projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
panfrost: Remove MALI_ATTR_INTERNAL
[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 /* Semantic logging type.
62 *
63 * Raw: for raw messages to be printed as is.
64 * Message: for helpful information to be commented out in replays.
65 * Property: for properties of a struct
66 *
67 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
68 */
69
70 enum pandecode_log_type {
71 PANDECODE_RAW,
72 PANDECODE_MESSAGE,
73 PANDECODE_PROPERTY
74 };
75
76 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
77 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
78 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
79
80 unsigned pandecode_indent = 0;
81
82 static void
83 pandecode_make_indent(void)
84 {
85 for (unsigned i = 0; i < pandecode_indent; ++i)
86 printf(" ");
87 }
88
89 static void
90 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
91 {
92 va_list ap;
93
94 pandecode_make_indent();
95
96 if (type == PANDECODE_MESSAGE)
97 printf("// ");
98 else if (type == PANDECODE_PROPERTY)
99 printf(".");
100
101 va_start(ap, format);
102 vprintf(format, ap);
103 va_end(ap);
104
105 if (type == PANDECODE_PROPERTY)
106 printf(",\n");
107 }
108
109 static void
110 pandecode_log_cont(const char *format, ...)
111 {
112 va_list ap;
113
114 va_start(ap, format);
115 vprintf(format, ap);
116 va_end(ap);
117 }
118
119 /* To check for memory safety issues, validates that the given pointer in GPU
120 * memory is valid, containing at least sz bytes. The goal is to eliminate
121 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
122 * overruns) by statically validating pointers.
123 */
124
125 static void
126 pandecode_validate_buffer(mali_ptr addr, size_t sz)
127 {
128 if (!addr) {
129 pandecode_msg("XXX: null pointer deref");
130 return;
131 }
132
133 /* Find a BO */
134
135 struct pandecode_mapped_memory *bo =
136 pandecode_find_mapped_gpu_mem_containing(addr);
137
138 if (!bo) {
139 pandecode_msg("XXX: invalid memory dereference\n");
140 return;
141 }
142
143 /* Bounds check */
144
145 unsigned offset = addr - bo->gpu_va;
146 unsigned total = offset + sz;
147
148 if (total > bo->length) {
149 pandecode_msg("XXX: buffer overrun. "
150 "Chunk of size %zu at offset %d in buffer of size %zu. "
151 "Overrun by %zu bytes. \n",
152 sz, offset, bo->length, total - bo->length);
153 return;
154 }
155 }
156
157 struct pandecode_flag_info {
158 u64 flag;
159 const char *name;
160 };
161
162 static void
163 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
164 u64 flags)
165 {
166 bool decodable_flags_found = false;
167
168 for (int i = 0; flag_info[i].name; i++) {
169 if ((flags & flag_info[i].flag) != flag_info[i].flag)
170 continue;
171
172 if (!decodable_flags_found) {
173 decodable_flags_found = true;
174 } else {
175 pandecode_log_cont(" | ");
176 }
177
178 pandecode_log_cont("%s", flag_info[i].name);
179
180 flags &= ~flag_info[i].flag;
181 }
182
183 if (decodable_flags_found) {
184 if (flags)
185 pandecode_log_cont(" | 0x%" PRIx64, flags);
186 } else {
187 pandecode_log_cont("0x%" PRIx64, flags);
188 }
189 }
190
191 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
192 static const struct pandecode_flag_info gl_enable_flag_info[] = {
193 FLAG_INFO(OCCLUSION_QUERY),
194 FLAG_INFO(OCCLUSION_PRECISE),
195 FLAG_INFO(FRONT_CCW_TOP),
196 FLAG_INFO(CULL_FACE_FRONT),
197 FLAG_INFO(CULL_FACE_BACK),
198 {}
199 };
200 #undef FLAG_INFO
201
202 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
203 static const struct pandecode_flag_info clear_flag_info[] = {
204 FLAG_INFO(FAST),
205 FLAG_INFO(SLOW),
206 FLAG_INFO(SLOW_STENCIL),
207 {}
208 };
209 #undef FLAG_INFO
210
211 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
212 static const struct pandecode_flag_info mask_flag_info[] = {
213 FLAG_INFO(R),
214 FLAG_INFO(G),
215 FLAG_INFO(B),
216 FLAG_INFO(A),
217 {}
218 };
219 #undef FLAG_INFO
220
221 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
222 static const struct pandecode_flag_info u3_flag_info[] = {
223 FLAG_INFO(HAS_MSAA),
224 FLAG_INFO(CAN_DISCARD),
225 FLAG_INFO(HAS_BLEND_SHADER),
226 FLAG_INFO(DEPTH_WRITEMASK),
227 {}
228 };
229
230 static const struct pandecode_flag_info u4_flag_info[] = {
231 FLAG_INFO(NO_MSAA),
232 FLAG_INFO(NO_DITHER),
233 FLAG_INFO(DEPTH_RANGE_A),
234 FLAG_INFO(DEPTH_RANGE_B),
235 FLAG_INFO(STENCIL_TEST),
236 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER),
237 {}
238 };
239 #undef FLAG_INFO
240
241 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
242 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
243 FLAG_INFO(MSAA),
244 FLAG_INFO(SRGB),
245 {}
246 };
247 #undef FLAG_INFO
248
249 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
250 static const struct pandecode_flag_info mfbd_extra_flag_info[] = {
251 FLAG_INFO(PRESENT),
252 FLAG_INFO(AFBC),
253 FLAG_INFO(ZS),
254 {}
255 };
256 #undef FLAG_INFO
257
258 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
259 static const struct pandecode_flag_info shader_midgard1_flag_info [] = {
260 FLAG_INFO(EARLY_Z),
261 FLAG_INFO(READS_TILEBUFFER),
262 FLAG_INFO(READS_ZS),
263 {}
264 };
265 #undef FLAG_INFO
266
267 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
268 static const struct pandecode_flag_info mfbd_flag_info [] = {
269 FLAG_INFO(DEPTH_WRITE),
270 FLAG_INFO(EXTRA),
271 {}
272 };
273 #undef FLAG_INFO
274
275 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
276 static const struct pandecode_flag_info sampler_flag_info [] = {
277 FLAG_INFO(MAG_NEAREST),
278 FLAG_INFO(MIN_NEAREST),
279 FLAG_INFO(MIP_LINEAR_1),
280 FLAG_INFO(MIP_LINEAR_2),
281 FLAG_INFO(NORM_COORDS),
282 {}
283 };
284 #undef FLAG_INFO
285
286 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
287 static const struct pandecode_flag_info sfbd_unk1_info [] = {
288 FLAG_INFO(MSAA_8),
289 FLAG_INFO(MSAA_A),
290 {}
291 };
292 #undef FLAG_INFO
293
294 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
295 static const struct pandecode_flag_info sfbd_unk2_info [] = {
296 FLAG_INFO(MSAA_B),
297 FLAG_INFO(SRGB),
298 {}
299 };
300 #undef FLAG_INFO
301
302 extern char *replace_fragment;
303 extern char *replace_vertex;
304
305 static char *
306 pandecode_job_type(enum mali_job_type type)
307 {
308 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
309
310 switch (type) {
311 DEFINE_CASE(NULL);
312 DEFINE_CASE(WRITE_VALUE);
313 DEFINE_CASE(CACHE_FLUSH);
314 DEFINE_CASE(COMPUTE);
315 DEFINE_CASE(VERTEX);
316 DEFINE_CASE(TILER);
317 DEFINE_CASE(FUSED);
318 DEFINE_CASE(FRAGMENT);
319
320 case JOB_NOT_STARTED:
321 return "NOT_STARTED";
322
323 default:
324 pandecode_log("Warning! Unknown job type %x\n", type);
325 return "!?!?!?";
326 }
327
328 #undef DEFINE_CASE
329 }
330
331 static char *
332 pandecode_draw_mode(enum mali_draw_mode mode)
333 {
334 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
335
336 switch (mode) {
337 DEFINE_CASE(DRAW_NONE);
338 DEFINE_CASE(POINTS);
339 DEFINE_CASE(LINES);
340 DEFINE_CASE(TRIANGLES);
341 DEFINE_CASE(TRIANGLE_STRIP);
342 DEFINE_CASE(TRIANGLE_FAN);
343 DEFINE_CASE(LINE_STRIP);
344 DEFINE_CASE(LINE_LOOP);
345 DEFINE_CASE(POLYGON);
346 DEFINE_CASE(QUADS);
347 DEFINE_CASE(QUAD_STRIP);
348
349 default:
350 pandecode_msg("XXX: invalid draw mode %X\n", mode);
351 return "";
352 }
353
354 #undef DEFINE_CASE
355 }
356
357 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
358 static char *
359 pandecode_func(enum mali_func mode)
360 {
361 switch (mode) {
362 DEFINE_CASE(NEVER);
363 DEFINE_CASE(LESS);
364 DEFINE_CASE(EQUAL);
365 DEFINE_CASE(LEQUAL);
366 DEFINE_CASE(GREATER);
367 DEFINE_CASE(NOTEQUAL);
368 DEFINE_CASE(GEQUAL);
369 DEFINE_CASE(ALWAYS);
370
371 default:
372 pandecode_msg("XXX: invalid func %X\n", mode);
373 return "";
374 }
375 }
376 #undef DEFINE_CASE
377
378 /* Why is this duplicated? Who knows... */
379 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
380 static char *
381 pandecode_alt_func(enum mali_alt_func mode)
382 {
383 switch (mode) {
384 DEFINE_CASE(NEVER);
385 DEFINE_CASE(LESS);
386 DEFINE_CASE(EQUAL);
387 DEFINE_CASE(LEQUAL);
388 DEFINE_CASE(GREATER);
389 DEFINE_CASE(NOTEQUAL);
390 DEFINE_CASE(GEQUAL);
391 DEFINE_CASE(ALWAYS);
392
393 default:
394 pandecode_msg("XXX: invalid alt func %X\n", mode);
395 return "";
396 }
397 }
398 #undef DEFINE_CASE
399
400 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
401 static char *
402 pandecode_stencil_op(enum mali_stencil_op op)
403 {
404 switch (op) {
405 DEFINE_CASE(KEEP);
406 DEFINE_CASE(REPLACE);
407 DEFINE_CASE(ZERO);
408 DEFINE_CASE(INVERT);
409 DEFINE_CASE(INCR_WRAP);
410 DEFINE_CASE(DECR_WRAP);
411 DEFINE_CASE(INCR);
412 DEFINE_CASE(DECR);
413
414 default:
415 pandecode_msg("XXX: invalid stencil op %X\n", op);
416 return "";
417 }
418 }
419
420 #undef DEFINE_CASE
421
422 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
423 {
424 switch(mode) {
425 /* TODO: Combine to just "instanced" once this can be done
426 * unambiguously in all known cases */
427 case MALI_ATTR_POT_DIVIDE:
428 return "instanced_pot";
429 case MALI_ATTR_MODULO:
430 return "instanced_mod";
431 case MALI_ATTR_NPOT_DIVIDE:
432 return "instanced_npot";
433 case MALI_ATTR_IMAGE:
434 return "image";
435 default:
436 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
437 return "";
438 }
439 }
440
441 static const char *
442 pandecode_special_record(uint64_t v, bool* attribute)
443 {
444 switch(v) {
445 case MALI_ATTR_VERTEXID:
446 *attribute = true;
447 return "gl_VertexID";
448 case MALI_ATTR_INSTANCEID:
449 *attribute = true;
450 return "gl_InstanceID";
451 case MALI_VARYING_FRAG_COORD:
452 return "gl_FragCoord";
453 case MALI_VARYING_FRONT_FACING:
454 return "gl_FrontFacing";
455 case MALI_VARYING_POINT_COORD:
456 return "gl_PointCoord";
457 default:
458 pandecode_msg("XXX: invalid special record %" PRIx64 "\n", v);
459 return "";
460 }
461 }
462
463 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
464 static char *
465 pandecode_wrap_mode(enum mali_wrap_mode op)
466 {
467 switch (op) {
468 DEFINE_CASE(REPEAT);
469 DEFINE_CASE(CLAMP_TO_EDGE);
470 DEFINE_CASE(CLAMP_TO_BORDER);
471 DEFINE_CASE(MIRRORED_REPEAT);
472
473 default:
474 pandecode_msg("XXX: invalid wrap mode %X\n", op);
475 return "";
476 }
477 }
478 #undef DEFINE_CASE
479
480 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
481 static char *
482 pandecode_block_format(enum mali_block_format fmt)
483 {
484 switch (fmt) {
485 DEFINE_CASE(TILED);
486 DEFINE_CASE(UNKNOWN);
487 DEFINE_CASE(LINEAR);
488 DEFINE_CASE(AFBC);
489
490 default:
491 unreachable("Invalid case");
492 }
493 }
494 #undef DEFINE_CASE
495
496 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
497 char *
498 pandecode_exception_access(enum mali_exception_access access)
499 {
500 switch (access) {
501 DEFINE_CASE(NONE);
502 DEFINE_CASE(EXECUTE);
503 DEFINE_CASE(READ);
504 DEFINE_CASE(WRITE);
505
506 default:
507 unreachable("Invalid case");
508 }
509 }
510 #undef DEFINE_CASE
511
512 /* Midgard's tiler descriptor is embedded within the
513 * larger FBD */
514
515 static void
516 pandecode_midgard_tiler_descriptor(
517 const struct midgard_tiler_descriptor *t,
518 unsigned width,
519 unsigned height,
520 bool is_fragment,
521 bool has_hierarchy)
522 {
523 pandecode_log(".tiler = {\n");
524 pandecode_indent++;
525
526 if (t->hierarchy_mask == MALI_TILER_DISABLED)
527 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
528 else
529 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
530
531 /* We know this name from the kernel, but we never see it nonzero */
532
533 if (t->flags)
534 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
535
536 MEMORY_PROP(t, polygon_list);
537
538 /* The body is offset from the base of the polygon list */
539 assert(t->polygon_list_body > t->polygon_list);
540 unsigned body_offset = t->polygon_list_body - t->polygon_list;
541
542 /* It needs to fit inside the reported size */
543 assert(t->polygon_list_size >= body_offset);
544
545 /* Check that we fit */
546 struct pandecode_mapped_memory *plist =
547 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
548
549 assert(t->polygon_list_size <= plist->length);
550
551 /* Now that we've sanity checked, we'll try to calculate the sizes
552 * ourselves for comparison */
553
554 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
555 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
556
557 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
558 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
559 ref_header, ref_size);
560 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
561 pandecode_msg("body offset %d\n", body_offset);
562 }
563
564 /* The tiler heap has a start and end specified -- it should be
565 * identical to what we have in the BO. The exception is if tiling is
566 * disabled. */
567
568 MEMORY_PROP(t, heap_start);
569 assert(t->heap_end >= t->heap_start);
570
571 struct pandecode_mapped_memory *heap =
572 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
573
574 unsigned heap_size = t->heap_end - t->heap_start;
575
576 /* Tiling is enabled with a special flag */
577 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
578 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
579
580 bool tiling_enabled = hierarchy_mask;
581
582 if (tiling_enabled) {
583 /* When tiling is enabled, the heap should be a tight fit */
584 unsigned heap_offset = t->heap_start - heap->gpu_va;
585 if ((heap_offset + heap_size) != heap->length) {
586 pandecode_msg("XXX: heap size %u (expected %zu)\n",
587 heap_size, heap->length - heap_offset);
588 }
589
590 /* We should also have no other flags */
591 if (tiler_flags)
592 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
593 } else {
594 /* When tiling is disabled, we should have that flag and no others */
595
596 if (tiler_flags != MALI_TILER_DISABLED) {
597 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
598 tiler_flags);
599 }
600
601 /* We should also have an empty heap */
602 if (heap_size) {
603 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
604 heap_size);
605 }
606
607 /* Disabled tiling is used only for clear-only jobs, which are
608 * purely FRAGMENT, so we should never see this for
609 * non-FRAGMENT descriptors. */
610
611 if (!is_fragment)
612 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
613 }
614
615 /* We've never seen weights used in practice, but we know from the
616 * kernel these fields is there */
617
618 bool nonzero_weights = false;
619
620 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
621 nonzero_weights |= t->weights[w] != 0x0;
622 }
623
624 if (nonzero_weights) {
625 pandecode_log(".weights = {");
626
627 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
628 pandecode_log("%d, ", t->weights[w]);
629 }
630
631 pandecode_log("},");
632 }
633
634 pandecode_indent--;
635 pandecode_log("}\n");
636 }
637
638 /* Information about the framebuffer passed back for
639 * additional analysis */
640
641 struct pandecode_fbd {
642 unsigned width;
643 unsigned height;
644 unsigned rt_count;
645 bool has_extra;
646 };
647
648 static void
649 pandecode_sfbd_format(struct mali_sfbd_format format)
650 {
651 pandecode_log(".format = {\n");
652 pandecode_indent++;
653
654 pandecode_log(".unk1 = ");
655 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
656 pandecode_log_cont(",\n");
657
658 /* TODO: Map formats so we can check swizzles and print nicely */
659 pandecode_log("swizzle");
660 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
661 pandecode_log_cont(",\n");
662
663 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
664 MALI_NEGATIVE(format.nr_channels));
665
666 pandecode_log(".unk2 = ");
667 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
668 pandecode_log_cont(",\n");
669
670 pandecode_prop("block = %s", pandecode_block_format(format.block));
671
672 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
673
674 pandecode_indent--;
675 pandecode_log("},\n");
676 }
677
678 static struct pandecode_fbd
679 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
680 {
681 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
682 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
683
684 struct pandecode_fbd info = {
685 .has_extra = false,
686 .rt_count = 1
687 };
688
689 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
690 pandecode_indent++;
691
692 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
693 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
694
695 pandecode_sfbd_format(s->format);
696
697 info.width = s->width + 1;
698 info.height = s->height + 1;
699
700 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
701 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
702
703 MEMORY_PROP(s, checksum);
704
705 if (s->checksum_stride)
706 pandecode_prop("checksum_stride = %d", s->checksum_stride);
707
708 MEMORY_PROP(s, framebuffer);
709 pandecode_prop("stride = %d", s->stride);
710
711 /* Earlier in the actual commandstream -- right before width -- but we
712 * delay to flow nicer */
713
714 pandecode_log(".clear_flags = ");
715 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
716 pandecode_log_cont(",\n");
717
718 if (s->depth_buffer) {
719 MEMORY_PROP(s, depth_buffer);
720 pandecode_prop("depth_stride = %d", s->depth_stride);
721 }
722
723 if (s->stencil_buffer) {
724 MEMORY_PROP(s, stencil_buffer);
725 pandecode_prop("stencil_stride = %d", s->stencil_stride);
726 }
727
728 if (s->depth_stride_zero ||
729 s->stencil_stride_zero ||
730 s->zero7 || s->zero8) {
731 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
732 pandecode_prop("depth_stride_zero = 0x%x",
733 s->depth_stride_zero);
734 pandecode_prop("stencil_stride_zero = 0x%x",
735 s->stencil_stride_zero);
736 pandecode_prop("zero7 = 0x%" PRIx32,
737 s->zero7);
738 pandecode_prop("zero8 = 0x%" PRIx32,
739 s->zero8);
740 }
741
742 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
743 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
744 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
745 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
746 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
747 }
748
749 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
750 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
751 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
752 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
753 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
754 }
755
756 if (s->clear_stencil) {
757 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
758 }
759
760 MEMORY_PROP(s, scratchpad);
761 const struct midgard_tiler_descriptor t = s->tiler;
762
763 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
764 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
765
766 pandecode_indent--;
767 pandecode_log("};\n");
768
769 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
770 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
771 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
772 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
773 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
774
775 printf(".zero3 = {");
776
777 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
778 printf("%X, ", s->zero3[i]);
779
780 printf("},\n");
781
782 printf(".zero6 = {");
783
784 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
785 printf("%X, ", s->zero6[i]);
786
787 printf("},\n");
788
789 return info;
790 }
791
792 static void
793 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
794 {
795 pandecode_log(".unknown%d = {", name);
796
797 for (int i = 0; i < count; ++i)
798 printf("%X, ", slide[i]);
799
800 pandecode_log("},\n");
801 }
802
803 #define SHORT_SLIDE(num) \
804 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
805
806 static void
807 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
808 {
809 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
810 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
811
812 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
813 pandecode_indent++;
814
815 SHORT_SLIDE(1);
816
817 pandecode_indent--;
818 printf("},\n");
819 }
820
821 /* Extracts the number of components associated with a Mali format */
822
823 static unsigned
824 pandecode_format_component_count(enum mali_format fmt)
825 {
826 /* Mask out the format class */
827 unsigned top = fmt & 0b11100000;
828
829 switch (top) {
830 case MALI_FORMAT_SNORM:
831 case MALI_FORMAT_UINT:
832 case MALI_FORMAT_UNORM:
833 case MALI_FORMAT_SINT:
834 return ((fmt >> 3) & 3) + 1;
835 default:
836 /* TODO: Validate */
837 return 4;
838 }
839 }
840
841 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
842
843 static unsigned
844 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
845 {
846 unsigned mask = 0;
847 assert(MALI_CHANNEL_RED == 0);
848
849 for (unsigned c = 0; c < 4; ++c) {
850 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
851
852 if (chan <= MALI_CHANNEL_ALPHA)
853 mask |= (1 << chan);
854 }
855
856 return mask;
857 }
858
859 /* Validates that a (format, swizzle) pair is valid, in the sense that the
860 * swizzle doesn't access any components that are undefined in the format.
861 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
862 * omitted */
863
864 static bool
865 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
866 {
867 unsigned nr_comp = pandecode_format_component_count(fmt);
868 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
869 unsigned valid_mask = (1 << nr_comp) - 1;
870 unsigned invalid_mask = ~valid_mask;
871
872 if (access_mask & invalid_mask) {
873 pandecode_msg("XXX: invalid components accessed\n");
874 return false;
875 }
876
877 /* Check for the default non-swizzling swizzle so we can suppress
878 * useless printing for the defaults */
879
880 unsigned default_swizzles[4] = {
881 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
882 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
883 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
884 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
885 };
886
887 return (swizzle == default_swizzles[nr_comp - 1]);
888 }
889
890 /* Maps MALI_RGBA32F to rgba32f, etc */
891
892 static void
893 pandecode_format_short(enum mali_format fmt, bool srgb)
894 {
895 /* We want a type-like format, so cut off the initial MALI_ */
896 char *format = pandecode_format(fmt);
897 format += strlen("MALI_");
898
899 unsigned len = strlen(format);
900 char *lower_format = calloc(1, len + 1);
901
902 for (unsigned i = 0; i < len; ++i)
903 lower_format[i] = tolower(format[i]);
904
905 /* Sanity check sRGB flag is applied to RGB, per the name */
906 if (srgb && lower_format[0] != 'r')
907 pandecode_msg("XXX: sRGB applied to non-colour format\n");
908
909 /* Just prefix with an s, so you get formats like srgba8_unorm */
910 if (srgb)
911 pandecode_log_cont("s");
912
913 pandecode_log_cont("%s", lower_format);
914 free(lower_format);
915 }
916
917 static void
918 pandecode_swizzle(unsigned swizzle, enum mali_format format)
919 {
920 /* First, do some validation */
921 bool trivial_swizzle = pandecode_validate_format_swizzle(
922 format, swizzle);
923
924 if (trivial_swizzle)
925 return;
926
927 /* Next, print the swizzle */
928 pandecode_log_cont(".");
929
930 static const char components[] = "rgba01";
931
932 for (unsigned c = 0; c < 4; ++c) {
933 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
934
935 if (chan >= MALI_CHANNEL_RESERVED_0) {
936 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
937 continue;
938 }
939 pandecode_log_cont("%c", components[chan]);
940 }
941 }
942
943 static void
944 pandecode_rt_format(struct mali_rt_format format)
945 {
946 pandecode_log(".format = {\n");
947 pandecode_indent++;
948
949 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
950 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
951 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
952
953 pandecode_prop("block = %s", pandecode_block_format(format.block));
954
955 /* TODO: Map formats so we can check swizzles and print nicely */
956 pandecode_log("swizzle");
957 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
958 pandecode_log_cont(",\n");
959
960 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
961 MALI_NEGATIVE(format.nr_channels));
962
963 pandecode_log(".flags = ");
964 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
965 pandecode_log_cont(",\n");
966
967 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
968 * which is a faster hardware-based alternative to the wallpaper method
969 * to preserve framebuffer contents across frames. In practice, MFBD
970 * preload is buggy on Midgard, and so this is a chicken bit. If this
971 * bit isn't set, most likely something broke unrelated to preload */
972
973 if (!format.no_preload) {
974 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
975 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
976 }
977
978 if (format.zero)
979 pandecode_prop("zero = 0x%" PRIx32, format.zero);
980
981 pandecode_indent--;
982 pandecode_log("},\n");
983 }
984
985 static void
986 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
987 {
988 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
989 pandecode_indent++;
990
991 for (int i = 0; i < MALI_NEGATIVE(fb->rt_count_1); i++) {
992 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
993 struct pandecode_mapped_memory *mem =
994 pandecode_find_mapped_gpu_mem_containing(rt_va);
995 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
996
997 pandecode_log("{\n");
998 pandecode_indent++;
999
1000 pandecode_rt_format(rt->format);
1001
1002 if (rt->format.block == MALI_BLOCK_AFBC) {
1003 pandecode_log(".afbc = {\n");
1004 pandecode_indent++;
1005
1006 char *a = pointer_as_memory_reference(rt->afbc.metadata);
1007 pandecode_prop("metadata = %s", a);
1008 free(a);
1009
1010 pandecode_prop("stride = %d", rt->afbc.stride);
1011 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
1012
1013 pandecode_indent--;
1014 pandecode_log("},\n");
1015 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
1016 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1017 rt->afbc.metadata,
1018 rt->afbc.stride,
1019 rt->afbc.unk);
1020 }
1021
1022 MEMORY_PROP(rt, framebuffer);
1023 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1024
1025 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1026 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1027 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1028 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1029 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1030 }
1031
1032 if (rt->zero1 || rt->zero2 || rt->zero3) {
1033 pandecode_msg("XXX: render target zeros tripped\n");
1034 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1035 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1036 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
1037 }
1038
1039 pandecode_indent--;
1040 pandecode_log("},\n");
1041 }
1042
1043 pandecode_indent--;
1044 pandecode_log("};\n");
1045 }
1046
1047 static struct pandecode_fbd
1048 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment)
1049 {
1050 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1051 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1052
1053 struct pandecode_fbd info;
1054
1055 if (fb->sample_locations) {
1056 /* The blob stores all possible sample locations in a single buffer
1057 * allocated on startup, and just switches the pointer when switching
1058 * MSAA state. For now, we just put the data into the cmdstream, but we
1059 * should do something like what the blob does with a real driver.
1060 *
1061 * There seem to be 32 slots for sample locations, followed by another
1062 * 16. The second 16 is just the center location followed by 15 zeros
1063 * in all the cases I've identified (maybe shader vs. depth/color
1064 * samples?).
1065 */
1066
1067 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
1068
1069 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
1070
1071 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
1072 pandecode_indent++;
1073
1074 for (int i = 0; i < 32 + 16; i++) {
1075 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1076 }
1077
1078 pandecode_indent--;
1079 pandecode_log("};\n");
1080 }
1081
1082 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1083 pandecode_indent++;
1084
1085 pandecode_prop("stack_shift = 0x%x", fb->stack_shift);
1086 pandecode_prop("unk0 = 0x%x", fb->unk0);
1087
1088 if (fb->sample_locations)
1089 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1090
1091 /* Assume that unknown1 was emitted in the last job for
1092 * now */
1093 MEMORY_PROP(fb, unknown1);
1094
1095 info.width = fb->width1 + 1;
1096 info.height = fb->height1 + 1;
1097 info.rt_count = fb->rt_count_1 + 1;
1098
1099 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1100 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1101 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1102 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1103
1104 pandecode_prop("unk1 = 0x%x", fb->unk1);
1105 pandecode_prop("unk2 = 0x%x", fb->unk2);
1106 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1107 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1108
1109 pandecode_log(".mfbd_flags = ");
1110 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1111 pandecode_log_cont(",\n");
1112
1113 if (fb->clear_stencil)
1114 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1115
1116 if (fb->clear_depth)
1117 pandecode_prop("clear_depth = %f", fb->clear_depth);
1118
1119 /* TODO: What is this? Let's not blow up.. */
1120 if (fb->unknown2 != 0x1F)
1121 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1122
1123 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1124 MEMORY_PROP(fb, scratchpad);
1125 const struct midgard_tiler_descriptor t = fb->tiler;
1126 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
1127
1128 if (fb->zero3 || fb->zero4) {
1129 pandecode_msg("XXX: framebuffer zeros tripped\n");
1130 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1131 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1132 }
1133
1134 pandecode_indent--;
1135 pandecode_log("};\n");
1136
1137 gpu_va += sizeof(struct bifrost_framebuffer);
1138
1139 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1140
1141 if (info.has_extra) {
1142 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1143 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1144
1145 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1146 pandecode_indent++;
1147
1148 MEMORY_PROP(fbx, checksum);
1149
1150 if (fbx->checksum_stride)
1151 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1152
1153 pandecode_log(".flags = ");
1154 pandecode_log_decoded_flags(mfbd_extra_flag_info, fbx->flags);
1155 pandecode_log_cont(",\n");
1156
1157 if (fbx->flags & MALI_EXTRA_AFBC_ZS) {
1158 pandecode_log(".ds_afbc = {\n");
1159 pandecode_indent++;
1160
1161 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1162 pandecode_prop("depth_stencil_afbc_stride = %d",
1163 fbx->ds_afbc.depth_stencil_afbc_stride);
1164 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1165
1166 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
1167 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1168 pandecode_prop("zero1 = 0x%" PRIx32,
1169 fbx->ds_afbc.zero1);
1170 pandecode_prop("padding = 0x%" PRIx64,
1171 fbx->ds_afbc.padding);
1172 }
1173
1174 pandecode_indent--;
1175 pandecode_log("},\n");
1176 } else {
1177 pandecode_log(".ds_linear = {\n");
1178 pandecode_indent++;
1179
1180 if (fbx->ds_linear.depth) {
1181 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1182 pandecode_prop("depth_stride = %d",
1183 fbx->ds_linear.depth_stride);
1184 }
1185
1186 if (fbx->ds_linear.stencil) {
1187 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1188 pandecode_prop("stencil_stride = %d",
1189 fbx->ds_linear.stencil_stride);
1190 }
1191
1192 if (fbx->ds_linear.depth_stride_zero ||
1193 fbx->ds_linear.stencil_stride_zero ||
1194 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1195 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1196 pandecode_prop("depth_stride_zero = 0x%x",
1197 fbx->ds_linear.depth_stride_zero);
1198 pandecode_prop("stencil_stride_zero = 0x%x",
1199 fbx->ds_linear.stencil_stride_zero);
1200 pandecode_prop("zero1 = 0x%" PRIx32,
1201 fbx->ds_linear.zero1);
1202 pandecode_prop("zero2 = 0x%" PRIx32,
1203 fbx->ds_linear.zero2);
1204 }
1205
1206 pandecode_indent--;
1207 pandecode_log("},\n");
1208 }
1209
1210 if (fbx->zero3 || fbx->zero4) {
1211 pandecode_msg("XXX: fb_extra zeros tripped\n");
1212 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1213 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1214 }
1215
1216 pandecode_indent--;
1217 pandecode_log("};\n");
1218
1219 gpu_va += sizeof(struct bifrost_fb_extra);
1220 }
1221
1222 if (is_fragment)
1223 pandecode_render_target(gpu_va, job_no, fb);
1224
1225 return info;
1226 }
1227
1228 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1229 * count */
1230
1231 static void
1232 pandecode_padded_vertices(unsigned shift, unsigned k)
1233 {
1234 unsigned odd = 2*k + 1;
1235 unsigned pot = 1 << shift;
1236 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1237 }
1238
1239 /* Given a magic divisor, recover what we were trying to divide by.
1240 *
1241 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1242 * 0 <= m < 2^N, for N bits in m.
1243 *
1244 * Let q represent the number we would like to divide by.
1245 *
1246 * By definition of a magic divisor for N-bit unsigned integers (a number you
1247 * multiply by to magically get division), m is a number such that:
1248 *
1249 * (m * x) & (2^N - 1) = floor(x/q).
1250 * for all x on Z where 0 <= x < 2^N
1251 *
1252 * Ignore the case where any of the above values equals zero; it is irrelevant
1253 * for our purposes (instanced arrays).
1254 *
1255 * Choose x = q. Then:
1256 *
1257 * (m * x) & (2^N - 1) = floor(x/q).
1258 * (m * q) & (2^N - 1) = floor(q/q).
1259 *
1260 * floor(q/q) = floor(1) = 1, therefore:
1261 *
1262 * (m * q) & (2^N - 1) = 1
1263 *
1264 * Recall the identity that the bitwise AND of one less than a power-of-two
1265 * equals the modulo with that power of two, i.e. for all x:
1266 *
1267 * x & (2^N - 1) = x % N
1268 *
1269 * Therefore:
1270 *
1271 * mq % (2^N) = 1
1272 *
1273 * By definition, a modular multiplicative inverse of a number m is the number
1274 * q such that with respect to a modulos M:
1275 *
1276 * mq % M = 1
1277 *
1278 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1279 *
1280 */
1281
1282 static void
1283 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1284 {
1285 #if 0
1286 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1287 * shift) the most lame way possible... just repeatedly add.
1288 * Asymptoptically slow but nobody cares in practice, unless you have
1289 * massive numbers of vertices or high divisors. */
1290
1291 unsigned inverse = 0;
1292
1293 /* Magic implicitly has the highest bit set */
1294 magic |= (1 << 31);
1295
1296 /* Depending on rounding direction */
1297 if (extra)
1298 magic++;
1299
1300 for (;;) {
1301 uint32_t product = magic * inverse;
1302
1303 if (shift) {
1304 product >>= shift;
1305 }
1306
1307 if (product == 1)
1308 break;
1309
1310 ++inverse;
1311 }
1312
1313 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1314
1315 /* Recall we're supposed to divide by (gl_level_divisor *
1316 * padded_num_vertices) */
1317
1318 unsigned padded_num_vertices = inverse / orig_divisor;
1319
1320 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1321 #endif
1322 }
1323
1324 static void
1325 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1326 mali_ptr addr, int job_no, char *suffix,
1327 int count, bool varying, enum mali_job_type job_type)
1328 {
1329 char *prefix = varying ? "varying" : "attribute";
1330 assert(addr);
1331
1332 if (!count) {
1333 pandecode_msg("warn: No %s records\n", prefix);
1334 return;
1335 }
1336
1337 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1338
1339 for (int i = 0; i < count; ++i) {
1340 /* First, check for special records */
1341 if (attr[i].elements < MALI_RECORD_SPECIAL) {
1342 if (attr[i].size)
1343 pandecode_msg("XXX: tripped size=%d\n", attr[i].size);
1344
1345 if (attr[i].stride) {
1346 /* gl_InstanceID passes a magic divisor in the
1347 * stride field to divide by the padded vertex
1348 * count. No other records should do so, so
1349 * stride should otherwise be zero. Note that
1350 * stride in the usual attribute sense doesn't
1351 * apply to special records. */
1352
1353 bool has_divisor = attr[i].elements == MALI_ATTR_INSTANCEID;
1354
1355 pandecode_log_cont("/* %smagic divisor = %X */ ",
1356 has_divisor ? "" : "XXX: ", attr[i].stride);
1357 }
1358
1359 if (attr[i].shift || attr[i].extra_flags) {
1360 /* Attributes use these fields for
1361 * instancing/padding/etc type issues, but
1362 * varyings don't */
1363
1364 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1365 varying ? "XXX: " : "",
1366 attr[i].shift, attr[i].extra_flags);
1367 }
1368
1369 /* Print the special record name */
1370 bool attribute = false;
1371 pandecode_log("%s_%d = %s;\n", prefix, i, pandecode_special_record(attr[i].elements, &attribute));
1372
1373 /* Sanity check */
1374 if (attribute == varying)
1375 pandecode_msg("XXX: mismatched special record\n");
1376
1377 continue;
1378 }
1379
1380 enum mali_attr_mode mode = attr[i].elements & 7;
1381
1382 if (mode == MALI_ATTR_UNUSED)
1383 pandecode_msg("XXX: unused attribute record\n");
1384
1385 /* For non-linear records, we need to print the type of record */
1386 if (mode != MALI_ATTR_LINEAR)
1387 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1388
1389 /* Print the name to link with attr_meta */
1390 pandecode_log_cont("%s_%d", prefix, i);
1391
1392 /* Print the stride and size */
1393 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1394
1395 /* TODO: Sanity check the quotient itself. It must be equal to
1396 * (or be greater than, if the driver added padding) the padded
1397 * vertex count. */
1398
1399 /* Finally, print the pointer */
1400 mali_ptr raw_elements = attr[i].elements & ~7;
1401 char *a = pointer_as_memory_reference(raw_elements);
1402 pandecode_log_cont(" = (%s);\n", a);
1403 free(a);
1404
1405 /* Check the pointer */
1406 pandecode_validate_buffer(raw_elements, attr[i].size);
1407
1408 /* shift/extra_flags exist only for instanced */
1409 if (attr[i].shift | attr[i].extra_flags) {
1410 /* These are set to random values by the blob for
1411 * varyings, most likely a symptom of uninitialized
1412 * memory where the hardware masked the bug. As such we
1413 * put this at a warning, not an error. */
1414
1415 if (mode == MALI_ATTR_LINEAR)
1416 pandecode_msg("warn: instancing fields set for linear\n");
1417
1418 pandecode_prop("shift = %d", attr[i].shift);
1419 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1420 }
1421
1422 /* Decode further where possible */
1423
1424 if (mode == MALI_ATTR_MODULO) {
1425 pandecode_padded_vertices(
1426 attr[i].shift,
1427 attr[i].extra_flags);
1428 }
1429
1430 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1431 i++;
1432 pandecode_log("{\n");
1433 pandecode_indent++;
1434 pandecode_prop("unk = 0x%x", attr[i].unk);
1435 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1436 if (attr[i].zero != 0)
1437 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1438 pandecode_prop("divisor = %d", attr[i].divisor);
1439 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1440 pandecode_indent--;
1441 pandecode_log("}, \n");
1442 }
1443
1444 }
1445
1446 pandecode_log("\n");
1447 }
1448
1449 static mali_ptr
1450 pandecode_shader_address(const char *name, mali_ptr ptr)
1451 {
1452 /* TODO: Decode flags */
1453 mali_ptr shader_ptr = ptr & ~15;
1454
1455 char *a = pointer_as_memory_reference(shader_ptr);
1456 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1457 free(a);
1458
1459 return shader_ptr;
1460 }
1461
1462 static void
1463 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1464 {
1465 unsigned any_nonzero =
1466 stencil->ref | stencil->mask | stencil->func |
1467 stencil->sfail | stencil->dpfail | stencil->dppass;
1468
1469 if (any_nonzero == 0)
1470 return;
1471
1472 const char *func = pandecode_func(stencil->func);
1473 const char *sfail = pandecode_stencil_op(stencil->sfail);
1474 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1475 const char *dppass = pandecode_stencil_op(stencil->dppass);
1476
1477 if (stencil->zero)
1478 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1479
1480 pandecode_log(".stencil_%s = {\n", name);
1481 pandecode_indent++;
1482 pandecode_prop("ref = %d", stencil->ref);
1483 pandecode_prop("mask = 0x%02X", stencil->mask);
1484 pandecode_prop("func = %s", func);
1485 pandecode_prop("sfail = %s", sfail);
1486 pandecode_prop("dpfail = %s", dpfail);
1487 pandecode_prop("dppass = %s", dppass);
1488 pandecode_indent--;
1489 pandecode_log("},\n");
1490 }
1491
1492 static void
1493 pandecode_blend_equation(const struct mali_blend_equation *blend)
1494 {
1495 if (blend->zero1)
1496 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1497
1498 pandecode_log(".equation = {\n");
1499 pandecode_indent++;
1500
1501 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1502 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1503
1504 pandecode_log(".color_mask = ");
1505 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1506 pandecode_log_cont(",\n");
1507
1508 pandecode_indent--;
1509 pandecode_log("},\n");
1510 }
1511
1512 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1513
1514 static unsigned
1515 decode_bifrost_constant(u16 constant)
1516 {
1517 float lo = (float) (constant & 0xFF);
1518 float hi = (float) (constant >> 8);
1519
1520 return (hi / 255.0) + (lo / 65535.0);
1521 }
1522
1523 static mali_ptr
1524 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1525 {
1526 struct bifrost_blend_rt *b =
1527 ((struct bifrost_blend_rt *) descs) + rt_no;
1528
1529 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1530 pandecode_indent++;
1531
1532 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1533 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1534 b->constant, decode_bifrost_constant(b->constant));
1535
1536 /* TODO figure out blend shader enable bit */
1537 pandecode_blend_equation(&b->equation);
1538 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1539 pandecode_prop("index = 0x%" PRIx16, b->index);
1540 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1541
1542 pandecode_indent--;
1543 pandecode_log("},\n");
1544
1545 return 0;
1546 }
1547
1548 static mali_ptr
1549 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1550 {
1551 /* constant/equation is in a union */
1552 if (!blend->shader)
1553 return 0;
1554
1555 pandecode_log(".blend = {\n");
1556 pandecode_indent++;
1557
1558 if (is_shader) {
1559 pandecode_shader_address("shader", blend->shader);
1560 } else {
1561 pandecode_blend_equation(&blend->equation);
1562 pandecode_prop("constant = %f", blend->constant);
1563 }
1564
1565 pandecode_indent--;
1566 pandecode_log("},\n");
1567
1568 /* Return blend shader to disassemble if present */
1569 return is_shader ? (blend->shader & ~0xF) : 0;
1570 }
1571
1572 static mali_ptr
1573 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1574 {
1575 struct midgard_blend_rt *b =
1576 ((struct midgard_blend_rt *) descs) + rt_no;
1577
1578 /* Flags determine presence of blend shader */
1579 bool is_shader = (b->flags & 0xF) >= 0x2;
1580
1581 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1582 pandecode_indent++;
1583
1584 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1585
1586 union midgard_blend blend = b->blend;
1587 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1588
1589 pandecode_indent--;
1590 pandecode_log("};\n");
1591
1592 return shader;
1593 }
1594
1595 /* Attributes and varyings have descriptor records, which contain information
1596 * about their format and ordering with the attribute/varying buffers. We'll
1597 * want to validate that the combinations specified are self-consistent.
1598 */
1599
1600 static int
1601 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1602 {
1603 char base[128];
1604 char *prefix = varying ? "varying" : "attribute";
1605 unsigned max_index = 0;
1606 snprintf(base, sizeof(base), "%s_meta", prefix);
1607
1608 struct mali_attr_meta *attr_meta;
1609 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1610
1611 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1612
1613 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1614 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1615 sizeof(*attr_mem));
1616
1617 /* If the record is discard, it should be zero for everything else */
1618
1619 if (attr_meta->format == MALI_VARYING_DISCARD) {
1620 uint64_t zero =
1621 attr_meta->index |
1622 attr_meta->unknown1 |
1623 attr_meta->unknown3 |
1624 attr_meta->src_offset;
1625
1626 if (zero)
1627 pandecode_msg("XXX: expected empty record for varying discard\n");
1628
1629 /* We want to look for a literal 0000 swizzle -- this
1630 * is not encoded with all zeroes, however */
1631
1632 enum mali_channel z = MALI_CHANNEL_ZERO;
1633 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1634 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1635
1636 if (!good_swizzle)
1637 pandecode_msg("XXX: expected zero swizzle for discard\n");
1638
1639 if (!varying)
1640 pandecode_msg("XXX: cannot discard attribute\n");
1641
1642 /* If we're all good, omit the record */
1643 if (!zero && varying && good_swizzle) {
1644 pandecode_log("/* discarded varying */\n");
1645 continue;
1646 }
1647 }
1648
1649 if (attr_meta->index > max_index)
1650 max_index = attr_meta->index;
1651
1652 if (attr_meta->unknown1 != 0x2) {
1653 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1654 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1655 }
1656
1657 if (attr_meta->unknown3) {
1658 pandecode_msg("XXX: unexpected unknown3 set\n");
1659 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1660 }
1661
1662 pandecode_format_short(attr_meta->format, false);
1663 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1664
1665 if (attr_meta->src_offset)
1666 pandecode_log_cont("[%u]", attr_meta->src_offset);
1667
1668 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1669
1670 pandecode_log_cont(";\n");
1671 }
1672
1673 pandecode_log("\n");
1674
1675 return count ? (max_index + 1) : 0;
1676 }
1677
1678 /* return bits [lo, hi) of word */
1679 static u32
1680 bits(u32 word, u32 lo, u32 hi)
1681 {
1682 if (hi - lo >= 32)
1683 return word; // avoid undefined behavior with the shift
1684
1685 return (word >> lo) & ((1 << (hi - lo)) - 1);
1686 }
1687
1688 static void
1689 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1690 {
1691 pandecode_log_cont("{\n");
1692 pandecode_indent++;
1693
1694 /* Decode invocation_count. See the comment before the definition of
1695 * invocation_count for an explanation.
1696 */
1697
1698 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1699 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1700 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1701 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1702 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1703 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1704
1705 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1706 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1707 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1708
1709 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1710 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1711 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1712
1713 /* Even though we have this decoded, we want to ensure that the
1714 * representation is "unique" so we don't lose anything by printing only
1715 * the final result. More specifically, we need to check that we were
1716 * passed something in canonical form, since the definition per the
1717 * hardware is inherently not unique. How? Well, take the resulting
1718 * decode and pack it ourselves! If it is bit exact with what we
1719 * decoded, we're good to go. */
1720
1721 struct mali_vertex_tiler_prefix ref;
1722 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1723
1724 bool canonical =
1725 (p->invocation_count == ref.invocation_count) &&
1726 (p->invocation_shifts == ref.invocation_shifts);
1727
1728 if (!canonical) {
1729 pandecode_msg("XXX: non-canonical workgroups packing\n");
1730 pandecode_msg("expected: %X, %X",
1731 ref.invocation_count,
1732 ref.invocation_shifts);
1733
1734 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1735 pandecode_prop("size_y_shift = %d", size_y_shift);
1736 pandecode_prop("size_z_shift = %d", size_z_shift);
1737 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
1738 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
1739 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
1740 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
1741 }
1742
1743 /* Regardless, print the decode */
1744 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1745 size_x, size_y, size_z,
1746 groups_x, groups_y, groups_z);
1747
1748 /* TODO: Decode */
1749 if (p->unknown_draw)
1750 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1751
1752 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1753
1754 if (p->draw_mode != MALI_DRAW_NONE)
1755 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1756
1757 /* Index count only exists for tiler jobs anyway */
1758
1759 if (p->index_count)
1760 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1761
1762
1763 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1764 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1765
1766 /* Validate an index buffer is present if we need one. TODO: verify
1767 * relationship between invocation_count and index_count */
1768
1769 if (p->indices) {
1770 unsigned count = p->index_count;
1771
1772 /* Grab the size */
1773 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1774
1775 /* Ensure we got a size, and if so, validate the index buffer
1776 * is large enough to hold a full set of indices of the given
1777 * size */
1778
1779 if (!index_raw_size)
1780 pandecode_msg("XXX: index size missing\n");
1781 else
1782 pandecode_validate_buffer(p->indices, count * size);
1783 } else if (index_raw_size)
1784 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1785
1786 if (p->offset_bias_correction)
1787 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1788
1789 /* TODO: Figure out what this is. It's not zero */
1790 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1791
1792 pandecode_indent--;
1793 pandecode_log("},\n");
1794 }
1795
1796 static void
1797 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1798 {
1799 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1800 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1801
1802 for (int i = 0; i < ubufs_count; i++) {
1803 unsigned size = (ubufs[i].size + 1) * 16;
1804 mali_ptr addr = ubufs[i].ptr << 2;
1805
1806 pandecode_validate_buffer(addr, size);
1807
1808 char *ptr = pointer_as_memory_reference(ubufs[i].ptr << 2);
1809 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1810 free(ptr);
1811 }
1812
1813 pandecode_log("\n");
1814 }
1815
1816 static void
1817 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1818 {
1819 pandecode_validate_buffer(uniforms, uniform_count * 16);
1820
1821 char *ptr = pointer_as_memory_reference(uniforms);
1822 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1823 free(ptr);
1824 }
1825
1826 static void
1827 pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1828 {
1829
1830 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1831
1832 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1833
1834 if (scratchpad->zero) {
1835 pandecode_msg("XXX: scratchpad zero tripped");
1836 pandecode_prop("zero = 0x%x\n", scratchpad->zero);
1837 }
1838
1839 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1840 pandecode_indent++;
1841
1842 pandecode_prop("flags = 0x%x", scratchpad->flags);
1843 MEMORY_PROP(scratchpad, gpu_scratchpad);
1844
1845 pandecode_indent--;
1846 pandecode_log("};\n");
1847 }
1848
1849 static const char *
1850 shader_type_for_job(unsigned type)
1851 {
1852 switch (type) {
1853 case JOB_TYPE_VERTEX: return "VERTEX";
1854 case JOB_TYPE_TILER: return "FRAGMENT";
1855 case JOB_TYPE_COMPUTE: return "COMPUTE";
1856 default:
1857 return "UNKNOWN";
1858 }
1859 }
1860
1861 static unsigned shader_id = 0;
1862
1863 static struct midgard_disasm_stats
1864 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1865 bool is_bifrost, unsigned gpu_id)
1866 {
1867 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1868 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1869
1870 /* Compute maximum possible size */
1871 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1872
1873 /* Print some boilerplate to clearly denote the assembly (which doesn't
1874 * obey indentation rules), and actually do the disassembly! */
1875
1876 printf("\n\n");
1877
1878 struct midgard_disasm_stats stats;
1879
1880 if (is_bifrost) {
1881 disassemble_bifrost(code, sz, false);
1882
1883 /* TODO: Extend stats to Bifrost */
1884 stats.texture_count = -128;
1885 stats.sampler_count = -128;
1886 stats.attribute_count = -128;
1887 stats.varying_count = -128;
1888 stats.uniform_count = -128;
1889 stats.uniform_buffer_count = -128;
1890 stats.work_count = -128;
1891
1892 stats.instruction_count = 0;
1893 stats.bundle_count = 0;
1894 stats.quadword_count = 0;
1895 stats.helper_invocations = false;
1896 } else {
1897 stats = disassemble_midgard(code, sz, gpu_id,
1898 type == JOB_TYPE_TILER ?
1899 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
1900 }
1901
1902 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1903 * driver-internal purposes with the blob and interfere */
1904
1905 bool should_shaderdb = type != JOB_TYPE_COMPUTE;
1906
1907 if (should_shaderdb) {
1908 unsigned nr_threads =
1909 (stats.work_count <= 4) ? 4 :
1910 (stats.work_count <= 8) ? 2 :
1911 1;
1912
1913 printf("shader%d - MESA_SHADER_%s shader: "
1914 "%u inst, %u bundles, %u quadwords, "
1915 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1916 shader_id++,
1917 shader_type_for_job(type),
1918 stats.instruction_count, stats.bundle_count, stats.quadword_count,
1919 stats.work_count, nr_threads);
1920 }
1921
1922
1923 return stats;
1924 }
1925
1926 static void
1927 pandecode_texture(mali_ptr u,
1928 struct pandecode_mapped_memory *tmem,
1929 unsigned job_no, unsigned tex)
1930 {
1931 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
1932
1933 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
1934 pandecode_indent++;
1935
1936 struct mali_texture_format f = t->format;
1937
1938 /* See the definiton of enum mali_texture_type */
1939
1940 bool is_cube = f.type == MALI_TEX_CUBE;
1941 unsigned dimension = is_cube ? 2 : f.type;
1942
1943 pandecode_make_indent();
1944
1945 /* TODO: Are there others? */
1946 bool is_zs = f.format == MALI_Z32_UNORM;
1947
1948 /* Recall Z/S switched the meaning of linear/tiled .. */
1949 if (is_zs && f.layout == MALI_TEXTURE_LINEAR)
1950 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1951
1952 /* Print the layout. Default is linear; a modifier can denote AFBC or
1953 * u-interleaved/tiled modes */
1954
1955 if (f.layout == MALI_TEXTURE_AFBC)
1956 pandecode_log_cont("afbc");
1957 else if (f.layout == MALI_TEXTURE_TILED)
1958 pandecode_log_cont(is_zs ? "linear" : "tiled");
1959 else if (f.layout == MALI_TEXTURE_LINEAR)
1960 pandecode_log_cont("linear");
1961 else
1962 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
1963
1964 pandecode_swizzle(t->swizzle, f.format);
1965 pandecode_log_cont(" ");
1966
1967 /* Distinguish cube/2D with modifier */
1968
1969 if (is_cube)
1970 pandecode_log_cont("cube ");
1971
1972 pandecode_format_short(f.format, f.srgb);
1973 pandecode_swizzle(f.swizzle, f.format);
1974
1975 /* All four width/height/depth/array_size dimensions are present
1976 * regardless of the type of texture, but it is an error to have
1977 * non-zero dimensions for unused dimensions. Verify this. array_size
1978 * can always be set, as can width. */
1979
1980 if (t->height && dimension < 2)
1981 pandecode_msg("XXX: nonzero height for <2D texture\n");
1982
1983 if (t->depth && dimension < 3)
1984 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1985
1986 /* Print only the dimensions that are actually there */
1987
1988 pandecode_log_cont(": %d", t->width + 1);
1989
1990 if (dimension >= 2)
1991 pandecode_log_cont("x%u", t->height + 1);
1992
1993 if (dimension >= 3)
1994 pandecode_log_cont("x%u", t->depth + 1);
1995
1996 if (t->array_size)
1997 pandecode_log_cont("[%u]", t->array_size + 1);
1998
1999 if (t->levels)
2000 pandecode_log_cont(" mip %u", t->levels);
2001
2002 pandecode_log_cont("\n");
2003
2004 if (f.unknown1 | f.zero) {
2005 pandecode_msg("XXX: texture format zero tripped\n");
2006 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
2007 pandecode_prop("zero = %" PRId32, f.zero);
2008 }
2009
2010 if (!f.unknown2) {
2011 pandecode_msg("XXX: expected unknown texture bit set\n");
2012 pandecode_prop("unknown2 = %" PRId32, f.unknown1);
2013 }
2014
2015 if (t->swizzle_zero) {
2016 pandecode_msg("XXX: swizzle zero tripped\n");
2017 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
2018 }
2019
2020 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
2021 pandecode_msg("XXX: texture zero tripped\n");
2022 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
2023 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
2024 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2025 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2026 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2027 }
2028
2029 pandecode_log(".payload = {\n");
2030 pandecode_indent++;
2031
2032 /* A bunch of bitmap pointers follow.
2033 * We work out the correct number,
2034 * based on the mipmap/cubemap
2035 * properties, but dump extra
2036 * possibilities to futureproof */
2037
2038 int bitmap_count = MALI_NEGATIVE(t->levels);
2039
2040 /* Miptree for each face */
2041 if (f.type == MALI_TEX_CUBE)
2042 bitmap_count *= 6;
2043
2044 /* Array of textures */
2045 bitmap_count *= MALI_NEGATIVE(t->array_size);
2046
2047 /* Stride for each element */
2048 if (f.manual_stride)
2049 bitmap_count *= 2;
2050
2051 /* Sanity check the size */
2052 int max_count = sizeof(t->payload) / sizeof(t->payload[0]);
2053 assert (bitmap_count <= max_count);
2054
2055 for (int i = 0; i < bitmap_count; ++i) {
2056 /* How we dump depends if this is a stride or a pointer */
2057
2058 if (f.manual_stride && (i & 1)) {
2059 /* signed 32-bit snuck in as a 64-bit pointer */
2060 uint64_t stride_set = t->payload[i];
2061 uint32_t clamped_stride = stride_set;
2062 int32_t stride = clamped_stride;
2063 assert(stride_set == clamped_stride);
2064 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2065 } else {
2066 char *a = pointer_as_memory_reference(t->payload[i]);
2067 pandecode_log("%s, \n", a);
2068 free(a);
2069 }
2070 }
2071
2072 pandecode_indent--;
2073 pandecode_log("},\n");
2074
2075 pandecode_indent--;
2076 pandecode_log("};\n");
2077 }
2078
2079 /* 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 */
2080
2081 static void
2082 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2083 {
2084 /* Nothing to do */
2085 if (claim == truth)
2086 return;
2087
2088 if (fuzzy)
2089 assert(truth >= 0);
2090
2091 if ((truth >= 0) && !fuzzy) {
2092 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2093 (truth < claim) ? "warn" : "XXX",
2094 name, truth, claim);
2095 } else if ((claim > -truth) && !fuzzy) {
2096 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2097 name, -truth, claim);
2098 } else if (fuzzy && (claim < truth))
2099 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2100 name, truth, claim);
2101
2102 pandecode_log(".%s = %" PRId16, name, claim);
2103
2104 if (fuzzy)
2105 pandecode_log_cont(" /* %u used */", truth);
2106
2107 pandecode_log_cont(",\n");
2108 }
2109
2110 static void
2111 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2112 bool is_bifrost, unsigned gpu_id)
2113 {
2114 struct midgard_disasm_stats stats =
2115 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
2116
2117 bool has_texture = (stats.texture_count > 0);
2118 bool has_sampler = (stats.sampler_count > 0);
2119 bool has_attribute = (stats.attribute_count > 0);
2120 bool has_varying = (stats.varying_count > 0);
2121 bool has_uniform = (stats.uniform_count > 0);
2122 bool has_ubo = (stats.uniform_buffer_count > 0);
2123
2124 if (has_texture || has_sampler)
2125 pandecode_msg("XXX: blend shader accessing textures\n");
2126
2127 if (has_attribute || has_varying)
2128 pandecode_msg("XXX: blend shader accessing interstage\n");
2129
2130 if (has_uniform || has_ubo)
2131 pandecode_msg("XXX: blend shader accessing uniforms\n");
2132 }
2133
2134 static void
2135 pandecode_vertex_tiler_postfix_pre(
2136 const struct mali_vertex_tiler_postfix *p,
2137 int job_no, enum mali_job_type job_type,
2138 char *suffix, bool is_bifrost, unsigned gpu_id)
2139 {
2140 struct pandecode_mapped_memory *attr_mem;
2141
2142 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2143 * are the only things actually needed from the FBD, vertex/tiler jobs
2144 * no longer reference the FBD -- instead, this field points to some
2145 * info about the scratchpad.
2146 */
2147
2148 struct pandecode_fbd fbd_info = {
2149 /* Default for Bifrost */
2150 .rt_count = 1
2151 };
2152
2153 if (is_bifrost)
2154 pandecode_scratchpad(p->framebuffer & ~1, job_no, suffix);
2155 else if (p->framebuffer & MALI_MFBD)
2156 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
2157 else if (job_type == JOB_TYPE_COMPUTE)
2158 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
2159 else
2160 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no, false, gpu_id);
2161
2162 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2163 int texture_count = 0, sampler_count = 0;
2164
2165 if (p->shader) {
2166 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2167 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2168
2169 /* Disassemble ahead-of-time to get stats. Initialize with
2170 * stats for the missing-shader case so we get validation
2171 * there, too */
2172
2173 struct midgard_disasm_stats info = {
2174 .texture_count = 0,
2175 .sampler_count = 0,
2176 .attribute_count = 0,
2177 .varying_count = 0,
2178 .work_count = 1,
2179
2180 .uniform_count = -128,
2181 .uniform_buffer_count = 0
2182 };
2183
2184 if (s->shader & ~0xF)
2185 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
2186
2187 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2188 pandecode_indent++;
2189
2190 /* Save for dumps */
2191 attribute_count = s->attribute_count;
2192 varying_count = s->varying_count;
2193 texture_count = s->texture_count;
2194 sampler_count = s->sampler_count;
2195
2196 if (is_bifrost) {
2197 uniform_count = s->bifrost2.uniform_count;
2198 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2199 } else {
2200 uniform_count = s->midgard1.uniform_count;
2201 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2202 }
2203
2204 pandecode_shader_address("shader", s->shader);
2205
2206 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2207 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2208 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2209 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2210 pandecode_shader_prop("uniform_buffer_count",
2211 uniform_buffer_count,
2212 info.uniform_buffer_count, true);
2213
2214 if (!is_bifrost) {
2215 pandecode_shader_prop("uniform_count",
2216 uniform_count,
2217 info.uniform_count, false);
2218
2219 pandecode_shader_prop("work_count",
2220 s->midgard1.work_count, info.work_count, false);
2221 }
2222
2223 if (is_bifrost) {
2224 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32, s->bifrost1.unk1);
2225 } else {
2226 bool helpers = s->midgard1.flags & MALI_HELPER_INVOCATIONS;
2227 s->midgard1.flags &= ~MALI_HELPER_INVOCATIONS;
2228
2229 if (helpers != info.helper_invocations) {
2230 pandecode_msg("XXX: expected helpers %u but got %u\n",
2231 info.helper_invocations, helpers);
2232 }
2233
2234 pandecode_log(".midgard1.flags = ");
2235 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
2236 pandecode_log_cont(",\n");
2237
2238 pandecode_prop("midgard1.unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
2239 }
2240
2241 if (s->depth_units || s->depth_factor) {
2242 pandecode_prop("depth_factor = %f", s->depth_factor);
2243 pandecode_prop("depth_units = %f", s->depth_units);
2244 }
2245
2246 if (s->alpha_coverage) {
2247 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
2248 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
2249
2250 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2251 invert_alpha_coverage ? "~" : "",
2252 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
2253 }
2254
2255 if (s->unknown2_3 || s->unknown2_4) {
2256 pandecode_log(".unknown2_3 = ");
2257
2258 int unknown2_3 = s->unknown2_3;
2259 int unknown2_4 = s->unknown2_4;
2260
2261 /* We're not quite sure what these flags mean without the depth test, if anything */
2262
2263 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2264 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2265 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2266
2267 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2268 }
2269
2270 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2271 pandecode_log_cont(",\n");
2272
2273 pandecode_log(".unknown2_4 = ");
2274 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2275 pandecode_log_cont(",\n");
2276 }
2277
2278 if (s->stencil_mask_front || s->stencil_mask_back) {
2279 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2280 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2281 }
2282
2283 pandecode_stencil("front", &s->stencil_front);
2284 pandecode_stencil("back", &s->stencil_back);
2285
2286 if (is_bifrost) {
2287 pandecode_log(".bifrost2 = {\n");
2288 pandecode_indent++;
2289
2290 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2291 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2292 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2293 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2294
2295 pandecode_indent--;
2296 pandecode_log("},\n");
2297 } else if (s->midgard2.unknown2_7) {
2298 pandecode_log(".midgard2 = {\n");
2299 pandecode_indent++;
2300
2301 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2302 pandecode_indent--;
2303 pandecode_log("},\n");
2304 }
2305
2306 if (s->unknown2_8)
2307 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
2308
2309 if (!is_bifrost) {
2310 /* TODO: Blend shaders routing/disasm */
2311 union midgard_blend blend = s->blend;
2312 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2313 if (shader & ~0xF)
2314 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2315 }
2316
2317 pandecode_indent--;
2318 pandecode_log("};\n");
2319
2320 /* MRT blend fields are used whenever MFBD is used, with
2321 * per-RT descriptors */
2322
2323 if (job_type == JOB_TYPE_TILER && p->framebuffer & MALI_MFBD) {
2324 void* blend_base = (void *) (s + 1);
2325
2326 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2327 mali_ptr shader = 0;
2328
2329 if (is_bifrost)
2330 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2331 else
2332 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2333
2334 if (shader & ~0xF)
2335 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2336
2337 }
2338 }
2339 } else
2340 pandecode_msg("XXX: missing shader descriptor\n");
2341
2342 if (p->viewport) {
2343 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2344 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2345
2346 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2347 pandecode_indent++;
2348
2349 pandecode_prop("clip_minx = %f", f->clip_minx);
2350 pandecode_prop("clip_miny = %f", f->clip_miny);
2351 pandecode_prop("clip_minz = %f", f->clip_minz);
2352 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2353 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2354 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2355
2356 /* Only the higher coordinates are MALI_POSITIVE scaled */
2357
2358 pandecode_prop("viewport0 = { %d, %d }",
2359 f->viewport0[0], f->viewport0[1]);
2360
2361 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2362 f->viewport1[0] + 1, f->viewport1[1] + 1);
2363
2364 pandecode_indent--;
2365 pandecode_log("};\n");
2366 }
2367
2368 unsigned max_attr_index = 0;
2369
2370 if (p->attribute_meta)
2371 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2372
2373 if (p->attributes) {
2374 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2375 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2376 }
2377
2378 /* Varyings are encoded like attributes but not actually sent; we just
2379 * pass a zero buffer with the right stride/size set, (or whatever)
2380 * since the GPU will write to it itself */
2381
2382 if (p->varying_meta) {
2383 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2384 }
2385
2386 if (p->varyings) {
2387 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2388
2389 /* Number of descriptors depends on whether there are
2390 * non-internal varyings */
2391
2392 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2393 }
2394
2395 if (p->uniform_buffers) {
2396 if (uniform_buffer_count)
2397 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2398 else
2399 pandecode_msg("warn: UBOs specified but not referenced\n");
2400 } else if (uniform_buffer_count)
2401 pandecode_msg("XXX: UBOs referenced but not specified\n");
2402
2403 /* We don't want to actually dump uniforms, but we do need to validate
2404 * that the counts we were given are sane */
2405
2406 if (p->uniforms) {
2407 if (uniform_count)
2408 pandecode_uniforms(p->uniforms, uniform_count);
2409 else
2410 pandecode_msg("warn: Uniforms specified but not referenced\n");
2411 } else if (uniform_count)
2412 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2413
2414 if (p->texture_trampoline) {
2415 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
2416
2417 if (mmem) {
2418 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
2419
2420 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
2421 pandecode_indent++;
2422
2423 for (int tex = 0; tex < texture_count; ++tex) {
2424 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2425 char *a = pointer_as_memory_reference(*u);
2426 pandecode_log("%s,\n", a);
2427 free(a);
2428 }
2429
2430 pandecode_indent--;
2431 pandecode_log("};\n");
2432
2433 /* Now, finally, descend down into the texture descriptor */
2434 for (unsigned tex = 0; tex < texture_count; ++tex) {
2435 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2436 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2437 if (tmem)
2438 pandecode_texture(*u, tmem, job_no, tex);
2439 }
2440 }
2441 }
2442
2443 if (p->sampler_descriptor) {
2444 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2445
2446 if (smem) {
2447 struct mali_sampler_descriptor *s;
2448
2449 mali_ptr d = p->sampler_descriptor;
2450
2451 for (int i = 0; i < sampler_count; ++i) {
2452 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2453
2454 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2455 pandecode_indent++;
2456
2457 pandecode_log(".filter_mode = ");
2458 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2459 pandecode_log_cont(",\n");
2460
2461 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2462 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2463
2464 if (s->lod_bias)
2465 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s->lod_bias));
2466
2467 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2468 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2469 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2470
2471 pandecode_prop("compare_func = %s", pandecode_alt_func(s->compare_func));
2472
2473 if (s->zero || s->zero2) {
2474 pandecode_msg("XXX: sampler zero tripped\n");
2475 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2476 }
2477
2478 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2479
2480 pandecode_prop("border_color = { %f, %f, %f, %f }",
2481 s->border_color[0],
2482 s->border_color[1],
2483 s->border_color[2],
2484 s->border_color[3]);
2485
2486 pandecode_indent--;
2487 pandecode_log("};\n");
2488 }
2489 }
2490 }
2491 }
2492
2493 static void
2494 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2495 {
2496 if (p->shader & 0xF)
2497 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2498
2499 if (!(p->position_varying || p->occlusion_counter))
2500 return;
2501
2502 pandecode_log(".postfix = {\n");
2503 pandecode_indent++;
2504
2505 MEMORY_PROP(p, position_varying);
2506 MEMORY_PROP(p, occlusion_counter);
2507
2508 pandecode_indent--;
2509 pandecode_log("},\n");
2510 }
2511
2512 static void
2513 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2514 {
2515 pandecode_log_cont("{\n");
2516 pandecode_indent++;
2517
2518 pandecode_prop("unk2 = 0x%x", v->unk2);
2519
2520 if (v->zero0 || v->zero1) {
2521 pandecode_msg("XXX: vertex only zero tripped");
2522 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2523 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2524 }
2525
2526 pandecode_indent--;
2527 pandecode_log("}\n");
2528 }
2529
2530 static void
2531 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2532 {
2533
2534 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2535 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2536
2537 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2538 pandecode_indent++;
2539
2540 if (h->zero) {
2541 pandecode_msg("XXX: tiler heap zero tripped\n");
2542 pandecode_prop("zero = 0x%x", h->zero);
2543 }
2544
2545 for (int i = 0; i < 12; i++) {
2546 if (h->zeros[i] != 0) {
2547 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2548 i, h->zeros[i]);
2549 }
2550 }
2551
2552 pandecode_prop("heap_size = 0x%x", h->heap_size);
2553 MEMORY_PROP(h, tiler_heap_start);
2554 MEMORY_PROP(h, tiler_heap_free);
2555
2556 /* this might point to the beginning of another buffer, when it's
2557 * really the end of the tiler heap buffer, so we have to be careful
2558 * here. but for zero length, we need the same pointer.
2559 */
2560
2561 if (h->tiler_heap_end == h->tiler_heap_start) {
2562 MEMORY_PROP(h, tiler_heap_start);
2563 } else {
2564 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2565 pandecode_prop("tiler_heap_end = %s + 1", a);
2566 free(a);
2567 }
2568
2569 pandecode_indent--;
2570 pandecode_log("};\n");
2571 }
2572
2573 static void
2574 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2575 {
2576 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2577 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2578
2579 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2580
2581 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2582 pandecode_indent++;
2583
2584 if (t->zero0 || t->zero1) {
2585 pandecode_msg("XXX: tiler meta zero tripped\n");
2586 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2587 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2588 }
2589
2590 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2591 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2592
2593 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2594 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2595
2596 for (int i = 0; i < 12; i++) {
2597 if (t->zeros[i] != 0) {
2598 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2599 i, t->zeros[i]);
2600 }
2601 }
2602
2603 pandecode_indent--;
2604 pandecode_log("};\n");
2605 }
2606
2607 static void
2608 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2609 {
2610 pandecode_log(".gl_enables = ");
2611
2612 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2613
2614 pandecode_log_cont(",\n");
2615 }
2616
2617 static void
2618 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2619 {
2620 if (u.pointer == 0x0)
2621 return;
2622
2623 pandecode_log(".primitive_size = {\n");
2624 pandecode_indent++;
2625
2626 if (constant) {
2627 pandecode_prop("constant = %f", u.constant);
2628 } else {
2629 MEMORY_PROP((&u), pointer);
2630 }
2631
2632 pandecode_indent--;
2633 pandecode_log("},\n");
2634 }
2635
2636 static void
2637 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2638 {
2639 pandecode_log_cont("{\n");
2640 pandecode_indent++;
2641
2642 /* TODO: gl_PointSize on Bifrost */
2643 pandecode_primitive_size(t->primitive_size, true);
2644
2645 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2646
2647 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2648 || t->zero6 || t->zero7 || t->zero8) {
2649 pandecode_msg("XXX: tiler only zero tripped\n");
2650 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2651 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2652 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2653 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2654 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2655 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2656 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2657 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2658 }
2659
2660 pandecode_indent--;
2661 pandecode_log("},\n");
2662 }
2663
2664 static int
2665 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2666 const struct pandecode_mapped_memory *mem,
2667 mali_ptr payload, int job_no, unsigned gpu_id)
2668 {
2669 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2670
2671 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
2672
2673 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2674 pandecode_indent++;
2675
2676 pandecode_log(".prefix = ");
2677 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2678
2679 pandecode_log(".vertex = ");
2680 pandecode_vertex_only_bfr(&v->vertex);
2681
2682 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2683
2684 pandecode_indent--;
2685 pandecode_log("};\n");
2686
2687 return sizeof(*v);
2688 }
2689
2690 static int
2691 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2692 const struct pandecode_mapped_memory *mem,
2693 mali_ptr payload, int job_no, unsigned gpu_id)
2694 {
2695 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2696
2697 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
2698 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2699
2700 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2701 pandecode_indent++;
2702
2703 pandecode_log(".prefix = ");
2704 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2705
2706 pandecode_log(".tiler = ");
2707 pandecode_tiler_only_bfr(&t->tiler, job_no);
2708
2709 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2710
2711 pandecode_indent--;
2712 pandecode_log("};\n");
2713
2714 return sizeof(*t);
2715 }
2716
2717 static int
2718 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2719 const struct pandecode_mapped_memory *mem,
2720 mali_ptr payload, int job_no, unsigned gpu_id)
2721 {
2722 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2723
2724 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
2725
2726 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2727 pandecode_indent++;
2728
2729 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2730 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2731
2732 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2733
2734 pandecode_log(".prefix = ");
2735 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
2736
2737 pandecode_gl_enables(v->gl_enables, h->job_type);
2738
2739 if (v->instance_shift || v->instance_odd) {
2740 pandecode_prop("instance_shift = 0x%d /* %d */",
2741 v->instance_shift, 1 << v->instance_shift);
2742 pandecode_prop("instance_odd = 0x%X /* %d */",
2743 v->instance_odd, (2 * v->instance_odd) + 1);
2744
2745 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2746 }
2747
2748 if (v->offset_start)
2749 pandecode_prop("offset_start = %d", v->offset_start);
2750
2751 if (v->zero5) {
2752 pandecode_msg("XXX: midgard payload zero tripped\n");
2753 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2754 }
2755
2756 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2757
2758 pandecode_indent--;
2759 pandecode_log("};\n");
2760
2761 return sizeof(*v);
2762 }
2763
2764 static int
2765 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2766 mali_ptr payload, int job_no,
2767 bool is_bifrost, unsigned gpu_id)
2768 {
2769 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2770
2771 bool is_mfbd = s->framebuffer & MALI_MFBD;
2772
2773 /* Bifrost theoretically may retain support for SFBD on compute jobs,
2774 * but for graphics workloads with a FRAGMENT payload, use MFBD */
2775
2776 if (!is_mfbd && is_bifrost)
2777 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2778
2779 struct pandecode_fbd info;
2780
2781 if (is_mfbd)
2782 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2783 else
2784 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
2785
2786 /* Compute the tag for the tagged pointer. This contains the type of
2787 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2788 * additional structures follow the MFBD header (an extra payload or
2789 * not, as well as a count of render targets) */
2790
2791 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
2792
2793 if (is_mfbd) {
2794 if (info.has_extra)
2795 expected_tag |= MALI_MFBD_TAG_EXTRA;
2796
2797 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
2798 }
2799
2800 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
2801 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2802 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
2803 pandecode_prop("max_tile_coord = 0x%X\n", s->min_tile_coord);
2804 }
2805
2806 /* Extract tile coordinates */
2807
2808 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
2809 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
2810
2811 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2812 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2813
2814 /* For the max, we also want the floored (rather than ceiled) version for checking */
2815
2816 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
2817 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
2818
2819 /* Validate the coordinates are well-ordered */
2820
2821 if (min_x == max_x)
2822 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2823 else if (min_x > max_x)
2824 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2825
2826 if (min_y == max_y)
2827 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2828 else if (min_y > max_y)
2829 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2830
2831 /* Validate the coordinates fit inside the framebuffer. We use floor,
2832 * rather than ceil, for the max coordinates, since the tile
2833 * coordinates for something like an 800x600 framebuffer will actually
2834 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2835
2836 if ((min_x > info.width) || (max_x_f > info.width))
2837 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2838
2839 if ((min_y > info.height) || (max_y_f > info.height))
2840 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2841
2842 /* After validation, we print */
2843
2844 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
2845
2846 /* The FBD is a tagged pointer */
2847
2848 unsigned tag = (s->framebuffer & ~FBD_MASK);
2849
2850 if (tag != expected_tag)
2851 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
2852
2853 return sizeof(*s);
2854 }
2855
2856 static int job_descriptor_number = 0;
2857
2858 int
2859 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id)
2860 {
2861 struct mali_job_descriptor_header *h;
2862
2863 int start_number = 0;
2864
2865 bool first = true;
2866 bool last_size;
2867
2868 do {
2869 struct pandecode_mapped_memory *mem =
2870 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2871
2872 void *payload;
2873
2874 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2875
2876 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2877 * high 32-bits of the 64-bit pointer are reused to store
2878 * something else.
2879 */
2880 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2881 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2882 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2883
2884 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
2885
2886 int job_no = job_descriptor_number++;
2887
2888 if (first)
2889 start_number = job_no;
2890
2891 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2892 pandecode_indent++;
2893
2894 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2895
2896 /* Save for next job fixing */
2897 last_size = h->job_descriptor_size;
2898
2899 if (h->job_descriptor_size)
2900 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2901
2902 if (h->exception_status && h->exception_status != 0x1)
2903 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2904 h->exception_status,
2905 (h->exception_status >> 16) & 0xFFFF,
2906 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2907 h->exception_status & 0xFF);
2908
2909 if (h->first_incomplete_task)
2910 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2911
2912 if (h->fault_pointer)
2913 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2914
2915 if (h->job_barrier)
2916 pandecode_prop("job_barrier = %d", h->job_barrier);
2917
2918 pandecode_prop("job_index = %d", h->job_index);
2919
2920 if (h->unknown_flags)
2921 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2922
2923 if (h->job_dependency_index_1)
2924 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2925
2926 if (h->job_dependency_index_2)
2927 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2928
2929 pandecode_indent--;
2930 pandecode_log("};\n");
2931
2932 /* Do not touch the field yet -- decode the payload first, and
2933 * don't touch that either. This is essential for the uploads
2934 * to occur in sequence and therefore be dynamically allocated
2935 * correctly. Do note the size, however, for that related
2936 * reason. */
2937
2938 switch (h->job_type) {
2939 case JOB_TYPE_WRITE_VALUE: {
2940 struct mali_payload_write_value *s = payload;
2941 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2942 pandecode_indent++;
2943 MEMORY_PROP(s, address);
2944
2945 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
2946 pandecode_msg("XXX: unknown value descriptor\n");
2947 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
2948 }
2949
2950 if (s->reserved) {
2951 pandecode_msg("XXX: set value tripped\n");
2952 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
2953 }
2954
2955 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
2956 pandecode_indent--;
2957 pandecode_log("};\n");
2958
2959 break;
2960 }
2961
2962 case JOB_TYPE_TILER:
2963 case JOB_TYPE_VERTEX:
2964 case JOB_TYPE_COMPUTE:
2965 if (bifrost) {
2966 if (h->job_type == JOB_TYPE_TILER)
2967 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2968 else
2969 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2970 } else
2971 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
2972
2973 break;
2974
2975 case JOB_TYPE_FRAGMENT:
2976 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
2977 break;
2978
2979 default:
2980 break;
2981 }
2982
2983 /* Handle linkage */
2984
2985 if (!first) {
2986 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2987
2988 if (last_size)
2989 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no);
2990 else
2991 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no);
2992 }
2993
2994 first = false;
2995
2996 } while ((jc_gpu_va = h->job_descriptor_size ? h->next_job_64 : h->next_job_32));
2997
2998 return start_number;
2999 }