projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
panfrost: Decode blend shaders for SFBD
[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 int pandecode_jc(mali_ptr jc_gpu_va, bool bifrost);
44
45 static void pandecode_swizzle(unsigned swizzle, enum mali_format format);
46
47 #define MEMORY_PROP(obj, p) {\
48 if (obj->p) { \
49 char *a = pointer_as_memory_reference(obj->p); \
50 pandecode_prop("%s = %s", #p, a); \
51 free(a); \
52 } \
53 }
54
55 #define MEMORY_PROP_DIR(obj, p) {\
56 if (obj.p) { \
57 char *a = pointer_as_memory_reference(obj.p); \
58 pandecode_prop("%s = %s", #p, a); \
59 free(a); \
60 } \
61 }
62
63 /* Semantic logging type.
64 *
65 * Raw: for raw messages to be printed as is.
66 * Message: for helpful information to be commented out in replays.
67 * Property: for properties of a struct
68 *
69 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
70 */
71
72 enum pandecode_log_type {
73 PANDECODE_RAW,
74 PANDECODE_MESSAGE,
75 PANDECODE_PROPERTY
76 };
77
78 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
79 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
80 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
81
82 unsigned pandecode_indent = 0;
83
84 static void
85 pandecode_make_indent(void)
86 {
87 for (unsigned i = 0; i < pandecode_indent; ++i)
88 printf(" ");
89 }
90
91 static void
92 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
93 {
94 va_list ap;
95
96 pandecode_make_indent();
97
98 if (type == PANDECODE_MESSAGE)
99 printf("// ");
100 else if (type == PANDECODE_PROPERTY)
101 printf(".");
102
103 va_start(ap, format);
104 vprintf(format, ap);
105 va_end(ap);
106
107 if (type == PANDECODE_PROPERTY)
108 printf(",\n");
109 }
110
111 static void
112 pandecode_log_cont(const char *format, ...)
113 {
114 va_list ap;
115
116 va_start(ap, format);
117 vprintf(format, ap);
118 va_end(ap);
119 }
120
121 /* To check for memory safety issues, validates that the given pointer in GPU
122 * memory is valid, containing at least sz bytes. The goal is to eliminate
123 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
124 * overruns) by statically validating pointers.
125 */
126
127 static void
128 pandecode_validate_buffer(mali_ptr addr, size_t sz)
129 {
130 if (!addr) {
131 pandecode_msg("XXX: null pointer deref");
132 return;
133 }
134
135 /* Find a BO */
136
137 struct pandecode_mapped_memory *bo =
138 pandecode_find_mapped_gpu_mem_containing(addr);
139
140 if (!bo) {
141 pandecode_msg("XXX: invalid memory dereference\n");
142 return;
143 }
144
145 /* Bounds check */
146
147 unsigned offset = addr - bo->gpu_va;
148 unsigned total = offset + sz;
149
150 if (total > bo->length) {
151 pandecode_msg("XXX: buffer overrun. "
152 "Chunk of size %zu at offset %d in buffer of size %zu. "
153 "Overrun by %zu bytes. \n",
154 sz, offset, bo->length, total - bo->length);
155 return;
156 }
157 }
158
159 struct pandecode_flag_info {
160 u64 flag;
161 const char *name;
162 };
163
164 static void
165 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
166 u64 flags)
167 {
168 bool decodable_flags_found = false;
169
170 for (int i = 0; flag_info[i].name; i++) {
171 if ((flags & flag_info[i].flag) != flag_info[i].flag)
172 continue;
173
174 if (!decodable_flags_found) {
175 decodable_flags_found = true;
176 } else {
177 pandecode_log_cont(" | ");
178 }
179
180 pandecode_log_cont("%s", flag_info[i].name);
181
182 flags &= ~flag_info[i].flag;
183 }
184
185 if (decodable_flags_found) {
186 if (flags)
187 pandecode_log_cont(" | 0x%" PRIx64, flags);
188 } else {
189 pandecode_log_cont("0x%" PRIx64, flags);
190 }
191 }
192
193 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
194 static const struct pandecode_flag_info gl_enable_flag_info[] = {
195 FLAG_INFO(OCCLUSION_QUERY),
196 FLAG_INFO(OCCLUSION_PRECISE),
197 FLAG_INFO(FRONT_CCW_TOP),
198 FLAG_INFO(CULL_FACE_FRONT),
199 FLAG_INFO(CULL_FACE_BACK),
200 {}
201 };
202 #undef FLAG_INFO
203
204 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
205 static const struct pandecode_flag_info clear_flag_info[] = {
206 FLAG_INFO(FAST),
207 FLAG_INFO(SLOW),
208 FLAG_INFO(SLOW_STENCIL),
209 {}
210 };
211 #undef FLAG_INFO
212
213 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
214 static const struct pandecode_flag_info mask_flag_info[] = {
215 FLAG_INFO(R),
216 FLAG_INFO(G),
217 FLAG_INFO(B),
218 FLAG_INFO(A),
219 {}
220 };
221 #undef FLAG_INFO
222
223 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
224 static const struct pandecode_flag_info u3_flag_info[] = {
225 FLAG_INFO(HAS_MSAA),
226 FLAG_INFO(CAN_DISCARD),
227 FLAG_INFO(HAS_BLEND_SHADER),
228 FLAG_INFO(DEPTH_WRITEMASK),
229 {}
230 };
231
232 static const struct pandecode_flag_info u4_flag_info[] = {
233 FLAG_INFO(NO_MSAA),
234 FLAG_INFO(NO_DITHER),
235 FLAG_INFO(DEPTH_RANGE_A),
236 FLAG_INFO(DEPTH_RANGE_B),
237 FLAG_INFO(STENCIL_TEST),
238 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER),
239 {}
240 };
241 #undef FLAG_INFO
242
243 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
244 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
245 FLAG_INFO(MSAA),
246 FLAG_INFO(SRGB),
247 {}
248 };
249 #undef FLAG_INFO
250
251 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
252 static const struct pandecode_flag_info mfbd_extra_flag_info[] = {
253 FLAG_INFO(PRESENT),
254 FLAG_INFO(AFBC),
255 FLAG_INFO(ZS),
256 {}
257 };
258 #undef FLAG_INFO
259
260 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
261 static const struct pandecode_flag_info shader_midgard1_flag_info [] = {
262 FLAG_INFO(EARLY_Z),
263 FLAG_INFO(READS_TILEBUFFER),
264 FLAG_INFO(READS_ZS),
265 {}
266 };
267 #undef FLAG_INFO
268
269 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
270 static const struct pandecode_flag_info mfbd_flag_info [] = {
271 FLAG_INFO(DEPTH_WRITE),
272 FLAG_INFO(EXTRA),
273 {}
274 };
275 #undef FLAG_INFO
276
277 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
278 static const struct pandecode_flag_info sampler_flag_info [] = {
279 FLAG_INFO(MAG_NEAREST),
280 FLAG_INFO(MIN_NEAREST),
281 FLAG_INFO(MIP_LINEAR_1),
282 FLAG_INFO(MIP_LINEAR_2),
283 FLAG_INFO(NORM_COORDS),
284 {}
285 };
286 #undef FLAG_INFO
287
288 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
289 static const struct pandecode_flag_info sfbd_unk1_info [] = {
290 FLAG_INFO(MSAA_8),
291 FLAG_INFO(MSAA_A),
292 {}
293 };
294 #undef FLAG_INFO
295
296 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
297 static const struct pandecode_flag_info sfbd_unk2_info [] = {
298 FLAG_INFO(MSAA_B),
299 FLAG_INFO(SRGB),
300 {}
301 };
302 #undef FLAG_INFO
303
304 extern char *replace_fragment;
305 extern char *replace_vertex;
306
307 static char *
308 pandecode_job_type(enum mali_job_type type)
309 {
310 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
311
312 switch (type) {
313 DEFINE_CASE(NULL);
314 DEFINE_CASE(SET_VALUE);
315 DEFINE_CASE(CACHE_FLUSH);
316 DEFINE_CASE(COMPUTE);
317 DEFINE_CASE(VERTEX);
318 DEFINE_CASE(TILER);
319 DEFINE_CASE(FUSED);
320 DEFINE_CASE(FRAGMENT);
321
322 case JOB_NOT_STARTED:
323 return "NOT_STARTED";
324
325 default:
326 pandecode_log("Warning! Unknown job type %x\n", type);
327 return "!?!?!?";
328 }
329
330 #undef DEFINE_CASE
331 }
332
333 static char *
334 pandecode_draw_mode(enum mali_draw_mode mode)
335 {
336 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
337
338 switch (mode) {
339 DEFINE_CASE(DRAW_NONE);
340 DEFINE_CASE(POINTS);
341 DEFINE_CASE(LINES);
342 DEFINE_CASE(TRIANGLES);
343 DEFINE_CASE(TRIANGLE_STRIP);
344 DEFINE_CASE(TRIANGLE_FAN);
345 DEFINE_CASE(LINE_STRIP);
346 DEFINE_CASE(LINE_LOOP);
347 DEFINE_CASE(POLYGON);
348 DEFINE_CASE(QUADS);
349 DEFINE_CASE(QUAD_STRIP);
350
351 default:
352 pandecode_msg("XXX: invalid draw mode %X\n", mode);
353 return "";
354 }
355
356 #undef DEFINE_CASE
357 }
358
359 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
360 static char *
361 pandecode_func(enum mali_func mode)
362 {
363 switch (mode) {
364 DEFINE_CASE(NEVER);
365 DEFINE_CASE(LESS);
366 DEFINE_CASE(EQUAL);
367 DEFINE_CASE(LEQUAL);
368 DEFINE_CASE(GREATER);
369 DEFINE_CASE(NOTEQUAL);
370 DEFINE_CASE(GEQUAL);
371 DEFINE_CASE(ALWAYS);
372
373 default:
374 pandecode_msg("XXX: invalid func %X\n", mode);
375 return "";
376 }
377 }
378 #undef DEFINE_CASE
379
380 /* Why is this duplicated? Who knows... */
381 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
382 static char *
383 pandecode_alt_func(enum mali_alt_func mode)
384 {
385 switch (mode) {
386 DEFINE_CASE(NEVER);
387 DEFINE_CASE(LESS);
388 DEFINE_CASE(EQUAL);
389 DEFINE_CASE(LEQUAL);
390 DEFINE_CASE(GREATER);
391 DEFINE_CASE(NOTEQUAL);
392 DEFINE_CASE(GEQUAL);
393 DEFINE_CASE(ALWAYS);
394
395 default:
396 pandecode_msg("XXX: invalid alt func %X\n", mode);
397 return "";
398 }
399 }
400 #undef DEFINE_CASE
401
402 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
403 static char *
404 pandecode_stencil_op(enum mali_stencil_op op)
405 {
406 switch (op) {
407 DEFINE_CASE(KEEP);
408 DEFINE_CASE(REPLACE);
409 DEFINE_CASE(ZERO);
410 DEFINE_CASE(INVERT);
411 DEFINE_CASE(INCR_WRAP);
412 DEFINE_CASE(DECR_WRAP);
413 DEFINE_CASE(INCR);
414 DEFINE_CASE(DECR);
415
416 default:
417 pandecode_msg("XXX: invalid stencil op %X\n", op);
418 return "";
419 }
420 }
421
422 #undef DEFINE_CASE
423
424 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
425 {
426 switch(mode) {
427 /* TODO: Combine to just "instanced" once this can be done
428 * unambiguously in all known cases */
429 case MALI_ATTR_POT_DIVIDE:
430 return "instanced_pot";
431 case MALI_ATTR_MODULO:
432 return "instanced_mod";
433 case MALI_ATTR_NPOT_DIVIDE:
434 return "instanced_npot";
435 case MALI_ATTR_IMAGE:
436 return "image";
437 case MALI_ATTR_INTERNAL:
438 return "internal";
439 default:
440 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
441 return "";
442 }
443 }
444
445 static const char *
446 pandecode_special_varying(uint64_t v)
447 {
448 switch(v) {
449 case MALI_VARYING_FRAG_COORD:
450 return "gl_FragCoord";
451 case MALI_VARYING_FRONT_FACING:
452 return "gl_FrontFacing";
453 case MALI_VARYING_POINT_COORD:
454 return "gl_PointCoord";
455 default:
456 pandecode_msg("XXX: invalid special varying %" PRIx64 "\n", v);
457 return "";
458 }
459 }
460
461 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
462 static char *
463 pandecode_wrap_mode(enum mali_wrap_mode op)
464 {
465 switch (op) {
466 DEFINE_CASE(REPEAT);
467 DEFINE_CASE(CLAMP_TO_EDGE);
468 DEFINE_CASE(CLAMP_TO_BORDER);
469 DEFINE_CASE(MIRRORED_REPEAT);
470
471 default:
472 pandecode_msg("XXX: invalid wrap mode %X\n", op);
473 return "";
474 }
475 }
476 #undef DEFINE_CASE
477
478 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
479 static char *
480 pandecode_block_format(enum mali_block_format fmt)
481 {
482 switch (fmt) {
483 DEFINE_CASE(TILED);
484 DEFINE_CASE(UNKNOWN);
485 DEFINE_CASE(LINEAR);
486 DEFINE_CASE(AFBC);
487
488 default:
489 unreachable("Invalid case");
490 }
491 }
492 #undef DEFINE_CASE
493
494 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
495 static char *
496 pandecode_exception_access(enum mali_exception_access access)
497 {
498 switch (access) {
499 DEFINE_CASE(NONE);
500 DEFINE_CASE(EXECUTE);
501 DEFINE_CASE(READ);
502 DEFINE_CASE(WRITE);
503
504 default:
505 unreachable("Invalid case");
506 }
507 }
508 #undef DEFINE_CASE
509
510 /* Midgard's tiler descriptor is embedded within the
511 * larger FBD */
512
513 static void
514 pandecode_midgard_tiler_descriptor(
515 const struct midgard_tiler_descriptor *t,
516 unsigned width,
517 unsigned height,
518 bool is_fragment)
519 {
520 pandecode_log(".tiler = {\n");
521 pandecode_indent++;
522
523 if (t->hierarchy_mask == MALI_TILER_DISABLED)
524 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
525 else
526 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
527
528 /* We know this name from the kernel, but we never see it nonzero */
529
530 if (t->flags)
531 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
532
533 MEMORY_PROP(t, polygon_list);
534
535 /* The body is offset from the base of the polygon list */
536 assert(t->polygon_list_body > t->polygon_list);
537 unsigned body_offset = t->polygon_list_body - t->polygon_list;
538
539 /* It needs to fit inside the reported size */
540 assert(t->polygon_list_size >= body_offset);
541
542 /* Check that we fit */
543 struct pandecode_mapped_memory *plist =
544 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
545
546 assert(t->polygon_list_size <= plist->length);
547
548 /* Now that we've sanity checked, we'll try to calculate the sizes
549 * ourselves for comparison */
550
551 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask);
552 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask);
553
554 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
555 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
556 ref_header, ref_size);
557 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
558 pandecode_msg("body offset %d\n", body_offset);
559 }
560
561 /* The tiler heap has a start and end specified -- it should be
562 * identical to what we have in the BO. The exception is if tiling is
563 * disabled. */
564
565 MEMORY_PROP(t, heap_start);
566 assert(t->heap_end >= t->heap_start);
567
568 struct pandecode_mapped_memory *heap =
569 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
570
571 unsigned heap_size = t->heap_end - t->heap_start;
572
573 /* Tiling is enabled with a special flag */
574 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
575 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
576
577 bool tiling_enabled = hierarchy_mask;
578
579 if (tiling_enabled) {
580 /* When tiling is enabled, the heap should be a tight fit */
581 unsigned heap_offset = t->heap_start - heap->gpu_va;
582 if ((heap_offset + heap_size) != heap->length) {
583 pandecode_msg("XXX: heap size %u (expected %zu)\n",
584 heap_size, heap->length - heap_offset);
585 }
586
587 /* We should also have no other flags */
588 if (tiler_flags)
589 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
590 } else {
591 /* When tiling is disabled, we should have that flag and no others */
592
593 if (tiler_flags != MALI_TILER_DISABLED) {
594 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
595 tiler_flags);
596 }
597
598 /* We should also have an empty heap */
599 if (heap_size) {
600 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
601 heap_size);
602 }
603
604 /* Disabled tiling is used only for clear-only jobs, which are
605 * purely FRAGMENT, so we should never see this for
606 * non-FRAGMENT descriptors. */
607
608 if (!is_fragment)
609 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
610 }
611
612 /* We've never seen weights used in practice, but we know from the
613 * kernel these fields is there */
614
615 bool nonzero_weights = false;
616
617 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
618 nonzero_weights |= t->weights[w] != 0x0;
619 }
620
621 if (nonzero_weights) {
622 pandecode_log(".weights = {");
623
624 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
625 pandecode_log("%d, ", t->weights[w]);
626 }
627
628 pandecode_log("},");
629 }
630
631 pandecode_indent--;
632 pandecode_log("}\n");
633 }
634
635 /* Information about the framebuffer passed back for
636 * additional analysis */
637
638 struct pandecode_fbd {
639 unsigned width;
640 unsigned height;
641 unsigned rt_count;
642 bool has_extra;
643 };
644
645 static void
646 pandecode_sfbd_format(struct mali_sfbd_format format)
647 {
648 pandecode_log(".format = {\n");
649 pandecode_indent++;
650
651 pandecode_log(".unk1 = ");
652 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
653 pandecode_log_cont(",\n");
654
655 /* TODO: Map formats so we can check swizzles and print nicely */
656 pandecode_log("swizzle");
657 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
658 pandecode_log_cont(",\n");
659
660 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
661 MALI_NEGATIVE(format.nr_channels));
662
663 pandecode_log(".unk2 = ");
664 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
665 pandecode_log_cont(",\n");
666
667 pandecode_prop("block = %s", pandecode_block_format(format.block));
668
669 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
670
671 pandecode_indent--;
672 pandecode_log("},\n");
673 }
674
675 static struct pandecode_fbd
676 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment)
677 {
678 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
679 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
680
681 struct pandecode_fbd info = {
682 .has_extra = false,
683 .rt_count = 1
684 };
685
686 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
687 pandecode_indent++;
688
689 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
690 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
691
692 pandecode_sfbd_format(s->format);
693
694 info.width = s->width + 1;
695 info.height = s->height + 1;
696
697 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
698 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
699
700 MEMORY_PROP(s, checksum);
701
702 if (s->checksum_stride)
703 pandecode_prop("checksum_stride = %d", s->checksum_stride);
704
705 MEMORY_PROP(s, framebuffer);
706 pandecode_prop("stride = %d", s->stride);
707
708 /* Earlier in the actual commandstream -- right before width -- but we
709 * delay to flow nicer */
710
711 pandecode_log(".clear_flags = ");
712 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
713 pandecode_log_cont(",\n");
714
715 if (s->depth_buffer) {
716 MEMORY_PROP(s, depth_buffer);
717 pandecode_prop("depth_stride = %d", s->depth_stride);
718 }
719
720 if (s->stencil_buffer) {
721 MEMORY_PROP(s, stencil_buffer);
722 pandecode_prop("stencil_stride = %d", s->stencil_stride);
723 }
724
725 if (s->depth_stride_zero ||
726 s->stencil_stride_zero ||
727 s->zero7 || s->zero8) {
728 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
729 pandecode_prop("depth_stride_zero = 0x%x",
730 s->depth_stride_zero);
731 pandecode_prop("stencil_stride_zero = 0x%x",
732 s->stencil_stride_zero);
733 pandecode_prop("zero7 = 0x%" PRIx32,
734 s->zero7);
735 pandecode_prop("zero8 = 0x%" PRIx32,
736 s->zero8);
737 }
738
739 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
740 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
741 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
742 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
743 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
744 }
745
746 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
747 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
748 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
749 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
750 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
751 }
752
753 if (s->clear_stencil) {
754 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
755 }
756
757 MEMORY_PROP(s, unknown_address_0);
758 const struct midgard_tiler_descriptor t = s->tiler;
759 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment);
760
761 pandecode_indent--;
762 pandecode_log("};\n");
763
764 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
765 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
766 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
767 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
768 pandecode_prop("zero5 = 0x%" PRIx32, s->zero4);
769
770 printf(".zero3 = {");
771
772 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
773 printf("%X, ", s->zero3[i]);
774
775 printf("},\n");
776
777 printf(".zero6 = {");
778
779 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
780 printf("%X, ", s->zero6[i]);
781
782 printf("},\n");
783
784 return info;
785 }
786
787 static void
788 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
789 {
790 pandecode_log(".unknown%d = {", name);
791
792 for (int i = 0; i < count; ++i)
793 printf("%X, ", slide[i]);
794
795 pandecode_log("},\n");
796 }
797
798 #define SHORT_SLIDE(num) \
799 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
800
801 static void
802 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
803 {
804 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
805 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
806
807 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
808 pandecode_indent++;
809
810 SHORT_SLIDE(1);
811
812 pandecode_indent--;
813 printf("},\n");
814 }
815
816 /* Extracts the number of components associated with a Mali format */
817
818 static unsigned
819 pandecode_format_component_count(enum mali_format fmt)
820 {
821 /* Mask out the format class */
822 unsigned top = fmt & 0b11100000;
823
824 switch (top) {
825 case MALI_FORMAT_SNORM:
826 case MALI_FORMAT_UINT:
827 case MALI_FORMAT_UNORM:
828 case MALI_FORMAT_SINT:
829 return ((fmt >> 3) & 3) + 1;
830 default:
831 /* TODO: Validate */
832 return 4;
833 }
834 }
835
836 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
837
838 static unsigned
839 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
840 {
841 unsigned mask = 0;
842 assert(MALI_CHANNEL_RED == 0);
843
844 for (unsigned c = 0; c < 4; ++c) {
845 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
846
847 if (chan <= MALI_CHANNEL_ALPHA)
848 mask |= (1 << chan);
849 }
850
851 return mask;
852 }
853
854 /* Validates that a (format, swizzle) pair is valid, in the sense that the
855 * swizzle doesn't access any components that are undefined in the format.
856 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
857 * omitted */
858
859 static bool
860 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
861 {
862 unsigned nr_comp = pandecode_format_component_count(fmt);
863 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
864 unsigned valid_mask = (1 << nr_comp) - 1;
865 unsigned invalid_mask = ~valid_mask;
866
867 if (access_mask & invalid_mask) {
868 pandecode_msg("XXX: invalid components accessed\n");
869 return false;
870 }
871
872 /* Check for the default non-swizzling swizzle so we can suppress
873 * useless printing for the defaults */
874
875 unsigned default_swizzles[4] = {
876 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
877 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
878 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
879 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
880 };
881
882 return (swizzle == default_swizzles[nr_comp - 1]);
883 }
884
885 /* Maps MALI_RGBA32F to rgba32f, etc */
886
887 static void
888 pandecode_format_short(enum mali_format fmt, bool srgb)
889 {
890 /* We want a type-like format, so cut off the initial MALI_ */
891 char *format = pandecode_format(fmt);
892 format += strlen("MALI_");
893
894 unsigned len = strlen(format);
895 char *lower_format = calloc(1, len + 1);
896
897 for (unsigned i = 0; i < len; ++i)
898 lower_format[i] = tolower(format[i]);
899
900 /* Sanity check sRGB flag is applied to RGB, per the name */
901 if (srgb && lower_format[0] != 'r')
902 pandecode_msg("XXX: sRGB applied to non-colour format\n");
903
904 /* Just prefix with an s, so you get formats like srgba8_unorm */
905 if (srgb)
906 pandecode_log_cont("s");
907
908 pandecode_log_cont("%s", lower_format);
909 free(lower_format);
910 }
911
912 static void
913 pandecode_swizzle(unsigned swizzle, enum mali_format format)
914 {
915 /* First, do some validation */
916 bool trivial_swizzle = pandecode_validate_format_swizzle(
917 format, swizzle);
918
919 if (trivial_swizzle)
920 return;
921
922 /* Next, print the swizzle */
923 pandecode_log_cont(".");
924
925 static const char components[] = "rgba01";
926
927 for (unsigned c = 0; c < 4; ++c) {
928 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
929
930 if (chan >= MALI_CHANNEL_RESERVED_0) {
931 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
932 continue;
933 }
934 pandecode_log_cont("%c", components[chan]);
935 }
936 }
937
938 static void
939 pandecode_rt_format(struct mali_rt_format format)
940 {
941 pandecode_log(".format = {\n");
942 pandecode_indent++;
943
944 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
945 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
946 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
947
948 pandecode_prop("block = %s", pandecode_block_format(format.block));
949
950 /* TODO: Map formats so we can check swizzles and print nicely */
951 pandecode_log("swizzle");
952 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
953 pandecode_log_cont(",\n");
954
955 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
956 MALI_NEGATIVE(format.nr_channels));
957
958 pandecode_log(".flags = ");
959 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
960 pandecode_log_cont(",\n");
961
962 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
963 * which is a faster hardware-based alternative to the wallpaper method
964 * to preserve framebuffer contents across frames. In practice, MFBD
965 * preload is buggy on Midgard, and so this is a chicken bit. If this
966 * bit isn't set, most likely something broke unrelated to preload */
967
968 if (!format.no_preload) {
969 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
970 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
971 }
972
973 if (format.zero)
974 pandecode_prop("zero = 0x%" PRIx32, format.zero);
975
976 pandecode_indent--;
977 pandecode_log("},\n");
978 }
979
980 static void
981 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
982 {
983 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
984 pandecode_indent++;
985
986 for (int i = 0; i < MALI_NEGATIVE(fb->rt_count_1); i++) {
987 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
988 struct pandecode_mapped_memory *mem =
989 pandecode_find_mapped_gpu_mem_containing(rt_va);
990 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
991
992 pandecode_log("{\n");
993 pandecode_indent++;
994
995 pandecode_rt_format(rt->format);
996
997 if (rt->format.block == MALI_BLOCK_AFBC) {
998 pandecode_log(".afbc = {\n");
999 pandecode_indent++;
1000
1001 char *a = pointer_as_memory_reference(rt->afbc.metadata);
1002 pandecode_prop("metadata = %s", a);
1003 free(a);
1004
1005 pandecode_prop("stride = %d", rt->afbc.stride);
1006 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
1007
1008 pandecode_indent--;
1009 pandecode_log("},\n");
1010 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
1011 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1012 rt->afbc.metadata,
1013 rt->afbc.stride,
1014 rt->afbc.unk);
1015 }
1016
1017 MEMORY_PROP(rt, framebuffer);
1018 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1019
1020 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1021 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1022 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1023 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1024 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1025 }
1026
1027 if (rt->zero1 || rt->zero2 || rt->zero3) {
1028 pandecode_msg("XXX: render target zeros tripped\n");
1029 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1030 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1031 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
1032 }
1033
1034 pandecode_indent--;
1035 pandecode_log("},\n");
1036 }
1037
1038 pandecode_indent--;
1039 pandecode_log("};\n");
1040 }
1041
1042 static struct pandecode_fbd
1043 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment)
1044 {
1045 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1046 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1047
1048 struct pandecode_fbd info;
1049
1050 if (fb->sample_locations) {
1051 /* The blob stores all possible sample locations in a single buffer
1052 * allocated on startup, and just switches the pointer when switching
1053 * MSAA state. For now, we just put the data into the cmdstream, but we
1054 * should do something like what the blob does with a real driver.
1055 *
1056 * There seem to be 32 slots for sample locations, followed by another
1057 * 16. The second 16 is just the center location followed by 15 zeros
1058 * in all the cases I've identified (maybe shader vs. depth/color
1059 * samples?).
1060 */
1061
1062 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
1063
1064 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
1065
1066 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
1067 pandecode_indent++;
1068
1069 for (int i = 0; i < 32 + 16; i++) {
1070 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1071 }
1072
1073 pandecode_indent--;
1074 pandecode_log("};\n");
1075 }
1076
1077 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1078 pandecode_indent++;
1079
1080 pandecode_prop("unk0 = 0x%x", fb->unk0);
1081
1082 if (fb->sample_locations)
1083 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1084
1085 /* Assume that unknown1 was emitted in the last job for
1086 * now */
1087 MEMORY_PROP(fb, unknown1);
1088
1089 info.width = fb->width1 + 1;
1090 info.height = fb->height1 + 1;
1091 info.rt_count = fb->rt_count_1 + 1;
1092
1093 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1094 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1095 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1096 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1097
1098 pandecode_prop("unk1 = 0x%x", fb->unk1);
1099 pandecode_prop("unk2 = 0x%x", fb->unk2);
1100 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1101 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1102
1103 pandecode_log(".mfbd_flags = ");
1104 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1105 pandecode_log_cont(",\n");
1106
1107 if (fb->clear_stencil)
1108 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1109
1110 if (fb->clear_depth)
1111 pandecode_prop("clear_depth = %f", fb->clear_depth);
1112
1113 /* TODO: What is this? Let's not blow up.. */
1114 if (fb->unknown2 != 0x1F)
1115 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1116
1117 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1118 MEMORY_PROP(fb, scratchpad);
1119 const struct midgard_tiler_descriptor t = fb->tiler;
1120 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment);
1121
1122 if (fb->zero3 || fb->zero4) {
1123 pandecode_msg("XXX: framebuffer zeros tripped\n");
1124 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1125 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1126 }
1127
1128 pandecode_indent--;
1129 pandecode_log("};\n");
1130
1131 gpu_va += sizeof(struct bifrost_framebuffer);
1132
1133 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1134
1135 if (info.has_extra) {
1136 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1137 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1138
1139 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1140 pandecode_indent++;
1141
1142 MEMORY_PROP(fbx, checksum);
1143
1144 if (fbx->checksum_stride)
1145 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1146
1147 pandecode_log(".flags = ");
1148 pandecode_log_decoded_flags(mfbd_extra_flag_info, fbx->flags);
1149 pandecode_log_cont(",\n");
1150
1151 if (fbx->flags & MALI_EXTRA_AFBC_ZS) {
1152 pandecode_log(".ds_afbc = {\n");
1153 pandecode_indent++;
1154
1155 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1156 pandecode_prop("depth_stencil_afbc_stride = %d",
1157 fbx->ds_afbc.depth_stencil_afbc_stride);
1158 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1159
1160 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
1161 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1162 pandecode_prop("zero1 = 0x%" PRIx32,
1163 fbx->ds_afbc.zero1);
1164 pandecode_prop("padding = 0x%" PRIx64,
1165 fbx->ds_afbc.padding);
1166 }
1167
1168 pandecode_indent--;
1169 pandecode_log("},\n");
1170 } else {
1171 pandecode_log(".ds_linear = {\n");
1172 pandecode_indent++;
1173
1174 if (fbx->ds_linear.depth) {
1175 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1176 pandecode_prop("depth_stride = %d",
1177 fbx->ds_linear.depth_stride);
1178 }
1179
1180 if (fbx->ds_linear.stencil) {
1181 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1182 pandecode_prop("stencil_stride = %d",
1183 fbx->ds_linear.stencil_stride);
1184 }
1185
1186 if (fbx->ds_linear.depth_stride_zero ||
1187 fbx->ds_linear.stencil_stride_zero ||
1188 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1189 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1190 pandecode_prop("depth_stride_zero = 0x%x",
1191 fbx->ds_linear.depth_stride_zero);
1192 pandecode_prop("stencil_stride_zero = 0x%x",
1193 fbx->ds_linear.stencil_stride_zero);
1194 pandecode_prop("zero1 = 0x%" PRIx32,
1195 fbx->ds_linear.zero1);
1196 pandecode_prop("zero2 = 0x%" PRIx32,
1197 fbx->ds_linear.zero2);
1198 }
1199
1200 pandecode_indent--;
1201 pandecode_log("},\n");
1202 }
1203
1204 if (fbx->zero3 || fbx->zero4) {
1205 pandecode_msg("XXX: fb_extra zeros tripped\n");
1206 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1207 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1208 }
1209
1210 pandecode_indent--;
1211 pandecode_log("};\n");
1212
1213 gpu_va += sizeof(struct bifrost_fb_extra);
1214 }
1215
1216 if (is_fragment)
1217 pandecode_render_target(gpu_va, job_no, fb);
1218
1219 return info;
1220 }
1221
1222 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1223 * count */
1224
1225 static void
1226 pandecode_padded_vertices(unsigned shift, unsigned k)
1227 {
1228 unsigned odd = 2*k + 1;
1229 unsigned pot = 1 << shift;
1230 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1231 }
1232
1233 /* Given a magic divisor, recover what we were trying to divide by.
1234 *
1235 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1236 * 0 <= m < 2^N, for N bits in m.
1237 *
1238 * Let q represent the number we would like to divide by.
1239 *
1240 * By definition of a magic divisor for N-bit unsigned integers (a number you
1241 * multiply by to magically get division), m is a number such that:
1242 *
1243 * (m * x) & (2^N - 1) = floor(x/q).
1244 * for all x on Z where 0 <= x < 2^N
1245 *
1246 * Ignore the case where any of the above values equals zero; it is irrelevant
1247 * for our purposes (instanced arrays).
1248 *
1249 * Choose x = q. Then:
1250 *
1251 * (m * x) & (2^N - 1) = floor(x/q).
1252 * (m * q) & (2^N - 1) = floor(q/q).
1253 *
1254 * floor(q/q) = floor(1) = 1, therefore:
1255 *
1256 * (m * q) & (2^N - 1) = 1
1257 *
1258 * Recall the identity that the bitwise AND of one less than a power-of-two
1259 * equals the modulo with that power of two, i.e. for all x:
1260 *
1261 * x & (2^N - 1) = x % N
1262 *
1263 * Therefore:
1264 *
1265 * mq % (2^N) = 1
1266 *
1267 * By definition, a modular multiplicative inverse of a number m is the number
1268 * q such that with respect to a modulos M:
1269 *
1270 * mq % M = 1
1271 *
1272 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1273 *
1274 */
1275
1276 static void
1277 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1278 {
1279 #if 0
1280 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1281 * shift) the most lame way possible... just repeatedly add.
1282 * Asymptoptically slow but nobody cares in practice, unless you have
1283 * massive numbers of vertices or high divisors. */
1284
1285 unsigned inverse = 0;
1286
1287 /* Magic implicitly has the highest bit set */
1288 magic |= (1 << 31);
1289
1290 /* Depending on rounding direction */
1291 if (extra)
1292 magic++;
1293
1294 for (;;) {
1295 uint32_t product = magic * inverse;
1296
1297 if (shift) {
1298 product >>= shift;
1299 }
1300
1301 if (product == 1)
1302 break;
1303
1304 ++inverse;
1305 }
1306
1307 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1308
1309 /* Recall we're supposed to divide by (gl_level_divisor *
1310 * padded_num_vertices) */
1311
1312 unsigned padded_num_vertices = inverse / orig_divisor;
1313
1314 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1315 #endif
1316 }
1317
1318 static void
1319 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1320 mali_ptr addr, int job_no, char *suffix,
1321 int count, bool varying, enum mali_job_type job_type)
1322 {
1323 char *prefix = varying ? "varying" : "attribute";
1324 assert(addr);
1325
1326 if (!count) {
1327 pandecode_msg("warn: No %s records\n", prefix);
1328 return;
1329 }
1330
1331 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1332
1333 for (int i = 0; i < count; ++i) {
1334 /* First, check for special records */
1335 if (attr[i].elements < MALI_VARYING_SPECIAL) {
1336 /* Special records are always varyings */
1337
1338 if (!varying)
1339 pandecode_msg("XXX: Special varying in attribute field\n");
1340
1341 if (job_type != JOB_TYPE_TILER)
1342 pandecode_msg("XXX: Special varying in non-FS\n");
1343
1344 /* We're special, so all fields should be zero */
1345 unsigned zero = attr[i].stride | attr[i].size;
1346 zero |= attr[i].shift | attr[i].extra_flags;
1347
1348 if (zero)
1349 pandecode_msg("XXX: Special varying has non-zero fields\n");
1350 else {
1351 /* Print the special varying name */
1352 pandecode_log("varying_%d = %s;", i, pandecode_special_varying(attr[i].elements));
1353 continue;
1354 }
1355 }
1356
1357 enum mali_attr_mode mode = attr[i].elements & 7;
1358
1359 if (mode == MALI_ATTR_UNUSED)
1360 pandecode_msg("XXX: unused attribute record\n");
1361
1362 /* For non-linear records, we need to print the type of record */
1363 if (mode != MALI_ATTR_LINEAR)
1364 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1365
1366 /* Print the name to link with attr_meta */
1367 pandecode_log_cont("%s_%d", prefix, i);
1368
1369 /* Print the stride and size */
1370 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1371
1372 /* TODO: Sanity check the quotient itself. It must be equal to
1373 * (or be greater than, if the driver added padding) the padded
1374 * vertex count. */
1375
1376 /* Finally, print the pointer */
1377 mali_ptr raw_elements = attr[i].elements & ~7;
1378 char *a = pointer_as_memory_reference(raw_elements);
1379 pandecode_log_cont(" = (%s);\n", a);
1380 free(a);
1381
1382 /* Check the pointer */
1383 pandecode_validate_buffer(raw_elements, attr[i].size);
1384
1385 /* shift/extra_flags exist only for instanced */
1386 if (attr[i].shift | attr[i].extra_flags) {
1387 /* These are set to random values by the blob for
1388 * varyings, most likely a symptom of uninitialized
1389 * memory where the hardware masked the bug. As such we
1390 * put this at a warning, not an error. */
1391
1392 if (mode == MALI_ATTR_LINEAR)
1393 pandecode_msg("warn: instancing fields set for linear\n");
1394
1395 pandecode_prop("shift = %d", attr[i].shift);
1396 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1397 }
1398
1399 /* Decode further where possible */
1400
1401 if (mode == MALI_ATTR_MODULO) {
1402 pandecode_padded_vertices(
1403 attr[i].shift,
1404 attr[i].extra_flags);
1405 }
1406
1407 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1408 i++;
1409 pandecode_log("{\n");
1410 pandecode_indent++;
1411 pandecode_prop("unk = 0x%x", attr[i].unk);
1412 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1413 if (attr[i].zero != 0)
1414 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1415 pandecode_prop("divisor = %d", attr[i].divisor);
1416 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1417 pandecode_indent--;
1418 pandecode_log("}, \n");
1419 }
1420
1421 }
1422
1423 pandecode_log("\n");
1424 }
1425
1426 static mali_ptr
1427 pandecode_shader_address(const char *name, mali_ptr ptr)
1428 {
1429 /* TODO: Decode flags */
1430 mali_ptr shader_ptr = ptr & ~15;
1431
1432 char *a = pointer_as_memory_reference(shader_ptr);
1433 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1434 free(a);
1435
1436 return shader_ptr;
1437 }
1438
1439 static void
1440 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1441 {
1442 unsigned any_nonzero =
1443 stencil->ref | stencil->mask | stencil->func |
1444 stencil->sfail | stencil->dpfail | stencil->dppass;
1445
1446 if (any_nonzero == 0)
1447 return;
1448
1449 const char *func = pandecode_func(stencil->func);
1450 const char *sfail = pandecode_stencil_op(stencil->sfail);
1451 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1452 const char *dppass = pandecode_stencil_op(stencil->dppass);
1453
1454 if (stencil->zero)
1455 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1456
1457 pandecode_log(".stencil_%s = {\n", name);
1458 pandecode_indent++;
1459 pandecode_prop("ref = %d", stencil->ref);
1460 pandecode_prop("mask = 0x%02X", stencil->mask);
1461 pandecode_prop("func = %s", func);
1462 pandecode_prop("sfail = %s", sfail);
1463 pandecode_prop("dpfail = %s", dpfail);
1464 pandecode_prop("dppass = %s", dppass);
1465 pandecode_indent--;
1466 pandecode_log("},\n");
1467 }
1468
1469 static void
1470 pandecode_blend_equation(const struct mali_blend_equation *blend)
1471 {
1472 if (blend->zero1)
1473 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1474
1475 pandecode_log(".equation = {\n");
1476 pandecode_indent++;
1477
1478 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1479 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1480
1481 pandecode_log(".color_mask = ");
1482 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1483 pandecode_log_cont(",\n");
1484
1485 pandecode_indent--;
1486 pandecode_log("},\n");
1487 }
1488
1489 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1490
1491 static unsigned
1492 decode_bifrost_constant(u16 constant)
1493 {
1494 float lo = (float) (constant & 0xFF);
1495 float hi = (float) (constant >> 8);
1496
1497 return (hi / 255.0) + (lo / 65535.0);
1498 }
1499
1500 static mali_ptr
1501 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1502 {
1503 struct bifrost_blend_rt *b =
1504 ((struct bifrost_blend_rt *) descs) + rt_no;
1505
1506 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1507 pandecode_indent++;
1508
1509 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1510 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1511 b->constant, decode_bifrost_constant(b->constant));
1512
1513 /* TODO figure out blend shader enable bit */
1514 pandecode_blend_equation(&b->equation);
1515 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1516 pandecode_prop("index = 0x%" PRIx16, b->index);
1517 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1518
1519 pandecode_indent--;
1520 pandecode_log("},\n");
1521
1522 return 0;
1523 }
1524
1525 static mali_ptr
1526 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1527 {
1528 /* constant/equation is in a union */
1529 if (!blend->shader)
1530 return 0;
1531
1532 pandecode_log(".blend = {\n");
1533 pandecode_indent++;
1534
1535 if (is_shader) {
1536 pandecode_shader_address("shader", blend->shader);
1537 } else {
1538 pandecode_blend_equation(&blend->equation);
1539 pandecode_prop("constant = %f", blend->constant);
1540 }
1541
1542 pandecode_indent--;
1543 pandecode_log("},\n");
1544
1545 /* Return blend shader to disassemble if present */
1546 return is_shader ? (blend->shader & ~0xF) : 0;
1547 }
1548
1549 static mali_ptr
1550 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1551 {
1552 struct midgard_blend_rt *b =
1553 ((struct midgard_blend_rt *) descs) + rt_no;
1554
1555 /* Flags determine presence of blend shader */
1556 bool is_shader = (b->flags & 0xF) >= 0x2;
1557
1558 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1559 pandecode_indent++;
1560
1561 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1562
1563 union midgard_blend blend = b->blend;
1564 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1565
1566 pandecode_indent--;
1567 pandecode_log("};\n");
1568
1569 return shader;
1570 }
1571
1572 /* Attributes and varyings have descriptor records, which contain information
1573 * about their format and ordering with the attribute/varying buffers. We'll
1574 * want to validate that the combinations specified are self-consistent.
1575 */
1576
1577 static int
1578 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1579 {
1580 char base[128];
1581 char *prefix = varying ? "varying" : "attribute";
1582 unsigned max_index = 0;
1583 snprintf(base, sizeof(base), "%s_meta", prefix);
1584
1585 struct mali_attr_meta *attr_meta;
1586 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1587
1588 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1589
1590 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1591 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1592 sizeof(*attr_mem));
1593
1594 /* If the record is discard, it should be zero for everything else */
1595
1596 if (attr_meta->format == MALI_VARYING_DISCARD) {
1597 uint64_t zero =
1598 attr_meta->index |
1599 attr_meta->unknown1 |
1600 attr_meta->unknown3 |
1601 attr_meta->src_offset;
1602
1603 if (zero)
1604 pandecode_msg("XXX: expected empty record for varying discard\n");
1605
1606 /* We want to look for a literal 0000 swizzle -- this
1607 * is not encoded with all zeroes, however */
1608
1609 enum mali_channel z = MALI_CHANNEL_ZERO;
1610 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1611 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1612
1613 if (!good_swizzle)
1614 pandecode_msg("XXX: expected zero swizzle for discard\n");
1615
1616 if (!varying)
1617 pandecode_msg("XXX: cannot discard attribute\n");
1618
1619 /* If we're all good, omit the record */
1620 if (!zero && varying && good_swizzle) {
1621 pandecode_log("/* discarded varying */\n");
1622 continue;
1623 }
1624 }
1625
1626 if (attr_meta->index > max_index)
1627 max_index = attr_meta->index;
1628
1629 if (attr_meta->unknown1 != 0x2) {
1630 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1631 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1632 }
1633
1634 if (attr_meta->unknown3) {
1635 pandecode_msg("XXX: unexpected unknown3 set\n");
1636 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1637 }
1638
1639 pandecode_format_short(attr_meta->format, false);
1640 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1641
1642 if (attr_meta->src_offset)
1643 pandecode_log_cont("[%u]", attr_meta->src_offset);
1644
1645 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1646
1647 pandecode_log_cont(";\n");
1648 }
1649
1650 pandecode_log("\n");
1651
1652 return count ? (max_index + 1) : 0;
1653 }
1654
1655 /* return bits [lo, hi) of word */
1656 static u32
1657 bits(u32 word, u32 lo, u32 hi)
1658 {
1659 if (hi - lo >= 32)
1660 return word; // avoid undefined behavior with the shift
1661
1662 return (word >> lo) & ((1 << (hi - lo)) - 1);
1663 }
1664
1665 static void
1666 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool noninstanced)
1667 {
1668 pandecode_log_cont("{\n");
1669 pandecode_indent++;
1670
1671 /* Decode invocation_count. See the comment before the definition of
1672 * invocation_count for an explanation.
1673 */
1674
1675 unsigned size_x = bits(p->invocation_count, 0, p->size_y_shift) + 1;
1676 unsigned size_y = bits(p->invocation_count, p->size_y_shift, p->size_z_shift) + 1;
1677 unsigned size_z = bits(p->invocation_count, p->size_z_shift, p->workgroups_x_shift) + 1;
1678
1679 unsigned groups_x = bits(p->invocation_count, p->workgroups_x_shift, p->workgroups_y_shift) + 1;
1680 unsigned groups_y = bits(p->invocation_count, p->workgroups_y_shift, p->workgroups_z_shift) + 1;
1681 unsigned groups_z = bits(p->invocation_count, p->workgroups_z_shift, 32) + 1;
1682
1683 /* Even though we have this decoded, we want to ensure that the
1684 * representation is "unique" so we don't lose anything by printing only
1685 * the final result. More specifically, we need to check that we were
1686 * passed something in canonical form, since the definition per the
1687 * hardware is inherently not unique. How? Well, take the resulting
1688 * decode and pack it ourselves! If it is bit exact with what we
1689 * decoded, we're good to go. */
1690
1691 struct mali_vertex_tiler_prefix ref;
1692 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, noninstanced);
1693
1694 bool canonical =
1695 (p->invocation_count == ref.invocation_count) &&
1696 (p->size_y_shift == ref.size_y_shift) &&
1697 (p->size_z_shift == ref.size_z_shift) &&
1698 (p->workgroups_x_shift == ref.workgroups_x_shift) &&
1699 (p->workgroups_y_shift == ref.workgroups_y_shift) &&
1700 (p->workgroups_z_shift == ref.workgroups_z_shift) &&
1701 (p->workgroups_x_shift_2 == ref.workgroups_x_shift_2);
1702
1703 if (!canonical) {
1704 pandecode_msg("XXX: non-canonical workgroups packing\n");
1705 pandecode_msg("expected: %X, %d, %d, %d, %d, %d, %d\n",
1706 ref.invocation_count,
1707 ref.size_y_shift,
1708 ref.size_z_shift,
1709 ref.workgroups_x_shift,
1710 ref.workgroups_y_shift,
1711 ref.workgroups_z_shift,
1712 ref.workgroups_x_shift_2);
1713
1714 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1715 pandecode_prop("size_y_shift = %d", p->size_y_shift);
1716 pandecode_prop("size_z_shift = %d", p->size_z_shift);
1717 pandecode_prop("workgroups_x_shift = %d", p->workgroups_x_shift);
1718 pandecode_prop("workgroups_y_shift = %d", p->workgroups_y_shift);
1719 pandecode_prop("workgroups_z_shift = %d", p->workgroups_z_shift);
1720 pandecode_prop("workgroups_x_shift_2 = %d", p->workgroups_x_shift_2);
1721 }
1722
1723 /* Regardless, print the decode */
1724 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1725 size_x, size_y, size_z,
1726 groups_x, groups_y, groups_z);
1727
1728 /* TODO: Decode */
1729 if (p->unknown_draw)
1730 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1731
1732 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1733
1734 if (p->draw_mode != MALI_DRAW_NONE)
1735 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1736
1737 /* Index count only exists for tiler jobs anyway */
1738
1739 if (p->index_count)
1740 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1741
1742
1743 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1744 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1745
1746 /* Validate an index buffer is present if we need one. TODO: verify
1747 * relationship between invocation_count and index_count */
1748
1749 if (p->indices) {
1750 unsigned count = p->index_count;
1751
1752 /* Grab the size */
1753 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1754
1755 /* Ensure we got a size, and if so, validate the index buffer
1756 * is large enough to hold a full set of indices of the given
1757 * size */
1758
1759 if (!index_raw_size)
1760 pandecode_msg("XXX: index size missing\n");
1761 else
1762 pandecode_validate_buffer(p->indices, count * size);
1763 } else if (index_raw_size)
1764 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1765
1766 if (p->offset_bias_correction)
1767 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1768
1769 /* TODO: Figure out what this is. It's not zero */
1770 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1771
1772 pandecode_indent--;
1773 pandecode_log("},\n");
1774 }
1775
1776 static void
1777 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1778 {
1779 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1780 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1781
1782 for (int i = 0; i < ubufs_count; i++) {
1783 unsigned size = (ubufs[i].size + 1) * 16;
1784 mali_ptr addr = ubufs[i].ptr << 2;
1785
1786 pandecode_validate_buffer(addr, size);
1787
1788 char *ptr = pointer_as_memory_reference(ubufs[i].ptr << 2);
1789 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1790 free(ptr);
1791 }
1792
1793 pandecode_log("\n");
1794 }
1795
1796 static void
1797 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1798 {
1799 pandecode_validate_buffer(uniforms, uniform_count * 16);
1800
1801 char *ptr = pointer_as_memory_reference(uniforms);
1802 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1803 free(ptr);
1804 }
1805
1806 static void
1807 pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1808 {
1809
1810 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1811
1812 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1813
1814 if (scratchpad->zero) {
1815 pandecode_msg("XXX: scratchpad zero tripped");
1816 pandecode_prop("zero = 0x%x\n", scratchpad->zero);
1817 }
1818
1819 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1820 pandecode_indent++;
1821
1822 pandecode_prop("flags = 0x%x", scratchpad->flags);
1823 MEMORY_PROP(scratchpad, gpu_scratchpad);
1824
1825 pandecode_indent--;
1826 pandecode_log("};\n");
1827 }
1828
1829 static unsigned shader_id = 0;
1830
1831 static struct midgard_disasm_stats
1832 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1833 bool is_bifrost)
1834 {
1835 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1836 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1837
1838 /* Compute maximum possible size */
1839 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1840
1841 /* Print some boilerplate to clearly denote the assembly (which doesn't
1842 * obey indentation rules), and actually do the disassembly! */
1843
1844 printf("\n\n");
1845
1846 struct midgard_disasm_stats stats;
1847
1848 if (is_bifrost) {
1849 disassemble_bifrost(code, sz, false);
1850
1851 /* TODO: Extend stats to Bifrost */
1852 stats.texture_count = -128;
1853 stats.sampler_count = -128;
1854 stats.attribute_count = -128;
1855 stats.varying_count = -128;
1856 stats.uniform_count = -128;
1857 stats.uniform_buffer_count = -128;
1858 stats.work_count = -128;
1859
1860 stats.instruction_count = 0;
1861 stats.bundle_count = 0;
1862 stats.quadword_count = 0;
1863 stats.helper_invocations = false;
1864 } else {
1865 stats = disassemble_midgard(code, sz);
1866 }
1867
1868 /* Print shader-db stats */
1869
1870 unsigned nr_threads =
1871 (stats.work_count <= 4) ? 4 :
1872 (stats.work_count <= 8) ? 2 :
1873 1;
1874
1875 printf("shader%d - %s shader: "
1876 "%u inst, %u bundles, %u quadwords, "
1877 "%u registers, %u threads, 0 loops\n\n\n",
1878 shader_id++,
1879 (type == JOB_TYPE_TILER) ? "FRAGMENT" : "VERTEX",
1880 stats.instruction_count, stats.bundle_count, stats.quadword_count,
1881 stats.work_count, nr_threads);
1882
1883
1884 return stats;
1885 }
1886
1887 static void
1888 pandecode_texture(mali_ptr u,
1889 struct pandecode_mapped_memory *tmem,
1890 unsigned job_no, unsigned tex)
1891 {
1892 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
1893
1894 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
1895 pandecode_indent++;
1896
1897 struct mali_texture_format f = t->format;
1898
1899 /* See the definiton of enum mali_texture_type */
1900
1901 bool is_cube = f.type == MALI_TEX_CUBE;
1902 unsigned dimension = is_cube ? 2 : f.type;
1903
1904 pandecode_make_indent();
1905
1906 /* TODO: Are there others? */
1907 bool is_zs = f.format == MALI_Z32_UNORM;
1908
1909 /* Recall Z/S switched the meaning of linear/tiled .. */
1910 if (is_zs && f.layout == MALI_TEXTURE_LINEAR)
1911 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1912
1913 /* Print the layout. Default is linear; a modifier can denote AFBC or
1914 * u-interleaved/tiled modes */
1915
1916 if (f.layout == MALI_TEXTURE_AFBC)
1917 pandecode_log_cont("afbc");
1918 else if (f.layout == MALI_TEXTURE_TILED)
1919 pandecode_log_cont(is_zs ? "linear" : "tiled");
1920 else if (f.layout == MALI_TEXTURE_LINEAR)
1921 pandecode_log_cont("linear");
1922 else
1923 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
1924
1925 pandecode_swizzle(t->swizzle, f.format);
1926 pandecode_log_cont(" ");
1927
1928 /* Distinguish cube/2D with modifier */
1929
1930 if (is_cube)
1931 pandecode_log_cont("cube ");
1932
1933 pandecode_format_short(f.format, f.srgb);
1934 pandecode_swizzle(f.swizzle, f.format);
1935
1936 /* All four width/height/depth/array_size dimensions are present
1937 * regardless of the type of texture, but it is an error to have
1938 * non-zero dimensions for unused dimensions. Verify this. array_size
1939 * can always be set, as can width. */
1940
1941 if (t->height && dimension < 2)
1942 pandecode_msg("XXX: nonzero height for <2D texture\n");
1943
1944 if (t->depth && dimension < 3)
1945 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1946
1947 /* Print only the dimensions that are actually there */
1948
1949 pandecode_log_cont(": %d", t->width + 1);
1950
1951 if (dimension >= 2)
1952 pandecode_log_cont("x%u", t->height + 1);
1953
1954 if (dimension >= 3)
1955 pandecode_log_cont("x%u", t->depth + 1);
1956
1957 if (t->array_size)
1958 pandecode_log_cont("[%u]", t->array_size + 1);
1959
1960 if (t->levels)
1961 pandecode_log_cont(" mip %u", t->levels);
1962
1963 pandecode_log_cont("\n");
1964
1965 if (f.unknown1 | f.zero) {
1966 pandecode_msg("XXX: texture format zero tripped\n");
1967 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
1968 pandecode_prop("zero = %" PRId32, f.zero);
1969 }
1970
1971 if (!f.unknown2) {
1972 pandecode_msg("XXX: expected unknown texture bit set\n");
1973 pandecode_prop("unknown2 = %" PRId32, f.unknown1);
1974 }
1975
1976 if (t->swizzle_zero) {
1977 pandecode_msg("XXX: swizzle zero tripped\n");
1978 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
1979 }
1980
1981 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
1982 pandecode_msg("XXX: texture zero tripped\n");
1983 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
1984 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
1985 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
1986 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
1987 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
1988 }
1989
1990 pandecode_log(".payload = {\n");
1991 pandecode_indent++;
1992
1993 /* A bunch of bitmap pointers follow.
1994 * We work out the correct number,
1995 * based on the mipmap/cubemap
1996 * properties, but dump extra
1997 * possibilities to futureproof */
1998
1999 int bitmap_count = MALI_NEGATIVE(t->levels);
2000
2001 /* Miptree for each face */
2002 if (f.type == MALI_TEX_CUBE)
2003 bitmap_count *= 6;
2004
2005 /* Array of textures */
2006 bitmap_count *= MALI_NEGATIVE(t->array_size);
2007
2008 /* Stride for each element */
2009 if (f.manual_stride)
2010 bitmap_count *= 2;
2011
2012 /* Sanity check the size */
2013 int max_count = sizeof(t->payload) / sizeof(t->payload[0]);
2014 assert (bitmap_count <= max_count);
2015
2016 for (int i = 0; i < bitmap_count; ++i) {
2017 /* How we dump depends if this is a stride or a pointer */
2018
2019 if (f.manual_stride && (i & 1)) {
2020 /* signed 32-bit snuck in as a 64-bit pointer */
2021 uint64_t stride_set = t->payload[i];
2022 uint32_t clamped_stride = stride_set;
2023 int32_t stride = clamped_stride;
2024 assert(stride_set == clamped_stride);
2025 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2026 } else {
2027 char *a = pointer_as_memory_reference(t->payload[i]);
2028 pandecode_log("%s, \n", a);
2029 free(a);
2030 }
2031 }
2032
2033 pandecode_indent--;
2034 pandecode_log("},\n");
2035
2036 pandecode_indent--;
2037 pandecode_log("};\n");
2038 }
2039
2040 /* 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 */
2041
2042 static void
2043 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2044 {
2045 /* Nothing to do */
2046 if (claim == truth)
2047 return;
2048
2049 if (fuzzy)
2050 assert(truth >= 0);
2051
2052 if ((truth >= 0) && !fuzzy) {
2053 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2054 (truth < claim) ? "warn" : "XXX",
2055 name, truth, claim);
2056 } else if ((claim > -truth) && !fuzzy) {
2057 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2058 name, -truth, claim);
2059 } else if (fuzzy && (claim < truth))
2060 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2061 name, truth, claim);
2062
2063 pandecode_log(".%s = %" PRId16, name, claim);
2064
2065 if (fuzzy)
2066 pandecode_log_cont(" /* %u used */", truth);
2067
2068 pandecode_log_cont(",\n");
2069 }
2070
2071 static void
2072 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2073 bool is_bifrost)
2074 {
2075 struct midgard_disasm_stats stats =
2076 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost);
2077
2078 bool has_texture = (stats.texture_count > 0);
2079 bool has_sampler = (stats.sampler_count > 0);
2080 bool has_attribute = (stats.attribute_count > 0);
2081 bool has_varying = (stats.varying_count > 0);
2082 bool has_uniform = (stats.uniform_count > 0);
2083 bool has_ubo = (stats.uniform_buffer_count > 0);
2084
2085 if (has_texture || has_sampler)
2086 pandecode_msg("XXX: blend shader accessing textures\n");
2087
2088 if (has_attribute || has_varying)
2089 pandecode_msg("XXX: blend shader accessing interstage\n");
2090
2091 if (has_uniform || has_ubo)
2092 pandecode_msg("XXX: blend shader accessing uniforms\n");
2093 }
2094
2095 static void
2096 pandecode_vertex_tiler_postfix_pre(
2097 const struct mali_vertex_tiler_postfix *p,
2098 int job_no, enum mali_job_type job_type,
2099 char *suffix, bool is_bifrost)
2100 {
2101 struct pandecode_mapped_memory *attr_mem;
2102
2103 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2104 * are the only things actually needed from the FBD, vertex/tiler jobs
2105 * no longer reference the FBD -- instead, this field points to some
2106 * info about the scratchpad.
2107 */
2108
2109 struct pandecode_fbd fbd_info = {
2110 /* Default for Bifrost */
2111 .rt_count = 1
2112 };
2113
2114 if (is_bifrost)
2115 pandecode_scratchpad(p->framebuffer & ~FBD_TYPE, job_no, suffix);
2116 else if (p->framebuffer & MALI_MFBD)
2117 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
2118 else if (job_type == JOB_TYPE_COMPUTE)
2119 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
2120 else
2121 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no, false);
2122
2123 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2124 int texture_count = 0, sampler_count = 0;
2125
2126 if (p->shader) {
2127 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2128 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2129
2130 /* Disassemble ahead-of-time to get stats. Initialize with
2131 * stats for the missing-shader case so we get validation
2132 * there, too */
2133
2134 struct midgard_disasm_stats info = {
2135 .texture_count = 0,
2136 .sampler_count = 0,
2137 .attribute_count = 0,
2138 .varying_count = 0,
2139 .work_count = 1,
2140
2141 .uniform_count = -128,
2142 .uniform_buffer_count = 0
2143 };
2144
2145 if (s->shader & ~0xF)
2146 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost);
2147
2148 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2149 pandecode_indent++;
2150
2151 /* Save for dumps */
2152 attribute_count = s->attribute_count;
2153 varying_count = s->varying_count;
2154 texture_count = s->texture_count;
2155 sampler_count = s->sampler_count;
2156
2157 if (is_bifrost) {
2158 uniform_count = s->bifrost2.uniform_count;
2159 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2160 } else {
2161 uniform_count = s->midgard1.uniform_count;
2162 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2163 }
2164
2165 pandecode_shader_address("shader", s->shader);
2166
2167 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2168 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2169 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2170 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2171 pandecode_shader_prop("uniform_buffer_count",
2172 uniform_buffer_count,
2173 info.uniform_buffer_count, true);
2174
2175 if (!is_bifrost) {
2176 pandecode_shader_prop("uniform_count",
2177 uniform_count,
2178 info.uniform_count, false);
2179
2180 pandecode_shader_prop("work_count",
2181 s->midgard1.work_count, info.work_count, false);
2182 }
2183
2184 if (is_bifrost) {
2185 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32, s->bifrost1.unk1);
2186 } else {
2187 bool helpers = s->midgard1.flags & MALI_HELPER_INVOCATIONS;
2188 s->midgard1.flags &= ~MALI_HELPER_INVOCATIONS;
2189
2190 if (helpers != info.helper_invocations) {
2191 pandecode_msg("XXX: expected helpers %u but got %u\n",
2192 info.helper_invocations, helpers);
2193 }
2194
2195 pandecode_log(".midgard1.flags = ");
2196 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
2197 pandecode_log_cont(",\n");
2198
2199 pandecode_prop("midgard1.unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
2200 }
2201
2202 if (s->depth_units || s->depth_factor) {
2203 pandecode_prop("depth_factor = %f", s->depth_factor);
2204 pandecode_prop("depth_units = %f", s->depth_units);
2205 }
2206
2207 if (s->alpha_coverage) {
2208 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
2209 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
2210
2211 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2212 invert_alpha_coverage ? "~" : "",
2213 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
2214 }
2215
2216 if (s->unknown2_3 || s->unknown2_4) {
2217 pandecode_log(".unknown2_3 = ");
2218
2219 int unknown2_3 = s->unknown2_3;
2220 int unknown2_4 = s->unknown2_4;
2221
2222 /* We're not quite sure what these flags mean without the depth test, if anything */
2223
2224 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2225 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2226 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2227
2228 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2229 }
2230
2231 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2232 pandecode_log_cont(",\n");
2233
2234 pandecode_log(".unknown2_4 = ");
2235 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2236 pandecode_log_cont(",\n");
2237 }
2238
2239 if (s->stencil_mask_front || s->stencil_mask_back) {
2240 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2241 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2242 }
2243
2244 pandecode_stencil("front", &s->stencil_front);
2245 pandecode_stencil("back", &s->stencil_back);
2246
2247 if (is_bifrost) {
2248 pandecode_log(".bifrost2 = {\n");
2249 pandecode_indent++;
2250
2251 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2252 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2253 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2254 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2255
2256 pandecode_indent--;
2257 pandecode_log("},\n");
2258 } else if (s->midgard2.unknown2_7) {
2259 pandecode_log(".midgard2 = {\n");
2260 pandecode_indent++;
2261
2262 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2263 pandecode_indent--;
2264 pandecode_log("},\n");
2265 }
2266
2267 if (s->unknown2_8)
2268 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
2269
2270 if (!is_bifrost) {
2271 /* TODO: Blend shaders routing/disasm */
2272 union midgard_blend blend = s->blend;
2273 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2274 if (shader & ~0xF)
2275 pandecode_blend_shader_disassemble(shader, job_no, job_type, false);
2276 }
2277
2278 pandecode_indent--;
2279 pandecode_log("};\n");
2280
2281 /* MRT blend fields are used whenever MFBD is used, with
2282 * per-RT descriptors */
2283
2284 if (job_type == JOB_TYPE_TILER) {
2285 void* blend_base = (void *) (s + 1);
2286
2287 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2288 mali_ptr shader = 0;
2289
2290 if (is_bifrost)
2291 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2292 else
2293 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2294
2295 if (shader & ~0xF)
2296 pandecode_blend_shader_disassemble(shader, job_no, job_type, false);
2297
2298 }
2299 }
2300 } else
2301 pandecode_msg("XXX: missing shader descriptor\n");
2302
2303 if (p->viewport) {
2304 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2305 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2306
2307 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2308 pandecode_indent++;
2309
2310 pandecode_prop("clip_minx = %f", f->clip_minx);
2311 pandecode_prop("clip_miny = %f", f->clip_miny);
2312 pandecode_prop("clip_minz = %f", f->clip_minz);
2313 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2314 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2315 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2316
2317 /* Only the higher coordinates are MALI_POSITIVE scaled */
2318
2319 pandecode_prop("viewport0 = { %d, %d }",
2320 f->viewport0[0], f->viewport0[1]);
2321
2322 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2323 f->viewport1[0] + 1, f->viewport1[1] + 1);
2324
2325 pandecode_indent--;
2326 pandecode_log("};\n");
2327 }
2328
2329 unsigned max_attr_index = 0;
2330
2331 if (p->attribute_meta)
2332 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2333
2334 if (p->attributes) {
2335 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2336 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2337 }
2338
2339 /* Varyings are encoded like attributes but not actually sent; we just
2340 * pass a zero buffer with the right stride/size set, (or whatever)
2341 * since the GPU will write to it itself */
2342
2343 if (p->varying_meta) {
2344 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2345 }
2346
2347 if (p->varyings) {
2348 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2349
2350 /* Number of descriptors depends on whether there are
2351 * non-internal varyings */
2352
2353 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2354 }
2355
2356 if (p->uniform_buffers) {
2357 if (uniform_buffer_count)
2358 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2359 else
2360 pandecode_msg("warn: UBOs specified but not referenced\n");
2361 } else if (uniform_buffer_count)
2362 pandecode_msg("XXX: UBOs referenced but not specified\n");
2363
2364 /* We don't want to actually dump uniforms, but we do need to validate
2365 * that the counts we were given are sane */
2366
2367 if (p->uniforms) {
2368 if (uniform_count)
2369 pandecode_uniforms(p->uniforms, uniform_count);
2370 else
2371 pandecode_msg("warn: Uniforms specified but not referenced\n");
2372 } else if (uniform_count)
2373 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2374
2375 if (p->texture_trampoline) {
2376 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
2377
2378 if (mmem) {
2379 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
2380
2381 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
2382 pandecode_indent++;
2383
2384 for (int tex = 0; tex < texture_count; ++tex) {
2385 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2386 char *a = pointer_as_memory_reference(*u);
2387 pandecode_log("%s,\n", a);
2388 free(a);
2389 }
2390
2391 pandecode_indent--;
2392 pandecode_log("};\n");
2393
2394 /* Now, finally, descend down into the texture descriptor */
2395 for (unsigned tex = 0; tex < texture_count; ++tex) {
2396 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2397 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2398 if (tmem)
2399 pandecode_texture(*u, tmem, job_no, tex);
2400 }
2401 }
2402 }
2403
2404 if (p->sampler_descriptor) {
2405 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2406
2407 if (smem) {
2408 struct mali_sampler_descriptor *s;
2409
2410 mali_ptr d = p->sampler_descriptor;
2411
2412 for (int i = 0; i < sampler_count; ++i) {
2413 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2414
2415 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2416 pandecode_indent++;
2417
2418 pandecode_log(".filter_mode = ");
2419 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2420 pandecode_log_cont(",\n");
2421
2422 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2423 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2424
2425 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2426 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2427 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2428
2429 pandecode_prop("compare_func = %s", pandecode_alt_func(s->compare_func));
2430
2431 if (s->zero || s->zero2) {
2432 pandecode_msg("XXX: sampler zero tripped\n");
2433 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2434 }
2435
2436 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2437
2438 pandecode_prop("border_color = { %f, %f, %f, %f }",
2439 s->border_color[0],
2440 s->border_color[1],
2441 s->border_color[2],
2442 s->border_color[3]);
2443
2444 pandecode_indent--;
2445 pandecode_log("};\n");
2446 }
2447 }
2448 }
2449 }
2450
2451 static void
2452 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2453 {
2454 if (p->shader & 0xF)
2455 pandecode_msg("warn: shader tagged %X\n", p->shader & 0xF);
2456
2457 if (!(p->position_varying || p->occlusion_counter))
2458 return;
2459
2460 pandecode_log(".postfix = {\n");
2461 pandecode_indent++;
2462
2463 MEMORY_PROP(p, position_varying);
2464 MEMORY_PROP(p, occlusion_counter);
2465
2466 pandecode_indent--;
2467 pandecode_log("},\n");
2468 }
2469
2470 static void
2471 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2472 {
2473 pandecode_log_cont("{\n");
2474 pandecode_indent++;
2475
2476 pandecode_prop("unk2 = 0x%x", v->unk2);
2477
2478 if (v->zero0 || v->zero1) {
2479 pandecode_msg("XXX: vertex only zero tripped");
2480 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2481 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2482 }
2483
2484 pandecode_indent--;
2485 pandecode_log("}\n");
2486 }
2487
2488 static void
2489 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2490 {
2491
2492 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2493 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2494
2495 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2496 pandecode_indent++;
2497
2498 if (h->zero) {
2499 pandecode_msg("XXX: tiler heap zero tripped\n");
2500 pandecode_prop("zero = 0x%x", h->zero);
2501 }
2502
2503 for (int i = 0; i < 12; i++) {
2504 if (h->zeros[i] != 0) {
2505 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2506 i, h->zeros[i]);
2507 }
2508 }
2509
2510 pandecode_prop("heap_size = 0x%x", h->heap_size);
2511 MEMORY_PROP(h, tiler_heap_start);
2512 MEMORY_PROP(h, tiler_heap_free);
2513
2514 /* this might point to the beginning of another buffer, when it's
2515 * really the end of the tiler heap buffer, so we have to be careful
2516 * here. but for zero length, we need the same pointer.
2517 */
2518
2519 if (h->tiler_heap_end == h->tiler_heap_start) {
2520 MEMORY_PROP(h, tiler_heap_start);
2521 } else {
2522 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2523 pandecode_prop("tiler_heap_end = %s + 1", a);
2524 free(a);
2525 }
2526
2527 pandecode_indent--;
2528 pandecode_log("};\n");
2529 }
2530
2531 static void
2532 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2533 {
2534 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2535 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2536
2537 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2538
2539 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2540 pandecode_indent++;
2541
2542 if (t->zero0 || t->zero1) {
2543 pandecode_msg("XXX: tiler meta zero tripped\n");
2544 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2545 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2546 }
2547
2548 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2549 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2550
2551 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2552 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2553
2554 for (int i = 0; i < 12; i++) {
2555 if (t->zeros[i] != 0) {
2556 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2557 i, t->zeros[i]);
2558 }
2559 }
2560
2561 pandecode_indent--;
2562 pandecode_log("};\n");
2563 }
2564
2565 static void
2566 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2567 {
2568 pandecode_log(".gl_enables = ");
2569
2570 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2571
2572 pandecode_log_cont(",\n");
2573 }
2574
2575 static void
2576 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2577 {
2578 if (u.pointer == 0x0)
2579 return;
2580
2581 pandecode_log(".primitive_size = {\n");
2582 pandecode_indent++;
2583
2584 if (constant) {
2585 pandecode_prop("constant = %f", u.constant);
2586 } else {
2587 MEMORY_PROP((&u), pointer);
2588 }
2589
2590 pandecode_indent--;
2591 pandecode_log("},\n");
2592 }
2593
2594 static void
2595 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2596 {
2597 pandecode_log_cont("{\n");
2598 pandecode_indent++;
2599
2600 /* TODO: gl_PointSize on Bifrost */
2601 pandecode_primitive_size(t->primitive_size, true);
2602
2603 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2604
2605 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2606 || t->zero6 || t->zero7 || t->zero8) {
2607 pandecode_msg("XXX: tiler only zero tripped\n");
2608 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2609 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2610 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2611 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2612 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2613 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2614 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2615 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2616 }
2617
2618 pandecode_indent--;
2619 pandecode_log("},\n");
2620 }
2621
2622 static int
2623 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2624 const struct pandecode_mapped_memory *mem,
2625 mali_ptr payload, int job_no)
2626 {
2627 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2628
2629 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true);
2630
2631 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2632 pandecode_indent++;
2633
2634 pandecode_log(".prefix = ");
2635 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2636
2637 pandecode_log(".vertex = ");
2638 pandecode_vertex_only_bfr(&v->vertex);
2639
2640 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2641
2642 pandecode_indent--;
2643 pandecode_log("};\n");
2644
2645 return sizeof(*v);
2646 }
2647
2648 static int
2649 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2650 const struct pandecode_mapped_memory *mem,
2651 mali_ptr payload, int job_no)
2652 {
2653 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2654
2655 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true);
2656 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2657
2658 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2659 pandecode_indent++;
2660
2661 pandecode_log(".prefix = ");
2662 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2663
2664 pandecode_log(".tiler = ");
2665 pandecode_tiler_only_bfr(&t->tiler, job_no);
2666
2667 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2668
2669 pandecode_indent--;
2670 pandecode_log("};\n");
2671
2672 return sizeof(*t);
2673 }
2674
2675 static int
2676 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2677 const struct pandecode_mapped_memory *mem,
2678 mali_ptr payload, int job_no)
2679 {
2680 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2681
2682 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false);
2683
2684 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2685 pandecode_indent++;
2686
2687 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2688 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2689
2690 bool instanced = v->instance_shift || v->instance_odd;
2691 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2692
2693 pandecode_log(".prefix = ");
2694 pandecode_vertex_tiler_prefix(&v->prefix, job_no, !instanced && is_graphics);
2695
2696 pandecode_gl_enables(v->gl_enables, h->job_type);
2697
2698 if (v->instance_shift || v->instance_odd) {
2699 pandecode_prop("instance_shift = 0x%d /* %d */",
2700 v->instance_shift, 1 << v->instance_shift);
2701 pandecode_prop("instance_odd = 0x%X /* %d */",
2702 v->instance_odd, (2 * v->instance_odd) + 1);
2703
2704 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2705 }
2706
2707 if (v->offset_start)
2708 pandecode_prop("offset_start = %d", v->offset_start);
2709
2710 if (v->zero5) {
2711 pandecode_msg("XXX: midgard payload zero tripped\n");
2712 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2713 }
2714
2715 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2716
2717 pandecode_indent--;
2718 pandecode_log("};\n");
2719
2720 return sizeof(*v);
2721 }
2722
2723 static int
2724 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2725 mali_ptr payload, int job_no,
2726 bool is_bifrost)
2727 {
2728 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2729
2730 bool is_mfbd = (s->framebuffer & FBD_TYPE) == MALI_MFBD;
2731
2732 /* Bifrost theoretically may retain support for SFBD on compute jobs,
2733 * but for graphics workloads with a FRAGMENT payload, use MFBD */
2734
2735 if (!is_mfbd && is_bifrost)
2736 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2737
2738 struct pandecode_fbd info;
2739
2740 if (is_mfbd)
2741 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2742 else
2743 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true);
2744
2745 /* Compute the tag for the tagged pointer. This contains the type of
2746 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2747 * additional structures follow the MFBD header (an extra payload or
2748 * not, as well as a count of render targets) */
2749
2750 unsigned expected_tag = is_mfbd ? MALI_MFBD : MALI_SFBD;
2751
2752 if (is_mfbd) {
2753 if (info.has_extra)
2754 expected_tag |= MALI_MFBD_TAG_EXTRA;
2755
2756 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
2757 }
2758
2759 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
2760 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2761 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
2762 pandecode_prop("max_tile_coord = 0x%X\n", s->min_tile_coord);
2763 }
2764
2765 /* Extract tile coordinates */
2766
2767 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
2768 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
2769
2770 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2771 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2772
2773 /* For the max, we also want the floored (rather than ceiled) version for checking */
2774
2775 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
2776 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
2777
2778 /* Validate the coordinates are well-ordered */
2779
2780 if (min_x == max_x)
2781 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2782 else if (min_x > max_x)
2783 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2784
2785 if (min_y == max_y)
2786 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2787 else if (min_y > max_y)
2788 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2789
2790 /* Validate the coordinates fit inside the framebuffer. We use floor,
2791 * rather than ceil, for the max coordinates, since the tile
2792 * coordinates for something like an 800x600 framebuffer will actually
2793 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2794
2795 if ((min_x > info.width) || (max_x_f > info.width))
2796 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2797
2798 if ((min_y > info.height) || (max_y_f > info.height))
2799 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2800
2801 /* After validation, we print */
2802
2803 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
2804
2805 /* The FBD is a tagged pointer */
2806
2807 unsigned tag = (s->framebuffer & ~FBD_MASK);
2808
2809 if (tag != expected_tag)
2810 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
2811
2812 return sizeof(*s);
2813 }
2814
2815 static int job_descriptor_number = 0;
2816
2817 int
2818 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost)
2819 {
2820 struct mali_job_descriptor_header *h;
2821
2822 int start_number = 0;
2823
2824 bool first = true;
2825 bool last_size;
2826
2827 do {
2828 struct pandecode_mapped_memory *mem =
2829 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2830
2831 void *payload;
2832
2833 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2834
2835 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2836 * high 32-bits of the 64-bit pointer are reused to store
2837 * something else.
2838 */
2839 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2840 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2841 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2842
2843 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
2844
2845 int job_no = job_descriptor_number++;
2846
2847 if (first)
2848 start_number = job_no;
2849
2850 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2851 pandecode_indent++;
2852
2853 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2854
2855 /* Save for next job fixing */
2856 last_size = h->job_descriptor_size;
2857
2858 if (h->job_descriptor_size)
2859 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2860
2861 if (h->exception_status && h->exception_status != 0x1)
2862 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2863 h->exception_status,
2864 (h->exception_status >> 16) & 0xFFFF,
2865 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2866 h->exception_status & 0xFF);
2867
2868 if (h->first_incomplete_task)
2869 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2870
2871 if (h->fault_pointer)
2872 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2873
2874 if (h->job_barrier)
2875 pandecode_prop("job_barrier = %d", h->job_barrier);
2876
2877 pandecode_prop("job_index = %d", h->job_index);
2878
2879 if (h->unknown_flags)
2880 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2881
2882 if (h->job_dependency_index_1)
2883 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2884
2885 if (h->job_dependency_index_2)
2886 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2887
2888 pandecode_indent--;
2889 pandecode_log("};\n");
2890
2891 /* Do not touch the field yet -- decode the payload first, and
2892 * don't touch that either. This is essential for the uploads
2893 * to occur in sequence and therefore be dynamically allocated
2894 * correctly. Do note the size, however, for that related
2895 * reason. */
2896
2897 switch (h->job_type) {
2898 case JOB_TYPE_SET_VALUE: {
2899 struct mali_payload_set_value *s = payload;
2900 pandecode_log("struct mali_payload_set_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2901 pandecode_indent++;
2902 MEMORY_PROP(s, out);
2903 pandecode_prop("unknown = 0x%" PRIX64, s->unknown);
2904 pandecode_indent--;
2905 pandecode_log("};\n");
2906
2907 break;
2908 }
2909
2910 case JOB_TYPE_TILER:
2911 case JOB_TYPE_VERTEX:
2912 case JOB_TYPE_COMPUTE:
2913 if (bifrost) {
2914 if (h->job_type == JOB_TYPE_TILER)
2915 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no);
2916 else
2917 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no);
2918 } else
2919 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no);
2920
2921 break;
2922
2923 case JOB_TYPE_FRAGMENT:
2924 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost);
2925 break;
2926
2927 default:
2928 break;
2929 }
2930
2931 /* Handle linkage */
2932
2933 if (!first) {
2934 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2935
2936 if (last_size)
2937 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no);
2938 else
2939 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no);
2940 }
2941
2942 first = false;
2943
2944 } while ((jc_gpu_va = h->job_descriptor_size ? h->next_job_64 : h->next_job_32));
2945
2946 return start_number;
2947 }