++;
1578
1579 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1580
1581 union midgard_blend blend = b->blend;
1582 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1583
1584 pandecode_indent--;
1585 pandecode_log("};\n");
1586
1587 return shader;
1588 }
1589
1590 /* Attributes and varyings have descriptor records, which contain information
1591 * about their format and ordering with the attribute/varying buffers. We'll
1592 * want to validate that the combinations specified are self-consistent.
1593 */
1594
1595 static int
1596 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1597 {
1598 char base[128];
1599 char *prefix = varying ? "varying" : "attribute";
1600 unsigned max_index = 0;
1601 snprintf(base, sizeof(base), "%s_meta", prefix);
1602
1603 struct mali_attr_meta *attr_meta;
1604 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1605
1606 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1607
1608 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1609 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1610 sizeof(*attr_mem));
1611
1612 /* If the record is discard, it should be zero for everything else */
1613
1614 if (attr_meta->format == MALI_VARYING_DISCARD) {
1615 uint64_t zero =
1616 attr_meta->index |
1617 attr_meta->unknown1 |
1618 attr_meta->unknown3 |
1619 attr_meta->src_offset;
1620
1621 if (zero)
1622 pandecode_msg("XXX: expected empty record for varying discard\n");
1623
1624 /* We want to look for a literal 0000 swizzle -- this
1625 * is not encoded with all zeroes, however */
1626
1627 enum mali_channel z = MALI_CHANNEL_ZERO;
1628 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1629 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1630
1631 if (!good_swizzle)
1632 pandecode_msg("XXX: expected zero swizzle for discard\n");
1633
1634 if (!varying)
1635 pandecode_msg("XXX: cannot discard attribute\n");
1636
1637 /* If we're all good, omit the record */
1638 if (!zero && varying && good_swizzle) {
1639 pandecode_log("/* discarded varying */\n");
1640 continue;
1641 }
1642 }
1643
1644 if (attr_meta->index > max_index)
1645 max_index = attr_meta->index;
1646
1647 if (attr_meta->unknown1 != 0x2) {
1648 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1649 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1650 }
1651
1652 if (attr_meta->unknown3) {
1653 pandecode_msg("XXX: unexpected unknown3 set\n");
1654 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1655 }
1656
1657 pandecode_format_short(attr_meta->format, false);
1658 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1659
1660 if (attr_meta->src_offset)
1661 pandecode_log_cont("[%u]", attr_meta->src_offset);
1662
1663 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1664
1665 pandecode_log_cont(";\n");
1666 }
1667
1668 pandecode_log("\n");
1669
1670 return count ? (max_index + 1) : 0;
1671 }
1672
1673 /* return bits [lo, hi) of word */
1674 static u32
1675 bits(u32 word, u32 lo, u32 hi)
1676 {
1677 if (hi - lo >= 32)
1678 return word; // avoid undefined behavior with the shift
1679
1680 return (word >> lo) & ((1 << (hi - lo)) - 1);
1681 }
1682
1683 static void
1684 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1685 {
1686 pandecode_log_cont("{\n");
1687 pandecode_indent++;
1688
1689 /* Decode invocation_count. See the comment before the definition of
1690 * invocation_count for an explanation.
1691 */
1692
1693 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1694 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1695 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1696 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1697 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1698 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1699
1700 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1701 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1702 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1703
1704 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1705 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1706 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1707
1708 /* Even though we have this decoded, we want to ensure that the
1709 * representation is "unique" so we don't lose anything by printing only
1710 * the final result. More specifically, we need to check that we were
1711 * passed something in canonical form, since the definition per the
1712 * hardware is inherently not unique. How? Well, take the resulting
1713 * decode and pack it ourselves! If it is bit exact with what we
1714 * decoded, we're good to go. */
1715
1716 struct mali_vertex_tiler_prefix ref;
1717 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1718
1719 bool canonical =
1720 (p->invocation_count == ref.invocation_count) &&
1721 (p->invocation_shifts == ref.invocation_shifts);
1722
1723 if (!canonical) {
1724 pandecode_msg("XXX: non-canonical workgroups packing\n");
1725 pandecode_msg("expected: %X, %X",
1726 ref.invocation_count,
1727 ref.invocation_shifts);
1728
1729 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1730 pandecode_prop("size_y_shift = %d", size_y_shift);
1731 pandecode_prop("size_z_shift = %d", size_z_shift);
1732 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
1733 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
1734 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
1735 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
1736 }
1737
1738 /* Regardless, print the decode */
1739 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1740 size_x, size_y, size_z,
1741 groups_x, groups_y, groups_z);
1742
1743 /* TODO: Decode */
1744 if (p->unknown_draw)
1745 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1746
1747 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1748
1749 if (p->draw_mode != MALI_DRAW_NONE)
1750 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1751
1752 /* Index count only exists for tiler jobs anyway */
1753
1754 if (p->index_count)
1755 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1756
1757
1758 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1759 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1760
1761 /* Validate an index buffer is present if we need one. TODO: verify
1762 * relationship between invocation_count and index_count */
1763
1764 if (p->indices) {
1765 unsigned count = p->index_count;
1766
1767 /* Grab the size */
1768 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1769
1770 /* Ensure we got a size, and if so, validate the index buffer
1771 * is large enough to hold a full set of indices of the given
1772 * size */
1773
1774 if (!index_raw_size)
1775 pandecode_msg("XXX: index size missing\n");
1776 else
1777 pandecode_validate_buffer(p->indices, count * size);
1778 } else if (index_raw_size)
1779 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1780
1781 if (p->offset_bias_correction)
1782 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1783
1784 /* TODO: Figure out what this is. It's not zero */
1785 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1786
1787 pandecode_indent--;
1788 pandecode_log("},\n");
1789 }
1790
1791 static void
1792 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1793 {
1794 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1795 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1796
1797 for (int i = 0; i < ubufs_count; i++) {
1798 unsigned size = (ubufs[i].size + 1) * 16;
1799 mali_ptr addr = ubufs[i].ptr << 2;
1800
1801 pandecode_validate_buffer(addr, size);
1802
1803 char *ptr = pointer_as_memory_reference(ubufs[i].ptr << 2);
1804 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1805 free(ptr);
1806 }
1807
1808 pandecode_log("\n");
1809 }
1810
1811 static void
1812 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1813 {
1814 pandecode_validate_buffer(uniforms, uniform_count * 16);
1815
1816 char *ptr = pointer_as_memory_reference(uniforms);
1817 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1818 free(ptr);
1819 }
1820
1821 static void
1822 pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1823 {
1824
1825 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1826
1827 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1828
1829 if (scratchpad->zero) {
1830 pandecode_msg("XXX: scratchpad zero tripped");
1831 pandecode_prop("zero = 0x%x\n", scratchpad->zero);
1832 }
1833
1834 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1835 pandecode_indent++;
1836
1837 pandecode_prop("flags = 0x%x", scratchpad->flags);
1838 MEMORY_PROP(scratchpad, gpu_scratchpad);
1839
1840 pandecode_indent--;
1841 pandecode_log("};\n");
1842 }
1843
1844 static const char *
1845 shader_type_for_job(unsigned type)
1846 {
1847 switch (type) {
1848 case JOB_TYPE_VERTEX: return "VERTEX";
1849 case JOB_TYPE_TILER: return "FRAGMENT";
1850 case JOB_TYPE_COMPUTE: return "COMPUTE";
1851 default:
1852 return "UNKNOWN";
1853 }
1854 }
1855
1856 static unsigned shader_id = 0;
1857
1858 static struct midgard_disasm_stats
1859 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1860 bool is_bifrost, unsigned gpu_id)
1861 {
1862 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1863 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1864
1865 /* Compute maximum possible size */
1866 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1867
1868 /* Print some boilerplate to clearly denote the assembly (which doesn't
1869 * obey indentation rules), and actually do the disassembly! */
1870
1871 pandecode_log_cont("\n\n");
1872
1873 struct midgard_disasm_stats stats;
1874
1875 if (is_bifrost) {
1876 disassemble_bifrost(pandecode_dump_stream, code, sz, false);
1877
1878 /* TODO: Extend stats to Bifrost */
1879 stats.texture_count = -128;
1880 stats.sampler_count = -128;
1881 stats.attribute_count = -128;
1882 stats.varying_count = -128;
1883 stats.uniform_count = -128;
1884 stats.uniform_buffer_count = -128;
1885 stats.work_count = -128;
1886
1887 stats.instruction_count = 0;
1888 stats.bundle_count = 0;
1889 stats.quadword_count = 0;
1890 stats.helper_invocations = false;
1891 } else {
1892 stats = disassemble_midgard(pandecode_dump_stream,
1893 code, sz, gpu_id,
1894 type == JOB_TYPE_TILER ?
1895 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
1896 }
1897
1898 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1899 * driver-internal purposes with the blob and interfere */
1900
1901 bool should_shaderdb = type != JOB_TYPE_COMPUTE;
1902
1903 if (should_shaderdb) {
1904 unsigned nr_threads =
1905 (stats.work_count <= 4) ? 4 :
1906 (stats.work_count <= 8) ? 2 :
1907 1;
1908
1909 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
1910 "%u inst, %u bundles, %u quadwords, "
1911 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1912 shader_id++,
1913 shader_type_for_job(type),
1914 stats.instruction_count, stats.bundle_count, stats.quadword_count,
1915 stats.work_count, nr_threads);
1916 }
1917
1918
1919 return stats;
1920 }
1921
1922 static void
1923 pandecode_texture(mali_ptr u,
1924 struct pandecode_mapped_memory *tmem,
1925 unsigned job_no, unsigned tex)
1926 {
1927 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
1928
1929 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
1930 pandecode_indent++;
1931
1932 struct mali_texture_format f = t->format;
1933
1934 /* See the definiton of enum mali_texture_type */
1935
1936 bool is_cube = f.type == MALI_TEX_CUBE;
1937 unsigned dimension = is_cube ? 2 : f.type;
1938
1939 pandecode_make_indent();
1940
1941 /* TODO: Are there others? */
1942 bool is_zs = f.format == MALI_Z32_UNORM;
1943
1944 /* Recall Z/S switched the meaning of linear/tiled .. */
1945 if (is_zs && f.layout == MALI_TEXTURE_LINEAR)
1946 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1947
1948 /* Print the layout. Default is linear; a modifier can denote AFBC or
1949 * u-interleaved/tiled modes */
1950
1951 if (f.layout == MALI_TEXTURE_AFBC)
1952 pandecode_log_cont("afbc");
1953 else if (f.layout == MALI_TEXTURE_TILED)
1954 pandecode_log_cont("tiled");
1955 else if (f.layout == MALI_TEXTURE_LINEAR)
1956 pandecode_log_cont("linear");
1957 else
1958 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
1959
1960 pandecode_swizzle(t->swizzle, f.format);
1961 pandecode_log_cont(" ");
1962
1963 /* Distinguish cube/2D with modifier */
1964
1965 if (is_cube)
1966 pandecode_log_cont("cube ");
1967
1968 pandecode_format_short(f.format, f.srgb);
1969 pandecode_swizzle(f.swizzle, f.format);
1970
1971 /* All four width/height/depth/array_size dimensions are present
1972 * regardless of the type of texture, but it is an error to have
1973 * non-zero dimensions for unused dimensions. Verify this. array_size
1974 * can always be set, as can width. */
1975
1976 if (t->height && dimension < 2)
1977 pandecode_msg("XXX: nonzero height for <2D texture\n");
1978
1979 if (t->depth && dimension < 3)
1980 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1981
1982 /* Print only the dimensions that are actually there */
1983
1984 pandecode_log_cont(": %d", t->width + 1);
1985
1986 if (dimension >= 2)
1987 pandecode_log_cont("x%u", t->height + 1);
1988
1989 if (dimension >= 3)
1990 pandecode_log_cont("x%u", t->depth + 1);
1991
1992 if (t->array_size)
1993 pandecode_log_cont("[%u]", t->array_size + 1);
1994
1995 if (t->levels)
1996 pandecode_log_cont(" mip %u", t->levels);
1997
1998 pandecode_log_cont("\n");
1999
2000 if (f.unknown1 | f.zero) {
2001 pandecode_msg("XXX: texture format zero tripped\n");
2002 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
2003 pandecode_prop("zero = %" PRId32, f.zero);
2004 }
2005
2006 if (!f.unknown2) {
2007 pandecode_msg("XXX: expected unknown texture bit set\n");
2008 pandecode_prop("unknown2 = %" PRId32, f.unknown1);
2009 }
2010
2011 if (t->swizzle_zero) {
2012 pandecode_msg("XXX: swizzle zero tripped\n");
2013 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
2014 }
2015
2016 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
2017 pandecode_msg("XXX: texture zero tripped\n");
2018 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
2019 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
2020 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2021 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2022 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2023 }
2024
2025 pandecode_log(".payload = {\n");
2026 pandecode_indent++;
2027
2028 /* A bunch of bitmap pointers follow.
2029 * We work out the correct number,
2030 * based on the mipmap/cubemap
2031 * properties, but dump extra
2032 * possibilities to futureproof */
2033
2034 int bitmap_count = t->levels + 1;
2035
2036 /* Miptree for each face */
2037 if (f.type == MALI_TEX_CUBE)
2038 bitmap_count *= 6;
2039 else if (f.type == MALI_TEX_3D)
2040 bitmap_count *= t->depth;
2041
2042 /* Array of textures */
2043 bitmap_count *= (t->array_size + 1);
2044
2045 /* Stride for each element */
2046 if (f.manual_stride)
2047 bitmap_count *= 2;
2048
2049 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
2050 u + sizeof(*t), sizeof(mali_ptr) * bitmap_count);
2051 for (int i = 0; i < bitmap_count; ++i) {
2052 /* How we dump depends if this is a stride or a pointer */
2053
2054 if (f.manual_stride && (i & 1)) {
2055 /* signed 32-bit snuck in as a 64-bit pointer */
2056 uint64_t stride_set = pointers_and_strides[i];
2057 uint32_t clamped_stride = stride_set;
2058 int32_t stride = clamped_stride;
2059 assert(stride_set == clamped_stride);
2060 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2061 } else {
2062 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
2063 pandecode_log("%s, \n", a);
2064 free(a);
2065 }
2066 }
2067
2068 pandecode_indent--;
2069 pandecode_log("},\n");
2070
2071 pandecode_indent--;
2072 pandecode_log("};\n");
2073 }
2074
2075 /* 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 */
2076
2077 static void
2078 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2079 {
2080 /* Nothing to do */
2081 if (claim == truth)
2082 return;
2083
2084 if (fuzzy)
2085 assert(truth >= 0);
2086
2087 if ((truth >= 0) && !fuzzy) {
2088 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2089 (truth < claim) ? "warn" : "XXX",
2090 name, truth, claim);
2091 } else if ((claim > -truth) && !fuzzy) {
2092 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2093 name, -truth, claim);
2094 } else if (fuzzy && (claim < truth))
2095 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2096 name, truth, claim);
2097
2098 pandecode_log(".%s = %" PRId16, name, claim);
2099
2100 if (fuzzy)
2101 pandecode_log_cont(" /* %u used */", truth);
2102
2103 pandecode_log_cont(",\n");
2104 }
2105
2106 static void
2107 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2108 bool is_bifrost, unsigned gpu_id)
2109 {
2110 struct midgard_disasm_stats stats =
2111 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
2112
2113 bool has_texture = (stats.texture_count > 0);
2114 bool has_sampler = (stats.sampler_count > 0);
2115 bool has_attribute = (stats.attribute_count > 0);
2116 bool has_varying = (stats.varying_count > 0);
2117 bool has_uniform = (stats.uniform_count > 0);
2118 bool has_ubo = (stats.uniform_buffer_count > 0);
2119
2120 if (has_texture || has_sampler)
2121 pandecode_msg("XXX: blend shader accessing textures\n");
2122
2123 if (has_attribute || has_varying)
2124 pandecode_msg("XXX: blend shader accessing interstage\n");
2125
2126 if (has_uniform || has_ubo)
2127 pandecode_msg("XXX: blend shader accessing uniforms\n");
2128 }
2129
2130 static void
2131 pandecode_vertex_tiler_postfix_pre(
2132 const struct mali_vertex_tiler_postfix *p,
2133 int job_no, enum mali_job_type job_type,
2134 char *suffix, bool is_bifrost, unsigned gpu_id)
2135 {
2136 struct pandecode_mapped_memory *attr_mem;
2137
2138 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2139 * are the only things actually needed from the FBD, vertex/tiler jobs
2140 * no longer reference the FBD -- instead, this field points to some
2141 * info about the scratchpad.
2142 */
2143
2144 struct pandecode_fbd fbd_info = {
2145 /* Default for Bifrost */
2146 .rt_count = 1
2147 };
2148
2149 if (is_bifrost)
2150 pandecode_scratchpad(p->framebuffer & ~1, job_no, suffix);
2151 else if (p->framebuffer & MALI_MFBD)
2152 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
2153 else if (job_type == JOB_TYPE_COMPUTE)
2154 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
2155 else
2156 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no, false, gpu_id);
2157
2158 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2159 int texture_count = 0, sampler_count = 0;
2160
2161 if (p->shader) {
2162 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2163 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2164
2165 /* Disassemble ahead-of-time to get stats. Initialize with
2166 * stats for the missing-shader case so we get validation
2167 * there, too */
2168
2169 struct midgard_disasm_stats info = {
2170 .texture_count = 0,
2171 .sampler_count = 0,
2172 .attribute_count = 0,
2173 .varying_count = 0,
2174 .work_count = 1,
2175
2176 .uniform_count = -128,
2177 .uniform_buffer_count = 0
2178 };
2179
2180 if (s->shader & ~0xF)
2181 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
2182
2183 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2184 pandecode_indent++;
2185
2186 /* Save for dumps */
2187 attribute_count = s->attribute_count;
2188 varying_count = s->varying_count;
2189 texture_count = s->texture_count;
2190 sampler_count = s->sampler_count;
2191
2192 if (is_bifrost) {
2193 uniform_count = s->bifrost2.uniform_count;
2194 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2195 } else {
2196 uniform_count = s->midgard1.uniform_count;
2197 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2198 }
2199
2200 pandecode_shader_address("shader", s->shader);
2201
2202 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2203 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2204 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2205 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2206 pandecode_shader_prop("uniform_buffer_count",
2207 uniform_buffer_count,
2208 info.uniform_buffer_count, true);
2209
2210 if (!is_bifrost) {
2211 pandecode_shader_prop("uniform_count",
2212 uniform_count,
2213 info.uniform_count, false);
2214
2215 pandecode_shader_prop("work_count",
2216 s->midgard1.work_count, info.work_count, false);
2217 }
2218
2219 if (is_bifrost) {
2220 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32, s->bifrost1.unk1);
2221 } else {
2222 bool helpers = s->midgard1.flags & MALI_HELPER_INVOCATIONS;
2223 s->midgard1.flags &= ~MALI_HELPER_INVOCATIONS;
2224
2225 if (helpers != info.helper_invocations) {
2226 pandecode_msg("XXX: expected helpers %u but got %u\n",
2227 info.helper_invocations, helpers);
2228 }
2229
2230 pandecode_log(".midgard1.flags = ");
2231 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
2232 pandecode_log_cont(",\n");
2233
2234 pandecode_prop("midgard1.unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
2235 }
2236
2237 if (s->depth_units || s->depth_factor) {
2238 pandecode_prop("depth_factor = %f", s->depth_factor);
2239 pandecode_prop("depth_units = %f", s->depth_units);
2240 }
2241
2242 if (s->alpha_coverage) {
2243 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
2244 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
2245
2246 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2247 invert_alpha_coverage ? "~" : "",
2248 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
2249 }
2250
2251 if (s->unknown2_3 || s->unknown2_4) {
2252 pandecode_log(".unknown2_3 = ");
2253
2254 int unknown2_3 = s->unknown2_3;
2255 int unknown2_4 = s->unknown2_4;
2256
2257 /* We're not quite sure what these flags mean without the depth test, if anything */
2258
2259 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2260 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2261 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2262
2263 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2264 }
2265
2266 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2267 pandecode_log_cont(",\n");
2268
2269 pandecode_log(".unknown2_4 = ");
2270 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2271 pandecode_log_cont(",\n");
2272 }
2273
2274 if (s->stencil_mask_front || s->stencil_mask_back) {
2275 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2276 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2277 }
2278
2279 pandecode_stencil("front", &s->stencil_front);
2280 pandecode_stencil("back", &s->stencil_back);
2281
2282 if (is_bifrost) {
2283 pandecode_log(".bifrost2 = {\n");
2284 pandecode_indent++;
2285
2286 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2287 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2288 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2289 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2290
2291 pandecode_indent--;
2292 pandecode_log("},\n");
2293 } else if (s->midgard2.unknown2_7) {
2294 pandecode_log(".midgard2 = {\n");
2295 pandecode_indent++;
2296
2297 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2298 pandecode_indent--;
2299 pandecode_log("},\n");
2300 }
2301
2302 if (s->unknown2_8)
2303 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
2304
2305 if (!is_bifrost) {
2306 /* TODO: Blend shaders routing/disasm */
2307 union midgard_blend blend = s->blend;
2308 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2309 if (shader & ~0xF)
2310 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2311 }
2312
2313 pandecode_indent--;
2314 pandecode_log("};\n");
2315
2316 /* MRT blend fields are used whenever MFBD is used, with
2317 * per-RT descriptors */
2318
2319 if (job_type == JOB_TYPE_TILER && p->framebuffer & MALI_MFBD) {
2320 void* blend_base = (void *) (s + 1);
2321
2322 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2323 mali_ptr shader = 0;
2324
2325 if (is_bifrost)
2326 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2327 else
2328 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2329
2330 if (shader & ~0xF)
2331 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2332
2333 }
2334 }
2335 } else
2336 pandecode_msg("XXX: missing shader descriptor\n");
2337
2338 if (p->viewport) {
2339 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2340 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2341
2342 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2343 pandecode_indent++;
2344
2345 pandecode_prop("clip_minx = %f", f->clip_minx);
2346 pandecode_prop("clip_miny = %f", f->clip_miny);
2347 pandecode_prop("clip_minz = %f", f->clip_minz);
2348 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2349 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2350 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2351
2352 /* Only the higher coordinates are MALI_POSITIVE scaled */
2353
2354 pandecode_prop("viewport0 = { %d, %d }",
2355 f->viewport0[0], f->viewport0[1]);
2356
2357 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2358 f->viewport1[0] + 1, f->viewport1[1] + 1);
2359
2360 pandecode_indent--;
2361 pandecode_log("};\n");
2362 }
2363
2364 unsigned max_attr_index = 0;
2365
2366 if (p->attribute_meta)
2367 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2368
2369 if (p->attributes) {
2370 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2371 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2372 }
2373
2374 /* Varyings are encoded like attributes but not actually sent; we just
2375 * pass a zero buffer with the right stride/size set, (or whatever)
2376 * since the GPU will write to it itself */
2377
2378 if (p->varying_meta) {
2379 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2380 }
2381
2382 if (p->varyings) {
2383 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2384
2385 /* Number of descriptors depends on whether there are
2386 * non-internal varyings */
2387
2388 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2389 }
2390
2391 if (p->uniform_buffers) {
2392 if (uniform_buffer_count)
2393 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2394 else
2395 pandecode_msg("warn: UBOs specified but not referenced\n");
2396 } else if (uniform_buffer_count)
2397 pandecode_msg("XXX: UBOs referenced but not specified\n");
2398
2399 /* We don't want to actually dump uniforms, but we do need to validate
2400 * that the counts we were given are sane */
2401
2402 if (p->uniforms) {
2403 if (uniform_count)
2404 pandecode_uniforms(p->uniforms, uniform_count);
2405 else
2406 pandecode_msg("warn: Uniforms specified but not referenced\n");
2407 } else if (uniform_count)
2408 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2409
2410 if (p->texture_trampoline) {
2411 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
2412
2413 if (mmem) {
2414 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
2415
2416 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
2417 pandecode_indent++;
2418
2419 for (int tex = 0; tex < texture_count; ++tex) {
2420 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2421 char *a = pointer_as_memory_reference(*u);
2422 pandecode_log("%s,\n", a);
2423 free(a);
2424 }
2425
2426 pandecode_indent--;
2427 pandecode_log("};\n");
2428
2429 /* Now, finally, descend down into the texture descriptor */
2430 for (unsigned tex = 0; tex < texture_count; ++tex) {
2431 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2432 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2433 if (tmem)
2434 pandecode_texture(*u, tmem, job_no, tex);
2435 }
2436 }
2437 }
2438
2439 if (p->sampler_descriptor) {
2440 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2441
2442 if (smem) {
2443 struct mali_sampler_descriptor *s;
2444
2445 mali_ptr d = p->sampler_descriptor;
2446
2447 for (int i = 0; i < sampler_count; ++i) {
2448 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2449
2450 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2451 pandecode_indent++;
2452
2453 pandecode_log(".filter_mode = ");
2454 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2455 pandecode_log_cont(",\n");
2456
2457 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2458 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2459
2460 if (s->lod_bias)
2461 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s->lod_bias));
2462
2463 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2464 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2465 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2466
2467 pandecode_prop("compare_func = %s", pandecode_func(s->compare_func));
2468
2469 if (s->zero || s->zero2) {
2470 pandecode_msg("XXX: sampler zero tripped\n");
2471 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2472 }
2473
2474 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2475
2476 pandecode_prop("border_color = { %f, %f, %f, %f }",
2477 s->border_color[0],
2478 s->border_color[1],
2479 s->border_color[2],
2480 s->border_color[3]);
2481
2482 pandecode_indent--;
2483 pandecode_log("};\n");
2484 }
2485 }
2486 }
2487 }
2488
2489 static void
2490 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2491 {
2492 if (p->shader & 0xF)
2493 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2494
2495 if (!(p->position_varying || p->occlusion_counter))
2496 return;
2497
2498 pandecode_log(".postfix = {\n");
2499 pandecode_indent++;
2500
2501 MEMORY_PROP(p, position_varying);
2502 MEMORY_PROP(p, occlusion_counter);
2503
2504 pandecode_indent--;
2505 pandecode_log("},\n");
2506 }
2507
2508 static void
2509 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2510 {
2511 pandecode_log_cont("{\n");
2512 pandecode_indent++;
2513
2514 pandecode_prop("unk2 = 0x%x", v->unk2);
2515
2516 if (v->zero0 || v->zero1) {
2517 pandecode_msg("XXX: vertex only zero tripped");
2518 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2519 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2520 }
2521
2522 pandecode_indent--;
2523 pandecode_log("}\n");
2524 }
2525
2526 static void
2527 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2528 {
2529
2530 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2531 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2532
2533 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2534 pandecode_indent++;
2535
2536 if (h->zero) {
2537 pandecode_msg("XXX: tiler heap zero tripped\n");
2538 pandecode_prop("zero = 0x%x", h->zero);
2539 }
2540
2541 for (int i = 0; i < 12; i++) {
2542 if (h->zeros[i] != 0) {
2543 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2544 i, h->zeros[i]);
2545 }
2546 }
2547
2548 pandecode_prop("heap_size = 0x%x", h->heap_size);
2549 MEMORY_PROP(h, tiler_heap_start);
2550 MEMORY_PROP(h, tiler_heap_free);
2551
2552 /* this might point to the beginning of another buffer, when it's
2553 * really the end of the tiler heap buffer, so we have to be careful
2554 * here. but for zero length, we need the same pointer.
2555 */
2556
2557 if (h->tiler_heap_end == h->tiler_heap_start) {
2558 MEMORY_PROP(h, tiler_heap_start);
2559 } else {
2560 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2561 pandecode_prop("tiler_heap_end = %s + 1", a);
2562 free(a);
2563 }
2564
2565 pandecode_indent--;
2566 pandecode_log("};\n");
2567 }
2568
2569 static void
2570 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2571 {
2572 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2573 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2574
2575 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2576
2577 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2578 pandecode_indent++;
2579
2580 if (t->zero0 || t->zero1) {
2581 pandecode_msg("XXX: tiler meta zero tripped\n");
2582 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2583 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2584 }
2585
2586 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2587 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2588
2589 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2590 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2591
2592 for (int i = 0; i < 12; i++) {
2593 if (t->zeros[i] != 0) {
2594 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2595 i, t->zeros[i]);
2596 }
2597 }
2598
2599 pandecode_indent--;
2600 pandecode_log("};\n");
2601 }
2602
2603 static void
2604 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2605 {
2606 pandecode_log(".gl_enables = ");
2607
2608 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2609
2610 pandecode_log_cont(",\n");
2611 }
2612
2613 static void
2614 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2615 {
2616 if (u.pointer == 0x0)
2617 return;
2618
2619 pandecode_log(".primitive_size = {\n");
2620 pandecode_indent++;
2621
2622 if (constant) {
2623 pandecode_prop("constant = %f", u.constant);
2624 } else {
2625 MEMORY_PROP((&u), pointer);
2626 }
2627
2628 pandecode_indent--;
2629 pandecode_log("},\n");
2630 }
2631
2632 static void
2633 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2634 {
2635 pandecode_log_cont("{\n");
2636 pandecode_indent++;
2637
2638 /* TODO: gl_PointSize on Bifrost */
2639 pandecode_primitive_size(t->primitive_size, true);
2640
2641 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2642
2643 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2644 || t->zero6 || t->zero7 || t->zero8) {
2645 pandecode_msg("XXX: tiler only zero tripped\n");
2646 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2647 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2648 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2649 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2650 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2651 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2652 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2653 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2654 }
2655
2656 pandecode_indent--;
2657 pandecode_log("},\n");
2658 }
2659
2660 static int
2661 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2662 const struct pandecode_mapped_memory *mem,
2663 mali_ptr payload, int job_no, unsigned gpu_id)
2664 {
2665 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2666
2667 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
2668
2669 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2670 pandecode_indent++;
2671
2672 pandecode_log(".prefix = ");
2673 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2674
2675 pandecode_log(".vertex = ");
2676 pandecode_vertex_only_bfr(&v->vertex);
2677
2678 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2679
2680 pandecode_indent--;
2681 pandecode_log("};\n");
2682
2683 return sizeof(*v);
2684 }
2685
2686 static int
2687 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2688 const struct pandecode_mapped_memory *mem,
2689 mali_ptr payload, int job_no, unsigned gpu_id)
2690 {
2691 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2692
2693 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
2694 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2695
2696 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2697 pandecode_indent++;
2698
2699 pandecode_log(".prefix = ");
2700 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2701
2702 pandecode_log(".tiler = ");
2703 pandecode_tiler_only_bfr(&t->tiler, job_no);
2704
2705 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2706
2707 pandecode_indent--;
2708 pandecode_log("};\n");
2709
2710 return sizeof(*t);
2711 }
2712
2713 static int
2714 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2715 const struct pandecode_mapped_memory *mem,
2716 mali_ptr payload, int job_no, unsigned gpu_id)
2717 {
2718 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2719
2720 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
2721
2722 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2723 pandecode_indent++;
2724
2725 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2726 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2727
2728 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2729
2730 pandecode_log(".prefix = ");
2731 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
2732
2733 pandecode_gl_enables(v->gl_enables, h->job_type);
2734
2735 if (v->instance_shift || v->instance_odd) {
2736 pandecode_prop("instance_shift = 0x%d /* %d */",
2737 v->instance_shift, 1 << v->instance_shift);
2738 pandecode_prop("instance_odd = 0x%X /* %d */",
2739 v->instance_odd, (2 * v->instance_odd) + 1);
2740
2741 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2742 }
2743
2744 if (v->offset_start)
2745 pandecode_prop("offset_start = %d", v->offset_start);
2746
2747 if (v->zero5) {
2748 pandecode_msg("XXX: midgard payload zero tripped\n");
2749 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2750 }
2751
2752 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2753
2754 pandecode_indent--;
2755 pandecode_log("};\n");
2756
2757 return sizeof(*v);
2758 }
2759
2760 static int
2761 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2762 mali_ptr payload, int job_no,
2763 bool is_bifrost, unsigned gpu_id)
2764 {
2765 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2766
2767 bool is_mfbd = s->framebuffer & MALI_MFBD;
2768
2769 /* Bifrost theoretically may retain support for SFBD on compute jobs,
2770 * but for graphics workloads with a FRAGMENT payload, use MFBD */
2771
2772 if (!is_mfbd && is_bifrost)
2773 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2774
2775 struct pandecode_fbd info;
2776
2777 if (is_mfbd)
2778 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2779 else
2780 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
2781
2782 /* Compute the tag for the tagged pointer. This contains the type of
2783 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2784 * additional structures follow the MFBD header (an extra payload or
2785 * not, as well as a count of render targets) */
2786
2787 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
2788
2789 if (is_mfbd) {
2790 if (info.has_extra)
2791 expected_tag |= MALI_MFBD_TAG_EXTRA;
2792
2793 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
2794 }
2795
2796 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
2797 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2798 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
2799 pandecode_prop("max_tile_coord = 0x%X\n", s->min_tile_coord);
2800 }
2801
2802 /* Extract tile coordinates */
2803
2804 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
2805 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
2806
2807 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2808 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2809
2810 /* For the max, we also want the floored (rather than ceiled) version for checking */
2811
2812 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
2813 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
2814
2815 /* Validate the coordinates are well-ordered */
2816
2817 if (min_x == max_x)
2818 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2819 else if (min_x > max_x)
2820 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2821
2822 if (min_y == max_y)
2823 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2824 else if (min_y > max_y)
2825 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2826
2827 /* Validate the coordinates fit inside the framebuffer. We use floor,
2828 * rather than ceil, for the max coordinates, since the tile
2829 * coordinates for something like an 800x600 framebuffer will actually
2830 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2831
2832 if ((min_x > info.width) || (max_x_f > info.width))
2833 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2834
2835 if ((min_y > info.height) || (max_y_f > info.height))
2836 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2837
2838 /* After validation, we print */
2839
2840 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
2841
2842 /* The FBD is a tagged pointer */
2843
2844 unsigned tag = (s->framebuffer & ~FBD_MASK);
2845
2846 if (tag != expected_tag)
2847 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
2848
2849 return sizeof(*s);
2850 }
2851
2852 static int job_descriptor_number = 0;
2853
2854 int
2855 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id)
2856 {
2857 struct mali_job_descriptor_header *h;
2858
2859 int start_number = 0;
2860
2861 bool first = true;
2862 bool last_size;
2863
2864 do {
2865 struct pandecode_mapped_memory *mem =
2866 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2867
2868 void *payload;
2869
2870 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2871
2872 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2873 * high 32-bits of the 64-bit pointer are reused to store
2874 * something else.
2875 */
2876 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2877 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2878 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2879
2880 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
2881
2882 int job_no = job_descriptor_number++;
2883
2884 if (first)
2885 start_number = job_no;
2886
2887 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2888 pandecode_indent++;
2889
2890 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2891
2892 /* Save for next job fixing */
2893 last_size = h->job_descriptor_size;
2894
2895 if (h->job_descriptor_size)
2896 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2897
2898 if (h->exception_status && h->exception_status != 0x1)
2899 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2900 h->exception_status,
2901 (h->exception_status >> 16) & 0xFFFF,
2902 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2903 h->exception_status & 0xFF);
2904
2905 if (h->first_incomplete_task)
2906 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2907
2908 if (h->fault_pointer)
2909 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2910
2911 if (h->job_barrier)
2912 pandecode_prop("job_barrier = %d", h->job_barrier);
2913
2914 pandecode_prop("job_index = %d", h->job_index);
2915
2916 if (h->unknown_flags)
2917 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2918
2919 if (h->job_dependency_index_1)
2920 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2921
2922 if (h->job_dependency_index_2)
2923 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2924
2925 pandecode_indent--;
2926 pandecode_log("};\n");
2927
2928 /* Do not touch the field yet -- decode the payload first, and
2929 * don't touch that either. This is essential for the uploads
2930 * to occur in sequence and therefore be dynamically allocated
2931 * correctly. Do note the size, however, for that related
2932 * reason. */
2933
2934 switch (h->job_type) {
2935 case JOB_TYPE_WRITE_VALUE: {
2936 struct mali_payload_write_value *s = payload;
2937 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2938 pandecode_indent++;
2939 MEMORY_PROP(s, address);
2940
2941 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
2942 pandecode_msg("XXX: unknown value descriptor\n");
2943 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
2944 }
2945
2946 if (s->reserved) {
2947 pandecode_msg("XXX: set value tripped\n");
2948 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
2949 }
2950
2951 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
2952 pandecode_indent--;
2953 pandecode_log("};\n");
2954
2955 break;
2956 }
2957
2958 case JOB_TYPE_TILER:
2959 case JOB_TYPE_VERTEX:
2960 case JOB_TYPE_COMPUTE:
2961 if (bifrost) {
2962 if (h->job_type == JOB_TYPE_TILER)
2963 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2964 else
2965 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2966 } else
2967 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
2968
2969 break;
2970
2971 case JOB_TYPE_FRAGMENT:
2972 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
2973 break;
2974
2975 default:
2976 break;
2977 }
2978
2979 /* Handle linkage */
2980
2981 if (!first) {
2982 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2983 pandecode_log_cont("next_job = job_%d_p;\n\n", job_no);
2984 }
2985
2986 first = false;
2987
2988 } while ((jc_gpu_va = h->next_job));
2989
2990 return start_number;
2991 }