projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
pan/decode: Guard texture unknowns as zero trips
[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 %d at offset %d in buffer of size %d. "
151 "Overrun by %d bytes.",
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_TEST),
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(HELPER_INVOCATIONS),
271 FLAG_INFO(READS_TILEBUFFER),
272 FLAG_INFO(READS_ZS),
273 {}
274 };
275 #undef FLAG_INFO
276
277 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
278 static const struct pandecode_flag_info mfbd_flag_info [] = {
279 FLAG_INFO(DEPTH_WRITE),
280 FLAG_INFO(EXTRA),
281 {}
282 };
283 #undef FLAG_INFO
284
285 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
286 static const struct pandecode_flag_info sampler_flag_info [] = {
287 FLAG_INFO(MAG_NEAREST),
288 FLAG_INFO(MIN_NEAREST),
289 FLAG_INFO(MIP_LINEAR_1),
290 FLAG_INFO(MIP_LINEAR_2),
291 FLAG_INFO(NORM_COORDS),
292 {}
293 };
294 #undef FLAG_INFO
295
296 extern char *replace_fragment;
297 extern char *replace_vertex;
298
299 static char *
300 pandecode_job_type(enum mali_job_type type)
301 {
302 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
303
304 switch (type) {
305 DEFINE_CASE(NULL);
306 DEFINE_CASE(SET_VALUE);
307 DEFINE_CASE(CACHE_FLUSH);
308 DEFINE_CASE(COMPUTE);
309 DEFINE_CASE(VERTEX);
310 DEFINE_CASE(TILER);
311 DEFINE_CASE(FUSED);
312 DEFINE_CASE(FRAGMENT);
313
314 case JOB_NOT_STARTED:
315 return "NOT_STARTED";
316
317 default:
318 pandecode_log("Warning! Unknown job type %x\n", type);
319 return "!?!?!?";
320 }
321
322 #undef DEFINE_CASE
323 }
324
325 static char *
326 pandecode_draw_mode(enum mali_draw_mode mode)
327 {
328 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
329
330 switch (mode) {
331 DEFINE_CASE(DRAW_NONE);
332 DEFINE_CASE(POINTS);
333 DEFINE_CASE(LINES);
334 DEFINE_CASE(TRIANGLES);
335 DEFINE_CASE(TRIANGLE_STRIP);
336 DEFINE_CASE(TRIANGLE_FAN);
337 DEFINE_CASE(LINE_STRIP);
338 DEFINE_CASE(LINE_LOOP);
339 DEFINE_CASE(POLYGON);
340 DEFINE_CASE(QUADS);
341 DEFINE_CASE(QUAD_STRIP);
342
343 default:
344 pandecode_msg("XXX: invalid draw mode %X\n", mode);
345 return "";
346 }
347
348 #undef DEFINE_CASE
349 }
350
351 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
352 static char *
353 pandecode_func(enum mali_func mode)
354 {
355 switch (mode) {
356 DEFINE_CASE(NEVER);
357 DEFINE_CASE(LESS);
358 DEFINE_CASE(EQUAL);
359 DEFINE_CASE(LEQUAL);
360 DEFINE_CASE(GREATER);
361 DEFINE_CASE(NOTEQUAL);
362 DEFINE_CASE(GEQUAL);
363 DEFINE_CASE(ALWAYS);
364
365 default:
366 pandecode_msg("XXX: invalid func %X\n", mode);
367 return "";
368 }
369 }
370 #undef DEFINE_CASE
371
372 /* Why is this duplicated? Who knows... */
373 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
374 static char *
375 pandecode_alt_func(enum mali_alt_func mode)
376 {
377 switch (mode) {
378 DEFINE_CASE(NEVER);
379 DEFINE_CASE(LESS);
380 DEFINE_CASE(EQUAL);
381 DEFINE_CASE(LEQUAL);
382 DEFINE_CASE(GREATER);
383 DEFINE_CASE(NOTEQUAL);
384 DEFINE_CASE(GEQUAL);
385 DEFINE_CASE(ALWAYS);
386
387 default:
388 pandecode_msg("XXX: invalid alt func %X\n", mode);
389 return "";
390 }
391 }
392 #undef DEFINE_CASE
393
394 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
395 static char *
396 pandecode_stencil_op(enum mali_stencil_op op)
397 {
398 switch (op) {
399 DEFINE_CASE(KEEP);
400 DEFINE_CASE(REPLACE);
401 DEFINE_CASE(ZERO);
402 DEFINE_CASE(INVERT);
403 DEFINE_CASE(INCR_WRAP);
404 DEFINE_CASE(DECR_WRAP);
405 DEFINE_CASE(INCR);
406 DEFINE_CASE(DECR);
407
408 default:
409 pandecode_msg("XXX: invalid stencil op %X\n", op);
410 return "";
411 }
412 }
413
414 #undef DEFINE_CASE
415
416 #define DEFINE_CASE(name) case MALI_ATTR_ ## name: return "MALI_ATTR_" #name
417 static char *pandecode_attr_mode(enum mali_attr_mode mode)
418 {
419 switch(mode) {
420 DEFINE_CASE(UNUSED);
421 DEFINE_CASE(LINEAR);
422 DEFINE_CASE(POT_DIVIDE);
423 DEFINE_CASE(MODULO);
424 DEFINE_CASE(NPOT_DIVIDE);
425 DEFINE_CASE(IMAGE);
426 DEFINE_CASE(INTERNAL);
427 default:
428 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
429 return "";
430 }
431 }
432
433 #undef DEFINE_CASE
434
435 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
436 static char *
437 pandecode_wrap_mode(enum mali_wrap_mode op)
438 {
439 switch (op) {
440 DEFINE_CASE(REPEAT);
441 DEFINE_CASE(CLAMP_TO_EDGE);
442 DEFINE_CASE(CLAMP_TO_BORDER);
443 DEFINE_CASE(MIRRORED_REPEAT);
444
445 default:
446 pandecode_msg("XXX: invalid wrap mode %X\n", op);
447 return "";
448 }
449 }
450 #undef DEFINE_CASE
451
452 #define DEFINE_CASE(name) case MALI_TEX_## name: return "MALI_TEX_" #name
453 static char *
454 pandecode_texture_type(enum mali_texture_type type)
455 {
456 switch (type) {
457 DEFINE_CASE(1D);
458 DEFINE_CASE(2D);
459 DEFINE_CASE(3D);
460 DEFINE_CASE(CUBE);
461
462 default:
463 unreachable("Unknown case");
464 }
465 }
466 #undef DEFINE_CASE
467
468 #define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
469 static char *
470 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt)
471 {
472 switch (fmt) {
473 DEFINE_CASE(TILED);
474 DEFINE_CASE(UNKNOWN);
475 DEFINE_CASE(LINEAR);
476 DEFINE_CASE(AFBC);
477
478 default:
479 unreachable("Invalid case");
480 }
481 }
482 #undef DEFINE_CASE
483
484 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
485 static char *
486 pandecode_exception_access(enum mali_exception_access access)
487 {
488 switch (access) {
489 DEFINE_CASE(NONE);
490 DEFINE_CASE(EXECUTE);
491 DEFINE_CASE(READ);
492 DEFINE_CASE(WRITE);
493
494 default:
495 unreachable("Invalid case");
496 }
497 }
498 #undef DEFINE_CASE
499
500 /* Midgard's tiler descriptor is embedded within the
501 * larger FBD */
502
503 static void
504 pandecode_midgard_tiler_descriptor(
505 const struct midgard_tiler_descriptor *t,
506 unsigned width,
507 unsigned height,
508 bool is_fragment)
509 {
510 pandecode_log(".tiler = {\n");
511 pandecode_indent++;
512
513 if (t->hierarchy_mask == MALI_TILER_DISABLED)
514 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
515 else
516 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
517
518 /* We know this name from the kernel, but we never see it nonzero */
519
520 if (t->flags)
521 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
522
523 MEMORY_PROP(t, polygon_list);
524
525 /* The body is offset from the base of the polygon list */
526 assert(t->polygon_list_body > t->polygon_list);
527 unsigned body_offset = t->polygon_list_body - t->polygon_list;
528
529 /* It needs to fit inside the reported size */
530 assert(t->polygon_list_size >= body_offset);
531
532 /* Check that we fit */
533 struct pandecode_mapped_memory *plist =
534 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
535
536 assert(t->polygon_list_size <= plist->length);
537
538 /* Now that we've sanity checked, we'll try to calculate the sizes
539 * ourselves for comparison */
540
541 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask);
542 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask);
543
544 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
545 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
546 ref_header, ref_size);
547 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
548 pandecode_msg("body offset %d\n", body_offset);
549 }
550
551 /* The tiler heap has a start and end specified -- it should be
552 * identical to what we have in the BO. The exception is if tiling is
553 * disabled. */
554
555 MEMORY_PROP(t, heap_start);
556 assert(t->heap_end >= t->heap_start);
557
558 struct pandecode_mapped_memory *heap =
559 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
560
561 unsigned heap_size = t->heap_end - t->heap_start;
562
563 /* Tiling is enabled with a special flag */
564 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
565 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
566
567 bool tiling_enabled = hierarchy_mask;
568
569 if (tiling_enabled) {
570 /* When tiling is enabled, the heap should be a tight fit */
571 unsigned heap_offset = t->heap_start - heap->gpu_va;
572 if ((heap_offset + heap_size) != heap->length) {
573 pandecode_msg("XXX: heap size %d (expected %d)\n",
574 heap_size, heap->length - heap_offset);
575 }
576
577 /* We should also have no other flags */
578 if (tiler_flags)
579 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
580 } else {
581 /* When tiling is disabled, we should have that flag and no others */
582
583 if (tiler_flags != MALI_TILER_DISABLED) {
584 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
585 tiler_flags);
586 }
587
588 /* We should also have an empty heap */
589 if (heap_size) {
590 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
591 heap_size);
592 }
593
594 /* Disabled tiling is used only for clear-only jobs, which are
595 * purely FRAGMENT, so we should never see this for
596 * non-FRAGMENT descriptors. */
597
598 if (!is_fragment)
599 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
600 }
601
602 /* We've never seen weights used in practice, but we know from the
603 * kernel these fields is there */
604
605 bool nonzero_weights = false;
606
607 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
608 nonzero_weights |= t->weights[w] != 0x0;
609 }
610
611 if (nonzero_weights) {
612 pandecode_log(".weights = {");
613
614 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
615 pandecode_log("%d, ", t->weights[w]);
616 }
617
618 pandecode_log("},");
619 }
620
621 pandecode_indent--;
622 pandecode_log("}\n");
623 }
624
625 static void
626 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment)
627 {
628 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
629 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
630
631 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
632 pandecode_indent++;
633
634 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
635 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
636
637 pandecode_log(".format = ");
638 pandecode_log_decoded_flags(fb_fmt_flag_info, s->format);
639 pandecode_log_cont(",\n");
640
641 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", s->width + 1);
642 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", s->height + 1);
643
644 MEMORY_PROP(s, framebuffer);
645 pandecode_prop("stride = %d", s->stride);
646
647 /* Earlier in the actual commandstream -- right before width -- but we
648 * delay to flow nicer */
649
650 pandecode_log(".clear_flags = ");
651 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
652 pandecode_log_cont(",\n");
653
654 if (s->depth_buffer | s->depth_buffer_enable) {
655 MEMORY_PROP(s, depth_buffer);
656 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s->depth_buffer_enable));
657 }
658
659 if (s->stencil_buffer | s->stencil_buffer_enable) {
660 MEMORY_PROP(s, stencil_buffer);
661 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s->stencil_buffer_enable));
662 }
663
664 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
665 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
666 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
667 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
668 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
669 }
670
671 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
672 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
673 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
674 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
675 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
676 }
677
678 if (s->clear_stencil) {
679 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
680 }
681
682 MEMORY_PROP(s, unknown_address_0);
683 const struct midgard_tiler_descriptor t = s->tiler;
684 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment);
685
686 pandecode_indent--;
687 pandecode_log("};\n");
688
689 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
690 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
691 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
692 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
693
694 printf(".zero3 = {");
695
696 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
697 printf("%X, ", s->zero3[i]);
698
699 printf("},\n");
700
701 printf(".zero6 = {");
702
703 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
704 printf("%X, ", s->zero6[i]);
705
706 printf("},\n");
707 }
708
709 static void
710 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
711 {
712 pandecode_log(".unknown%d = {", name);
713
714 for (int i = 0; i < count; ++i)
715 printf("%X, ", slide[i]);
716
717 pandecode_log("},\n");
718 }
719
720 #define SHORT_SLIDE(num) \
721 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
722
723 static void
724 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
725 {
726 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
727 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
728
729 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
730 pandecode_indent++;
731
732 SHORT_SLIDE(1);
733
734 pandecode_indent--;
735 printf("},\n");
736 }
737
738 /* Extracts the number of components associated with a Mali format */
739
740 static unsigned
741 pandecode_format_component_count(enum mali_format fmt)
742 {
743 /* Mask out the format class */
744 unsigned top = fmt & 0b11100000;
745
746 switch (top) {
747 case MALI_FORMAT_SNORM:
748 case MALI_FORMAT_UINT:
749 case MALI_FORMAT_UNORM:
750 case MALI_FORMAT_SINT:
751 return ((fmt >> 3) & 3) + 1;
752 default:
753 /* TODO: Validate */
754 return 4;
755 }
756 }
757
758 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
759
760 static unsigned
761 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
762 {
763 unsigned mask = 0;
764 assert(MALI_CHANNEL_RED == 0);
765
766 for (unsigned c = 0; c < 4; ++c) {
767 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
768
769 if (chan <= MALI_CHANNEL_ALPHA)
770 mask |= (1 << chan);
771 }
772
773 return mask;
774 }
775
776 /* Validates that a (format, swizzle) pair is valid, in the sense that the
777 * swizzle doesn't access any components that are undefined in the format.
778 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
779 * omitted */
780
781 static bool
782 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
783 {
784 unsigned nr_comp = pandecode_format_component_count(fmt);
785 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
786 unsigned valid_mask = (1 << nr_comp) - 1;
787 unsigned invalid_mask = ~valid_mask;
788
789 if (access_mask & invalid_mask) {
790 pandecode_msg("XXX: invalid components accessed\n");
791 return false;
792 }
793
794 /* Check for the default non-swizzling swizzle so we can suppress
795 * useless printing for the defaults */
796
797 unsigned default_swizzles[4] = {
798 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
799 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
800 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
801 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
802 };
803
804 return (swizzle == default_swizzles[nr_comp - 1]);
805 }
806
807 /* Maps MALI_RGBA32F to rgba32f, etc */
808
809 static void
810 pandecode_format_short(enum mali_format fmt)
811 {
812 /* We want a type-like format, so cut off the initial MALI_ */
813 char *format = pandecode_format(fmt);
814 format += strlen("MALI_");
815
816 unsigned len = strlen(format);
817 char *lower_format = calloc(1, len + 1);
818
819 for (unsigned i = 0; i < len; ++i)
820 lower_format[i] = tolower(format[i]);
821
822 pandecode_log_cont("%s", lower_format);
823 free(lower_format);
824 }
825
826 static void
827 pandecode_swizzle(unsigned swizzle, enum mali_format format)
828 {
829 /* First, do some validation */
830 bool trivial_swizzle = pandecode_validate_format_swizzle(
831 format, swizzle);
832
833 if (trivial_swizzle)
834 return;
835
836 /* Next, print the swizzle */
837 pandecode_log_cont(".");
838
839 static const char components[] = "rgba01";
840
841 for (unsigned c = 0; c < 4; ++c) {
842 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
843
844 if (chan >= MALI_CHANNEL_RESERVED_0) {
845 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
846 continue;
847 }
848 pandecode_log_cont("%c", components[chan]);
849 }
850 }
851
852 static void
853 pandecode_rt_format(struct mali_rt_format format)
854 {
855 pandecode_log(".format = {\n");
856 pandecode_indent++;
857
858 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
859 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
860 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
861
862 pandecode_prop("block = %s",
863 pandecode_mfbd_block_format(format.block));
864
865 /* TODO: Map formats so we can check swizzles and print nicely */
866 pandecode_log("swizzle");
867 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
868 pandecode_log_cont(",\n");
869
870 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
871 MALI_NEGATIVE(format.nr_channels));
872
873 pandecode_log(".flags = ");
874 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
875 pandecode_log_cont(",\n");
876
877 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
878 * which is a faster hardware-based alternative to the wallpaper method
879 * to preserve framebuffer contents across frames. In practice, MFBD
880 * preload is buggy on Midgard, and so this is a chicken bit. If this
881 * bit isn't set, most likely something broke unrelated to preload */
882
883 if (!format.no_preload) {
884 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
885 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
886 }
887
888 if (format.zero)
889 pandecode_prop("zero = 0x%" PRIx32, format.zero);
890
891 pandecode_indent--;
892 pandecode_log("},\n");
893 }
894
895 static void
896 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
897 {
898 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
899 pandecode_indent++;
900
901 for (int i = 0; i < MALI_NEGATIVE(fb->rt_count_1); i++) {
902 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
903 struct pandecode_mapped_memory *mem =
904 pandecode_find_mapped_gpu_mem_containing(rt_va);
905 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
906
907 pandecode_log("{\n");
908 pandecode_indent++;
909
910 pandecode_rt_format(rt->format);
911
912 if (rt->format.block == MALI_MFBD_BLOCK_AFBC) {
913 pandecode_log(".afbc = {\n");
914 pandecode_indent++;
915
916 char *a = pointer_as_memory_reference(rt->afbc.metadata);
917 pandecode_prop("metadata = %s", a);
918 free(a);
919
920 pandecode_prop("stride = %d", rt->afbc.stride);
921 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
922
923 pandecode_indent--;
924 pandecode_log("},\n");
925 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
926 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
927 rt->afbc.metadata,
928 rt->afbc.stride,
929 rt->afbc.unk);
930 }
931
932 MEMORY_PROP(rt, framebuffer);
933 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
934
935 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
936 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
937 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
938 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
939 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
940 }
941
942 if (rt->zero1 || rt->zero2 || rt->zero3) {
943 pandecode_msg("XXX: render target zeros tripped\n");
944 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
945 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
946 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
947 }
948
949 pandecode_indent--;
950 pandecode_log("},\n");
951 }
952
953 pandecode_indent--;
954 pandecode_log("};\n");
955 }
956
957 static unsigned
958 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment)
959 {
960 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
961 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
962
963 if (fb->sample_locations) {
964 /* The blob stores all possible sample locations in a single buffer
965 * allocated on startup, and just switches the pointer when switching
966 * MSAA state. For now, we just put the data into the cmdstream, but we
967 * should do something like what the blob does with a real driver.
968 *
969 * There seem to be 32 slots for sample locations, followed by another
970 * 16. The second 16 is just the center location followed by 15 zeros
971 * in all the cases I've identified (maybe shader vs. depth/color
972 * samples?).
973 */
974
975 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
976
977 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
978
979 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
980 pandecode_indent++;
981
982 for (int i = 0; i < 32 + 16; i++) {
983 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
984 }
985
986 pandecode_indent--;
987 pandecode_log("};\n");
988 }
989
990 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
991 pandecode_indent++;
992
993 pandecode_prop("unk0 = 0x%x", fb->unk0);
994
995 if (fb->sample_locations)
996 pandecode_prop("sample_locations = sample_locations_%d", job_no);
997
998 /* Assume that unknown1 was emitted in the last job for
999 * now */
1000 MEMORY_PROP(fb, unknown1);
1001
1002 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1003 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1004 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1005 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1006
1007 pandecode_prop("unk1 = 0x%x", fb->unk1);
1008 pandecode_prop("unk2 = 0x%x", fb->unk2);
1009 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1010 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1011
1012 pandecode_log(".mfbd_flags = ");
1013 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1014 pandecode_log_cont(",\n");
1015
1016 if (fb->clear_stencil)
1017 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1018
1019 if (fb->clear_depth)
1020 pandecode_prop("clear_depth = %f", fb->clear_depth);
1021
1022 /* TODO: What is this? Let's not blow up.. */
1023 if (fb->unknown2 != 0x1F)
1024 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1025
1026 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1027 MEMORY_PROP(fb, scratchpad);
1028 const struct midgard_tiler_descriptor t = fb->tiler;
1029 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment);
1030
1031 if (fb->zero3 || fb->zero4) {
1032 pandecode_msg("XXX: framebuffer zeros tripped\n");
1033 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1034 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1035 }
1036
1037 pandecode_indent--;
1038 pandecode_log("};\n");
1039
1040 gpu_va += sizeof(struct bifrost_framebuffer);
1041
1042 if ((fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment) {
1043 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1044 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1045
1046 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1047 pandecode_indent++;
1048
1049 MEMORY_PROP(fbx, checksum);
1050
1051 if (fbx->checksum_stride)
1052 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1053
1054 pandecode_log(".flags = ");
1055 pandecode_log_decoded_flags(mfbd_extra_flag_info, fbx->flags);
1056 pandecode_log_cont(",\n");
1057
1058 if (fbx->flags & MALI_EXTRA_AFBC_ZS) {
1059 pandecode_log(".ds_afbc = {\n");
1060 pandecode_indent++;
1061
1062 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1063 pandecode_prop("depth_stencil_afbc_stride = %d",
1064 fbx->ds_afbc.depth_stencil_afbc_stride);
1065 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1066
1067 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
1068 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1069 pandecode_prop("zero1 = 0x%" PRIx32,
1070 fbx->ds_afbc.zero1);
1071 pandecode_prop("padding = 0x%" PRIx64,
1072 fbx->ds_afbc.padding);
1073 }
1074
1075 pandecode_indent--;
1076 pandecode_log("},\n");
1077 } else {
1078 pandecode_log(".ds_linear = {\n");
1079 pandecode_indent++;
1080
1081 if (fbx->ds_linear.depth) {
1082 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1083 pandecode_prop("depth_stride = %d",
1084 fbx->ds_linear.depth_stride);
1085 }
1086
1087 if (fbx->ds_linear.stencil) {
1088 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1089 pandecode_prop("stencil_stride = %d",
1090 fbx->ds_linear.stencil_stride);
1091 }
1092
1093 if (fbx->ds_linear.depth_stride_zero ||
1094 fbx->ds_linear.stencil_stride_zero ||
1095 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1096 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1097 pandecode_prop("depth_stride_zero = 0x%x",
1098 fbx->ds_linear.depth_stride_zero);
1099 pandecode_prop("stencil_stride_zero = 0x%x",
1100 fbx->ds_linear.stencil_stride_zero);
1101 pandecode_prop("zero1 = 0x%" PRIx32,
1102 fbx->ds_linear.zero1);
1103 pandecode_prop("zero2 = 0x%" PRIx32,
1104 fbx->ds_linear.zero2);
1105 }
1106
1107 pandecode_indent--;
1108 pandecode_log("},\n");
1109 }
1110
1111 if (fbx->zero3 || fbx->zero4) {
1112 pandecode_msg("XXX: fb_extra zeros tripped\n");
1113 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1114 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1115 }
1116
1117 pandecode_indent--;
1118 pandecode_log("};\n");
1119
1120 gpu_va += sizeof(struct bifrost_fb_extra);
1121 }
1122
1123 if (is_fragment)
1124 pandecode_render_target(gpu_va, job_no, fb);
1125
1126 /* Passback the render target count */
1127 return MALI_NEGATIVE(fb->rt_count_1);
1128 }
1129
1130 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1131 * count */
1132
1133 static void
1134 pandecode_padded_vertices(unsigned shift, unsigned k)
1135 {
1136 unsigned odd = 2*k + 1;
1137 unsigned pot = 1 << shift;
1138 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1139 }
1140
1141 /* Given a magic divisor, recover what we were trying to divide by.
1142 *
1143 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1144 * 0 <= m < 2^N, for N bits in m.
1145 *
1146 * Let q represent the number we would like to divide by.
1147 *
1148 * By definition of a magic divisor for N-bit unsigned integers (a number you
1149 * multiply by to magically get division), m is a number such that:
1150 *
1151 * (m * x) & (2^N - 1) = floor(x/q).
1152 * for all x on Z where 0 <= x < 2^N
1153 *
1154 * Ignore the case where any of the above values equals zero; it is irrelevant
1155 * for our purposes (instanced arrays).
1156 *
1157 * Choose x = q. Then:
1158 *
1159 * (m * x) & (2^N - 1) = floor(x/q).
1160 * (m * q) & (2^N - 1) = floor(q/q).
1161 *
1162 * floor(q/q) = floor(1) = 1, therefore:
1163 *
1164 * (m * q) & (2^N - 1) = 1
1165 *
1166 * Recall the identity that the bitwise AND of one less than a power-of-two
1167 * equals the modulo with that power of two, i.e. for all x:
1168 *
1169 * x & (2^N - 1) = x % N
1170 *
1171 * Therefore:
1172 *
1173 * mq % (2^N) = 1
1174 *
1175 * By definition, a modular multiplicative inverse of a number m is the number
1176 * q such that with respect to a modulos M:
1177 *
1178 * mq % M = 1
1179 *
1180 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1181 *
1182 */
1183
1184 static void
1185 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1186 {
1187 #if 0
1188 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1189 * shift) the most lame way possible... just repeatedly add.
1190 * Asymptoptically slow but nobody cares in practice, unless you have
1191 * massive numbers of vertices or high divisors. */
1192
1193 unsigned inverse = 0;
1194
1195 /* Magic implicitly has the highest bit set */
1196 magic |= (1 << 31);
1197
1198 /* Depending on rounding direction */
1199 if (extra)
1200 magic++;
1201
1202 for (;;) {
1203 uint32_t product = magic * inverse;
1204
1205 if (shift) {
1206 product >>= shift;
1207 }
1208
1209 if (product == 1)
1210 break;
1211
1212 ++inverse;
1213 }
1214
1215 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1216
1217 /* Recall we're supposed to divide by (gl_level_divisor *
1218 * padded_num_vertices) */
1219
1220 unsigned padded_num_vertices = inverse / orig_divisor;
1221
1222 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1223 #endif
1224 }
1225
1226 static void
1227 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1228 mali_ptr addr, int job_no, char *suffix,
1229 int count, bool varying)
1230 {
1231 char *prefix = varying ? "varyings" : "attributes";
1232
1233 if (!addr) {
1234 pandecode_msg("no %s\n", prefix);
1235 return;
1236 }
1237
1238 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1239
1240 char base[128];
1241 snprintf(base, sizeof(base), "%s_data_%d%s", prefix, job_no, suffix);
1242
1243 for (int i = 0; i < count; ++i) {
1244 enum mali_attr_mode mode = attr[i].elements & 7;
1245
1246 if (mode == MALI_ATTR_UNUSED)
1247 continue;
1248
1249 mali_ptr raw_elements = attr[i].elements & ~7;
1250
1251 /* TODO: Do we maybe want to dump the attribute values
1252 * themselves given the specified format? Or is that too hard?
1253 * */
1254
1255 char *a = pointer_as_memory_reference(raw_elements);
1256 pandecode_log("mali_ptr %s_%d_p = %s;\n", base, i, a);
1257 free(a);
1258 }
1259
1260 pandecode_log("union mali_attr %s_%d[] = {\n", prefix, job_no);
1261 pandecode_indent++;
1262
1263 for (int i = 0; i < count; ++i) {
1264 pandecode_log("{\n");
1265 pandecode_indent++;
1266
1267 unsigned mode = attr[i].elements & 7;
1268 pandecode_prop("elements = (%s_%d_p) | %s", base, i, pandecode_attr_mode(mode));
1269 pandecode_prop("shift = %d", attr[i].shift);
1270 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1271 pandecode_prop("stride = 0x%" PRIx32, attr[i].stride);
1272 pandecode_prop("size = 0x%" PRIx32, attr[i].size);
1273
1274 /* Decode further where possible */
1275
1276 if (mode == MALI_ATTR_MODULO) {
1277 pandecode_padded_vertices(
1278 attr[i].shift,
1279 attr[i].extra_flags);
1280 }
1281
1282 pandecode_indent--;
1283 pandecode_log("}, \n");
1284
1285 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1286 i++;
1287 pandecode_log("{\n");
1288 pandecode_indent++;
1289 pandecode_prop("unk = 0x%x", attr[i].unk);
1290 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1291 if (attr[i].zero != 0)
1292 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1293 pandecode_prop("divisor = %d", attr[i].divisor);
1294 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1295 pandecode_indent--;
1296 pandecode_log("}, \n");
1297 }
1298
1299 }
1300
1301 pandecode_indent--;
1302 pandecode_log("};\n");
1303 }
1304
1305 static mali_ptr
1306 pandecode_shader_address(const char *name, mali_ptr ptr)
1307 {
1308 /* TODO: Decode flags */
1309 mali_ptr shader_ptr = ptr & ~15;
1310
1311 char *a = pointer_as_memory_reference(shader_ptr);
1312 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1313 free(a);
1314
1315 return shader_ptr;
1316 }
1317
1318 static bool
1319 all_zero(unsigned *buffer, unsigned count)
1320 {
1321 for (unsigned i = 0; i < count; ++i) {
1322 if (buffer[i])
1323 return false;
1324 }
1325
1326 return true;
1327 }
1328
1329 static void
1330 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1331 {
1332 if (all_zero((unsigned *) stencil, sizeof(stencil) / sizeof(unsigned)))
1333 return;
1334
1335 const char *func = pandecode_func(stencil->func);
1336 const char *sfail = pandecode_stencil_op(stencil->sfail);
1337 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1338 const char *dppass = pandecode_stencil_op(stencil->dppass);
1339
1340 if (stencil->zero)
1341 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1342
1343 pandecode_log(".stencil_%s = {\n", name);
1344 pandecode_indent++;
1345 pandecode_prop("ref = %d", stencil->ref);
1346 pandecode_prop("mask = 0x%02X", stencil->mask);
1347 pandecode_prop("func = %s", func);
1348 pandecode_prop("sfail = %s", sfail);
1349 pandecode_prop("dpfail = %s", dpfail);
1350 pandecode_prop("dppass = %s", dppass);
1351 pandecode_indent--;
1352 pandecode_log("},\n");
1353 }
1354
1355 static void
1356 pandecode_blend_equation(const struct mali_blend_equation *blend)
1357 {
1358 if (blend->zero1)
1359 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1360
1361 pandecode_log(".equation = {\n");
1362 pandecode_indent++;
1363
1364 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1365 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1366
1367 pandecode_log(".color_mask = ");
1368 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1369 pandecode_log_cont(",\n");
1370
1371 pandecode_indent--;
1372 pandecode_log("},\n");
1373 }
1374
1375 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1376
1377 static unsigned
1378 decode_bifrost_constant(u16 constant)
1379 {
1380 float lo = (float) (constant & 0xFF);
1381 float hi = (float) (constant >> 8);
1382
1383 return (hi / 255.0) + (lo / 65535.0);
1384 }
1385
1386 static mali_ptr
1387 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1388 {
1389 struct bifrost_blend_rt *b =
1390 ((struct bifrost_blend_rt *) descs) + rt_no;
1391
1392 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1393 pandecode_indent++;
1394
1395 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1396 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1397 b->constant, decode_bifrost_constant(b->constant));
1398
1399 /* TODO figure out blend shader enable bit */
1400 pandecode_blend_equation(&b->equation);
1401 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1402 pandecode_prop("index = 0x%" PRIx16, b->index);
1403 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1404
1405 pandecode_indent--;
1406 pandecode_log("},\n");
1407
1408 return 0;
1409 }
1410
1411 static mali_ptr
1412 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1413 {
1414 if (all_zero((unsigned *) blend, sizeof(blend) / sizeof(unsigned)))
1415 return 0;
1416
1417 pandecode_log(".blend = {\n");
1418 pandecode_indent++;
1419
1420 if (is_shader) {
1421 pandecode_shader_address("shader", blend->shader);
1422 } else {
1423 pandecode_blend_equation(&blend->equation);
1424 pandecode_prop("constant = %f", blend->constant);
1425 }
1426
1427 pandecode_indent--;
1428 pandecode_log("},\n");
1429
1430 /* Return blend shader to disassemble if present */
1431 return is_shader ? (blend->shader & ~0xF) : 0;
1432 }
1433
1434 static mali_ptr
1435 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1436 {
1437 struct midgard_blend_rt *b =
1438 ((struct midgard_blend_rt *) descs) + rt_no;
1439
1440 /* Flags determine presence of blend shader */
1441 bool is_shader = (b->flags & 0xF) >= 0x2;
1442
1443 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1444 pandecode_indent++;
1445
1446 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1447
1448 union midgard_blend blend = b->blend;
1449 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1450
1451 pandecode_indent--;
1452 pandecode_log("};\n");
1453
1454 return shader;
1455 }
1456
1457 /* Attributes and varyings have descriptor records, which contain information
1458 * about their format and ordering with the attribute/varying buffers. We'll
1459 * want to validate that the combinations specified are self-consistent.
1460 */
1461
1462 static int
1463 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1464 {
1465 char base[128];
1466 char *prefix = varying ? "varying" : "attribute";
1467 unsigned max_index = 0;
1468 snprintf(base, sizeof(base), "%s_meta", prefix);
1469
1470 pandecode_log("struct mali_attr_meta %s_%d%s[] = {\n", base, job_no, suffix);
1471 pandecode_indent++;
1472
1473 struct mali_attr_meta *attr_meta;
1474 mali_ptr p = varying ? (v->varying_meta & ~0xF) : v->attribute_meta;
1475
1476 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1477
1478 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1479 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1480 sizeof(*attr_mem));
1481
1482 /* If the record is discard, it should be zero for everything else */
1483
1484 if (attr_meta->format == MALI_VARYING_DISCARD) {
1485 uint64_t zero =
1486 attr_meta->index |
1487 attr_meta->unknown1 |
1488 attr_meta->unknown3 |
1489 attr_meta->src_offset;
1490
1491 if (zero)
1492 pandecode_msg("XXX: expected empty record for varying discard\n");
1493
1494 /* We want to look for a literal 0000 swizzle -- this
1495 * is not encoded with all zeroes, however */
1496
1497 enum mali_channel z = MALI_CHANNEL_ZERO;
1498 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1499 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1500
1501 if (!good_swizzle)
1502 pandecode_msg("XXX: expected zero swizzle for discard\n");
1503
1504 if (!varying)
1505 pandecode_msg("XXX: cannot discard attribute\n");
1506
1507 /* If we're all good, omit the record */
1508 if (!zero && varying && good_swizzle) {
1509 pandecode_log("/* discarded varying */\n");
1510 continue;
1511 }
1512 }
1513
1514 if (attr_meta->index > max_index)
1515 max_index = attr_meta->index;
1516
1517 if (attr_meta->unknown1 != 0x2) {
1518 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1519 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1520 }
1521
1522 if (attr_meta->unknown3) {
1523 pandecode_msg("XXX: unexpected unknown3 set\n");
1524 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1525 }
1526
1527 pandecode_make_indent();
1528 pandecode_format_short(attr_meta->format);
1529 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1530
1531 if (attr_meta->src_offset)
1532 pandecode_log_cont("[%u]", attr_meta->src_offset);
1533
1534 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1535
1536 pandecode_log_cont(";\n");
1537 }
1538
1539 pandecode_indent--;
1540 pandecode_log("};\n");
1541
1542 return count ? (max_index + 1) : 0;
1543 }
1544
1545 static void
1546 pandecode_indices(uintptr_t pindices, uint32_t index_count, int job_no)
1547 {
1548 struct pandecode_mapped_memory *imem = pandecode_find_mapped_gpu_mem_containing(pindices);
1549
1550 if (imem) {
1551 /* Indices are literally just a u32 array :) */
1552
1553 uint32_t *PANDECODE_PTR_VAR(indices, imem, pindices);
1554
1555 pandecode_log("uint32_t indices_%d[] = {\n", job_no);
1556 pandecode_indent++;
1557
1558 for (unsigned i = 0; i < (index_count + 1); i += 3)
1559 pandecode_log("%d, %d, %d,\n",
1560 indices[i],
1561 indices[i + 1],
1562 indices[i + 2]);
1563
1564 pandecode_indent--;
1565 pandecode_log("};\n");
1566 }
1567 }
1568
1569 /* return bits [lo, hi) of word */
1570 static u32
1571 bits(u32 word, u32 lo, u32 hi)
1572 {
1573 if (hi - lo >= 32)
1574 return word; // avoid undefined behavior with the shift
1575
1576 return (word >> lo) & ((1 << (hi - lo)) - 1);
1577 }
1578
1579 static void
1580 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool noninstanced)
1581 {
1582 pandecode_log_cont("{\n");
1583 pandecode_indent++;
1584
1585 /* Decode invocation_count. See the comment before the definition of
1586 * invocation_count for an explanation.
1587 */
1588
1589 unsigned size_x = bits(p->invocation_count, 0, p->size_y_shift) + 1;
1590 unsigned size_y = bits(p->invocation_count, p->size_y_shift, p->size_z_shift) + 1;
1591 unsigned size_z = bits(p->invocation_count, p->size_z_shift, p->workgroups_x_shift) + 1;
1592
1593 unsigned groups_x = bits(p->invocation_count, p->workgroups_x_shift, p->workgroups_y_shift) + 1;
1594 unsigned groups_y = bits(p->invocation_count, p->workgroups_y_shift, p->workgroups_z_shift) + 1;
1595 unsigned groups_z = bits(p->invocation_count, p->workgroups_z_shift, 32) + 1;
1596
1597 /* Even though we have this decoded, we want to ensure that the
1598 * representation is "unique" so we don't lose anything by printing only
1599 * the final result. More specifically, we need to check that we were
1600 * passed something in canonical form, since the definition per the
1601 * hardware is inherently not unique. How? Well, take the resulting
1602 * decode and pack it ourselves! If it is bit exact with what we
1603 * decoded, we're good to go. */
1604
1605 struct mali_vertex_tiler_prefix ref;
1606 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, noninstanced);
1607
1608 bool canonical =
1609 (p->invocation_count == ref.invocation_count) &&
1610 (p->size_y_shift == ref.size_y_shift) &&
1611 (p->size_z_shift == ref.size_z_shift) &&
1612 (p->workgroups_x_shift == ref.workgroups_x_shift) &&
1613 (p->workgroups_y_shift == ref.workgroups_y_shift) &&
1614 (p->workgroups_z_shift == ref.workgroups_z_shift) &&
1615 (p->workgroups_x_shift_2 == ref.workgroups_x_shift_2);
1616
1617 if (!canonical) {
1618 pandecode_msg("XXX: non-canonical workgroups packing\n");
1619 pandecode_msg("expected: %X, %d, %d, %d, %d, %d\n",
1620 ref.invocation_count,
1621 ref.size_y_shift,
1622 ref.size_z_shift,
1623 ref.workgroups_x_shift,
1624 ref.workgroups_y_shift,
1625 ref.workgroups_z_shift,
1626 ref.workgroups_x_shift_2);
1627
1628 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1629 pandecode_prop("size_y_shift = %d", p->size_y_shift);
1630 pandecode_prop("size_z_shift = %d", p->size_z_shift);
1631 pandecode_prop("workgroups_x_shift = %d", p->workgroups_x_shift);
1632 pandecode_prop("workgroups_y_shift = %d", p->workgroups_y_shift);
1633 pandecode_prop("workgroups_z_shift = %d", p->workgroups_z_shift);
1634 pandecode_prop("workgroups_x_shift_2 = %d", p->workgroups_x_shift_2);
1635 }
1636
1637 /* Regardless, print the decode */
1638 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1639 size_x, size_y, size_z,
1640 groups_x, groups_y, groups_z);
1641
1642 /* TODO: Decode */
1643 if (p->unknown_draw)
1644 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1645
1646 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1647
1648 if (p->draw_mode != MALI_DRAW_NONE)
1649 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1650
1651 /* Index count only exists for tiler jobs anyway */
1652
1653 if (p->index_count)
1654 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1655
1656 if (p->offset_bias_correction)
1657 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1658
1659 /* TODO: Figure out what this is. It's not zero */
1660 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1661
1662 pandecode_indent--;
1663 pandecode_log("},\n");
1664 }
1665
1666 static void
1667 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1668 {
1669 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1670 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1671
1672 pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%"PRIx64"_%d[] = {\n",
1673 pubufs, job_no);
1674 pandecode_indent++;
1675
1676 for (int i = 0; i < ubufs_count; i++) {
1677 pandecode_log("{\n");
1678 pandecode_indent++;
1679
1680 unsigned size = (ubufs[i].size + 1) * 16;
1681 mali_ptr addr = ubufs[i].ptr << 2;
1682
1683 pandecode_validate_buffer(addr, size);
1684
1685 char *ptr = pointer_as_memory_reference(ubufs[i].ptr << 2);
1686 pandecode_prop("size = %u", size);
1687 pandecode_prop("ptr = (%s) >> 2", ptr);
1688 pandecode_indent--;
1689 pandecode_log("},\n");
1690 free(ptr);
1691 }
1692
1693 pandecode_indent--;
1694 pandecode_log("};\n");
1695 }
1696
1697 static void
1698 pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1699 {
1700
1701 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1702
1703 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1704
1705 if (scratchpad->zero) {
1706 pandecode_msg("XXX: scratchpad zero tripped");
1707 pandecode_prop("zero = 0x%x\n", scratchpad->zero);
1708 }
1709
1710 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1711 pandecode_indent++;
1712
1713 pandecode_prop("flags = 0x%x", scratchpad->flags);
1714 MEMORY_PROP(scratchpad, gpu_scratchpad);
1715
1716 pandecode_indent--;
1717 pandecode_log("};\n");
1718 }
1719
1720 static unsigned shader_id = 0;
1721
1722 static void
1723 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1724 bool is_bifrost, unsigned nr_regs)
1725 {
1726 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1727 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1728
1729 /* Compute maximum possible size */
1730 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1731
1732 /* Print some boilerplate to clearly denote the assembly (which doesn't
1733 * obey indentation rules), and actually do the disassembly! */
1734
1735 printf("\n\n");
1736
1737 char prefix[512];
1738
1739 snprintf(prefix, sizeof(prefix) - 1, "shader%d - %s shader: ",
1740 shader_id++,
1741 (type == JOB_TYPE_TILER) ? "FRAGMENT" : "VERTEX");
1742
1743 if (is_bifrost) {
1744 disassemble_bifrost(code, sz, false);
1745 } else {
1746 disassemble_midgard(code, sz, true, nr_regs, prefix);
1747 }
1748
1749 printf("\n\n");
1750 }
1751
1752 static void
1753 pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p,
1754 int job_no, enum mali_job_type job_type,
1755 char *suffix, bool is_bifrost)
1756 {
1757 mali_ptr shader_meta_ptr = (u64) (uintptr_t) (p->_shader_upper << 4);
1758 struct pandecode_mapped_memory *attr_mem;
1759
1760 unsigned rt_count = 1;
1761
1762 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1763 * are the only things actually needed from the FBD, vertex/tiler jobs
1764 * no longer reference the FBD -- instead, this field points to some
1765 * info about the scratchpad.
1766 */
1767 if (is_bifrost)
1768 pandecode_scratchpad(p->framebuffer & ~FBD_TYPE, job_no, suffix);
1769 else if (p->framebuffer & MALI_MFBD)
1770 rt_count = pandecode_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
1771 else if (job_type == JOB_TYPE_COMPUTE)
1772 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
1773 else
1774 pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no, false);
1775
1776 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
1777 int texture_count = 0, sampler_count = 0;
1778
1779 if (shader_meta_ptr) {
1780 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr);
1781 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, shader_meta_ptr);
1782
1783 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", shader_meta_ptr, job_no, suffix);
1784 pandecode_indent++;
1785
1786 /* Save for dumps */
1787 attribute_count = s->attribute_count;
1788 varying_count = s->varying_count;
1789 texture_count = s->texture_count;
1790 sampler_count = s->sampler_count;
1791
1792 if (is_bifrost) {
1793 uniform_count = s->bifrost2.uniform_count;
1794 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
1795 } else {
1796 uniform_count = s->midgard1.uniform_buffer_count;
1797 uniform_buffer_count = s->midgard1.uniform_buffer_count;
1798 }
1799
1800 mali_ptr shader_ptr = pandecode_shader_address("shader", s->shader);
1801
1802 pandecode_prop("texture_count = %" PRId16, s->texture_count);
1803 pandecode_prop("sampler_count = %" PRId16, s->sampler_count);
1804 pandecode_prop("attribute_count = %" PRId16, s->attribute_count);
1805 pandecode_prop("varying_count = %" PRId16, s->varying_count);
1806
1807 unsigned nr_registers = 0;
1808
1809 if (is_bifrost) {
1810 pandecode_log(".bifrost1 = {\n");
1811 pandecode_indent++;
1812
1813 pandecode_prop("uniform_buffer_count = %" PRId32, s->bifrost1.uniform_buffer_count);
1814 pandecode_prop("unk1 = 0x%" PRIx32, s->bifrost1.unk1);
1815
1816 pandecode_indent--;
1817 pandecode_log("},\n");
1818 } else {
1819 pandecode_log(".midgard1 = {\n");
1820 pandecode_indent++;
1821
1822 pandecode_prop("uniform_count = %" PRId16, s->midgard1.uniform_count);
1823 pandecode_prop("uniform_buffer_count = %" PRId16, s->midgard1.uniform_buffer_count);
1824 pandecode_prop("work_count = %" PRId16, s->midgard1.work_count);
1825 nr_registers = s->midgard1.work_count;
1826
1827 pandecode_log(".flags = ");
1828 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
1829 pandecode_log_cont(",\n");
1830
1831 pandecode_prop("unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
1832
1833 pandecode_indent--;
1834 pandecode_log("},\n");
1835 }
1836
1837 if (s->depth_units || s->depth_factor) {
1838 pandecode_prop("depth_factor = %f", s->depth_factor);
1839 pandecode_prop("depth_units = %f", s->depth_units);
1840 }
1841
1842 if (s->alpha_coverage) {
1843 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
1844 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
1845
1846 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
1847 invert_alpha_coverage ? "~" : "",
1848 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
1849 }
1850
1851 if (s->unknown2_3 || s->unknown2_4) {
1852 pandecode_log(".unknown2_3 = ");
1853
1854 int unknown2_3 = s->unknown2_3;
1855 int unknown2_4 = s->unknown2_4;
1856
1857 /* We're not quite sure what these flags mean without the depth test, if anything */
1858
1859 if (unknown2_3 & (MALI_DEPTH_TEST | MALI_DEPTH_FUNC_MASK)) {
1860 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
1861 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
1862
1863 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
1864 }
1865
1866 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
1867 pandecode_log_cont(",\n");
1868
1869 pandecode_log(".unknown2_4 = ");
1870 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
1871 pandecode_log_cont(",\n");
1872 }
1873
1874 if (s->stencil_mask_front || s->stencil_mask_back) {
1875 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
1876 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
1877 }
1878
1879 pandecode_stencil("front", &s->stencil_front);
1880 pandecode_stencil("back", &s->stencil_back);
1881
1882 if (is_bifrost) {
1883 pandecode_log(".bifrost2 = {\n");
1884 pandecode_indent++;
1885
1886 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
1887 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
1888 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
1889 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
1890
1891 pandecode_indent--;
1892 pandecode_log("},\n");
1893 } else if (s->midgard2.unknown2_7) {
1894 pandecode_log(".midgard2 = {\n");
1895 pandecode_indent++;
1896
1897 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
1898 pandecode_indent--;
1899 pandecode_log("},\n");
1900 }
1901
1902 if (s->unknown2_8)
1903 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
1904
1905 if (!is_bifrost) {
1906 /* TODO: Blend shaders routing/disasm */
1907
1908 union midgard_blend blend = s->blend;
1909 pandecode_midgard_blend(&blend, false);
1910 }
1911
1912 pandecode_indent--;
1913 pandecode_log("};\n");
1914
1915 /* MRT blend fields are used whenever MFBD is used, with
1916 * per-RT descriptors */
1917
1918 if (job_type == JOB_TYPE_TILER) {
1919 void* blend_base = (void *) (s + 1);
1920
1921 for (unsigned i = 0; i < rt_count; i++) {
1922 mali_ptr shader = 0;
1923
1924 if (is_bifrost)
1925 shader = pandecode_bifrost_blend(blend_base, job_no, i);
1926 else
1927 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
1928
1929 if (shader & ~0xF)
1930 pandecode_shader_disassemble(shader, job_no, job_type, false, 0);
1931 }
1932 }
1933
1934 if (shader_ptr & ~0xF)
1935 pandecode_shader_disassemble(shader_ptr, job_no, job_type, is_bifrost, nr_registers);
1936 } else
1937 pandecode_msg("<no shader>\n");
1938
1939 if (p->viewport) {
1940 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
1941 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
1942
1943 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
1944 pandecode_indent++;
1945
1946 pandecode_prop("clip_minx = %f", f->clip_minx);
1947 pandecode_prop("clip_miny = %f", f->clip_miny);
1948 pandecode_prop("clip_minz = %f", f->clip_minz);
1949 pandecode_prop("clip_maxx = %f", f->clip_maxx);
1950 pandecode_prop("clip_maxy = %f", f->clip_maxy);
1951 pandecode_prop("clip_maxz = %f", f->clip_maxz);
1952
1953 /* Only the higher coordinates are MALI_POSITIVE scaled */
1954
1955 pandecode_prop("viewport0 = { %d, %d }",
1956 f->viewport0[0], f->viewport0[1]);
1957
1958 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
1959 f->viewport1[0] + 1, f->viewport1[1] + 1);
1960
1961 pandecode_indent--;
1962 pandecode_log("};\n");
1963 }
1964
1965 if (p->attribute_meta) {
1966 unsigned max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
1967
1968 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
1969 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false);
1970 }
1971
1972 /* Varyings are encoded like attributes but not actually sent; we just
1973 * pass a zero buffer with the right stride/size set, (or whatever)
1974 * since the GPU will write to it itself */
1975
1976 if (p->varying_meta) {
1977 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
1978 }
1979
1980 if (p->varyings) {
1981 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
1982
1983 /* Number of descriptors depends on whether there are
1984 * non-internal varyings */
1985
1986 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true);
1987 }
1988
1989 if (p->uniform_buffers) {
1990 if (uniform_buffer_count)
1991 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
1992 else
1993 pandecode_msg("XXX: UBOs specified but not referenced\n");
1994 } else if (uniform_buffer_count)
1995 pandecode_msg("XXX: UBOs referenced but not specified\n");
1996
1997 /* We don't want to actually dump uniforms, but we do need to validate
1998 * that the counts we were given are sane */
1999
2000 if (p->uniforms) {
2001 if (uniform_count)
2002 pandecode_validate_buffer(p->uniforms, uniform_count * 16);
2003 else
2004 pandecode_msg("XXX: Uniforms specified but not referenced");
2005 } else if (uniform_count)
2006 pandecode_msg("XXX: UBOs referenced but not specified\n");
2007
2008 if (p->texture_trampoline) {
2009 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
2010
2011 if (mmem) {
2012 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
2013
2014 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
2015 pandecode_indent++;
2016
2017 for (int tex = 0; tex < texture_count; ++tex) {
2018 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2019 char *a = pointer_as_memory_reference(*u);
2020 pandecode_log("%s,\n", a);
2021 free(a);
2022 }
2023
2024 pandecode_indent--;
2025 pandecode_log("};\n");
2026
2027 /* Now, finally, descend down into the texture descriptor */
2028 for (int tex = 0; tex < texture_count; ++tex) {
2029 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2030 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2031
2032 if (tmem) {
2033 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, *u);
2034
2035 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", *u, job_no, tex);
2036 pandecode_indent++;
2037
2038 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", t->width + 1);
2039 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", t->height + 1);
2040 pandecode_prop("depth = MALI_POSITIVE(%" PRId16 ")", t->depth + 1);
2041 pandecode_prop("array_size = MALI_POSITIVE(%" PRId16 ")", t->array_size + 1);
2042 pandecode_prop("nr_mipmap_levels = %" PRId8, t->nr_mipmap_levels);
2043
2044 struct mali_texture_format f = t->format;
2045
2046 pandecode_log(".format = {\n");
2047 pandecode_indent++;
2048
2049 pandecode_log(".format = ");
2050 pandecode_format_short(f.format);
2051 pandecode_swizzle(f.swizzle, f.format);
2052 pandecode_log_cont(",\n");
2053
2054 pandecode_prop("type = %s", pandecode_texture_type(f.type));
2055 pandecode_prop("srgb = %" PRId32, f.srgb);
2056 pandecode_prop("usage2 = 0x%" PRIx32, f.usage2);
2057
2058 if (f.unknown1) {
2059 pandecode_msg("XXX: texture format zero tripped\n");
2060 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
2061 }
2062
2063 pandecode_indent--;
2064 pandecode_log("},\n");
2065
2066 if (t->swizzle_zero) {
2067 pandecode_msg("XXX: swizzle zero tripped\n");
2068 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
2069 }
2070
2071 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
2072 pandecode_msg("XXX: texture zero tripped\n");
2073 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
2074 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
2075 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2076 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2077 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2078 }
2079
2080 pandecode_log(".payload = {\n");
2081 pandecode_indent++;
2082
2083 /* A bunch of bitmap pointers follow.
2084 * We work out the correct number,
2085 * based on the mipmap/cubemap
2086 * properties, but dump extra
2087 * possibilities to futureproof */
2088
2089 int bitmap_count = MALI_NEGATIVE(t->nr_mipmap_levels);
2090 bool manual_stride = f.usage2 & MALI_TEX_MANUAL_STRIDE;
2091
2092 /* Miptree for each face */
2093 if (f.type == MALI_TEX_CUBE)
2094 bitmap_count *= 6;
2095
2096 /* Array of textures */
2097 bitmap_count *= MALI_NEGATIVE(t->array_size);
2098
2099 /* Stride for each element */
2100 if (manual_stride)
2101 bitmap_count *= 2;
2102
2103 /* Sanity check the size */
2104 int max_count = sizeof(t->payload) / sizeof(t->payload[0]);
2105 assert (bitmap_count <= max_count);
2106
2107 for (int i = 0; i < bitmap_count; ++i) {
2108 /* How we dump depends if this is a stride or a pointer */
2109
2110 if ((f.usage2 & MALI_TEX_MANUAL_STRIDE) && (i & 1)) {
2111 /* signed 32-bit snuck in as a 64-bit pointer */
2112 uint64_t stride_set = t->payload[i];
2113 uint32_t clamped_stride = stride_set;
2114 int32_t stride = clamped_stride;
2115 assert(stride_set == clamped_stride);
2116 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2117 } else {
2118 char *a = pointer_as_memory_reference(t->payload[i]);
2119 pandecode_log("%s, \n", a);
2120 free(a);
2121 }
2122 }
2123
2124 pandecode_indent--;
2125 pandecode_log("},\n");
2126
2127 pandecode_indent--;
2128 pandecode_log("};\n");
2129 }
2130 }
2131 }
2132 }
2133
2134 if (p->sampler_descriptor) {
2135 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2136
2137 if (smem) {
2138 struct mali_sampler_descriptor *s;
2139
2140 mali_ptr d = p->sampler_descriptor;
2141
2142 for (int i = 0; i < sampler_count; ++i) {
2143 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2144
2145 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2146 pandecode_indent++;
2147
2148 pandecode_log(".filter_mode = ");
2149 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2150 pandecode_log_cont(",\n");
2151
2152 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2153 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2154
2155 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2156 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2157 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2158
2159 pandecode_prop("compare_func = %s", pandecode_alt_func(s->compare_func));
2160
2161 if (s->zero || s->zero2) {
2162 pandecode_msg("XXX: sampler zero tripped\n");
2163 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2164 }
2165
2166 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2167
2168 pandecode_prop("border_color = { %f, %f, %f, %f }",
2169 s->border_color[0],
2170 s->border_color[1],
2171 s->border_color[2],
2172 s->border_color[3]);
2173
2174 pandecode_indent--;
2175 pandecode_log("};\n");
2176 }
2177 }
2178 }
2179 }
2180
2181 static void
2182 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2183 {
2184 if (!(p->position_varying || p->occlusion_counter || p->flags))
2185 return;
2186
2187 pandecode_log(".postfix = {\n");
2188 pandecode_indent++;
2189
2190 MEMORY_PROP(p, position_varying);
2191 MEMORY_PROP(p, occlusion_counter);
2192
2193 if (p->flags)
2194 pandecode_prop("flags = %d", p->flags);
2195
2196 pandecode_indent--;
2197 pandecode_log("},\n");
2198 }
2199
2200 static void
2201 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2202 {
2203 pandecode_log_cont("{\n");
2204 pandecode_indent++;
2205
2206 pandecode_prop("unk2 = 0x%x", v->unk2);
2207
2208 if (v->zero0 || v->zero1) {
2209 pandecode_msg("XXX: vertex only zero tripped");
2210 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2211 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2212 }
2213
2214 pandecode_indent--;
2215 pandecode_log("}\n");
2216 }
2217
2218 static void
2219 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2220 {
2221
2222 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2223 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2224
2225 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2226 pandecode_indent++;
2227
2228 if (h->zero) {
2229 pandecode_msg("XXX: tiler heap zero tripped\n");
2230 pandecode_prop("zero = 0x%x", h->zero);
2231 }
2232
2233 for (int i = 0; i < 12; i++) {
2234 if (h->zeros[i] != 0) {
2235 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2236 i, h->zeros[i]);
2237 }
2238 }
2239
2240 pandecode_prop("heap_size = 0x%x", h->heap_size);
2241 MEMORY_PROP(h, tiler_heap_start);
2242 MEMORY_PROP(h, tiler_heap_free);
2243
2244 /* this might point to the beginning of another buffer, when it's
2245 * really the end of the tiler heap buffer, so we have to be careful
2246 * here. but for zero length, we need the same pointer.
2247 */
2248
2249 if (h->tiler_heap_end == h->tiler_heap_start) {
2250 MEMORY_PROP(h, tiler_heap_start);
2251 } else {
2252 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2253 pandecode_prop("tiler_heap_end = %s + 1", a);
2254 free(a);
2255 }
2256
2257 pandecode_indent--;
2258 pandecode_log("};\n");
2259 }
2260
2261 static void
2262 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2263 {
2264 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2265 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2266
2267 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2268
2269 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2270 pandecode_indent++;
2271
2272 if (t->zero0 || t->zero1) {
2273 pandecode_msg("XXX: tiler meta zero tripped\n");
2274 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2275 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2276 }
2277
2278 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2279 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2280
2281 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2282 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2283
2284 for (int i = 0; i < 12; i++) {
2285 if (t->zeros[i] != 0) {
2286 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2287 i, t->zeros[i]);
2288 }
2289 }
2290
2291 pandecode_indent--;
2292 pandecode_log("};\n");
2293 }
2294
2295 static void
2296 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2297 {
2298 pandecode_log(".gl_enables = ");
2299
2300 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2301
2302 pandecode_log_cont(",\n");
2303 }
2304
2305 static void
2306 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2307 {
2308 if (u.pointer == 0x0)
2309 return;
2310
2311 pandecode_log(".primitive_size = {\n");
2312 pandecode_indent++;
2313
2314 if (constant) {
2315 pandecode_prop("constant = %f", u.constant);
2316 } else {
2317 MEMORY_PROP((&u), pointer);
2318 }
2319
2320 pandecode_indent--;
2321 pandecode_log("},\n");
2322 }
2323
2324 static void
2325 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2326 {
2327 pandecode_log_cont("{\n");
2328 pandecode_indent++;
2329
2330 /* TODO: gl_PointSize on Bifrost */
2331 pandecode_primitive_size(t->primitive_size, true);
2332
2333 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2334
2335 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2336 || t->zero6 || t->zero7 || t->zero8) {
2337 pandecode_msg("XXX: tiler only zero tripped\n");
2338 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2339 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2340 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2341 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2342 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2343 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2344 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2345 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2346 }
2347
2348 pandecode_indent--;
2349 pandecode_log("},\n");
2350 }
2351
2352 static int
2353 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2354 const struct pandecode_mapped_memory *mem,
2355 mali_ptr payload, int job_no)
2356 {
2357 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2358
2359 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true);
2360
2361 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2362 pandecode_indent++;
2363
2364 pandecode_log(".prefix = ");
2365 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2366
2367 pandecode_log(".vertex = ");
2368 pandecode_vertex_only_bfr(&v->vertex);
2369
2370 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2371
2372 pandecode_indent--;
2373 pandecode_log("};\n");
2374
2375 return sizeof(*v);
2376 }
2377
2378 static int
2379 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2380 const struct pandecode_mapped_memory *mem,
2381 mali_ptr payload, int job_no)
2382 {
2383 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2384
2385 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true);
2386
2387 pandecode_indices(t->prefix.indices, t->prefix.index_count, job_no);
2388 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2389
2390 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2391 pandecode_indent++;
2392
2393 pandecode_log(".prefix = ");
2394 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2395
2396 pandecode_log(".tiler = ");
2397 pandecode_tiler_only_bfr(&t->tiler, job_no);
2398
2399 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2400
2401 pandecode_indent--;
2402 pandecode_log("};\n");
2403
2404 return sizeof(*t);
2405 }
2406
2407 static int
2408 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2409 const struct pandecode_mapped_memory *mem,
2410 mali_ptr payload, int job_no)
2411 {
2412 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2413
2414 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false);
2415
2416 pandecode_indices(v->prefix.indices, v->prefix.index_count, job_no);
2417
2418 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2419 pandecode_indent++;
2420
2421 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2422 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2423
2424 bool instanced = v->instance_shift || v->instance_odd;
2425 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2426
2427 pandecode_log(".prefix = ");
2428 pandecode_vertex_tiler_prefix(&v->prefix, job_no, !instanced && is_graphics);
2429
2430 pandecode_gl_enables(v->gl_enables, h->job_type);
2431
2432 if (v->instance_shift || v->instance_odd) {
2433 pandecode_prop("instance_shift = 0x%d /* %d */",
2434 v->instance_shift, 1 << v->instance_shift);
2435 pandecode_prop("instance_odd = 0x%X /* %d */",
2436 v->instance_odd, (2 * v->instance_odd) + 1);
2437
2438 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2439 }
2440
2441 if (v->offset_start)
2442 pandecode_prop("offset_start = %d", v->offset_start);
2443
2444 if (v->zero5) {
2445 pandecode_msg("XXX: midgard payload zero tripped\n");
2446 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2447 }
2448
2449 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2450
2451 pandecode_indent--;
2452 pandecode_log("};\n");
2453
2454 return sizeof(*v);
2455 }
2456
2457 static int
2458 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2459 mali_ptr payload, int job_no,
2460 bool is_bifrost)
2461 {
2462 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2463
2464 bool fbd_dumped = false;
2465
2466 if (!is_bifrost && (s->framebuffer & FBD_TYPE) == MALI_SFBD) {
2467 /* Only SFBDs are understood, not MFBDs. We're speculating,
2468 * based on the versioning, kernel code, etc, that the
2469 * difference is between Single FrameBuffer Descriptor and
2470 * Multiple FrmaeBuffer Descriptor; the change apparently lines
2471 * up with multi-framebuffer support being added (T7xx onwards,
2472 * including Gxx). In any event, there's some field shuffling
2473 * that we haven't looked into yet. */
2474
2475 pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true);
2476 fbd_dumped = true;
2477 } else if ((s->framebuffer & FBD_TYPE) == MALI_MFBD) {
2478 /* We don't know if Bifrost supports SFBD's at all, since the
2479 * driver never uses them. And the format is different from
2480 * Midgard anyways, due to the tiler heap and scratchpad being
2481 * moved out into separate structures, so it's not clear what a
2482 * Bifrost SFBD would even look like without getting an actual
2483 * trace, which appears impossible.
2484 */
2485
2486 pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2487 fbd_dumped = true;
2488 }
2489
2490 uintptr_t p = (uintptr_t) s->framebuffer & FBD_MASK;
2491 pandecode_log("struct mali_payload_fragment payload_%"PRIx64"_%d = {\n", payload, job_no);
2492 pandecode_indent++;
2493
2494 /* See the comments by the macro definitions for mathematical context
2495 * on why this is so weird */
2496
2497 if (MALI_TILE_COORD_FLAGS(s->max_tile_coord) || MALI_TILE_COORD_FLAGS(s->min_tile_coord))
2498 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
2499
2500 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
2501 MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT,
2502 MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT);
2503
2504 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
2505 (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT,
2506 (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT);
2507
2508 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2509 * we have to fallback on offsets. */
2510
2511 const char *fbd_type = s->framebuffer & MALI_MFBD ? "MALI_MFBD" : "MALI_SFBD";
2512
2513 /* TODO: Decode */
2514 unsigned extra_flags = (s->framebuffer & ~FBD_MASK) & ~MALI_MFBD;
2515
2516 if (fbd_dumped)
2517 pandecode_prop("framebuffer = framebuffer_%d_p | %s | 0x%X", job_no,
2518 fbd_type, extra_flags);
2519 else
2520 pandecode_prop("framebuffer = %s | %s | 0x%X", pointer_as_memory_reference(p),
2521 fbd_type, extra_flags);
2522
2523 pandecode_indent--;
2524 pandecode_log("};\n");
2525
2526 return sizeof(*s);
2527 }
2528
2529 static int job_descriptor_number = 0;
2530
2531 int
2532 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost)
2533 {
2534 struct mali_job_descriptor_header *h;
2535
2536 int start_number = 0;
2537
2538 bool first = true;
2539 bool last_size;
2540
2541 do {
2542 struct pandecode_mapped_memory *mem =
2543 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2544
2545 void *payload;
2546
2547 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2548
2549 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2550 * high 32-bits of the 64-bit pointer are reused to store
2551 * something else.
2552 */
2553 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2554 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2555 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2556
2557 payload = pandecode_fetch_gpu_mem(mem, payload_ptr,
2558 MALI_PAYLOAD_SIZE);
2559
2560 int job_no = job_descriptor_number++;
2561
2562 if (first)
2563 start_number = job_no;
2564
2565 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2566 pandecode_indent++;
2567
2568 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2569
2570 /* Save for next job fixing */
2571 last_size = h->job_descriptor_size;
2572
2573 if (h->job_descriptor_size)
2574 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2575
2576 if (h->exception_status && h->exception_status != 0x1)
2577 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2578 h->exception_status,
2579 (h->exception_status >> 16) & 0xFFFF,
2580 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2581 h->exception_status & 0xFF);
2582
2583 if (h->first_incomplete_task)
2584 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2585
2586 if (h->fault_pointer)
2587 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2588
2589 if (h->job_barrier)
2590 pandecode_prop("job_barrier = %d", h->job_barrier);
2591
2592 pandecode_prop("job_index = %d", h->job_index);
2593
2594 if (h->unknown_flags)
2595 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2596
2597 if (h->job_dependency_index_1)
2598 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2599
2600 if (h->job_dependency_index_2)
2601 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2602
2603 pandecode_indent--;
2604 pandecode_log("};\n");
2605
2606 /* Do not touch the field yet -- decode the payload first, and
2607 * don't touch that either. This is essential for the uploads
2608 * to occur in sequence and therefore be dynamically allocated
2609 * correctly. Do note the size, however, for that related
2610 * reason. */
2611
2612 switch (h->job_type) {
2613 case JOB_TYPE_SET_VALUE: {
2614 struct mali_payload_set_value *s = payload;
2615 pandecode_log("struct mali_payload_set_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2616 pandecode_indent++;
2617 MEMORY_PROP(s, out);
2618 pandecode_prop("unknown = 0x%" PRIX64, s->unknown);
2619 pandecode_indent--;
2620 pandecode_log("};\n");
2621
2622 break;
2623 }
2624
2625 case JOB_TYPE_TILER:
2626 case JOB_TYPE_VERTEX:
2627 case JOB_TYPE_COMPUTE:
2628 if (bifrost) {
2629 if (h->job_type == JOB_TYPE_TILER)
2630 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no);
2631 else
2632 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no);
2633 } else
2634 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no);
2635
2636 break;
2637
2638 case JOB_TYPE_FRAGMENT:
2639 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost);
2640 break;
2641
2642 default:
2643 break;
2644 }
2645
2646 /* Handle linkage */
2647
2648 if (!first) {
2649 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2650
2651 if (last_size)
2652 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no);
2653 else
2654 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no);
2655 }
2656
2657 first = false;
2658
2659 } while ((jc_gpu_va = h->job_descriptor_size ? h->next_job_64 : h->next_job_32));
2660
2661 return start_number;
2662 }