projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
panfrost: Add checksum fields to SFBD descriptor
[mesa.git] / src / panfrost / pandecode / decode.c
1 /*
2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 */
25
26 #include <panfrost-job.h>
27 #include <stdio.h>
28 #include <stdlib.h>
29 #include <memory.h>
30 #include <stdbool.h>
31 #include <stdarg.h>
32 #include <ctype.h>
33 #include "decode.h"
34 #include "util/macros.h"
35 #include "util/u_math.h"
36
37 #include "pan_pretty_print.h"
38 #include "midgard/disassemble.h"
39 #include "bifrost/disassemble.h"
40
41 #include "pan_encoder.h"
42
43 int pandecode_jc(mali_ptr jc_gpu_va, bool bifrost);
44
45 #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_FRAMEBUFFER_##flag, "MALI_FRAMEBUFFER_" #flag }
242 static const struct pandecode_flag_info fb_fmt_flag_info[] = {
243 FLAG_INFO(MSAA_A),
244 FLAG_INFO(MSAA_B),
245 FLAG_INFO(MSAA_8),
246 {}
247 };
248 #undef FLAG_INFO
249
250 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
251 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
252 FLAG_INFO(MSAA),
253 FLAG_INFO(SRGB),
254 {}
255 };
256 #undef FLAG_INFO
257
258 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
259 static const struct pandecode_flag_info mfbd_extra_flag_info[] = {
260 FLAG_INFO(PRESENT),
261 FLAG_INFO(AFBC),
262 FLAG_INFO(ZS),
263 {}
264 };
265 #undef FLAG_INFO
266
267 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
268 static const struct pandecode_flag_info shader_midgard1_flag_info [] = {
269 FLAG_INFO(EARLY_Z),
270 FLAG_INFO(READS_TILEBUFFER),
271 FLAG_INFO(READS_ZS),
272 {}
273 };
274 #undef FLAG_INFO
275
276 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
277 static const struct pandecode_flag_info mfbd_flag_info [] = {
278 FLAG_INFO(DEPTH_WRITE),
279 FLAG_INFO(EXTRA),
280 {}
281 };
282 #undef FLAG_INFO
283
284 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
285 static const struct pandecode_flag_info sampler_flag_info [] = {
286 FLAG_INFO(MAG_NEAREST),
287 FLAG_INFO(MIN_NEAREST),
288 FLAG_INFO(MIP_LINEAR_1),
289 FLAG_INFO(MIP_LINEAR_2),
290 FLAG_INFO(NORM_COORDS),
291 {}
292 };
293 #undef FLAG_INFO
294
295 extern char *replace_fragment;
296 extern char *replace_vertex;
297
298 static char *
299 pandecode_job_type(enum mali_job_type type)
300 {
301 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
302
303 switch (type) {
304 DEFINE_CASE(NULL);
305 DEFINE_CASE(SET_VALUE);
306 DEFINE_CASE(CACHE_FLUSH);
307 DEFINE_CASE(COMPUTE);
308 DEFINE_CASE(VERTEX);
309 DEFINE_CASE(TILER);
310 DEFINE_CASE(FUSED);
311 DEFINE_CASE(FRAGMENT);
312
313 case JOB_NOT_STARTED:
314 return "NOT_STARTED";
315
316 default:
317 pandecode_log("Warning! Unknown job type %x\n", type);
318 return "!?!?!?";
319 }
320
321 #undef DEFINE_CASE
322 }
323
324 static char *
325 pandecode_draw_mode(enum mali_draw_mode mode)
326 {
327 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
328
329 switch (mode) {
330 DEFINE_CASE(DRAW_NONE);
331 DEFINE_CASE(POINTS);
332 DEFINE_CASE(LINES);
333 DEFINE_CASE(TRIANGLES);
334 DEFINE_CASE(TRIANGLE_STRIP);
335 DEFINE_CASE(TRIANGLE_FAN);
336 DEFINE_CASE(LINE_STRIP);
337 DEFINE_CASE(LINE_LOOP);
338 DEFINE_CASE(POLYGON);
339 DEFINE_CASE(QUADS);
340 DEFINE_CASE(QUAD_STRIP);
341
342 default:
343 pandecode_msg("XXX: invalid draw mode %X\n", mode);
344 return "";
345 }
346
347 #undef DEFINE_CASE
348 }
349
350 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
351 static char *
352 pandecode_func(enum mali_func mode)
353 {
354 switch (mode) {
355 DEFINE_CASE(NEVER);
356 DEFINE_CASE(LESS);
357 DEFINE_CASE(EQUAL);
358 DEFINE_CASE(LEQUAL);
359 DEFINE_CASE(GREATER);
360 DEFINE_CASE(NOTEQUAL);
361 DEFINE_CASE(GEQUAL);
362 DEFINE_CASE(ALWAYS);
363
364 default:
365 pandecode_msg("XXX: invalid func %X\n", mode);
366 return "";
367 }
368 }
369 #undef DEFINE_CASE
370
371 /* Why is this duplicated? Who knows... */
372 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
373 static char *
374 pandecode_alt_func(enum mali_alt_func mode)
375 {
376 switch (mode) {
377 DEFINE_CASE(NEVER);
378 DEFINE_CASE(LESS);
379 DEFINE_CASE(EQUAL);
380 DEFINE_CASE(LEQUAL);
381 DEFINE_CASE(GREATER);
382 DEFINE_CASE(NOTEQUAL);
383 DEFINE_CASE(GEQUAL);
384 DEFINE_CASE(ALWAYS);
385
386 default:
387 pandecode_msg("XXX: invalid alt func %X\n", mode);
388 return "";
389 }
390 }
391 #undef DEFINE_CASE
392
393 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
394 static char *
395 pandecode_stencil_op(enum mali_stencil_op op)
396 {
397 switch (op) {
398 DEFINE_CASE(KEEP);
399 DEFINE_CASE(REPLACE);
400 DEFINE_CASE(ZERO);
401 DEFINE_CASE(INVERT);
402 DEFINE_CASE(INCR_WRAP);
403 DEFINE_CASE(DECR_WRAP);
404 DEFINE_CASE(INCR);
405 DEFINE_CASE(DECR);
406
407 default:
408 pandecode_msg("XXX: invalid stencil op %X\n", op);
409 return "";
410 }
411 }
412
413 #undef DEFINE_CASE
414
415 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
416 {
417 switch(mode) {
418 /* TODO: Combine to just "instanced" once this can be done
419 * unambiguously in all known cases */
420 case MALI_ATTR_POT_DIVIDE:
421 return "instanced_pot";
422 case MALI_ATTR_MODULO:
423 return "instanced_mod";
424 case MALI_ATTR_NPOT_DIVIDE:
425 return "instanced_npot";
426 case MALI_ATTR_IMAGE:
427 return "image";
428 case MALI_ATTR_INTERNAL:
429 return "internal";
430 default:
431 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
432 return "";
433 }
434 }
435
436 static const char *
437 pandecode_special_varying(uint64_t v)
438 {
439 switch(v) {
440 case MALI_VARYING_FRAG_COORD:
441 return "gl_FragCoord";
442 case MALI_VARYING_FRONT_FACING:
443 return "gl_FrontFacing";
444 case MALI_VARYING_POINT_COORD:
445 return "gl_PointCoord";
446 default:
447 pandecode_msg("XXX: invalid special varying %" PRIx64 "\n", v);
448 return "";
449 }
450 }
451
452 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
453 static char *
454 pandecode_wrap_mode(enum mali_wrap_mode op)
455 {
456 switch (op) {
457 DEFINE_CASE(REPEAT);
458 DEFINE_CASE(CLAMP_TO_EDGE);
459 DEFINE_CASE(CLAMP_TO_BORDER);
460 DEFINE_CASE(MIRRORED_REPEAT);
461
462 default:
463 pandecode_msg("XXX: invalid wrap mode %X\n", op);
464 return "";
465 }
466 }
467 #undef DEFINE_CASE
468
469 #define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
470 static char *
471 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt)
472 {
473 switch (fmt) {
474 DEFINE_CASE(TILED);
475 DEFINE_CASE(UNKNOWN);
476 DEFINE_CASE(LINEAR);
477 DEFINE_CASE(AFBC);
478
479 default:
480 unreachable("Invalid case");
481 }
482 }
483 #undef DEFINE_CASE
484
485 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
486 static char *
487 pandecode_exception_access(enum mali_exception_access access)
488 {
489 switch (access) {
490 DEFINE_CASE(NONE);
491 DEFINE_CASE(EXECUTE);
492 DEFINE_CASE(READ);
493 DEFINE_CASE(WRITE);
494
495 default:
496 unreachable("Invalid case");
497 }
498 }
499 #undef DEFINE_CASE
500
501 /* Midgard's tiler descriptor is embedded within the
502 * larger FBD */
503
504 static void
505 pandecode_midgard_tiler_descriptor(
506 const struct midgard_tiler_descriptor *t,
507 unsigned width,
508 unsigned height,
509 bool is_fragment)
510 {
511 pandecode_log(".tiler = {\n");
512 pandecode_indent++;
513
514 if (t->hierarchy_mask == MALI_TILER_DISABLED)
515 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
516 else
517 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
518
519 /* We know this name from the kernel, but we never see it nonzero */
520
521 if (t->flags)
522 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
523
524 MEMORY_PROP(t, polygon_list);
525
526 /* The body is offset from the base of the polygon list */
527 assert(t->polygon_list_body > t->polygon_list);
528 unsigned body_offset = t->polygon_list_body - t->polygon_list;
529
530 /* It needs to fit inside the reported size */
531 assert(t->polygon_list_size >= body_offset);
532
533 /* Check that we fit */
534 struct pandecode_mapped_memory *plist =
535 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
536
537 assert(t->polygon_list_size <= plist->length);
538
539 /* Now that we've sanity checked, we'll try to calculate the sizes
540 * ourselves for comparison */
541
542 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask);
543 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask);
544
545 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
546 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
547 ref_header, ref_size);
548 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
549 pandecode_msg("body offset %d\n", body_offset);
550 }
551
552 /* The tiler heap has a start and end specified -- it should be
553 * identical to what we have in the BO. The exception is if tiling is
554 * disabled. */
555
556 MEMORY_PROP(t, heap_start);
557 assert(t->heap_end >= t->heap_start);
558
559 struct pandecode_mapped_memory *heap =
560 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
561
562 unsigned heap_size = t->heap_end - t->heap_start;
563
564 /* Tiling is enabled with a special flag */
565 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
566 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
567
568 bool tiling_enabled = hierarchy_mask;
569
570 if (tiling_enabled) {
571 /* When tiling is enabled, the heap should be a tight fit */
572 unsigned heap_offset = t->heap_start - heap->gpu_va;
573 if ((heap_offset + heap_size) != heap->length) {
574 pandecode_msg("XXX: heap size %u (expected %zu)\n",
575 heap_size, heap->length - heap_offset);
576 }
577
578 /* We should also have no other flags */
579 if (tiler_flags)
580 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
581 } else {
582 /* When tiling is disabled, we should have that flag and no others */
583
584 if (tiler_flags != MALI_TILER_DISABLED) {
585 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
586 tiler_flags);
587 }
588
589 /* We should also have an empty heap */
590 if (heap_size) {
591 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
592 heap_size);
593 }
594
595 /* Disabled tiling is used only for clear-only jobs, which are
596 * purely FRAGMENT, so we should never see this for
597 * non-FRAGMENT descriptors. */
598
599 if (!is_fragment)
600 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
601 }
602
603 /* We've never seen weights used in practice, but we know from the
604 * kernel these fields is there */
605
606 bool nonzero_weights = false;
607
608 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
609 nonzero_weights |= t->weights[w] != 0x0;
610 }
611
612 if (nonzero_weights) {
613 pandecode_log(".weights = {");
614
615 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
616 pandecode_log("%d, ", t->weights[w]);
617 }
618
619 pandecode_log("},");
620 }
621
622 pandecode_indent--;
623 pandecode_log("}\n");
624 }
625
626 /* Information about the framebuffer passed back for
627 * additional analysis */
628
629 struct pandecode_fbd {
630 unsigned width;
631 unsigned height;
632 unsigned rt_count;
633 bool has_extra;
634 };
635
636 static struct pandecode_fbd
637 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment)
638 {
639 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
640 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
641
642 struct pandecode_fbd info = {
643 .has_extra = false,
644 .rt_count = 1
645 };
646
647 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
648 pandecode_indent++;
649
650 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
651 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
652
653 pandecode_log(".format = ");
654 pandecode_log_decoded_flags(fb_fmt_flag_info, s->format);
655 pandecode_log_cont(",\n");
656
657 info.width = s->width + 1;
658 info.height = s->height + 1;
659
660 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
661 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
662
663 MEMORY_PROP(s, checksum);
664
665 if (s->checksum_stride)
666 pandecode_prop("checksum_stride = %d", s->checksum_stride);
667
668 MEMORY_PROP(s, framebuffer);
669 pandecode_prop("stride = %d", s->stride);
670
671 /* Earlier in the actual commandstream -- right before width -- but we
672 * delay to flow nicer */
673
674 pandecode_log(".clear_flags = ");
675 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
676 pandecode_log_cont(",\n");
677
678 if (s->depth_buffer | s->depth_buffer_enable) {
679 MEMORY_PROP(s, depth_buffer);
680 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s->depth_buffer_enable));
681 }
682
683 if (s->stencil_buffer | s->stencil_buffer_enable) {
684 MEMORY_PROP(s, stencil_buffer);
685 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s->stencil_buffer_enable));
686 }
687
688 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
689 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
690 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
691 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
692 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
693 }
694
695 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
696 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
697 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
698 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
699 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
700 }
701
702 if (s->clear_stencil) {
703 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
704 }
705
706 MEMORY_PROP(s, unknown_address_0);
707 const struct midgard_tiler_descriptor t = s->tiler;
708 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment);
709
710 pandecode_indent--;
711 pandecode_log("};\n");
712
713 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
714 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
715 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
716 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
717 pandecode_prop("zero5 = 0x%" PRIx32, s->zero4);
718
719 printf(".zero3 = {");
720
721 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
722 printf("%X, ", s->zero3[i]);
723
724 printf("},\n");
725
726 printf(".zero6 = {");
727
728 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
729 printf("%X, ", s->zero6[i]);
730
731 printf("},\n");
732
733 return info;
734 }
735
736 static void
737 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
738 {
739 pandecode_log(".unknown%d = {", name);
740
741 for (int i = 0; i < count; ++i)
742 printf("%X, ", slide[i]);
743
744 pandecode_log("},\n");
745 }
746
747 #define SHORT_SLIDE(num) \
748 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
749
750 static void
751 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
752 {
753 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
754 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
755
756 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
757 pandecode_indent++;
758
759 SHORT_SLIDE(1);
760
761 pandecode_indent--;
762 printf("},\n");
763 }
764
765 /* Extracts the number of components associated with a Mali format */
766
767 static unsigned
768 pandecode_format_component_count(enum mali_format fmt)
769 {
770 /* Mask out the format class */
771 unsigned top = fmt & 0b11100000;
772
773 switch (top) {
774 case MALI_FORMAT_SNORM:
775 case MALI_FORMAT_UINT:
776 case MALI_FORMAT_UNORM:
777 case MALI_FORMAT_SINT:
778 return ((fmt >> 3) & 3) + 1;
779 default:
780 /* TODO: Validate */
781 return 4;
782 }
783 }
784
785 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
786
787 static unsigned
788 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
789 {
790 unsigned mask = 0;
791 assert(MALI_CHANNEL_RED == 0);
792
793 for (unsigned c = 0; c < 4; ++c) {
794 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
795
796 if (chan <= MALI_CHANNEL_ALPHA)
797 mask |= (1 << chan);
798 }
799
800 return mask;
801 }
802
803 /* Validates that a (format, swizzle) pair is valid, in the sense that the
804 * swizzle doesn't access any components that are undefined in the format.
805 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
806 * omitted */
807
808 static bool
809 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
810 {
811 unsigned nr_comp = pandecode_format_component_count(fmt);
812 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
813 unsigned valid_mask = (1 << nr_comp) - 1;
814 unsigned invalid_mask = ~valid_mask;
815
816 if (access_mask & invalid_mask) {
817 pandecode_msg("XXX: invalid components accessed\n");
818 return false;
819 }
820
821 /* Check for the default non-swizzling swizzle so we can suppress
822 * useless printing for the defaults */
823
824 unsigned default_swizzles[4] = {
825 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
826 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
827 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
828 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
829 };
830
831 return (swizzle == default_swizzles[nr_comp - 1]);
832 }
833
834 /* Maps MALI_RGBA32F to rgba32f, etc */
835
836 static void
837 pandecode_format_short(enum mali_format fmt, bool srgb)
838 {
839 /* We want a type-like format, so cut off the initial MALI_ */
840 char *format = pandecode_format(fmt);
841 format += strlen("MALI_");
842
843 unsigned len = strlen(format);
844 char *lower_format = calloc(1, len + 1);
845
846 for (unsigned i = 0; i < len; ++i)
847 lower_format[i] = tolower(format[i]);
848
849 /* Sanity check sRGB flag is applied to RGB, per the name */
850 if (srgb && lower_format[0] != 'r')
851 pandecode_msg("XXX: sRGB applied to non-colour format\n");
852
853 /* Just prefix with an s, so you get formats like srgba8_unorm */
854 if (srgb)
855 pandecode_log_cont("s");
856
857 pandecode_log_cont("%s", lower_format);
858 free(lower_format);
859 }
860
861 static void
862 pandecode_swizzle(unsigned swizzle, enum mali_format format)
863 {
864 /* First, do some validation */
865 bool trivial_swizzle = pandecode_validate_format_swizzle(
866 format, swizzle);
867
868 if (trivial_swizzle)
869 return;
870
871 /* Next, print the swizzle */
872 pandecode_log_cont(".");
873
874 static const char components[] = "rgba01";
875
876 for (unsigned c = 0; c < 4; ++c) {
877 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
878
879 if (chan >= MALI_CHANNEL_RESERVED_0) {
880 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
881 continue;
882 }
883 pandecode_log_cont("%c", components[chan]);
884 }
885 }
886
887 static void
888 pandecode_rt_format(struct mali_rt_format format)
889 {
890 pandecode_log(".format = {\n");
891 pandecode_indent++;
892
893 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
894 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
895 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
896
897 pandecode_prop("block = %s",
898 pandecode_mfbd_block_format(format.block));
899
900 /* TODO: Map formats so we can check swizzles and print nicely */
901 pandecode_log("swizzle");
902 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
903 pandecode_log_cont(",\n");
904
905 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
906 MALI_NEGATIVE(format.nr_channels));
907
908 pandecode_log(".flags = ");
909 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
910 pandecode_log_cont(",\n");
911
912 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
913 * which is a faster hardware-based alternative to the wallpaper method
914 * to preserve framebuffer contents across frames. In practice, MFBD
915 * preload is buggy on Midgard, and so this is a chicken bit. If this
916 * bit isn't set, most likely something broke unrelated to preload */
917
918 if (!format.no_preload) {
919 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
920 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
921 }
922
923 if (format.zero)
924 pandecode_prop("zero = 0x%" PRIx32, format.zero);
925
926 pandecode_indent--;
927 pandecode_log("},\n");
928 }
929
930 static void
931 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
932 {
933 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
934 pandecode_indent++;
935
936 for (int i = 0; i < MALI_NEGATIVE(fb->rt_count_1); i++) {
937 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
938 struct pandecode_mapped_memory *mem =
939 pandecode_find_mapped_gpu_mem_containing(rt_va);
940 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
941
942 pandecode_log("{\n");
943 pandecode_indent++;
944
945 pandecode_rt_format(rt->format);
946
947 if (rt->format.block == MALI_MFBD_BLOCK_AFBC) {
948 pandecode_log(".afbc = {\n");
949 pandecode_indent++;
950
951 char *a = pointer_as_memory_reference(rt->afbc.metadata);
952 pandecode_prop("metadata = %s", a);
953 free(a);
954
955 pandecode_prop("stride = %d", rt->afbc.stride);
956 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
957
958 pandecode_indent--;
959 pandecode_log("},\n");
960 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
961 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
962 rt->afbc.metadata,
963 rt->afbc.stride,
964 rt->afbc.unk);
965 }
966
967 MEMORY_PROP(rt, framebuffer);
968 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
969
970 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
971 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
972 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
973 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
974 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
975 }
976
977 if (rt->zero1 || rt->zero2 || rt->zero3) {
978 pandecode_msg("XXX: render target zeros tripped\n");
979 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
980 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
981 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
982 }
983
984 pandecode_indent--;
985 pandecode_log("},\n");
986 }
987
988 pandecode_indent--;
989 pandecode_log("};\n");
990 }
991
992 static struct pandecode_fbd
993 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment)
994 {
995 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
996 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
997
998 struct pandecode_fbd info;
999
1000 if (fb->sample_locations) {
1001 /* The blob stores all possible sample locations in a single buffer
1002 * allocated on startup, and just switches the pointer when switching
1003 * MSAA state. For now, we just put the data into the cmdstream, but we
1004 * should do something like what the blob does with a real driver.
1005 *
1006 * There seem to be 32 slots for sample locations, followed by another
1007 * 16. The second 16 is just the center location followed by 15 zeros
1008 * in all the cases I've identified (maybe shader vs. depth/color
1009 * samples?).
1010 */
1011
1012 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
1013
1014 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
1015
1016 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
1017 pandecode_indent++;
1018
1019 for (int i = 0; i < 32 + 16; i++) {
1020 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1021 }
1022
1023 pandecode_indent--;
1024 pandecode_log("};\n");
1025 }
1026
1027 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1028 pandecode_indent++;
1029
1030 pandecode_prop("unk0 = 0x%x", fb->unk0);
1031
1032 if (fb->sample_locations)
1033 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1034
1035 /* Assume that unknown1 was emitted in the last job for
1036 * now */
1037 MEMORY_PROP(fb, unknown1);
1038
1039 info.width = fb->width1 + 1;
1040 info.height = fb->height1 + 1;
1041 info.rt_count = fb->rt_count_1 + 1;
1042
1043 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1044 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1045 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1046 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1047
1048 pandecode_prop("unk1 = 0x%x", fb->unk1);
1049 pandecode_prop("unk2 = 0x%x", fb->unk2);
1050 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1051 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1052
1053 pandecode_log(".mfbd_flags = ");
1054 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1055 pandecode_log_cont(",\n");
1056
1057 if (fb->clear_stencil)
1058 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1059
1060 if (fb->clear_depth)
1061 pandecode_prop("clear_depth = %f", fb->clear_depth);
1062
1063 /* TODO: What is this? Let's not blow up.. */
1064 if (fb->unknown2 != 0x1F)
1065 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1066
1067 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1068 MEMORY_PROP(fb, scratchpad);
1069 const struct midgard_tiler_descriptor t = fb->tiler;
1070 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment);
1071
1072 if (fb->zero3 || fb->zero4) {
1073 pandecode_msg("XXX: framebuffer zeros tripped\n");
1074 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1075 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1076 }
1077
1078 pandecode_indent--;
1079 pandecode_log("};\n");
1080
1081 gpu_va += sizeof(struct bifrost_framebuffer);
1082
1083 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1084
1085 if (info.has_extra) {
1086 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1087 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1088
1089 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1090 pandecode_indent++;
1091
1092 MEMORY_PROP(fbx, checksum);
1093
1094 if (fbx->checksum_stride)
1095 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1096
1097 pandecode_log(".flags = ");
1098 pandecode_log_decoded_flags(mfbd_extra_flag_info, fbx->flags);
1099 pandecode_log_cont(",\n");
1100
1101 if (fbx->flags & MALI_EXTRA_AFBC_ZS) {
1102 pandecode_log(".ds_afbc = {\n");
1103 pandecode_indent++;
1104
1105 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1106 pandecode_prop("depth_stencil_afbc_stride = %d",
1107 fbx->ds_afbc.depth_stencil_afbc_stride);
1108 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1109
1110 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
1111 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1112 pandecode_prop("zero1 = 0x%" PRIx32,
1113 fbx->ds_afbc.zero1);
1114 pandecode_prop("padding = 0x%" PRIx64,
1115 fbx->ds_afbc.padding);
1116 }
1117
1118 pandecode_indent--;
1119 pandecode_log("},\n");
1120 } else {
1121 pandecode_log(".ds_linear = {\n");
1122 pandecode_indent++;
1123
1124 if (fbx->ds_linear.depth) {
1125 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1126 pandecode_prop("depth_stride = %d",
1127 fbx->ds_linear.depth_stride);
1128 }
1129
1130 if (fbx->ds_linear.stencil) {
1131 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1132 pandecode_prop("stencil_stride = %d",
1133 fbx->ds_linear.stencil_stride);
1134 }
1135
1136 if (fbx->ds_linear.depth_stride_zero ||
1137 fbx->ds_linear.stencil_stride_zero ||
1138 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1139 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1140 pandecode_prop("depth_stride_zero = 0x%x",
1141 fbx->ds_linear.depth_stride_zero);
1142 pandecode_prop("stencil_stride_zero = 0x%x",
1143 fbx->ds_linear.stencil_stride_zero);
1144 pandecode_prop("zero1 = 0x%" PRIx32,
1145 fbx->ds_linear.zero1);
1146 pandecode_prop("zero2 = 0x%" PRIx32,
1147 fbx->ds_linear.zero2);
1148 }
1149
1150 pandecode_indent--;
1151 pandecode_log("},\n");
1152 }
1153
1154 if (fbx->zero3 || fbx->zero4) {
1155 pandecode_msg("XXX: fb_extra zeros tripped\n");
1156 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1157 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1158 }
1159
1160 pandecode_indent--;
1161 pandecode_log("};\n");
1162
1163 gpu_va += sizeof(struct bifrost_fb_extra);
1164 }
1165
1166 if (is_fragment)
1167 pandecode_render_target(gpu_va, job_no, fb);
1168
1169 return info;
1170 }
1171
1172 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1173 * count */
1174
1175 static void
1176 pandecode_padded_vertices(unsigned shift, unsigned k)
1177 {
1178 unsigned odd = 2*k + 1;
1179 unsigned pot = 1 << shift;
1180 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1181 }
1182
1183 /* Given a magic divisor, recover what we were trying to divide by.
1184 *
1185 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1186 * 0 <= m < 2^N, for N bits in m.
1187 *
1188 * Let q represent the number we would like to divide by.
1189 *
1190 * By definition of a magic divisor for N-bit unsigned integers (a number you
1191 * multiply by to magically get division), m is a number such that:
1192 *
1193 * (m * x) & (2^N - 1) = floor(x/q).
1194 * for all x on Z where 0 <= x < 2^N
1195 *
1196 * Ignore the case where any of the above values equals zero; it is irrelevant
1197 * for our purposes (instanced arrays).
1198 *
1199 * Choose x = q. Then:
1200 *
1201 * (m * x) & (2^N - 1) = floor(x/q).
1202 * (m * q) & (2^N - 1) = floor(q/q).
1203 *
1204 * floor(q/q) = floor(1) = 1, therefore:
1205 *
1206 * (m * q) & (2^N - 1) = 1
1207 *
1208 * Recall the identity that the bitwise AND of one less than a power-of-two
1209 * equals the modulo with that power of two, i.e. for all x:
1210 *
1211 * x & (2^N - 1) = x % N
1212 *
1213 * Therefore:
1214 *
1215 * mq % (2^N) = 1
1216 *
1217 * By definition, a modular multiplicative inverse of a number m is the number
1218 * q such that with respect to a modulos M:
1219 *
1220 * mq % M = 1
1221 *
1222 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1223 *
1224 */
1225
1226 static void
1227 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1228 {
1229 #if 0
1230 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1231 * shift) the most lame way possible... just repeatedly add.
1232 * Asymptoptically slow but nobody cares in practice, unless you have
1233 * massive numbers of vertices or high divisors. */
1234
1235 unsigned inverse = 0;
1236
1237 /* Magic implicitly has the highest bit set */
1238 magic |= (1 << 31);
1239
1240 /* Depending on rounding direction */
1241 if (extra)
1242 magic++;
1243
1244 for (;;) {
1245 uint32_t product = magic * inverse;
1246
1247 if (shift) {
1248 product >>= shift;
1249 }
1250
1251 if (product == 1)
1252 break;
1253
1254 ++inverse;
1255 }
1256
1257 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1258
1259 /* Recall we're supposed to divide by (gl_level_divisor *
1260 * padded_num_vertices) */
1261
1262 unsigned padded_num_vertices = inverse / orig_divisor;
1263
1264 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1265 #endif
1266 }
1267
1268 static void
1269 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1270 mali_ptr addr, int job_no, char *suffix,
1271 int count, bool varying, enum mali_job_type job_type)
1272 {
1273 char *prefix = varying ? "varying" : "attribute";
1274 assert(addr);
1275
1276 if (!count) {
1277 pandecode_msg("warn: No %s records\n", prefix);
1278 return;
1279 }
1280
1281 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1282
1283 for (int i = 0; i < count; ++i) {
1284 /* First, check for special records */
1285 if (attr[i].elements < MALI_VARYING_SPECIAL) {
1286 /* Special records are always varyings */
1287
1288 if (!varying)
1289 pandecode_msg("XXX: Special varying in attribute field\n");
1290
1291 if (job_type != JOB_TYPE_TILER)
1292 pandecode_msg("XXX: Special varying in non-FS\n");
1293
1294 /* We're special, so all fields should be zero */
1295 unsigned zero = attr[i].stride | attr[i].size;
1296 zero |= attr[i].shift | attr[i].extra_flags;
1297
1298 if (zero)
1299 pandecode_msg("XXX: Special varying has non-zero fields\n");
1300 else {
1301 /* Print the special varying name */
1302 pandecode_log("varying_%d = %s;", i, pandecode_special_varying(attr[i].elements));
1303 continue;
1304 }
1305 }
1306
1307 enum mali_attr_mode mode = attr[i].elements & 7;
1308
1309 if (mode == MALI_ATTR_UNUSED)
1310 pandecode_msg("XXX: unused attribute record\n");
1311
1312 /* For non-linear records, we need to print the type of record */
1313 if (mode != MALI_ATTR_LINEAR)
1314 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1315
1316 /* Print the name to link with attr_meta */
1317 pandecode_log_cont("%s_%d", prefix, i);
1318
1319 /* Print the stride and size */
1320 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1321
1322 /* TODO: Sanity check the quotient itself. It must be equal to
1323 * (or be greater than, if the driver added padding) the padded
1324 * vertex count. */
1325
1326 /* Finally, print the pointer */
1327 mali_ptr raw_elements = attr[i].elements & ~7;
1328 char *a = pointer_as_memory_reference(raw_elements);
1329 pandecode_log_cont(" = (%s);\n", a);
1330 free(a);
1331
1332 /* Check the pointer */
1333 pandecode_validate_buffer(raw_elements, attr[i].size);
1334
1335 /* shift/extra_flags exist only for instanced */
1336 if (attr[i].shift | attr[i].extra_flags) {
1337 /* These are set to random values by the blob for
1338 * varyings, most likely a symptom of uninitialized
1339 * memory where the hardware masked the bug. As such we
1340 * put this at a warning, not an error. */
1341
1342 if (mode == MALI_ATTR_LINEAR)
1343 pandecode_msg("warn: instancing fields set for linear\n");
1344
1345 pandecode_prop("shift = %d", attr[i].shift);
1346 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1347 }
1348
1349 /* Decode further where possible */
1350
1351 if (mode == MALI_ATTR_MODULO) {
1352 pandecode_padded_vertices(
1353 attr[i].shift,
1354 attr[i].extra_flags);
1355 }
1356
1357 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1358 i++;
1359 pandecode_log("{\n");
1360 pandecode_indent++;
1361 pandecode_prop("unk = 0x%x", attr[i].unk);
1362 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1363 if (attr[i].zero != 0)
1364 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1365 pandecode_prop("divisor = %d", attr[i].divisor);
1366 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1367 pandecode_indent--;
1368 pandecode_log("}, \n");
1369 }
1370
1371 }
1372
1373 pandecode_log("\n");
1374 }
1375
1376 static mali_ptr
1377 pandecode_shader_address(const char *name, mali_ptr ptr)
1378 {
1379 /* TODO: Decode flags */
1380 mali_ptr shader_ptr = ptr & ~15;
1381
1382 char *a = pointer_as_memory_reference(shader_ptr);
1383 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1384 free(a);
1385
1386 return shader_ptr;
1387 }
1388
1389 static void
1390 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1391 {
1392 unsigned any_nonzero =
1393 stencil->ref | stencil->mask | stencil->func |
1394 stencil->sfail | stencil->dpfail | stencil->dppass;
1395
1396 if (any_nonzero == 0)
1397 return;
1398
1399 const char *func = pandecode_func(stencil->func);
1400 const char *sfail = pandecode_stencil_op(stencil->sfail);
1401 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1402 const char *dppass = pandecode_stencil_op(stencil->dppass);
1403
1404 if (stencil->zero)
1405 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1406
1407 pandecode_log(".stencil_%s = {\n", name);
1408 pandecode_indent++;
1409 pandecode_prop("ref = %d", stencil->ref);
1410 pandecode_prop("mask = 0x%02X", stencil->mask);
1411 pandecode_prop("func = %s", func);
1412 pandecode_prop("sfail = %s", sfail);
1413 pandecode_prop("dpfail = %s", dpfail);
1414 pandecode_prop("dppass = %s", dppass);
1415 pandecode_indent--;
1416 pandecode_log("},\n");
1417 }
1418
1419 static void
1420 pandecode_blend_equation(const struct mali_blend_equation *blend)
1421 {
1422 if (blend->zero1)
1423 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1424
1425 pandecode_log(".equation = {\n");
1426 pandecode_indent++;
1427
1428 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1429 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1430
1431 pandecode_log(".color_mask = ");
1432 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1433 pandecode_log_cont(",\n");
1434
1435 pandecode_indent--;
1436 pandecode_log("},\n");
1437 }
1438
1439 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1440
1441 static unsigned
1442 decode_bifrost_constant(u16 constant)
1443 {
1444 float lo = (float) (constant & 0xFF);
1445 float hi = (float) (constant >> 8);
1446
1447 return (hi / 255.0) + (lo / 65535.0);
1448 }
1449
1450 static mali_ptr
1451 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1452 {
1453 struct bifrost_blend_rt *b =
1454 ((struct bifrost_blend_rt *) descs) + rt_no;
1455
1456 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1457 pandecode_indent++;
1458
1459 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1460 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1461 b->constant, decode_bifrost_constant(b->constant));
1462
1463 /* TODO figure out blend shader enable bit */
1464 pandecode_blend_equation(&b->equation);
1465 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1466 pandecode_prop("index = 0x%" PRIx16, b->index);
1467 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1468
1469 pandecode_indent--;
1470 pandecode_log("},\n");
1471
1472 return 0;
1473 }
1474
1475 static mali_ptr
1476 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1477 {
1478 /* constant/equation is in a union */
1479 if (!blend->shader)
1480 return 0;
1481
1482 pandecode_log(".blend = {\n");
1483 pandecode_indent++;
1484
1485 if (is_shader) {
1486 pandecode_shader_address("shader", blend->shader);
1487 } else {
1488 pandecode_blend_equation(&blend->equation);
1489 pandecode_prop("constant = %f", blend->constant);
1490 }
1491
1492 pandecode_indent--;
1493 pandecode_log("},\n");
1494
1495 /* Return blend shader to disassemble if present */
1496 return is_shader ? (blend->shader & ~0xF) : 0;
1497 }
1498
1499 static mali_ptr
1500 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1501 {
1502 struct midgard_blend_rt *b =
1503 ((struct midgard_blend_rt *) descs) + rt_no;
1504
1505 /* Flags determine presence of blend shader */
1506 bool is_shader = (b->flags & 0xF) >= 0x2;
1507
1508 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1509 pandecode_indent++;
1510
1511 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1512
1513 union midgard_blend blend = b->blend;
1514 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1515
1516 pandecode_indent--;
1517 pandecode_log("};\n");
1518
1519 return shader;
1520 }
1521
1522 /* Attributes and varyings have descriptor records, which contain information
1523 * about their format and ordering with the attribute/varying buffers. We'll
1524 * want to validate that the combinations specified are self-consistent.
1525 */
1526
1527 static int
1528 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1529 {
1530 char base[128];
1531 char *prefix = varying ? "varying" : "attribute";
1532 unsigned max_index = 0;
1533 snprintf(base, sizeof(base), "%s_meta", prefix);
1534
1535 struct mali_attr_meta *attr_meta;
1536 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1537
1538 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1539
1540 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1541 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1542 sizeof(*attr_mem));
1543
1544 /* If the record is discard, it should be zero for everything else */
1545
1546 if (attr_meta->format == MALI_VARYING_DISCARD) {
1547 uint64_t zero =
1548 attr_meta->index |
1549 attr_meta->unknown1 |
1550 attr_meta->unknown3 |
1551 attr_meta->src_offset;
1552
1553 if (zero)
1554 pandecode_msg("XXX: expected empty record for varying discard\n");
1555
1556 /* We want to look for a literal 0000 swizzle -- this
1557 * is not encoded with all zeroes, however */
1558
1559 enum mali_channel z = MALI_CHANNEL_ZERO;
1560 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1561 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1562
1563 if (!good_swizzle)
1564 pandecode_msg("XXX: expected zero swizzle for discard\n");
1565
1566 if (!varying)
1567 pandecode_msg("XXX: cannot discard attribute\n");
1568
1569 /* If we're all good, omit the record */
1570 if (!zero && varying && good_swizzle) {
1571 pandecode_log("/* discarded varying */\n");
1572 continue;
1573 }
1574 }
1575
1576 if (attr_meta->index > max_index)
1577 max_index = attr_meta->index;
1578
1579 if (attr_meta->unknown1 != 0x2) {
1580 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1581 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1582 }
1583
1584 if (attr_meta->unknown3) {
1585 pandecode_msg("XXX: unexpected unknown3 set\n");
1586 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1587 }
1588
1589 pandecode_format_short(attr_meta->format, false);
1590 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1591
1592 if (attr_meta->src_offset)
1593 pandecode_log_cont("[%u]", attr_meta->src_offset);
1594
1595 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1596
1597 pandecode_log_cont(";\n");
1598 }
1599
1600 pandecode_log("\n");
1601
1602 return count ? (max_index + 1) : 0;
1603 }
1604
1605 /* return bits [lo, hi) of word */
1606 static u32
1607 bits(u32 word, u32 lo, u32 hi)
1608 {
1609 if (hi - lo >= 32)
1610 return word; // avoid undefined behavior with the shift
1611
1612 return (word >> lo) & ((1 << (hi - lo)) - 1);
1613 }
1614
1615 static void
1616 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool noninstanced)
1617 {
1618 pandecode_log_cont("{\n");
1619 pandecode_indent++;
1620
1621 /* Decode invocation_count. See the comment before the definition of
1622 * invocation_count for an explanation.
1623 */
1624
1625 unsigned size_x = bits(p->invocation_count, 0, p->size_y_shift) + 1;
1626 unsigned size_y = bits(p->invocation_count, p->size_y_shift, p->size_z_shift) + 1;
1627 unsigned size_z = bits(p->invocation_count, p->size_z_shift, p->workgroups_x_shift) + 1;
1628
1629 unsigned groups_x = bits(p->invocation_count, p->workgroups_x_shift, p->workgroups_y_shift) + 1;
1630 unsigned groups_y = bits(p->invocation_count, p->workgroups_y_shift, p->workgroups_z_shift) + 1;
1631 unsigned groups_z = bits(p->invocation_count, p->workgroups_z_shift, 32) + 1;
1632
1633 /* Even though we have this decoded, we want to ensure that the
1634 * representation is "unique" so we don't lose anything by printing only
1635 * the final result. More specifically, we need to check that we were
1636 * passed something in canonical form, since the definition per the
1637 * hardware is inherently not unique. How? Well, take the resulting
1638 * decode and pack it ourselves! If it is bit exact with what we
1639 * decoded, we're good to go. */
1640
1641 struct mali_vertex_tiler_prefix ref;
1642 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, noninstanced);
1643
1644 bool canonical =
1645 (p->invocation_count == ref.invocation_count) &&
1646 (p->size_y_shift == ref.size_y_shift) &&
1647 (p->size_z_shift == ref.size_z_shift) &&
1648 (p->workgroups_x_shift == ref.workgroups_x_shift) &&
1649 (p->workgroups_y_shift == ref.workgroups_y_shift) &&
1650 (p->workgroups_z_shift == ref.workgroups_z_shift) &&
1651 (p->workgroups_x_shift_2 == ref.workgroups_x_shift_2);
1652
1653 if (!canonical) {
1654 pandecode_msg("XXX: non-canonical workgroups packing\n");
1655 pandecode_msg("expected: %X, %d, %d, %d, %d, %d, %d\n",
1656 ref.invocation_count,
1657 ref.size_y_shift,
1658 ref.size_z_shift,
1659 ref.workgroups_x_shift,
1660 ref.workgroups_y_shift,
1661 ref.workgroups_z_shift,
1662 ref.workgroups_x_shift_2);
1663
1664 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1665 pandecode_prop("size_y_shift = %d", p->size_y_shift);
1666 pandecode_prop("size_z_shift = %d", p->size_z_shift);
1667 pandecode_prop("workgroups_x_shift = %d", p->workgroups_x_shift);
1668 pandecode_prop("workgroups_y_shift = %d", p->workgroups_y_shift);
1669 pandecode_prop("workgroups_z_shift = %d", p->workgroups_z_shift);
1670 pandecode_prop("workgroups_x_shift_2 = %d", p->workgroups_x_shift_2);
1671 }
1672
1673 /* Regardless, print the decode */
1674 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1675 size_x, size_y, size_z,
1676 groups_x, groups_y, groups_z);
1677
1678 /* TODO: Decode */
1679 if (p->unknown_draw)
1680 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1681
1682 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1683
1684 if (p->draw_mode != MALI_DRAW_NONE)
1685 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1686
1687 /* Index count only exists for tiler jobs anyway */
1688
1689 if (p->index_count)
1690 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1691
1692
1693 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1694 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1695
1696 /* Validate an index buffer is present if we need one. TODO: verify
1697 * relationship between invocation_count and index_count */
1698
1699 if (p->indices) {
1700 unsigned count = p->index_count;
1701
1702 /* Grab the size */
1703 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1704
1705 /* Ensure we got a size, and if so, validate the index buffer
1706 * is large enough to hold a full set of indices of the given
1707 * size */
1708
1709 if (!index_raw_size)
1710 pandecode_msg("XXX: index size missing\n");
1711 else
1712 pandecode_validate_buffer(p->indices, count * size);
1713 } else if (index_raw_size)
1714 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1715
1716 if (p->offset_bias_correction)
1717 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1718
1719 /* TODO: Figure out what this is. It's not zero */
1720 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1721
1722 pandecode_indent--;
1723 pandecode_log("},\n");
1724 }
1725
1726 static void
1727 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1728 {
1729 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1730 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1731
1732 for (int i = 0; i < ubufs_count; i++) {
1733 unsigned size = (ubufs[i].size + 1) * 16;
1734 mali_ptr addr = ubufs[i].ptr << 2;
1735
1736 pandecode_validate_buffer(addr, size);
1737
1738 char *ptr = pointer_as_memory_reference(ubufs[i].ptr << 2);
1739 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1740 free(ptr);
1741 }
1742
1743 pandecode_log("\n");
1744 }
1745
1746 static void
1747 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1748 {
1749 pandecode_validate_buffer(uniforms, uniform_count * 16);
1750
1751 char *ptr = pointer_as_memory_reference(uniforms);
1752 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1753 free(ptr);
1754 }
1755
1756 static void
1757 pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1758 {
1759
1760 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1761
1762 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1763
1764 if (scratchpad->zero) {
1765 pandecode_msg("XXX: scratchpad zero tripped");
1766 pandecode_prop("zero = 0x%x\n", scratchpad->zero);
1767 }
1768
1769 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1770 pandecode_indent++;
1771
1772 pandecode_prop("flags = 0x%x", scratchpad->flags);
1773 MEMORY_PROP(scratchpad, gpu_scratchpad);
1774
1775 pandecode_indent--;
1776 pandecode_log("};\n");
1777 }
1778
1779 static unsigned shader_id = 0;
1780
1781 static struct midgard_disasm_stats
1782 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1783 bool is_bifrost)
1784 {
1785 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1786 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1787
1788 /* Compute maximum possible size */
1789 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1790
1791 /* Print some boilerplate to clearly denote the assembly (which doesn't
1792 * obey indentation rules), and actually do the disassembly! */
1793
1794 printf("\n\n");
1795
1796 struct midgard_disasm_stats stats;
1797
1798 if (is_bifrost) {
1799 disassemble_bifrost(code, sz, false);
1800
1801 /* TODO: Extend stats to Bifrost */
1802 stats.texture_count = -128;
1803 stats.sampler_count = -128;
1804 stats.attribute_count = -128;
1805 stats.varying_count = -128;
1806 stats.uniform_count = -128;
1807 stats.uniform_buffer_count = -128;
1808 stats.work_count = -128;
1809
1810 stats.instruction_count = 0;
1811 stats.bundle_count = 0;
1812 stats.quadword_count = 0;
1813 stats.helper_invocations = false;
1814 } else {
1815 stats = disassemble_midgard(code, sz);
1816 }
1817
1818 /* Print shader-db stats */
1819
1820 unsigned nr_threads =
1821 (stats.work_count <= 4) ? 4 :
1822 (stats.work_count <= 8) ? 2 :
1823 1;
1824
1825 printf("shader%d - %s shader: "
1826 "%u inst, %u bundles, %u quadwords, "
1827 "%u registers, %u threads, 0 loops\n\n\n",
1828 shader_id++,
1829 (type == JOB_TYPE_TILER) ? "FRAGMENT" : "VERTEX",
1830 stats.instruction_count, stats.bundle_count, stats.quadword_count,
1831 stats.work_count, nr_threads);
1832
1833
1834 return stats;
1835 }
1836
1837 static void
1838 pandecode_texture(mali_ptr u,
1839 struct pandecode_mapped_memory *tmem,
1840 unsigned job_no, unsigned tex)
1841 {
1842 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
1843
1844 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
1845 pandecode_indent++;
1846
1847 struct mali_texture_format f = t->format;
1848
1849 /* See the definiton of enum mali_texture_type */
1850
1851 bool is_cube = f.type == MALI_TEX_CUBE;
1852 unsigned dimension = is_cube ? 2 : f.type;
1853
1854 pandecode_make_indent();
1855
1856 /* TODO: Are there others? */
1857 bool is_zs = f.format == MALI_Z32_UNORM;
1858
1859 /* Recall Z/S switched the meaning of linear/tiled .. */
1860 if (is_zs && f.layout == MALI_TEXTURE_LINEAR)
1861 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1862
1863 /* Print the layout. Default is linear; a modifier can denote AFBC or
1864 * u-interleaved/tiled modes */
1865
1866 if (f.layout == MALI_TEXTURE_AFBC)
1867 pandecode_log_cont("afbc");
1868 else if (f.layout == MALI_TEXTURE_TILED)
1869 pandecode_log_cont(is_zs ? "linear" : "tiled");
1870 else if (f.layout == MALI_TEXTURE_LINEAR)
1871 pandecode_log_cont("linear");
1872 else
1873 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
1874
1875 pandecode_swizzle(t->swizzle, f.format);
1876 pandecode_log_cont(" ");
1877
1878 /* Distinguish cube/2D with modifier */
1879
1880 if (is_cube)
1881 pandecode_log_cont("cube ");
1882
1883 pandecode_format_short(f.format, f.srgb);
1884 pandecode_swizzle(f.swizzle, f.format);
1885
1886 /* All four width/height/depth/array_size dimensions are present
1887 * regardless of the type of texture, but it is an error to have
1888 * non-zero dimensions for unused dimensions. Verify this. array_size
1889 * can always be set, as can width. */
1890
1891 if (t->height && dimension < 2)
1892 pandecode_msg("XXX: nonzero height for <2D texture\n");
1893
1894 if (t->depth && dimension < 3)
1895 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1896
1897 /* Print only the dimensions that are actually there */
1898
1899 pandecode_log_cont(": %d", t->width + 1);
1900
1901 if (dimension >= 2)
1902 pandecode_log_cont("x%u", t->height + 1);
1903
1904 if (dimension >= 3)
1905 pandecode_log_cont("x%u", t->depth + 1);
1906
1907 if (t->array_size)
1908 pandecode_log_cont("[%u]", t->array_size + 1);
1909
1910 if (t->levels)
1911 pandecode_log_cont(" mip %u", t->levels);
1912
1913 pandecode_log_cont("\n");
1914
1915 if (f.unknown1 | f.zero) {
1916 pandecode_msg("XXX: texture format zero tripped\n");
1917 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
1918 pandecode_prop("zero = %" PRId32, f.zero);
1919 }
1920
1921 if (!f.unknown2) {
1922 pandecode_msg("XXX: expected unknown texture bit set\n");
1923 pandecode_prop("unknown2 = %" PRId32, f.unknown1);
1924 }
1925
1926 if (t->swizzle_zero) {
1927 pandecode_msg("XXX: swizzle zero tripped\n");
1928 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
1929 }
1930
1931 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
1932 pandecode_msg("XXX: texture zero tripped\n");
1933 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
1934 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
1935 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
1936 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
1937 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
1938 }
1939
1940 pandecode_log(".payload = {\n");
1941 pandecode_indent++;
1942
1943 /* A bunch of bitmap pointers follow.
1944 * We work out the correct number,
1945 * based on the mipmap/cubemap
1946 * properties, but dump extra
1947 * possibilities to futureproof */
1948
1949 int bitmap_count = MALI_NEGATIVE(t->levels);
1950
1951 /* Miptree for each face */
1952 if (f.type == MALI_TEX_CUBE)
1953 bitmap_count *= 6;
1954
1955 /* Array of textures */
1956 bitmap_count *= MALI_NEGATIVE(t->array_size);
1957
1958 /* Stride for each element */
1959 if (f.manual_stride)
1960 bitmap_count *= 2;
1961
1962 /* Sanity check the size */
1963 int max_count = sizeof(t->payload) / sizeof(t->payload[0]);
1964 assert (bitmap_count <= max_count);
1965
1966 for (int i = 0; i < bitmap_count; ++i) {
1967 /* How we dump depends if this is a stride or a pointer */
1968
1969 if (f.manual_stride && (i & 1)) {
1970 /* signed 32-bit snuck in as a 64-bit pointer */
1971 uint64_t stride_set = t->payload[i];
1972 uint32_t clamped_stride = stride_set;
1973 int32_t stride = clamped_stride;
1974 assert(stride_set == clamped_stride);
1975 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
1976 } else {
1977 char *a = pointer_as_memory_reference(t->payload[i]);
1978 pandecode_log("%s, \n", a);
1979 free(a);
1980 }
1981 }
1982
1983 pandecode_indent--;
1984 pandecode_log("},\n");
1985
1986 pandecode_indent--;
1987 pandecode_log("};\n");
1988 }
1989
1990 /* 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 */
1991
1992 static void
1993 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
1994 {
1995 /* Nothing to do */
1996 if (claim == truth)
1997 return;
1998
1999 if (fuzzy)
2000 assert(truth >= 0);
2001
2002 if ((truth >= 0) && !fuzzy) {
2003 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2004 (truth < claim) ? "warn" : "XXX",
2005 name, truth, claim);
2006 } else if ((claim > -truth) && !fuzzy) {
2007 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2008 name, -truth, claim);
2009 } else if (fuzzy && (claim < truth))
2010 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2011 name, truth, claim);
2012
2013 pandecode_log(".%s = %" PRId16, name, claim);
2014
2015 if (fuzzy)
2016 pandecode_log_cont(" /* %u used */", truth);
2017
2018 pandecode_log_cont(",\n");
2019 }
2020
2021 static void
2022 pandecode_vertex_tiler_postfix_pre(
2023 const struct mali_vertex_tiler_postfix *p,
2024 int job_no, enum mali_job_type job_type,
2025 char *suffix, bool is_bifrost)
2026 {
2027 struct pandecode_mapped_memory *attr_mem;
2028
2029 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2030 * are the only things actually needed from the FBD, vertex/tiler jobs
2031 * no longer reference the FBD -- instead, this field points to some
2032 * info about the scratchpad.
2033 */
2034
2035 struct pandecode_fbd fbd_info = {
2036 /* Default for Bifrost */
2037 .rt_count = 1
2038 };
2039
2040 if (is_bifrost)
2041 pandecode_scratchpad(p->framebuffer & ~FBD_TYPE, job_no, suffix);
2042 else if (p->framebuffer & MALI_MFBD)
2043 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
2044 else if (job_type == JOB_TYPE_COMPUTE)
2045 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
2046 else
2047 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no, false);
2048
2049 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2050 int texture_count = 0, sampler_count = 0;
2051
2052 if (p->shader) {
2053 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2054 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2055
2056 /* Disassemble ahead-of-time to get stats. Initialize with
2057 * stats for the missing-shader case so we get validation
2058 * there, too */
2059
2060 struct midgard_disasm_stats info = {
2061 .texture_count = 0,
2062 .sampler_count = 0,
2063 .attribute_count = 0,
2064 .varying_count = 0,
2065 .work_count = 1,
2066
2067 .uniform_count = -128,
2068 .uniform_buffer_count = 0
2069 };
2070
2071 if (s->shader & ~0xF)
2072 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost);
2073
2074 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2075 pandecode_indent++;
2076
2077 /* Save for dumps */
2078 attribute_count = s->attribute_count;
2079 varying_count = s->varying_count;
2080 texture_count = s->texture_count;
2081 sampler_count = s->sampler_count;
2082
2083 if (is_bifrost) {
2084 uniform_count = s->bifrost2.uniform_count;
2085 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2086 } else {
2087 uniform_count = s->midgard1.uniform_count;
2088 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2089 }
2090
2091 pandecode_shader_address("shader", s->shader);
2092
2093 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2094 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2095 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2096 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2097 pandecode_shader_prop("uniform_buffer_count",
2098 uniform_buffer_count,
2099 info.uniform_buffer_count, true);
2100
2101 if (!is_bifrost) {
2102 pandecode_shader_prop("uniform_count",
2103 uniform_count,
2104 info.uniform_count, false);
2105
2106 pandecode_shader_prop("work_count",
2107 s->midgard1.work_count, info.work_count, false);
2108 }
2109
2110 if (is_bifrost) {
2111 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32, s->bifrost1.unk1);
2112 } else {
2113 bool helpers = s->midgard1.flags & MALI_HELPER_INVOCATIONS;
2114 s->midgard1.flags &= ~MALI_HELPER_INVOCATIONS;
2115
2116 if (helpers != info.helper_invocations) {
2117 pandecode_msg("XXX: expected helpers %u but got %u\n",
2118 info.helper_invocations, helpers);
2119 }
2120
2121 pandecode_log(".midgard1.flags = ");
2122 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
2123 pandecode_log_cont(",\n");
2124
2125 pandecode_prop("midgard1.unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
2126 }
2127
2128 if (s->depth_units || s->depth_factor) {
2129 pandecode_prop("depth_factor = %f", s->depth_factor);
2130 pandecode_prop("depth_units = %f", s->depth_units);
2131 }
2132
2133 if (s->alpha_coverage) {
2134 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
2135 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
2136
2137 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2138 invert_alpha_coverage ? "~" : "",
2139 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
2140 }
2141
2142 if (s->unknown2_3 || s->unknown2_4) {
2143 pandecode_log(".unknown2_3 = ");
2144
2145 int unknown2_3 = s->unknown2_3;
2146 int unknown2_4 = s->unknown2_4;
2147
2148 /* We're not quite sure what these flags mean without the depth test, if anything */
2149
2150 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2151 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2152 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2153
2154 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2155 }
2156
2157 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2158 pandecode_log_cont(",\n");
2159
2160 pandecode_log(".unknown2_4 = ");
2161 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2162 pandecode_log_cont(",\n");
2163 }
2164
2165 if (s->stencil_mask_front || s->stencil_mask_back) {
2166 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2167 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2168 }
2169
2170 pandecode_stencil("front", &s->stencil_front);
2171 pandecode_stencil("back", &s->stencil_back);
2172
2173 if (is_bifrost) {
2174 pandecode_log(".bifrost2 = {\n");
2175 pandecode_indent++;
2176
2177 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2178 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2179 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2180 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2181
2182 pandecode_indent--;
2183 pandecode_log("},\n");
2184 } else if (s->midgard2.unknown2_7) {
2185 pandecode_log(".midgard2 = {\n");
2186 pandecode_indent++;
2187
2188 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2189 pandecode_indent--;
2190 pandecode_log("},\n");
2191 }
2192
2193 if (s->unknown2_8)
2194 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
2195
2196 if (!is_bifrost) {
2197 /* TODO: Blend shaders routing/disasm */
2198
2199 union midgard_blend blend = s->blend;
2200 pandecode_midgard_blend(&blend, false);
2201 }
2202
2203 pandecode_indent--;
2204 pandecode_log("};\n");
2205
2206 /* MRT blend fields are used whenever MFBD is used, with
2207 * per-RT descriptors */
2208
2209 if (job_type == JOB_TYPE_TILER) {
2210 void* blend_base = (void *) (s + 1);
2211
2212 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2213 mali_ptr shader = 0;
2214
2215 if (is_bifrost)
2216 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2217 else
2218 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2219
2220 if (shader & ~0xF) {
2221 struct midgard_disasm_stats stats =
2222 pandecode_shader_disassemble(shader, job_no, job_type, false);
2223
2224 bool has_texture = (stats.texture_count > 0);
2225 bool has_sampler = (stats.sampler_count > 0);
2226 bool has_attribute = (stats.attribute_count > 0);
2227 bool has_varying = (stats.varying_count > 0);
2228 bool has_uniform = (stats.uniform_count > 0);
2229 bool has_ubo = (stats.uniform_buffer_count > 0);
2230
2231 if (has_texture || has_sampler)
2232 pandecode_msg("XXX: blend shader accessing textures\n");
2233
2234 if (has_attribute || has_varying)
2235 pandecode_msg("XXX: blend shader accessing interstage\n");
2236
2237 if (has_uniform || has_ubo)
2238 pandecode_msg("XXX: blend shader accessing uniforms\n");
2239 }
2240
2241 }
2242 }
2243 } else
2244 pandecode_msg("XXX: missing shader descriptor\n");
2245
2246 if (p->viewport) {
2247 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2248 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2249
2250 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2251 pandecode_indent++;
2252
2253 pandecode_prop("clip_minx = %f", f->clip_minx);
2254 pandecode_prop("clip_miny = %f", f->clip_miny);
2255 pandecode_prop("clip_minz = %f", f->clip_minz);
2256 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2257 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2258 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2259
2260 /* Only the higher coordinates are MALI_POSITIVE scaled */
2261
2262 pandecode_prop("viewport0 = { %d, %d }",
2263 f->viewport0[0], f->viewport0[1]);
2264
2265 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2266 f->viewport1[0] + 1, f->viewport1[1] + 1);
2267
2268 pandecode_indent--;
2269 pandecode_log("};\n");
2270 }
2271
2272 unsigned max_attr_index = 0;
2273
2274 if (p->attribute_meta)
2275 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2276
2277 if (p->attributes) {
2278 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2279 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2280 }
2281
2282 /* Varyings are encoded like attributes but not actually sent; we just
2283 * pass a zero buffer with the right stride/size set, (or whatever)
2284 * since the GPU will write to it itself */
2285
2286 if (p->varying_meta) {
2287 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2288 }
2289
2290 if (p->varyings) {
2291 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2292
2293 /* Number of descriptors depends on whether there are
2294 * non-internal varyings */
2295
2296 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2297 }
2298
2299 if (p->uniform_buffers) {
2300 if (uniform_buffer_count)
2301 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2302 else
2303 pandecode_msg("warn: UBOs specified but not referenced\n");
2304 } else if (uniform_buffer_count)
2305 pandecode_msg("XXX: UBOs referenced but not specified\n");
2306
2307 /* We don't want to actually dump uniforms, but we do need to validate
2308 * that the counts we were given are sane */
2309
2310 if (p->uniforms) {
2311 if (uniform_count)
2312 pandecode_uniforms(p->uniforms, uniform_count);
2313 else
2314 pandecode_msg("warn: Uniforms specified but not referenced\n");
2315 } else if (uniform_count)
2316 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2317
2318 if (p->texture_trampoline) {
2319 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
2320
2321 if (mmem) {
2322 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
2323
2324 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
2325 pandecode_indent++;
2326
2327 for (int tex = 0; tex < texture_count; ++tex) {
2328 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2329 char *a = pointer_as_memory_reference(*u);
2330 pandecode_log("%s,\n", a);
2331 free(a);
2332 }
2333
2334 pandecode_indent--;
2335 pandecode_log("};\n");
2336
2337 /* Now, finally, descend down into the texture descriptor */
2338 for (unsigned tex = 0; tex < texture_count; ++tex) {
2339 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2340 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2341 if (tmem)
2342 pandecode_texture(*u, tmem, job_no, tex);
2343 }
2344 }
2345 }
2346
2347 if (p->sampler_descriptor) {
2348 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2349
2350 if (smem) {
2351 struct mali_sampler_descriptor *s;
2352
2353 mali_ptr d = p->sampler_descriptor;
2354
2355 for (int i = 0; i < sampler_count; ++i) {
2356 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2357
2358 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2359 pandecode_indent++;
2360
2361 pandecode_log(".filter_mode = ");
2362 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2363 pandecode_log_cont(",\n");
2364
2365 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2366 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2367
2368 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2369 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2370 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2371
2372 pandecode_prop("compare_func = %s", pandecode_alt_func(s->compare_func));
2373
2374 if (s->zero || s->zero2) {
2375 pandecode_msg("XXX: sampler zero tripped\n");
2376 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2377 }
2378
2379 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2380
2381 pandecode_prop("border_color = { %f, %f, %f, %f }",
2382 s->border_color[0],
2383 s->border_color[1],
2384 s->border_color[2],
2385 s->border_color[3]);
2386
2387 pandecode_indent--;
2388 pandecode_log("};\n");
2389 }
2390 }
2391 }
2392 }
2393
2394 static void
2395 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2396 {
2397 if (p->shader & 0xF)
2398 pandecode_msg("warn: shader tagged %X\n", p->shader & 0xF);
2399
2400 if (!(p->position_varying || p->occlusion_counter))
2401 return;
2402
2403 pandecode_log(".postfix = {\n");
2404 pandecode_indent++;
2405
2406 MEMORY_PROP(p, position_varying);
2407 MEMORY_PROP(p, occlusion_counter);
2408
2409 pandecode_indent--;
2410 pandecode_log("},\n");
2411 }
2412
2413 static void
2414 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2415 {
2416 pandecode_log_cont("{\n");
2417 pandecode_indent++;
2418
2419 pandecode_prop("unk2 = 0x%x", v->unk2);
2420
2421 if (v->zero0 || v->zero1) {
2422 pandecode_msg("XXX: vertex only zero tripped");
2423 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2424 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2425 }
2426
2427 pandecode_indent--;
2428 pandecode_log("}\n");
2429 }
2430
2431 static void
2432 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2433 {
2434
2435 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2436 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2437
2438 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2439 pandecode_indent++;
2440
2441 if (h->zero) {
2442 pandecode_msg("XXX: tiler heap zero tripped\n");
2443 pandecode_prop("zero = 0x%x", h->zero);
2444 }
2445
2446 for (int i = 0; i < 12; i++) {
2447 if (h->zeros[i] != 0) {
2448 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2449 i, h->zeros[i]);
2450 }
2451 }
2452
2453 pandecode_prop("heap_size = 0x%x", h->heap_size);
2454 MEMORY_PROP(h, tiler_heap_start);
2455 MEMORY_PROP(h, tiler_heap_free);
2456
2457 /* this might point to the beginning of another buffer, when it's
2458 * really the end of the tiler heap buffer, so we have to be careful
2459 * here. but for zero length, we need the same pointer.
2460 */
2461
2462 if (h->tiler_heap_end == h->tiler_heap_start) {
2463 MEMORY_PROP(h, tiler_heap_start);
2464 } else {
2465 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2466 pandecode_prop("tiler_heap_end = %s + 1", a);
2467 free(a);
2468 }
2469
2470 pandecode_indent--;
2471 pandecode_log("};\n");
2472 }
2473
2474 static void
2475 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2476 {
2477 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2478 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2479
2480 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2481
2482 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2483 pandecode_indent++;
2484
2485 if (t->zero0 || t->zero1) {
2486 pandecode_msg("XXX: tiler meta zero tripped\n");
2487 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2488 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2489 }
2490
2491 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2492 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2493
2494 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2495 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2496
2497 for (int i = 0; i < 12; i++) {
2498 if (t->zeros[i] != 0) {
2499 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2500 i, t->zeros[i]);
2501 }
2502 }
2503
2504 pandecode_indent--;
2505 pandecode_log("};\n");
2506 }
2507
2508 static void
2509 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2510 {
2511 pandecode_log(".gl_enables = ");
2512
2513 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2514
2515 pandecode_log_cont(",\n");
2516 }
2517
2518 static void
2519 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2520 {
2521 if (u.pointer == 0x0)
2522 return;
2523
2524 pandecode_log(".primitive_size = {\n");
2525 pandecode_indent++;
2526
2527 if (constant) {
2528 pandecode_prop("constant = %f", u.constant);
2529 } else {
2530 MEMORY_PROP((&u), pointer);
2531 }
2532
2533 pandecode_indent--;
2534 pandecode_log("},\n");
2535 }
2536
2537 static void
2538 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2539 {
2540 pandecode_log_cont("{\n");
2541 pandecode_indent++;
2542
2543 /* TODO: gl_PointSize on Bifrost */
2544 pandecode_primitive_size(t->primitive_size, true);
2545
2546 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2547
2548 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2549 || t->zero6 || t->zero7 || t->zero8) {
2550 pandecode_msg("XXX: tiler only zero tripped\n");
2551 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2552 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2553 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2554 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2555 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2556 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2557 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2558 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2559 }
2560
2561 pandecode_indent--;
2562 pandecode_log("},\n");
2563 }
2564
2565 static int
2566 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2567 const struct pandecode_mapped_memory *mem,
2568 mali_ptr payload, int job_no)
2569 {
2570 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2571
2572 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true);
2573
2574 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2575 pandecode_indent++;
2576
2577 pandecode_log(".prefix = ");
2578 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2579
2580 pandecode_log(".vertex = ");
2581 pandecode_vertex_only_bfr(&v->vertex);
2582
2583 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2584
2585 pandecode_indent--;
2586 pandecode_log("};\n");
2587
2588 return sizeof(*v);
2589 }
2590
2591 static int
2592 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2593 const struct pandecode_mapped_memory *mem,
2594 mali_ptr payload, int job_no)
2595 {
2596 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2597
2598 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true);
2599 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2600
2601 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2602 pandecode_indent++;
2603
2604 pandecode_log(".prefix = ");
2605 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2606
2607 pandecode_log(".tiler = ");
2608 pandecode_tiler_only_bfr(&t->tiler, job_no);
2609
2610 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2611
2612 pandecode_indent--;
2613 pandecode_log("};\n");
2614
2615 return sizeof(*t);
2616 }
2617
2618 static int
2619 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2620 const struct pandecode_mapped_memory *mem,
2621 mali_ptr payload, int job_no)
2622 {
2623 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2624
2625 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false);
2626
2627 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2628 pandecode_indent++;
2629
2630 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2631 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2632
2633 bool instanced = v->instance_shift || v->instance_odd;
2634 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2635
2636 pandecode_log(".prefix = ");
2637 pandecode_vertex_tiler_prefix(&v->prefix, job_no, !instanced && is_graphics);
2638
2639 pandecode_gl_enables(v->gl_enables, h->job_type);
2640
2641 if (v->instance_shift || v->instance_odd) {
2642 pandecode_prop("instance_shift = 0x%d /* %d */",
2643 v->instance_shift, 1 << v->instance_shift);
2644 pandecode_prop("instance_odd = 0x%X /* %d */",
2645 v->instance_odd, (2 * v->instance_odd) + 1);
2646
2647 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2648 }
2649
2650 if (v->offset_start)
2651 pandecode_prop("offset_start = %d", v->offset_start);
2652
2653 if (v->zero5) {
2654 pandecode_msg("XXX: midgard payload zero tripped\n");
2655 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2656 }
2657
2658 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2659
2660 pandecode_indent--;
2661 pandecode_log("};\n");
2662
2663 return sizeof(*v);
2664 }
2665
2666 static int
2667 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2668 mali_ptr payload, int job_no,
2669 bool is_bifrost)
2670 {
2671 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2672
2673 bool is_mfbd = (s->framebuffer & FBD_TYPE) == MALI_MFBD;
2674
2675 /* Bifrost theoretically may retain support for SFBD on compute jobs,
2676 * but for graphics workloads with a FRAGMENT payload, use MFBD */
2677
2678 if (!is_mfbd && is_bifrost)
2679 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2680
2681 struct pandecode_fbd info;
2682
2683 if (is_mfbd)
2684 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2685 else
2686 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true);
2687
2688 /* Compute the tag for the tagged pointer. This contains the type of
2689 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2690 * additional structures follow the MFBD header (an extra payload or
2691 * not, as well as a count of render targets) */
2692
2693 unsigned expected_tag = is_mfbd ? MALI_MFBD : MALI_SFBD;
2694
2695 if (is_mfbd) {
2696 if (info.has_extra)
2697 expected_tag |= MALI_MFBD_TAG_EXTRA;
2698
2699 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
2700 }
2701
2702 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
2703 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2704 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
2705 pandecode_prop("max_tile_coord = 0x%X\n", s->min_tile_coord);
2706 }
2707
2708 /* Extract tile coordinates */
2709
2710 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
2711 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
2712
2713 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2714 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2715
2716 /* For the max, we also want the floored (rather than ceiled) version for checking */
2717
2718 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
2719 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
2720
2721 /* Validate the coordinates are well-ordered */
2722
2723 if (min_x == max_x)
2724 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2725 else if (min_x > max_x)
2726 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2727
2728 if (min_y == max_y)
2729 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2730 else if (min_y > max_y)
2731 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2732
2733 /* Validate the coordinates fit inside the framebuffer. We use floor,
2734 * rather than ceil, for the max coordinates, since the tile
2735 * coordinates for something like an 800x600 framebuffer will actually
2736 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2737
2738 if ((min_x > info.width) || (max_x_f > info.width))
2739 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2740
2741 if ((min_y > info.height) || (max_y_f > info.height))
2742 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2743
2744 /* After validation, we print */
2745
2746 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
2747
2748 /* The FBD is a tagged pointer */
2749
2750 unsigned tag = (s->framebuffer & ~FBD_MASK);
2751
2752 if (tag != expected_tag)
2753 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
2754
2755 return sizeof(*s);
2756 }
2757
2758 static int job_descriptor_number = 0;
2759
2760 int
2761 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost)
2762 {
2763 struct mali_job_descriptor_header *h;
2764
2765 int start_number = 0;
2766
2767 bool first = true;
2768 bool last_size;
2769
2770 do {
2771 struct pandecode_mapped_memory *mem =
2772 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2773
2774 void *payload;
2775
2776 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2777
2778 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2779 * high 32-bits of the 64-bit pointer are reused to store
2780 * something else.
2781 */
2782 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2783 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2784 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2785
2786 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
2787
2788 int job_no = job_descriptor_number++;
2789
2790 if (first)
2791 start_number = job_no;
2792
2793 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2794 pandecode_indent++;
2795
2796 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2797
2798 /* Save for next job fixing */
2799 last_size = h->job_descriptor_size;
2800
2801 if (h->job_descriptor_size)
2802 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2803
2804 if (h->exception_status && h->exception_status != 0x1)
2805 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2806 h->exception_status,
2807 (h->exception_status >> 16) & 0xFFFF,
2808 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2809 h->exception_status & 0xFF);
2810
2811 if (h->first_incomplete_task)
2812 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2813
2814 if (h->fault_pointer)
2815 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2816
2817 if (h->job_barrier)
2818 pandecode_prop("job_barrier = %d", h->job_barrier);
2819
2820 pandecode_prop("job_index = %d", h->job_index);
2821
2822 if (h->unknown_flags)
2823 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2824
2825 if (h->job_dependency_index_1)
2826 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2827
2828 if (h->job_dependency_index_2)
2829 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2830
2831 pandecode_indent--;
2832 pandecode_log("};\n");
2833
2834 /* Do not touch the field yet -- decode the payload first, and
2835 * don't touch that either. This is essential for the uploads
2836 * to occur in sequence and therefore be dynamically allocated
2837 * correctly. Do note the size, however, for that related
2838 * reason. */
2839
2840 switch (h->job_type) {
2841 case JOB_TYPE_SET_VALUE: {
2842 struct mali_payload_set_value *s = payload;
2843 pandecode_log("struct mali_payload_set_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2844 pandecode_indent++;
2845 MEMORY_PROP(s, out);
2846 pandecode_prop("unknown = 0x%" PRIX64, s->unknown);
2847 pandecode_indent--;
2848 pandecode_log("};\n");
2849
2850 break;
2851 }
2852
2853 case JOB_TYPE_TILER:
2854 case JOB_TYPE_VERTEX:
2855 case JOB_TYPE_COMPUTE:
2856 if (bifrost) {
2857 if (h->job_type == JOB_TYPE_TILER)
2858 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no);
2859 else
2860 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no);
2861 } else
2862 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no);
2863
2864 break;
2865
2866 case JOB_TYPE_FRAGMENT:
2867 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost);
2868 break;
2869
2870 default:
2871 break;
2872 }
2873
2874 /* Handle linkage */
2875
2876 if (!first) {
2877 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2878
2879 if (last_size)
2880 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no);
2881 else
2882 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no);
2883 }
2884
2885 first = false;
2886
2887 } while ((jc_gpu_va = h->job_descriptor_size ? h->next_job_64 : h->next_job_32));
2888
2889 return start_number;
2890 }