projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
panfrost: Pack invocation_shifts manually instead of a bit field
[mesa.git] / src / panfrost / pandecode / decode.c
1 /*
2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 */
25
26 #include <panfrost-job.h>
27 #include <stdio.h>
28 #include <stdlib.h>
29 #include <memory.h>
30 #include <stdbool.h>
31 #include <stdarg.h>
32 #include <ctype.h>
33 #include "decode.h"
34 #include "util/macros.h"
35 #include "util/u_math.h"
36
37 #include "pan_pretty_print.h"
38 #include "midgard/disassemble.h"
39 #include "bifrost/disassemble.h"
40
41 #include "pan_encoder.h"
42
43 static void pandecode_swizzle(unsigned swizzle, enum mali_format format);
44
45 #define MEMORY_PROP(obj, p) {\
46 if (obj->p) { \
47 char *a = pointer_as_memory_reference(obj->p); \
48 pandecode_prop("%s = %s", #p, a); \
49 free(a); \
50 } \
51 }
52
53 #define MEMORY_PROP_DIR(obj, p) {\
54 if (obj.p) { \
55 char *a = pointer_as_memory_reference(obj.p); \
56 pandecode_prop("%s = %s", #p, a); \
57 free(a); \
58 } \
59 }
60
61 /* Semantic logging type.
62 *
63 * Raw: for raw messages to be printed as is.
64 * Message: for helpful information to be commented out in replays.
65 * Property: for properties of a struct
66 *
67 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
68 */
69
70 enum pandecode_log_type {
71 PANDECODE_RAW,
72 PANDECODE_MESSAGE,
73 PANDECODE_PROPERTY
74 };
75
76 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
77 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
78 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
79
80 unsigned pandecode_indent = 0;
81
82 static void
83 pandecode_make_indent(void)
84 {
85 for (unsigned i = 0; i < pandecode_indent; ++i)
86 printf(" ");
87 }
88
89 static void
90 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
91 {
92 va_list ap;
93
94 pandecode_make_indent();
95
96 if (type == PANDECODE_MESSAGE)
97 printf("// ");
98 else if (type == PANDECODE_PROPERTY)
99 printf(".");
100
101 va_start(ap, format);
102 vprintf(format, ap);
103 va_end(ap);
104
105 if (type == PANDECODE_PROPERTY)
106 printf(",\n");
107 }
108
109 static void
110 pandecode_log_cont(const char *format, ...)
111 {
112 va_list ap;
113
114 va_start(ap, format);
115 vprintf(format, ap);
116 va_end(ap);
117 }
118
119 /* To check for memory safety issues, validates that the given pointer in GPU
120 * memory is valid, containing at least sz bytes. The goal is to eliminate
121 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
122 * overruns) by statically validating pointers.
123 */
124
125 static void
126 pandecode_validate_buffer(mali_ptr addr, size_t sz)
127 {
128 if (!addr) {
129 pandecode_msg("XXX: null pointer deref");
130 return;
131 }
132
133 /* Find a BO */
134
135 struct pandecode_mapped_memory *bo =
136 pandecode_find_mapped_gpu_mem_containing(addr);
137
138 if (!bo) {
139 pandecode_msg("XXX: invalid memory dereference\n");
140 return;
141 }
142
143 /* Bounds check */
144
145 unsigned offset = addr - bo->gpu_va;
146 unsigned total = offset + sz;
147
148 if (total > bo->length) {
149 pandecode_msg("XXX: buffer overrun. "
150 "Chunk of size %zu at offset %d in buffer of size %zu. "
151 "Overrun by %zu bytes. \n",
152 sz, offset, bo->length, total - bo->length);
153 return;
154 }
155 }
156
157 struct pandecode_flag_info {
158 u64 flag;
159 const char *name;
160 };
161
162 static void
163 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
164 u64 flags)
165 {
166 bool decodable_flags_found = false;
167
168 for (int i = 0; flag_info[i].name; i++) {
169 if ((flags & flag_info[i].flag) != flag_info[i].flag)
170 continue;
171
172 if (!decodable_flags_found) {
173 decodable_flags_found = true;
174 } else {
175 pandecode_log_cont(" | ");
176 }
177
178 pandecode_log_cont("%s", flag_info[i].name);
179
180 flags &= ~flag_info[i].flag;
181 }
182
183 if (decodable_flags_found) {
184 if (flags)
185 pandecode_log_cont(" | 0x%" PRIx64, flags);
186 } else {
187 pandecode_log_cont("0x%" PRIx64, flags);
188 }
189 }
190
191 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
192 static const struct pandecode_flag_info gl_enable_flag_info[] = {
193 FLAG_INFO(OCCLUSION_QUERY),
194 FLAG_INFO(OCCLUSION_PRECISE),
195 FLAG_INFO(FRONT_CCW_TOP),
196 FLAG_INFO(CULL_FACE_FRONT),
197 FLAG_INFO(CULL_FACE_BACK),
198 {}
199 };
200 #undef FLAG_INFO
201
202 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
203 static const struct pandecode_flag_info clear_flag_info[] = {
204 FLAG_INFO(FAST),
205 FLAG_INFO(SLOW),
206 FLAG_INFO(SLOW_STENCIL),
207 {}
208 };
209 #undef FLAG_INFO
210
211 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
212 static const struct pandecode_flag_info mask_flag_info[] = {
213 FLAG_INFO(R),
214 FLAG_INFO(G),
215 FLAG_INFO(B),
216 FLAG_INFO(A),
217 {}
218 };
219 #undef FLAG_INFO
220
221 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
222 static const struct pandecode_flag_info u3_flag_info[] = {
223 FLAG_INFO(HAS_MSAA),
224 FLAG_INFO(CAN_DISCARD),
225 FLAG_INFO(HAS_BLEND_SHADER),
226 FLAG_INFO(DEPTH_WRITEMASK),
227 {}
228 };
229
230 static const struct pandecode_flag_info u4_flag_info[] = {
231 FLAG_INFO(NO_MSAA),
232 FLAG_INFO(NO_DITHER),
233 FLAG_INFO(DEPTH_RANGE_A),
234 FLAG_INFO(DEPTH_RANGE_B),
235 FLAG_INFO(STENCIL_TEST),
236 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER),
237 {}
238 };
239 #undef FLAG_INFO
240
241 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
242 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
243 FLAG_INFO(MSAA),
244 FLAG_INFO(SRGB),
245 {}
246 };
247 #undef FLAG_INFO
248
249 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
250 static const struct pandecode_flag_info mfbd_extra_flag_info[] = {
251 FLAG_INFO(PRESENT),
252 FLAG_INFO(AFBC),
253 FLAG_INFO(ZS),
254 {}
255 };
256 #undef FLAG_INFO
257
258 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
259 static const struct pandecode_flag_info shader_midgard1_flag_info [] = {
260 FLAG_INFO(EARLY_Z),
261 FLAG_INFO(READS_TILEBUFFER),
262 FLAG_INFO(READS_ZS),
263 {}
264 };
265 #undef FLAG_INFO
266
267 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
268 static const struct pandecode_flag_info mfbd_flag_info [] = {
269 FLAG_INFO(DEPTH_WRITE),
270 FLAG_INFO(EXTRA),
271 {}
272 };
273 #undef FLAG_INFO
274
275 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
276 static const struct pandecode_flag_info sampler_flag_info [] = {
277 FLAG_INFO(MAG_NEAREST),
278 FLAG_INFO(MIN_NEAREST),
279 FLAG_INFO(MIP_LINEAR_1),
280 FLAG_INFO(MIP_LINEAR_2),
281 FLAG_INFO(NORM_COORDS),
282 {}
283 };
284 #undef FLAG_INFO
285
286 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
287 static const struct pandecode_flag_info sfbd_unk1_info [] = {
288 FLAG_INFO(MSAA_8),
289 FLAG_INFO(MSAA_A),
290 {}
291 };
292 #undef FLAG_INFO
293
294 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
295 static const struct pandecode_flag_info sfbd_unk2_info [] = {
296 FLAG_INFO(MSAA_B),
297 FLAG_INFO(SRGB),
298 {}
299 };
300 #undef FLAG_INFO
301
302 extern char *replace_fragment;
303 extern char *replace_vertex;
304
305 static char *
306 pandecode_job_type(enum mali_job_type type)
307 {
308 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
309
310 switch (type) {
311 DEFINE_CASE(NULL);
312 DEFINE_CASE(WRITE_VALUE);
313 DEFINE_CASE(CACHE_FLUSH);
314 DEFINE_CASE(COMPUTE);
315 DEFINE_CASE(VERTEX);
316 DEFINE_CASE(TILER);
317 DEFINE_CASE(FUSED);
318 DEFINE_CASE(FRAGMENT);
319
320 case JOB_NOT_STARTED:
321 return "NOT_STARTED";
322
323 default:
324 pandecode_log("Warning! Unknown job type %x\n", type);
325 return "!?!?!?";
326 }
327
328 #undef DEFINE_CASE
329 }
330
331 static char *
332 pandecode_draw_mode(enum mali_draw_mode mode)
333 {
334 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
335
336 switch (mode) {
337 DEFINE_CASE(DRAW_NONE);
338 DEFINE_CASE(POINTS);
339 DEFINE_CASE(LINES);
340 DEFINE_CASE(TRIANGLES);
341 DEFINE_CASE(TRIANGLE_STRIP);
342 DEFINE_CASE(TRIANGLE_FAN);
343 DEFINE_CASE(LINE_STRIP);
344 DEFINE_CASE(LINE_LOOP);
345 DEFINE_CASE(POLYGON);
346 DEFINE_CASE(QUADS);
347 DEFINE_CASE(QUAD_STRIP);
348
349 default:
350 pandecode_msg("XXX: invalid draw mode %X\n", mode);
351 return "";
352 }
353
354 #undef DEFINE_CASE
355 }
356
357 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
358 static char *
359 pandecode_func(enum mali_func mode)
360 {
361 switch (mode) {
362 DEFINE_CASE(NEVER);
363 DEFINE_CASE(LESS);
364 DEFINE_CASE(EQUAL);
365 DEFINE_CASE(LEQUAL);
366 DEFINE_CASE(GREATER);
367 DEFINE_CASE(NOTEQUAL);
368 DEFINE_CASE(GEQUAL);
369 DEFINE_CASE(ALWAYS);
370
371 default:
372 pandecode_msg("XXX: invalid func %X\n", mode);
373 return "";
374 }
375 }
376 #undef DEFINE_CASE
377
378 /* Why is this duplicated? Who knows... */
379 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
380 static char *
381 pandecode_alt_func(enum mali_alt_func mode)
382 {
383 switch (mode) {
384 DEFINE_CASE(NEVER);
385 DEFINE_CASE(LESS);
386 DEFINE_CASE(EQUAL);
387 DEFINE_CASE(LEQUAL);
388 DEFINE_CASE(GREATER);
389 DEFINE_CASE(NOTEQUAL);
390 DEFINE_CASE(GEQUAL);
391 DEFINE_CASE(ALWAYS);
392
393 default:
394 pandecode_msg("XXX: invalid alt func %X\n", mode);
395 return "";
396 }
397 }
398 #undef DEFINE_CASE
399
400 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
401 static char *
402 pandecode_stencil_op(enum mali_stencil_op op)
403 {
404 switch (op) {
405 DEFINE_CASE(KEEP);
406 DEFINE_CASE(REPLACE);
407 DEFINE_CASE(ZERO);
408 DEFINE_CASE(INVERT);
409 DEFINE_CASE(INCR_WRAP);
410 DEFINE_CASE(DECR_WRAP);
411 DEFINE_CASE(INCR);
412 DEFINE_CASE(DECR);
413
414 default:
415 pandecode_msg("XXX: invalid stencil op %X\n", op);
416 return "";
417 }
418 }
419
420 #undef DEFINE_CASE
421
422 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
423 {
424 switch(mode) {
425 /* TODO: Combine to just "instanced" once this can be done
426 * unambiguously in all known cases */
427 case MALI_ATTR_POT_DIVIDE:
428 return "instanced_pot";
429 case MALI_ATTR_MODULO:
430 return "instanced_mod";
431 case MALI_ATTR_NPOT_DIVIDE:
432 return "instanced_npot";
433 case MALI_ATTR_IMAGE:
434 return "image";
435 case MALI_ATTR_INTERNAL:
436 return "internal";
437 default:
438 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
439 return "";
440 }
441 }
442
443 static const char *
444 pandecode_special_varying(uint64_t v)
445 {
446 switch(v) {
447 case MALI_VARYING_FRAG_COORD:
448 return "gl_FragCoord";
449 case MALI_VARYING_FRONT_FACING:
450 return "gl_FrontFacing";
451 case MALI_VARYING_POINT_COORD:
452 return "gl_PointCoord";
453 default:
454 pandecode_msg("XXX: invalid special varying %" PRIx64 "\n", v);
455 return "";
456 }
457 }
458
459 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
460 static char *
461 pandecode_wrap_mode(enum mali_wrap_mode op)
462 {
463 switch (op) {
464 DEFINE_CASE(REPEAT);
465 DEFINE_CASE(CLAMP_TO_EDGE);
466 DEFINE_CASE(CLAMP_TO_BORDER);
467 DEFINE_CASE(MIRRORED_REPEAT);
468
469 default:
470 pandecode_msg("XXX: invalid wrap mode %X\n", op);
471 return "";
472 }
473 }
474 #undef DEFINE_CASE
475
476 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
477 static char *
478 pandecode_block_format(enum mali_block_format fmt)
479 {
480 switch (fmt) {
481 DEFINE_CASE(TILED);
482 DEFINE_CASE(UNKNOWN);
483 DEFINE_CASE(LINEAR);
484 DEFINE_CASE(AFBC);
485
486 default:
487 unreachable("Invalid case");
488 }
489 }
490 #undef DEFINE_CASE
491
492 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
493 char *
494 pandecode_exception_access(enum mali_exception_access access)
495 {
496 switch (access) {
497 DEFINE_CASE(NONE);
498 DEFINE_CASE(EXECUTE);
499 DEFINE_CASE(READ);
500 DEFINE_CASE(WRITE);
501
502 default:
503 unreachable("Invalid case");
504 }
505 }
506 #undef DEFINE_CASE
507
508 /* Midgard's tiler descriptor is embedded within the
509 * larger FBD */
510
511 static void
512 pandecode_midgard_tiler_descriptor(
513 const struct midgard_tiler_descriptor *t,
514 unsigned width,
515 unsigned height,
516 bool is_fragment,
517 bool has_hierarchy)
518 {
519 pandecode_log(".tiler = {\n");
520 pandecode_indent++;
521
522 if (t->hierarchy_mask == MALI_TILER_DISABLED)
523 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
524 else
525 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
526
527 /* We know this name from the kernel, but we never see it nonzero */
528
529 if (t->flags)
530 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
531
532 MEMORY_PROP(t, polygon_list);
533
534 /* The body is offset from the base of the polygon list */
535 assert(t->polygon_list_body > t->polygon_list);
536 unsigned body_offset = t->polygon_list_body - t->polygon_list;
537
538 /* It needs to fit inside the reported size */
539 assert(t->polygon_list_size >= body_offset);
540
541 /* Check that we fit */
542 struct pandecode_mapped_memory *plist =
543 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
544
545 assert(t->polygon_list_size <= plist->length);
546
547 /* Now that we've sanity checked, we'll try to calculate the sizes
548 * ourselves for comparison */
549
550 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
551 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
552
553 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
554 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
555 ref_header, ref_size);
556 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
557 pandecode_msg("body offset %d\n", body_offset);
558 }
559
560 /* The tiler heap has a start and end specified -- it should be
561 * identical to what we have in the BO. The exception is if tiling is
562 * disabled. */
563
564 MEMORY_PROP(t, heap_start);
565 assert(t->heap_end >= t->heap_start);
566
567 struct pandecode_mapped_memory *heap =
568 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
569
570 unsigned heap_size = t->heap_end - t->heap_start;
571
572 /* Tiling is enabled with a special flag */
573 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
574 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
575
576 bool tiling_enabled = hierarchy_mask;
577
578 if (tiling_enabled) {
579 /* When tiling is enabled, the heap should be a tight fit */
580 unsigned heap_offset = t->heap_start - heap->gpu_va;
581 if ((heap_offset + heap_size) != heap->length) {
582 pandecode_msg("XXX: heap size %u (expected %zu)\n",
583 heap_size, heap->length - heap_offset);
584 }
585
586 /* We should also have no other flags */
587 if (tiler_flags)
588 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
589 } else {
590 /* When tiling is disabled, we should have that flag and no others */
591
592 if (tiler_flags != MALI_TILER_DISABLED) {
593 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
594 tiler_flags);
595 }
596
597 /* We should also have an empty heap */
598 if (heap_size) {
599 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
600 heap_size);
601 }
602
603 /* Disabled tiling is used only for clear-only jobs, which are
604 * purely FRAGMENT, so we should never see this for
605 * non-FRAGMENT descriptors. */
606
607 if (!is_fragment)
608 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
609 }
610
611 /* We've never seen weights used in practice, but we know from the
612 * kernel these fields is there */
613
614 bool nonzero_weights = false;
615
616 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
617 nonzero_weights |= t->weights[w] != 0x0;
618 }
619
620 if (nonzero_weights) {
621 pandecode_log(".weights = {");
622
623 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
624 pandecode_log("%d, ", t->weights[w]);
625 }
626
627 pandecode_log("},");
628 }
629
630 pandecode_indent--;
631 pandecode_log("}\n");
632 }
633
634 /* Information about the framebuffer passed back for
635 * additional analysis */
636
637 struct pandecode_fbd {
638 unsigned width;
639 unsigned height;
640 unsigned rt_count;
641 bool has_extra;
642 };
643
644 static void
645 pandecode_sfbd_format(struct mali_sfbd_format format)
646 {
647 pandecode_log(".format = {\n");
648 pandecode_indent++;
649
650 pandecode_log(".unk1 = ");
651 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
652 pandecode_log_cont(",\n");
653
654 /* TODO: Map formats so we can check swizzles and print nicely */
655 pandecode_log("swizzle");
656 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
657 pandecode_log_cont(",\n");
658
659 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
660 MALI_NEGATIVE(format.nr_channels));
661
662 pandecode_log(".unk2 = ");
663 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
664 pandecode_log_cont(",\n");
665
666 pandecode_prop("block = %s", pandecode_block_format(format.block));
667
668 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
669
670 pandecode_indent--;
671 pandecode_log("},\n");
672 }
673
674 static struct pandecode_fbd
675 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
676 {
677 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
678 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
679
680 struct pandecode_fbd info = {
681 .has_extra = false,
682 .rt_count = 1
683 };
684
685 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
686 pandecode_indent++;
687
688 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
689 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
690
691 pandecode_sfbd_format(s->format);
692
693 info.width = s->width + 1;
694 info.height = s->height + 1;
695
696 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
697 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
698
699 MEMORY_PROP(s, checksum);
700
701 if (s->checksum_stride)
702 pandecode_prop("checksum_stride = %d", s->checksum_stride);
703
704 MEMORY_PROP(s, framebuffer);
705 pandecode_prop("stride = %d", s->stride);
706
707 /* Earlier in the actual commandstream -- right before width -- but we
708 * delay to flow nicer */
709
710 pandecode_log(".clear_flags = ");
711 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
712 pandecode_log_cont(",\n");
713
714 if (s->depth_buffer) {
715 MEMORY_PROP(s, depth_buffer);
716 pandecode_prop("depth_stride = %d", s->depth_stride);
717 }
718
719 if (s->stencil_buffer) {
720 MEMORY_PROP(s, stencil_buffer);
721 pandecode_prop("stencil_stride = %d", s->stencil_stride);
722 }
723
724 if (s->depth_stride_zero ||
725 s->stencil_stride_zero ||
726 s->zero7 || s->zero8) {
727 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
728 pandecode_prop("depth_stride_zero = 0x%x",
729 s->depth_stride_zero);
730 pandecode_prop("stencil_stride_zero = 0x%x",
731 s->stencil_stride_zero);
732 pandecode_prop("zero7 = 0x%" PRIx32,
733 s->zero7);
734 pandecode_prop("zero8 = 0x%" PRIx32,
735 s->zero8);
736 }
737
738 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
739 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
740 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
741 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
742 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
743 }
744
745 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
746 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
747 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
748 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
749 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
750 }
751
752 if (s->clear_stencil) {
753 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
754 }
755
756 MEMORY_PROP(s, scratchpad);
757 const struct midgard_tiler_descriptor t = s->tiler;
758
759 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
760 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
761
762 pandecode_indent--;
763 pandecode_log("};\n");
764
765 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
766 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
767 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
768 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
769 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
770
771 printf(".zero3 = {");
772
773 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
774 printf("%X, ", s->zero3[i]);
775
776 printf("},\n");
777
778 printf(".zero6 = {");
779
780 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
781 printf("%X, ", s->zero6[i]);
782
783 printf("},\n");
784
785 return info;
786 }
787
788 static void
789 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
790 {
791 pandecode_log(".unknown%d = {", name);
792
793 for (int i = 0; i < count; ++i)
794 printf("%X, ", slide[i]);
795
796 pandecode_log("},\n");
797 }
798
799 #define SHORT_SLIDE(num) \
800 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
801
802 static void
803 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
804 {
805 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
806 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
807
808 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
809 pandecode_indent++;
810
811 SHORT_SLIDE(1);
812
813 pandecode_indent--;
814 printf("},\n");
815 }
816
817 /* Extracts the number of components associated with a Mali format */
818
819 static unsigned
820 pandecode_format_component_count(enum mali_format fmt)
821 {
822 /* Mask out the format class */
823 unsigned top = fmt & 0b11100000;
824
825 switch (top) {
826 case MALI_FORMAT_SNORM:
827 case MALI_FORMAT_UINT:
828 case MALI_FORMAT_UNORM:
829 case MALI_FORMAT_SINT:
830 return ((fmt >> 3) & 3) + 1;
831 default:
832 /* TODO: Validate */
833 return 4;
834 }
835 }
836
837 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
838
839 static unsigned
840 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
841 {
842 unsigned mask = 0;
843 assert(MALI_CHANNEL_RED == 0);
844
845 for (unsigned c = 0; c < 4; ++c) {
846 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
847
848 if (chan <= MALI_CHANNEL_ALPHA)
849 mask |= (1 << chan);
850 }
851
852 return mask;
853 }
854
855 /* Validates that a (format, swizzle) pair is valid, in the sense that the
856 * swizzle doesn't access any components that are undefined in the format.
857 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
858 * omitted */
859
860 static bool
861 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
862 {
863 unsigned nr_comp = pandecode_format_component_count(fmt);
864 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
865 unsigned valid_mask = (1 << nr_comp) - 1;
866 unsigned invalid_mask = ~valid_mask;
867
868 if (access_mask & invalid_mask) {
869 pandecode_msg("XXX: invalid components accessed\n");
870 return false;
871 }
872
873 /* Check for the default non-swizzling swizzle so we can suppress
874 * useless printing for the defaults */
875
876 unsigned default_swizzles[4] = {
877 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
878 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
879 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
880 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
881 };
882
883 return (swizzle == default_swizzles[nr_comp - 1]);
884 }
885
886 /* Maps MALI_RGBA32F to rgba32f, etc */
887
888 static void
889 pandecode_format_short(enum mali_format fmt, bool srgb)
890 {
891 /* We want a type-like format, so cut off the initial MALI_ */
892 char *format = pandecode_format(fmt);
893 format += strlen("MALI_");
894
895 unsigned len = strlen(format);
896 char *lower_format = calloc(1, len + 1);
897
898 for (unsigned i = 0; i < len; ++i)
899 lower_format[i] = tolower(format[i]);
900
901 /* Sanity check sRGB flag is applied to RGB, per the name */
902 if (srgb && lower_format[0] != 'r')
903 pandecode_msg("XXX: sRGB applied to non-colour format\n");
904
905 /* Just prefix with an s, so you get formats like srgba8_unorm */
906 if (srgb)
907 pandecode_log_cont("s");
908
909 pandecode_log_cont("%s", lower_format);
910 free(lower_format);
911 }
912
913 static void
914 pandecode_swizzle(unsigned swizzle, enum mali_format format)
915 {
916 /* First, do some validation */
917 bool trivial_swizzle = pandecode_validate_format_swizzle(
918 format, swizzle);
919
920 if (trivial_swizzle)
921 return;
922
923 /* Next, print the swizzle */
924 pandecode_log_cont(".");
925
926 static const char components[] = "rgba01";
927
928 for (unsigned c = 0; c < 4; ++c) {
929 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
930
931 if (chan >= MALI_CHANNEL_RESERVED_0) {
932 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
933 continue;
934 }
935 pandecode_log_cont("%c", components[chan]);
936 }
937 }
938
939 static void
940 pandecode_rt_format(struct mali_rt_format format)
941 {
942 pandecode_log(".format = {\n");
943 pandecode_indent++;
944
945 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
946 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
947 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
948
949 pandecode_prop("block = %s", pandecode_block_format(format.block));
950
951 /* TODO: Map formats so we can check swizzles and print nicely */
952 pandecode_log("swizzle");
953 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
954 pandecode_log_cont(",\n");
955
956 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
957 MALI_NEGATIVE(format.nr_channels));
958
959 pandecode_log(".flags = ");
960 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
961 pandecode_log_cont(",\n");
962
963 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
964 * which is a faster hardware-based alternative to the wallpaper method
965 * to preserve framebuffer contents across frames. In practice, MFBD
966 * preload is buggy on Midgard, and so this is a chicken bit. If this
967 * bit isn't set, most likely something broke unrelated to preload */
968
969 if (!format.no_preload) {
970 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
971 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
972 }
973
974 if (format.zero)
975 pandecode_prop("zero = 0x%" PRIx32, format.zero);
976
977 pandecode_indent--;
978 pandecode_log("},\n");
979 }
980
981 static void
982 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
983 {
984 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
985 pandecode_indent++;
986
987 for (int i = 0; i < MALI_NEGATIVE(fb->rt_count_1); i++) {
988 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
989 struct pandecode_mapped_memory *mem =
990 pandecode_find_mapped_gpu_mem_containing(rt_va);
991 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
992
993 pandecode_log("{\n");
994 pandecode_indent++;
995
996 pandecode_rt_format(rt->format);
997
998 if (rt->format.block == MALI_BLOCK_AFBC) {
999 pandecode_log(".afbc = {\n");
1000 pandecode_indent++;
1001
1002 char *a = pointer_as_memory_reference(rt->afbc.metadata);
1003 pandecode_prop("metadata = %s", a);
1004 free(a);
1005
1006 pandecode_prop("stride = %d", rt->afbc.stride);
1007 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
1008
1009 pandecode_indent--;
1010 pandecode_log("},\n");
1011 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
1012 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1013 rt->afbc.metadata,
1014 rt->afbc.stride,
1015 rt->afbc.unk);
1016 }
1017
1018 MEMORY_PROP(rt, framebuffer);
1019 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1020
1021 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1022 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1023 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1024 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1025 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1026 }
1027
1028 if (rt->zero1 || rt->zero2 || rt->zero3) {
1029 pandecode_msg("XXX: render target zeros tripped\n");
1030 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1031 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1032 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
1033 }
1034
1035 pandecode_indent--;
1036 pandecode_log("},\n");
1037 }
1038
1039 pandecode_indent--;
1040 pandecode_log("};\n");
1041 }
1042
1043 static struct pandecode_fbd
1044 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment)
1045 {
1046 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1047 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1048
1049 struct pandecode_fbd info;
1050
1051 if (fb->sample_locations) {
1052 /* The blob stores all possible sample locations in a single buffer
1053 * allocated on startup, and just switches the pointer when switching
1054 * MSAA state. For now, we just put the data into the cmdstream, but we
1055 * should do something like what the blob does with a real driver.
1056 *
1057 * There seem to be 32 slots for sample locations, followed by another
1058 * 16. The second 16 is just the center location followed by 15 zeros
1059 * in all the cases I've identified (maybe shader vs. depth/color
1060 * samples?).
1061 */
1062
1063 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
1064
1065 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
1066
1067 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
1068 pandecode_indent++;
1069
1070 for (int i = 0; i < 32 + 16; i++) {
1071 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1072 }
1073
1074 pandecode_indent--;
1075 pandecode_log("};\n");
1076 }
1077
1078 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1079 pandecode_indent++;
1080
1081 pandecode_prop("stack_shift = 0x%x", fb->stack_shift);
1082 pandecode_prop("unk0 = 0x%x", fb->unk0);
1083
1084 if (fb->sample_locations)
1085 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1086
1087 /* Assume that unknown1 was emitted in the last job for
1088 * now */
1089 MEMORY_PROP(fb, unknown1);
1090
1091 info.width = fb->width1 + 1;
1092 info.height = fb->height1 + 1;
1093 info.rt_count = fb->rt_count_1 + 1;
1094
1095 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1096 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1097 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1098 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1099
1100 pandecode_prop("unk1 = 0x%x", fb->unk1);
1101 pandecode_prop("unk2 = 0x%x", fb->unk2);
1102 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1103 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1104
1105 pandecode_log(".mfbd_flags = ");
1106 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1107 pandecode_log_cont(",\n");
1108
1109 if (fb->clear_stencil)
1110 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1111
1112 if (fb->clear_depth)
1113 pandecode_prop("clear_depth = %f", fb->clear_depth);
1114
1115 /* TODO: What is this? Let's not blow up.. */
1116 if (fb->unknown2 != 0x1F)
1117 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1118
1119 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1120 MEMORY_PROP(fb, scratchpad);
1121 const struct midgard_tiler_descriptor t = fb->tiler;
1122 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
1123
1124 if (fb->zero3 || fb->zero4) {
1125 pandecode_msg("XXX: framebuffer zeros tripped\n");
1126 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1127 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1128 }
1129
1130 pandecode_indent--;
1131 pandecode_log("};\n");
1132
1133 gpu_va += sizeof(struct bifrost_framebuffer);
1134
1135 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1136
1137 if (info.has_extra) {
1138 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1139 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1140
1141 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1142 pandecode_indent++;
1143
1144 MEMORY_PROP(fbx, checksum);
1145
1146 if (fbx->checksum_stride)
1147 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1148
1149 pandecode_log(".flags = ");
1150 pandecode_log_decoded_flags(mfbd_extra_flag_info, fbx->flags);
1151 pandecode_log_cont(",\n");
1152
1153 if (fbx->flags & MALI_EXTRA_AFBC_ZS) {
1154 pandecode_log(".ds_afbc = {\n");
1155 pandecode_indent++;
1156
1157 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1158 pandecode_prop("depth_stencil_afbc_stride = %d",
1159 fbx->ds_afbc.depth_stencil_afbc_stride);
1160 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1161
1162 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
1163 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1164 pandecode_prop("zero1 = 0x%" PRIx32,
1165 fbx->ds_afbc.zero1);
1166 pandecode_prop("padding = 0x%" PRIx64,
1167 fbx->ds_afbc.padding);
1168 }
1169
1170 pandecode_indent--;
1171 pandecode_log("},\n");
1172 } else {
1173 pandecode_log(".ds_linear = {\n");
1174 pandecode_indent++;
1175
1176 if (fbx->ds_linear.depth) {
1177 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1178 pandecode_prop("depth_stride = %d",
1179 fbx->ds_linear.depth_stride);
1180 }
1181
1182 if (fbx->ds_linear.stencil) {
1183 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1184 pandecode_prop("stencil_stride = %d",
1185 fbx->ds_linear.stencil_stride);
1186 }
1187
1188 if (fbx->ds_linear.depth_stride_zero ||
1189 fbx->ds_linear.stencil_stride_zero ||
1190 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1191 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1192 pandecode_prop("depth_stride_zero = 0x%x",
1193 fbx->ds_linear.depth_stride_zero);
1194 pandecode_prop("stencil_stride_zero = 0x%x",
1195 fbx->ds_linear.stencil_stride_zero);
1196 pandecode_prop("zero1 = 0x%" PRIx32,
1197 fbx->ds_linear.zero1);
1198 pandecode_prop("zero2 = 0x%" PRIx32,
1199 fbx->ds_linear.zero2);
1200 }
1201
1202 pandecode_indent--;
1203 pandecode_log("},\n");
1204 }
1205
1206 if (fbx->zero3 || fbx->zero4) {
1207 pandecode_msg("XXX: fb_extra zeros tripped\n");
1208 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1209 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1210 }
1211
1212 pandecode_indent--;
1213 pandecode_log("};\n");
1214
1215 gpu_va += sizeof(struct bifrost_fb_extra);
1216 }
1217
1218 if (is_fragment)
1219 pandecode_render_target(gpu_va, job_no, fb);
1220
1221 return info;
1222 }
1223
1224 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1225 * count */
1226
1227 static void
1228 pandecode_padded_vertices(unsigned shift, unsigned k)
1229 {
1230 unsigned odd = 2*k + 1;
1231 unsigned pot = 1 << shift;
1232 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1233 }
1234
1235 /* Given a magic divisor, recover what we were trying to divide by.
1236 *
1237 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1238 * 0 <= m < 2^N, for N bits in m.
1239 *
1240 * Let q represent the number we would like to divide by.
1241 *
1242 * By definition of a magic divisor for N-bit unsigned integers (a number you
1243 * multiply by to magically get division), m is a number such that:
1244 *
1245 * (m * x) & (2^N - 1) = floor(x/q).
1246 * for all x on Z where 0 <= x < 2^N
1247 *
1248 * Ignore the case where any of the above values equals zero; it is irrelevant
1249 * for our purposes (instanced arrays).
1250 *
1251 * Choose x = q. Then:
1252 *
1253 * (m * x) & (2^N - 1) = floor(x/q).
1254 * (m * q) & (2^N - 1) = floor(q/q).
1255 *
1256 * floor(q/q) = floor(1) = 1, therefore:
1257 *
1258 * (m * q) & (2^N - 1) = 1
1259 *
1260 * Recall the identity that the bitwise AND of one less than a power-of-two
1261 * equals the modulo with that power of two, i.e. for all x:
1262 *
1263 * x & (2^N - 1) = x % N
1264 *
1265 * Therefore:
1266 *
1267 * mq % (2^N) = 1
1268 *
1269 * By definition, a modular multiplicative inverse of a number m is the number
1270 * q such that with respect to a modulos M:
1271 *
1272 * mq % M = 1
1273 *
1274 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1275 *
1276 */
1277
1278 static void
1279 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1280 {
1281 #if 0
1282 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1283 * shift) the most lame way possible... just repeatedly add.
1284 * Asymptoptically slow but nobody cares in practice, unless you have
1285 * massive numbers of vertices or high divisors. */
1286
1287 unsigned inverse = 0;
1288
1289 /* Magic implicitly has the highest bit set */
1290 magic |= (1 << 31);
1291
1292 /* Depending on rounding direction */
1293 if (extra)
1294 magic++;
1295
1296 for (;;) {
1297 uint32_t product = magic * inverse;
1298
1299 if (shift) {
1300 product >>= shift;
1301 }
1302
1303 if (product == 1)
1304 break;
1305
1306 ++inverse;
1307 }
1308
1309 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1310
1311 /* Recall we're supposed to divide by (gl_level_divisor *
1312 * padded_num_vertices) */
1313
1314 unsigned padded_num_vertices = inverse / orig_divisor;
1315
1316 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1317 #endif
1318 }
1319
1320 static void
1321 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1322 mali_ptr addr, int job_no, char *suffix,
1323 int count, bool varying, enum mali_job_type job_type)
1324 {
1325 char *prefix = varying ? "varying" : "attribute";
1326 assert(addr);
1327
1328 if (!count) {
1329 pandecode_msg("warn: No %s records\n", prefix);
1330 return;
1331 }
1332
1333 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1334
1335 for (int i = 0; i < count; ++i) {
1336 /* First, check for special records */
1337 if (attr[i].elements < MALI_VARYING_SPECIAL) {
1338 /* Special records are always varyings */
1339
1340 if (!varying)
1341 pandecode_msg("XXX: Special varying in attribute field\n");
1342
1343 if (job_type != JOB_TYPE_TILER)
1344 pandecode_msg("XXX: Special varying in non-FS\n");
1345
1346 /* We're special, so all fields should be zero */
1347 unsigned zero = attr[i].stride | attr[i].size;
1348 zero |= attr[i].shift | attr[i].extra_flags;
1349
1350 if (zero)
1351 pandecode_msg("XXX: Special varying has non-zero fields\n");
1352 else {
1353 /* Print the special varying name */
1354 pandecode_log("varying_%d = %s;", i, pandecode_special_varying(attr[i].elements));
1355 continue;
1356 }
1357 }
1358
1359 enum mali_attr_mode mode = attr[i].elements & 7;
1360
1361 if (mode == MALI_ATTR_UNUSED)
1362 pandecode_msg("XXX: unused attribute record\n");
1363
1364 /* For non-linear records, we need to print the type of record */
1365 if (mode != MALI_ATTR_LINEAR)
1366 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1367
1368 /* Print the name to link with attr_meta */
1369 pandecode_log_cont("%s_%d", prefix, i);
1370
1371 /* Print the stride and size */
1372 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1373
1374 /* TODO: Sanity check the quotient itself. It must be equal to
1375 * (or be greater than, if the driver added padding) the padded
1376 * vertex count. */
1377
1378 /* Finally, print the pointer */
1379 mali_ptr raw_elements = attr[i].elements & ~7;
1380 char *a = pointer_as_memory_reference(raw_elements);
1381 pandecode_log_cont(" = (%s);\n", a);
1382 free(a);
1383
1384 /* Check the pointer */
1385 pandecode_validate_buffer(raw_elements, attr[i].size);
1386
1387 /* shift/extra_flags exist only for instanced */
1388 if (attr[i].shift | attr[i].extra_flags) {
1389 /* These are set to random values by the blob for
1390 * varyings, most likely a symptom of uninitialized
1391 * memory where the hardware masked the bug. As such we
1392 * put this at a warning, not an error. */
1393
1394 if (mode == MALI_ATTR_LINEAR)
1395 pandecode_msg("warn: instancing fields set for linear\n");
1396
1397 pandecode_prop("shift = %d", attr[i].shift);
1398 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1399 }
1400
1401 /* Decode further where possible */
1402
1403 if (mode == MALI_ATTR_MODULO) {
1404 pandecode_padded_vertices(
1405 attr[i].shift,
1406 attr[i].extra_flags);
1407 }
1408
1409 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1410 i++;
1411 pandecode_log("{\n");
1412 pandecode_indent++;
1413 pandecode_prop("unk = 0x%x", attr[i].unk);
1414 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1415 if (attr[i].zero != 0)
1416 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1417 pandecode_prop("divisor = %d", attr[i].divisor);
1418 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1419 pandecode_indent--;
1420 pandecode_log("}, \n");
1421 }
1422
1423 }
1424
1425 pandecode_log("\n");
1426 }
1427
1428 static mali_ptr
1429 pandecode_shader_address(const char *name, mali_ptr ptr)
1430 {
1431 /* TODO: Decode flags */
1432 mali_ptr shader_ptr = ptr & ~15;
1433
1434 char *a = pointer_as_memory_reference(shader_ptr);
1435 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1436 free(a);
1437
1438 return shader_ptr;
1439 }
1440
1441 static void
1442 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1443 {
1444 unsigned any_nonzero =
1445 stencil->ref | stencil->mask | stencil->func |
1446 stencil->sfail | stencil->dpfail | stencil->dppass;
1447
1448 if (any_nonzero == 0)
1449 return;
1450
1451 const char *func = pandecode_func(stencil->func);
1452 const char *sfail = pandecode_stencil_op(stencil->sfail);
1453 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1454 const char *dppass = pandecode_stencil_op(stencil->dppass);
1455
1456 if (stencil->zero)
1457 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1458
1459 pandecode_log(".stencil_%s = {\n", name);
1460 pandecode_indent++;
1461 pandecode_prop("ref = %d", stencil->ref);
1462 pandecode_prop("mask = 0x%02X", stencil->mask);
1463 pandecode_prop("func = %s", func);
1464 pandecode_prop("sfail = %s", sfail);
1465 pandecode_prop("dpfail = %s", dpfail);
1466 pandecode_prop("dppass = %s", dppass);
1467 pandecode_indent--;
1468 pandecode_log("},\n");
1469 }
1470
1471 static void
1472 pandecode_blend_equation(const struct mali_blend_equation *blend)
1473 {
1474 if (blend->zero1)
1475 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1476
1477 pandecode_log(".equation = {\n");
1478 pandecode_indent++;
1479
1480 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1481 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1482
1483 pandecode_log(".color_mask = ");
1484 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1485 pandecode_log_cont(",\n");
1486
1487 pandecode_indent--;
1488 pandecode_log("},\n");
1489 }
1490
1491 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1492
1493 static unsigned
1494 decode_bifrost_constant(u16 constant)
1495 {
1496 float lo = (float) (constant & 0xFF);
1497 float hi = (float) (constant >> 8);
1498
1499 return (hi / 255.0) + (lo / 65535.0);
1500 }
1501
1502 static mali_ptr
1503 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1504 {
1505 struct bifrost_blend_rt *b =
1506 ((struct bifrost_blend_rt *) descs) + rt_no;
1507
1508 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1509 pandecode_indent++;
1510
1511 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1512 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1513 b->constant, decode_bifrost_constant(b->constant));
1514
1515 /* TODO figure out blend shader enable bit */
1516 pandecode_blend_equation(&b->equation);
1517 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1518 pandecode_prop("index = 0x%" PRIx16, b->index);
1519 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1520
1521 pandecode_indent--;
1522 pandecode_log("},\n");
1523
1524 return 0;
1525 }
1526
1527 static mali_ptr
1528 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1529 {
1530 /* constant/equation is in a union */
1531 if (!blend->shader)
1532 return 0;
1533
1534 pandecode_log(".blend = {\n");
1535 pandecode_indent++;
1536
1537 if (is_shader) {
1538 pandecode_shader_address("shader", blend->shader);
1539 } else {
1540 pandecode_blend_equation(&blend->equation);
1541 pandecode_prop("constant = %f", blend->constant);
1542 }
1543
1544 pandecode_indent--;
1545 pandecode_log("},\n");
1546
1547 /* Return blend shader to disassemble if present */
1548 return is_shader ? (blend->shader & ~0xF) : 0;
1549 }
1550
1551 static mali_ptr
1552 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1553 {
1554 struct midgard_blend_rt *b =
1555 ((struct midgard_blend_rt *) descs) + rt_no;
1556
1557 /* Flags determine presence of blend shader */
1558 bool is_shader = (b->flags & 0xF) >= 0x2;
1559
1560 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1561 pandecode_indent++;
1562
1563 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1564
1565 union midgard_blend blend = b->blend;
1566 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1567
1568 pandecode_indent--;
1569 pandecode_log("};\n");
1570
1571 return shader;
1572 }
1573
1574 /* Attributes and varyings have descriptor records, which contain information
1575 * about their format and ordering with the attribute/varying buffers. We'll
1576 * want to validate that the combinations specified are self-consistent.
1577 */
1578
1579 static int
1580 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1581 {
1582 char base[128];
1583 char *prefix = varying ? "varying" : "attribute";
1584 unsigned max_index = 0;
1585 snprintf(base, sizeof(base), "%s_meta", prefix);
1586
1587 struct mali_attr_meta *attr_meta;
1588 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1589
1590 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1591
1592 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1593 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1594 sizeof(*attr_mem));
1595
1596 /* If the record is discard, it should be zero for everything else */
1597
1598 if (attr_meta->format == MALI_VARYING_DISCARD) {
1599 uint64_t zero =
1600 attr_meta->index |
1601 attr_meta->unknown1 |
1602 attr_meta->unknown3 |
1603 attr_meta->src_offset;
1604
1605 if (zero)
1606 pandecode_msg("XXX: expected empty record for varying discard\n");
1607
1608 /* We want to look for a literal 0000 swizzle -- this
1609 * is not encoded with all zeroes, however */
1610
1611 enum mali_channel z = MALI_CHANNEL_ZERO;
1612 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1613 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1614
1615 if (!good_swizzle)
1616 pandecode_msg("XXX: expected zero swizzle for discard\n");
1617
1618 if (!varying)
1619 pandecode_msg("XXX: cannot discard attribute\n");
1620
1621 /* If we're all good, omit the record */
1622 if (!zero && varying && good_swizzle) {
1623 pandecode_log("/* discarded varying */\n");
1624 continue;
1625 }
1626 }
1627
1628 if (attr_meta->index > max_index)
1629 max_index = attr_meta->index;
1630
1631 if (attr_meta->unknown1 != 0x2) {
1632 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1633 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1634 }
1635
1636 if (attr_meta->unknown3) {
1637 pandecode_msg("XXX: unexpected unknown3 set\n");
1638 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1639 }
1640
1641 pandecode_format_short(attr_meta->format, false);
1642 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1643
1644 if (attr_meta->src_offset)
1645 pandecode_log_cont("[%u]", attr_meta->src_offset);
1646
1647 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1648
1649 pandecode_log_cont(";\n");
1650 }
1651
1652 pandecode_log("\n");
1653
1654 return count ? (max_index + 1) : 0;
1655 }
1656
1657 /* return bits [lo, hi) of word */
1658 static u32
1659 bits(u32 word, u32 lo, u32 hi)
1660 {
1661 if (hi - lo >= 32)
1662 return word; // avoid undefined behavior with the shift
1663
1664 return (word >> lo) & ((1 << (hi - lo)) - 1);
1665 }
1666
1667 static void
1668 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool noninstanced)
1669 {
1670 pandecode_log_cont("{\n");
1671 pandecode_indent++;
1672
1673 /* Decode invocation_count. See the comment before the definition of
1674 * invocation_count for an explanation.
1675 */
1676
1677 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1678 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1679 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1680 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1681 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1682 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1683
1684 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1685 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1686 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1687
1688 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1689 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1690 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1691
1692 /* Even though we have this decoded, we want to ensure that the
1693 * representation is "unique" so we don't lose anything by printing only
1694 * the final result. More specifically, we need to check that we were
1695 * passed something in canonical form, since the definition per the
1696 * hardware is inherently not unique. How? Well, take the resulting
1697 * decode and pack it ourselves! If it is bit exact with what we
1698 * decoded, we're good to go. */
1699
1700 struct mali_vertex_tiler_prefix ref;
1701 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, noninstanced);
1702
1703 bool canonical =
1704 (p->invocation_count == ref.invocation_count) &&
1705 (p->invocation_shifts == ref.invocation_shifts);
1706
1707 if (!canonical) {
1708 pandecode_msg("XXX: non-canonical workgroups packing\n");
1709 pandecode_msg("expected: %X, %X",
1710 ref.invocation_count,
1711 ref.invocation_shifts);
1712
1713 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1714 pandecode_prop("size_y_shift = %d", size_y_shift);
1715 pandecode_prop("size_z_shift = %d", size_z_shift);
1716 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
1717 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
1718 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
1719 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
1720 }
1721
1722 /* Regardless, print the decode */
1723 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1724 size_x, size_y, size_z,
1725 groups_x, groups_y, groups_z);
1726
1727 /* TODO: Decode */
1728 if (p->unknown_draw)
1729 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1730
1731 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1732
1733 if (p->draw_mode != MALI_DRAW_NONE)
1734 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1735
1736 /* Index count only exists for tiler jobs anyway */
1737
1738 if (p->index_count)
1739 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1740
1741
1742 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1743 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1744
1745 /* Validate an index buffer is present if we need one. TODO: verify
1746 * relationship between invocation_count and index_count */
1747
1748 if (p->indices) {
1749 unsigned count = p->index_count;
1750
1751 /* Grab the size */
1752 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1753
1754 /* Ensure we got a size, and if so, validate the index buffer
1755 * is large enough to hold a full set of indices of the given
1756 * size */
1757
1758 if (!index_raw_size)
1759 pandecode_msg("XXX: index size missing\n");
1760 else
1761 pandecode_validate_buffer(p->indices, count * size);
1762 } else if (index_raw_size)
1763 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1764
1765 if (p->offset_bias_correction)
1766 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1767
1768 /* TODO: Figure out what this is. It's not zero */
1769 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1770
1771 pandecode_indent--;
1772 pandecode_log("},\n");
1773 }
1774
1775 static void
1776 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1777 {
1778 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1779 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1780
1781 for (int i = 0; i < ubufs_count; i++) {
1782 unsigned size = (ubufs[i].size + 1) * 16;
1783 mali_ptr addr = ubufs[i].ptr << 2;
1784
1785 pandecode_validate_buffer(addr, size);
1786
1787 char *ptr = pointer_as_memory_reference(ubufs[i].ptr << 2);
1788 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1789 free(ptr);
1790 }
1791
1792 pandecode_log("\n");
1793 }
1794
1795 static void
1796 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1797 {
1798 pandecode_validate_buffer(uniforms, uniform_count * 16);
1799
1800 char *ptr = pointer_as_memory_reference(uniforms);
1801 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1802 free(ptr);
1803 }
1804
1805 static void
1806 pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1807 {
1808
1809 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1810
1811 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1812
1813 if (scratchpad->zero) {
1814 pandecode_msg("XXX: scratchpad zero tripped");
1815 pandecode_prop("zero = 0x%x\n", scratchpad->zero);
1816 }
1817
1818 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1819 pandecode_indent++;
1820
1821 pandecode_prop("flags = 0x%x", scratchpad->flags);
1822 MEMORY_PROP(scratchpad, gpu_scratchpad);
1823
1824 pandecode_indent--;
1825 pandecode_log("};\n");
1826 }
1827
1828 static unsigned shader_id = 0;
1829
1830 static struct midgard_disasm_stats
1831 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1832 bool is_bifrost, unsigned gpu_id)
1833 {
1834 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1835 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1836
1837 /* Compute maximum possible size */
1838 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1839
1840 /* Print some boilerplate to clearly denote the assembly (which doesn't
1841 * obey indentation rules), and actually do the disassembly! */
1842
1843 printf("\n\n");
1844
1845 struct midgard_disasm_stats stats;
1846
1847 if (is_bifrost) {
1848 disassemble_bifrost(code, sz, false);
1849
1850 /* TODO: Extend stats to Bifrost */
1851 stats.texture_count = -128;
1852 stats.sampler_count = -128;
1853 stats.attribute_count = -128;
1854 stats.varying_count = -128;
1855 stats.uniform_count = -128;
1856 stats.uniform_buffer_count = -128;
1857 stats.work_count = -128;
1858
1859 stats.instruction_count = 0;
1860 stats.bundle_count = 0;
1861 stats.quadword_count = 0;
1862 stats.helper_invocations = false;
1863 } else {
1864 stats = disassemble_midgard(code, sz, gpu_id,
1865 type == JOB_TYPE_TILER ?
1866 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
1867 }
1868
1869 /* Print shader-db stats */
1870
1871 unsigned nr_threads =
1872 (stats.work_count <= 4) ? 4 :
1873 (stats.work_count <= 8) ? 2 :
1874 1;
1875
1876 printf("shader%d - %s shader: "
1877 "%u inst, %u bundles, %u quadwords, "
1878 "%u registers, %u threads, 0 loops\n\n\n",
1879 shader_id++,
1880 (type == JOB_TYPE_TILER) ? "FRAGMENT" : "VERTEX",
1881 stats.instruction_count, stats.bundle_count, stats.quadword_count,
1882 stats.work_count, nr_threads);
1883
1884
1885 return stats;
1886 }
1887
1888 static void
1889 pandecode_texture(mali_ptr u,
1890 struct pandecode_mapped_memory *tmem,
1891 unsigned job_no, unsigned tex)
1892 {
1893 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
1894
1895 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
1896 pandecode_indent++;
1897
1898 struct mali_texture_format f = t->format;
1899
1900 /* See the definiton of enum mali_texture_type */
1901
1902 bool is_cube = f.type == MALI_TEX_CUBE;
1903 unsigned dimension = is_cube ? 2 : f.type;
1904
1905 pandecode_make_indent();
1906
1907 /* TODO: Are there others? */
1908 bool is_zs = f.format == MALI_Z32_UNORM;
1909
1910 /* Recall Z/S switched the meaning of linear/tiled .. */
1911 if (is_zs && f.layout == MALI_TEXTURE_LINEAR)
1912 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1913
1914 /* Print the layout. Default is linear; a modifier can denote AFBC or
1915 * u-interleaved/tiled modes */
1916
1917 if (f.layout == MALI_TEXTURE_AFBC)
1918 pandecode_log_cont("afbc");
1919 else if (f.layout == MALI_TEXTURE_TILED)
1920 pandecode_log_cont(is_zs ? "linear" : "tiled");
1921 else if (f.layout == MALI_TEXTURE_LINEAR)
1922 pandecode_log_cont("linear");
1923 else
1924 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
1925
1926 pandecode_swizzle(t->swizzle, f.format);
1927 pandecode_log_cont(" ");
1928
1929 /* Distinguish cube/2D with modifier */
1930
1931 if (is_cube)
1932 pandecode_log_cont("cube ");
1933
1934 pandecode_format_short(f.format, f.srgb);
1935 pandecode_swizzle(f.swizzle, f.format);
1936
1937 /* All four width/height/depth/array_size dimensions are present
1938 * regardless of the type of texture, but it is an error to have
1939 * non-zero dimensions for unused dimensions. Verify this. array_size
1940 * can always be set, as can width. */
1941
1942 if (t->height && dimension < 2)
1943 pandecode_msg("XXX: nonzero height for <2D texture\n");
1944
1945 if (t->depth && dimension < 3)
1946 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1947
1948 /* Print only the dimensions that are actually there */
1949
1950 pandecode_log_cont(": %d", t->width + 1);
1951
1952 if (dimension >= 2)
1953 pandecode_log_cont("x%u", t->height + 1);
1954
1955 if (dimension >= 3)
1956 pandecode_log_cont("x%u", t->depth + 1);
1957
1958 if (t->array_size)
1959 pandecode_log_cont("[%u]", t->array_size + 1);
1960
1961 if (t->levels)
1962 pandecode_log_cont(" mip %u", t->levels);
1963
1964 pandecode_log_cont("\n");
1965
1966 if (f.unknown1 | f.zero) {
1967 pandecode_msg("XXX: texture format zero tripped\n");
1968 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
1969 pandecode_prop("zero = %" PRId32, f.zero);
1970 }
1971
1972 if (!f.unknown2) {
1973 pandecode_msg("XXX: expected unknown texture bit set\n");
1974 pandecode_prop("unknown2 = %" PRId32, f.unknown1);
1975 }
1976
1977 if (t->swizzle_zero) {
1978 pandecode_msg("XXX: swizzle zero tripped\n");
1979 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
1980 }
1981
1982 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
1983 pandecode_msg("XXX: texture zero tripped\n");
1984 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
1985 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
1986 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
1987 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
1988 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
1989 }
1990
1991 pandecode_log(".payload = {\n");
1992 pandecode_indent++;
1993
1994 /* A bunch of bitmap pointers follow.
1995 * We work out the correct number,
1996 * based on the mipmap/cubemap
1997 * properties, but dump extra
1998 * possibilities to futureproof */
1999
2000 int bitmap_count = MALI_NEGATIVE(t->levels);
2001
2002 /* Miptree for each face */
2003 if (f.type == MALI_TEX_CUBE)
2004 bitmap_count *= 6;
2005
2006 /* Array of textures */
2007 bitmap_count *= MALI_NEGATIVE(t->array_size);
2008
2009 /* Stride for each element */
2010 if (f.manual_stride)
2011 bitmap_count *= 2;
2012
2013 /* Sanity check the size */
2014 int max_count = sizeof(t->payload) / sizeof(t->payload[0]);
2015 assert (bitmap_count <= max_count);
2016
2017 for (int i = 0; i < bitmap_count; ++i) {
2018 /* How we dump depends if this is a stride or a pointer */
2019
2020 if (f.manual_stride && (i & 1)) {
2021 /* signed 32-bit snuck in as a 64-bit pointer */
2022 uint64_t stride_set = t->payload[i];
2023 uint32_t clamped_stride = stride_set;
2024 int32_t stride = clamped_stride;
2025 assert(stride_set == clamped_stride);
2026 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2027 } else {
2028 char *a = pointer_as_memory_reference(t->payload[i]);
2029 pandecode_log("%s, \n", a);
2030 free(a);
2031 }
2032 }
2033
2034 pandecode_indent--;
2035 pandecode_log("},\n");
2036
2037 pandecode_indent--;
2038 pandecode_log("};\n");
2039 }
2040
2041 /* 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 */
2042
2043 static void
2044 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2045 {
2046 /* Nothing to do */
2047 if (claim == truth)
2048 return;
2049
2050 if (fuzzy)
2051 assert(truth >= 0);
2052
2053 if ((truth >= 0) && !fuzzy) {
2054 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2055 (truth < claim) ? "warn" : "XXX",
2056 name, truth, claim);
2057 } else if ((claim > -truth) && !fuzzy) {
2058 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2059 name, -truth, claim);
2060 } else if (fuzzy && (claim < truth))
2061 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2062 name, truth, claim);
2063
2064 pandecode_log(".%s = %" PRId16, name, claim);
2065
2066 if (fuzzy)
2067 pandecode_log_cont(" /* %u used */", truth);
2068
2069 pandecode_log_cont(",\n");
2070 }
2071
2072 static void
2073 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2074 bool is_bifrost, unsigned gpu_id)
2075 {
2076 struct midgard_disasm_stats stats =
2077 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
2078
2079 bool has_texture = (stats.texture_count > 0);
2080 bool has_sampler = (stats.sampler_count > 0);
2081 bool has_attribute = (stats.attribute_count > 0);
2082 bool has_varying = (stats.varying_count > 0);
2083 bool has_uniform = (stats.uniform_count > 0);
2084 bool has_ubo = (stats.uniform_buffer_count > 0);
2085
2086 if (has_texture || has_sampler)
2087 pandecode_msg("XXX: blend shader accessing textures\n");
2088
2089 if (has_attribute || has_varying)
2090 pandecode_msg("XXX: blend shader accessing interstage\n");
2091
2092 if (has_uniform || has_ubo)
2093 pandecode_msg("XXX: blend shader accessing uniforms\n");
2094 }
2095
2096 static void
2097 pandecode_vertex_tiler_postfix_pre(
2098 const struct mali_vertex_tiler_postfix *p,
2099 int job_no, enum mali_job_type job_type,
2100 char *suffix, bool is_bifrost, unsigned gpu_id)
2101 {
2102 struct pandecode_mapped_memory *attr_mem;
2103
2104 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2105 * are the only things actually needed from the FBD, vertex/tiler jobs
2106 * no longer reference the FBD -- instead, this field points to some
2107 * info about the scratchpad.
2108 */
2109
2110 struct pandecode_fbd fbd_info = {
2111 /* Default for Bifrost */
2112 .rt_count = 1
2113 };
2114
2115 if (is_bifrost)
2116 pandecode_scratchpad(p->framebuffer & ~1, job_no, suffix);
2117 else if (p->framebuffer & MALI_MFBD)
2118 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
2119 else if (job_type == JOB_TYPE_COMPUTE)
2120 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
2121 else
2122 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no, false, gpu_id);
2123
2124 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2125 int texture_count = 0, sampler_count = 0;
2126
2127 if (p->shader) {
2128 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2129 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2130
2131 /* Disassemble ahead-of-time to get stats. Initialize with
2132 * stats for the missing-shader case so we get validation
2133 * there, too */
2134
2135 struct midgard_disasm_stats info = {
2136 .texture_count = 0,
2137 .sampler_count = 0,
2138 .attribute_count = 0,
2139 .varying_count = 0,
2140 .work_count = 1,
2141
2142 .uniform_count = -128,
2143 .uniform_buffer_count = 0
2144 };
2145
2146 if (s->shader & ~0xF)
2147 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
2148
2149 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2150 pandecode_indent++;
2151
2152 /* Save for dumps */
2153 attribute_count = s->attribute_count;
2154 varying_count = s->varying_count;
2155 texture_count = s->texture_count;
2156 sampler_count = s->sampler_count;
2157
2158 if (is_bifrost) {
2159 uniform_count = s->bifrost2.uniform_count;
2160 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2161 } else {
2162 uniform_count = s->midgard1.uniform_count;
2163 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2164 }
2165
2166 pandecode_shader_address("shader", s->shader);
2167
2168 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2169 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2170 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2171 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2172 pandecode_shader_prop("uniform_buffer_count",
2173 uniform_buffer_count,
2174 info.uniform_buffer_count, true);
2175
2176 if (!is_bifrost) {
2177 pandecode_shader_prop("uniform_count",
2178 uniform_count,
2179 info.uniform_count, false);
2180
2181 pandecode_shader_prop("work_count",
2182 s->midgard1.work_count, info.work_count, false);
2183 }
2184
2185 if (is_bifrost) {
2186 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32, s->bifrost1.unk1);
2187 } else {
2188 bool helpers = s->midgard1.flags & MALI_HELPER_INVOCATIONS;
2189 s->midgard1.flags &= ~MALI_HELPER_INVOCATIONS;
2190
2191 if (helpers != info.helper_invocations) {
2192 pandecode_msg("XXX: expected helpers %u but got %u\n",
2193 info.helper_invocations, helpers);
2194 }
2195
2196 pandecode_log(".midgard1.flags = ");
2197 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
2198 pandecode_log_cont(",\n");
2199
2200 pandecode_prop("midgard1.unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
2201 }
2202
2203 if (s->depth_units || s->depth_factor) {
2204 pandecode_prop("depth_factor = %f", s->depth_factor);
2205 pandecode_prop("depth_units = %f", s->depth_units);
2206 }
2207
2208 if (s->alpha_coverage) {
2209 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
2210 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
2211
2212 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2213 invert_alpha_coverage ? "~" : "",
2214 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
2215 }
2216
2217 if (s->unknown2_3 || s->unknown2_4) {
2218 pandecode_log(".unknown2_3 = ");
2219
2220 int unknown2_3 = s->unknown2_3;
2221 int unknown2_4 = s->unknown2_4;
2222
2223 /* We're not quite sure what these flags mean without the depth test, if anything */
2224
2225 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2226 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2227 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2228
2229 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2230 }
2231
2232 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2233 pandecode_log_cont(",\n");
2234
2235 pandecode_log(".unknown2_4 = ");
2236 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2237 pandecode_log_cont(",\n");
2238 }
2239
2240 if (s->stencil_mask_front || s->stencil_mask_back) {
2241 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2242 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2243 }
2244
2245 pandecode_stencil("front", &s->stencil_front);
2246 pandecode_stencil("back", &s->stencil_back);
2247
2248 if (is_bifrost) {
2249 pandecode_log(".bifrost2 = {\n");
2250 pandecode_indent++;
2251
2252 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2253 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2254 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2255 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2256
2257 pandecode_indent--;
2258 pandecode_log("},\n");
2259 } else if (s->midgard2.unknown2_7) {
2260 pandecode_log(".midgard2 = {\n");
2261 pandecode_indent++;
2262
2263 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2264 pandecode_indent--;
2265 pandecode_log("},\n");
2266 }
2267
2268 if (s->unknown2_8)
2269 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
2270
2271 if (!is_bifrost) {
2272 /* TODO: Blend shaders routing/disasm */
2273 union midgard_blend blend = s->blend;
2274 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2275 if (shader & ~0xF)
2276 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2277 }
2278
2279 pandecode_indent--;
2280 pandecode_log("};\n");
2281
2282 /* MRT blend fields are used whenever MFBD is used, with
2283 * per-RT descriptors */
2284
2285 if (job_type == JOB_TYPE_TILER && p->framebuffer & MALI_MFBD) {
2286 void* blend_base = (void *) (s + 1);
2287
2288 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2289 mali_ptr shader = 0;
2290
2291 if (is_bifrost)
2292 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2293 else
2294 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2295
2296 if (shader & ~0xF)
2297 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2298
2299 }
2300 }
2301 } else
2302 pandecode_msg("XXX: missing shader descriptor\n");
2303
2304 if (p->viewport) {
2305 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2306 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2307
2308 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2309 pandecode_indent++;
2310
2311 pandecode_prop("clip_minx = %f", f->clip_minx);
2312 pandecode_prop("clip_miny = %f", f->clip_miny);
2313 pandecode_prop("clip_minz = %f", f->clip_minz);
2314 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2315 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2316 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2317
2318 /* Only the higher coordinates are MALI_POSITIVE scaled */
2319
2320 pandecode_prop("viewport0 = { %d, %d }",
2321 f->viewport0[0], f->viewport0[1]);
2322
2323 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2324 f->viewport1[0] + 1, f->viewport1[1] + 1);
2325
2326 pandecode_indent--;
2327 pandecode_log("};\n");
2328 }
2329
2330 unsigned max_attr_index = 0;
2331
2332 if (p->attribute_meta)
2333 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2334
2335 if (p->attributes) {
2336 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2337 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2338 }
2339
2340 /* Varyings are encoded like attributes but not actually sent; we just
2341 * pass a zero buffer with the right stride/size set, (or whatever)
2342 * since the GPU will write to it itself */
2343
2344 if (p->varying_meta) {
2345 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2346 }
2347
2348 if (p->varyings) {
2349 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2350
2351 /* Number of descriptors depends on whether there are
2352 * non-internal varyings */
2353
2354 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2355 }
2356
2357 if (p->uniform_buffers) {
2358 if (uniform_buffer_count)
2359 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2360 else
2361 pandecode_msg("warn: UBOs specified but not referenced\n");
2362 } else if (uniform_buffer_count)
2363 pandecode_msg("XXX: UBOs referenced but not specified\n");
2364
2365 /* We don't want to actually dump uniforms, but we do need to validate
2366 * that the counts we were given are sane */
2367
2368 if (p->uniforms) {
2369 if (uniform_count)
2370 pandecode_uniforms(p->uniforms, uniform_count);
2371 else
2372 pandecode_msg("warn: Uniforms specified but not referenced\n");
2373 } else if (uniform_count)
2374 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2375
2376 if (p->texture_trampoline) {
2377 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
2378
2379 if (mmem) {
2380 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
2381
2382 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
2383 pandecode_indent++;
2384
2385 for (int tex = 0; tex < texture_count; ++tex) {
2386 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2387 char *a = pointer_as_memory_reference(*u);
2388 pandecode_log("%s,\n", a);
2389 free(a);
2390 }
2391
2392 pandecode_indent--;
2393 pandecode_log("};\n");
2394
2395 /* Now, finally, descend down into the texture descriptor */
2396 for (unsigned tex = 0; tex < texture_count; ++tex) {
2397 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2398 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2399 if (tmem)
2400 pandecode_texture(*u, tmem, job_no, tex);
2401 }
2402 }
2403 }
2404
2405 if (p->sampler_descriptor) {
2406 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2407
2408 if (smem) {
2409 struct mali_sampler_descriptor *s;
2410
2411 mali_ptr d = p->sampler_descriptor;
2412
2413 for (int i = 0; i < sampler_count; ++i) {
2414 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2415
2416 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2417 pandecode_indent++;
2418
2419 pandecode_log(".filter_mode = ");
2420 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2421 pandecode_log_cont(",\n");
2422
2423 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2424 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2425
2426 if (s->lod_bias)
2427 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s->lod_bias));
2428
2429 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2430 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2431 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2432
2433 pandecode_prop("compare_func = %s", pandecode_alt_func(s->compare_func));
2434
2435 if (s->zero || s->zero2) {
2436 pandecode_msg("XXX: sampler zero tripped\n");
2437 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2438 }
2439
2440 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2441
2442 pandecode_prop("border_color = { %f, %f, %f, %f }",
2443 s->border_color[0],
2444 s->border_color[1],
2445 s->border_color[2],
2446 s->border_color[3]);
2447
2448 pandecode_indent--;
2449 pandecode_log("};\n");
2450 }
2451 }
2452 }
2453 }
2454
2455 static void
2456 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2457 {
2458 if (p->shader & 0xF)
2459 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2460
2461 if (!(p->position_varying || p->occlusion_counter))
2462 return;
2463
2464 pandecode_log(".postfix = {\n");
2465 pandecode_indent++;
2466
2467 MEMORY_PROP(p, position_varying);
2468 MEMORY_PROP(p, occlusion_counter);
2469
2470 pandecode_indent--;
2471 pandecode_log("},\n");
2472 }
2473
2474 static void
2475 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2476 {
2477 pandecode_log_cont("{\n");
2478 pandecode_indent++;
2479
2480 pandecode_prop("unk2 = 0x%x", v->unk2);
2481
2482 if (v->zero0 || v->zero1) {
2483 pandecode_msg("XXX: vertex only zero tripped");
2484 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2485 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2486 }
2487
2488 pandecode_indent--;
2489 pandecode_log("}\n");
2490 }
2491
2492 static void
2493 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2494 {
2495
2496 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2497 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2498
2499 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2500 pandecode_indent++;
2501
2502 if (h->zero) {
2503 pandecode_msg("XXX: tiler heap zero tripped\n");
2504 pandecode_prop("zero = 0x%x", h->zero);
2505 }
2506
2507 for (int i = 0; i < 12; i++) {
2508 if (h->zeros[i] != 0) {
2509 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2510 i, h->zeros[i]);
2511 }
2512 }
2513
2514 pandecode_prop("heap_size = 0x%x", h->heap_size);
2515 MEMORY_PROP(h, tiler_heap_start);
2516 MEMORY_PROP(h, tiler_heap_free);
2517
2518 /* this might point to the beginning of another buffer, when it's
2519 * really the end of the tiler heap buffer, so we have to be careful
2520 * here. but for zero length, we need the same pointer.
2521 */
2522
2523 if (h->tiler_heap_end == h->tiler_heap_start) {
2524 MEMORY_PROP(h, tiler_heap_start);
2525 } else {
2526 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2527 pandecode_prop("tiler_heap_end = %s + 1", a);
2528 free(a);
2529 }
2530
2531 pandecode_indent--;
2532 pandecode_log("};\n");
2533 }
2534
2535 static void
2536 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2537 {
2538 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2539 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2540
2541 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2542
2543 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2544 pandecode_indent++;
2545
2546 if (t->zero0 || t->zero1) {
2547 pandecode_msg("XXX: tiler meta zero tripped\n");
2548 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2549 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2550 }
2551
2552 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2553 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2554
2555 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2556 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2557
2558 for (int i = 0; i < 12; i++) {
2559 if (t->zeros[i] != 0) {
2560 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2561 i, t->zeros[i]);
2562 }
2563 }
2564
2565 pandecode_indent--;
2566 pandecode_log("};\n");
2567 }
2568
2569 static void
2570 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2571 {
2572 pandecode_log(".gl_enables = ");
2573
2574 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2575
2576 pandecode_log_cont(",\n");
2577 }
2578
2579 static void
2580 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2581 {
2582 if (u.pointer == 0x0)
2583 return;
2584
2585 pandecode_log(".primitive_size = {\n");
2586 pandecode_indent++;
2587
2588 if (constant) {
2589 pandecode_prop("constant = %f", u.constant);
2590 } else {
2591 MEMORY_PROP((&u), pointer);
2592 }
2593
2594 pandecode_indent--;
2595 pandecode_log("},\n");
2596 }
2597
2598 static void
2599 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2600 {
2601 pandecode_log_cont("{\n");
2602 pandecode_indent++;
2603
2604 /* TODO: gl_PointSize on Bifrost */
2605 pandecode_primitive_size(t->primitive_size, true);
2606
2607 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2608
2609 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2610 || t->zero6 || t->zero7 || t->zero8) {
2611 pandecode_msg("XXX: tiler only zero tripped\n");
2612 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2613 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2614 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2615 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2616 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2617 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2618 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2619 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2620 }
2621
2622 pandecode_indent--;
2623 pandecode_log("},\n");
2624 }
2625
2626 static int
2627 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2628 const struct pandecode_mapped_memory *mem,
2629 mali_ptr payload, int job_no, unsigned gpu_id)
2630 {
2631 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2632
2633 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
2634
2635 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2636 pandecode_indent++;
2637
2638 pandecode_log(".prefix = ");
2639 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2640
2641 pandecode_log(".vertex = ");
2642 pandecode_vertex_only_bfr(&v->vertex);
2643
2644 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2645
2646 pandecode_indent--;
2647 pandecode_log("};\n");
2648
2649 return sizeof(*v);
2650 }
2651
2652 static int
2653 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2654 const struct pandecode_mapped_memory *mem,
2655 mali_ptr payload, int job_no, unsigned gpu_id)
2656 {
2657 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2658
2659 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
2660 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2661
2662 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2663 pandecode_indent++;
2664
2665 pandecode_log(".prefix = ");
2666 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2667
2668 pandecode_log(".tiler = ");
2669 pandecode_tiler_only_bfr(&t->tiler, job_no);
2670
2671 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2672
2673 pandecode_indent--;
2674 pandecode_log("};\n");
2675
2676 return sizeof(*t);
2677 }
2678
2679 static int
2680 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2681 const struct pandecode_mapped_memory *mem,
2682 mali_ptr payload, int job_no, unsigned gpu_id)
2683 {
2684 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2685
2686 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
2687
2688 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2689 pandecode_indent++;
2690
2691 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2692 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2693
2694 bool instanced = v->instance_shift || v->instance_odd;
2695 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2696
2697 pandecode_log(".prefix = ");
2698 pandecode_vertex_tiler_prefix(&v->prefix, job_no, !instanced && is_graphics);
2699
2700 pandecode_gl_enables(v->gl_enables, h->job_type);
2701
2702 if (v->instance_shift || v->instance_odd) {
2703 pandecode_prop("instance_shift = 0x%d /* %d */",
2704 v->instance_shift, 1 << v->instance_shift);
2705 pandecode_prop("instance_odd = 0x%X /* %d */",
2706 v->instance_odd, (2 * v->instance_odd) + 1);
2707
2708 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2709 }
2710
2711 if (v->offset_start)
2712 pandecode_prop("offset_start = %d", v->offset_start);
2713
2714 if (v->zero5) {
2715 pandecode_msg("XXX: midgard payload zero tripped\n");
2716 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2717 }
2718
2719 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2720
2721 pandecode_indent--;
2722 pandecode_log("};\n");
2723
2724 return sizeof(*v);
2725 }
2726
2727 static int
2728 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2729 mali_ptr payload, int job_no,
2730 bool is_bifrost, unsigned gpu_id)
2731 {
2732 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2733
2734 bool is_mfbd = s->framebuffer & MALI_MFBD;
2735
2736 /* Bifrost theoretically may retain support for SFBD on compute jobs,
2737 * but for graphics workloads with a FRAGMENT payload, use MFBD */
2738
2739 if (!is_mfbd && is_bifrost)
2740 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2741
2742 struct pandecode_fbd info;
2743
2744 if (is_mfbd)
2745 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2746 else
2747 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
2748
2749 /* Compute the tag for the tagged pointer. This contains the type of
2750 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2751 * additional structures follow the MFBD header (an extra payload or
2752 * not, as well as a count of render targets) */
2753
2754 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
2755
2756 if (is_mfbd) {
2757 if (info.has_extra)
2758 expected_tag |= MALI_MFBD_TAG_EXTRA;
2759
2760 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
2761 }
2762
2763 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
2764 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2765 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
2766 pandecode_prop("max_tile_coord = 0x%X\n", s->min_tile_coord);
2767 }
2768
2769 /* Extract tile coordinates */
2770
2771 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
2772 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
2773
2774 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2775 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2776
2777 /* For the max, we also want the floored (rather than ceiled) version for checking */
2778
2779 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
2780 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
2781
2782 /* Validate the coordinates are well-ordered */
2783
2784 if (min_x == max_x)
2785 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2786 else if (min_x > max_x)
2787 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2788
2789 if (min_y == max_y)
2790 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2791 else if (min_y > max_y)
2792 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2793
2794 /* Validate the coordinates fit inside the framebuffer. We use floor,
2795 * rather than ceil, for the max coordinates, since the tile
2796 * coordinates for something like an 800x600 framebuffer will actually
2797 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2798
2799 if ((min_x > info.width) || (max_x_f > info.width))
2800 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2801
2802 if ((min_y > info.height) || (max_y_f > info.height))
2803 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2804
2805 /* After validation, we print */
2806
2807 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
2808
2809 /* The FBD is a tagged pointer */
2810
2811 unsigned tag = (s->framebuffer & ~FBD_MASK);
2812
2813 if (tag != expected_tag)
2814 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
2815
2816 return sizeof(*s);
2817 }
2818
2819 static int job_descriptor_number = 0;
2820
2821 int
2822 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id)
2823 {
2824 struct mali_job_descriptor_header *h;
2825
2826 int start_number = 0;
2827
2828 bool first = true;
2829 bool last_size;
2830
2831 do {
2832 struct pandecode_mapped_memory *mem =
2833 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2834
2835 void *payload;
2836
2837 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2838
2839 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2840 * high 32-bits of the 64-bit pointer are reused to store
2841 * something else.
2842 */
2843 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2844 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2845 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2846
2847 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
2848
2849 int job_no = job_descriptor_number++;
2850
2851 if (first)
2852 start_number = job_no;
2853
2854 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2855 pandecode_indent++;
2856
2857 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2858
2859 /* Save for next job fixing */
2860 last_size = h->job_descriptor_size;
2861
2862 if (h->job_descriptor_size)
2863 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2864
2865 if (h->exception_status && h->exception_status != 0x1)
2866 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2867 h->exception_status,
2868 (h->exception_status >> 16) & 0xFFFF,
2869 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2870 h->exception_status & 0xFF);
2871
2872 if (h->first_incomplete_task)
2873 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2874
2875 if (h->fault_pointer)
2876 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2877
2878 if (h->job_barrier)
2879 pandecode_prop("job_barrier = %d", h->job_barrier);
2880
2881 pandecode_prop("job_index = %d", h->job_index);
2882
2883 if (h->unknown_flags)
2884 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2885
2886 if (h->job_dependency_index_1)
2887 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2888
2889 if (h->job_dependency_index_2)
2890 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2891
2892 pandecode_indent--;
2893 pandecode_log("};\n");
2894
2895 /* Do not touch the field yet -- decode the payload first, and
2896 * don't touch that either. This is essential for the uploads
2897 * to occur in sequence and therefore be dynamically allocated
2898 * correctly. Do note the size, however, for that related
2899 * reason. */
2900
2901 switch (h->job_type) {
2902 case JOB_TYPE_WRITE_VALUE: {
2903 struct mali_payload_write_value *s = payload;
2904 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2905 pandecode_indent++;
2906 MEMORY_PROP(s, address);
2907
2908 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
2909 pandecode_msg("XXX: unknown value descriptor\n");
2910 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
2911 }
2912
2913 if (s->reserved) {
2914 pandecode_msg("XXX: set value tripped\n");
2915 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
2916 }
2917
2918 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
2919 pandecode_indent--;
2920 pandecode_log("};\n");
2921
2922 break;
2923 }
2924
2925 case JOB_TYPE_TILER:
2926 case JOB_TYPE_VERTEX:
2927 case JOB_TYPE_COMPUTE:
2928 if (bifrost) {
2929 if (h->job_type == JOB_TYPE_TILER)
2930 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2931 else
2932 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2933 } else
2934 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
2935
2936 break;
2937
2938 case JOB_TYPE_FRAGMENT:
2939 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
2940 break;
2941
2942 default:
2943 break;
2944 }
2945
2946 /* Handle linkage */
2947
2948 if (!first) {
2949 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2950
2951 if (last_size)
2952 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no);
2953 else
2954 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no);
2955 }
2956
2957 first = false;
2958
2959 } while ((jc_gpu_va = h->job_descriptor_size ? h->next_job_64 : h->next_job_32));
2960
2961 return start_number;
2962 }