projects / mesa.git / blob
? search:


148be10fdc47bd5a95397825821d8e3c00bc2c82
[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_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
453 static char *
454 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt)
455 {
456 switch (fmt) {
457 DEFINE_CASE(TILED);
458 DEFINE_CASE(UNKNOWN);
459 DEFINE_CASE(LINEAR);
460 DEFINE_CASE(AFBC);
461
462 default:
463 unreachable("Invalid case");
464 }
465 }
466 #undef DEFINE_CASE
467
468 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
469 static char *
470 pandecode_exception_access(enum mali_exception_access access)
471 {
472 switch (access) {
473 DEFINE_CASE(NONE);
474 DEFINE_CASE(EXECUTE);
475 DEFINE_CASE(READ);
476 DEFINE_CASE(WRITE);
477
478 default:
479 unreachable("Invalid case");
480 }
481 }
482 #undef DEFINE_CASE
483
484 /* Midgard's tiler descriptor is embedded within the
485 * larger FBD */
486
487 static void
488 pandecode_midgard_tiler_descriptor(
489 const struct midgard_tiler_descriptor *t,
490 unsigned width,
491 unsigned height,
492 bool is_fragment)
493 {
494 pandecode_log(".tiler = {\n");
495 pandecode_indent++;
496
497 if (t->hierarchy_mask == MALI_TILER_DISABLED)
498 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
499 else
500 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
501
502 /* We know this name from the kernel, but we never see it nonzero */
503
504 if (t->flags)
505 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
506
507 MEMORY_PROP(t, polygon_list);
508
509 /* The body is offset from the base of the polygon list */
510 assert(t->polygon_list_body > t->polygon_list);
511 unsigned body_offset = t->polygon_list_body - t->polygon_list;
512
513 /* It needs to fit inside the reported size */
514 assert(t->polygon_list_size >= body_offset);
515
516 /* Check that we fit */
517 struct pandecode_mapped_memory *plist =
518 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
519
520 assert(t->polygon_list_size <= plist->length);
521
522 /* Now that we've sanity checked, we'll try to calculate the sizes
523 * ourselves for comparison */
524
525 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask);
526 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask);
527
528 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
529 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
530 ref_header, ref_size);
531 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
532 pandecode_msg("body offset %d\n", body_offset);
533 }
534
535 /* The tiler heap has a start and end specified -- it should be
536 * identical to what we have in the BO. The exception is if tiling is
537 * disabled. */
538
539 MEMORY_PROP(t, heap_start);
540 assert(t->heap_end >= t->heap_start);
541
542 struct pandecode_mapped_memory *heap =
543 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
544
545 unsigned heap_size = t->heap_end - t->heap_start;
546
547 /* Tiling is enabled with a special flag */
548 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
549 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
550
551 bool tiling_enabled = hierarchy_mask;
552
553 if (tiling_enabled) {
554 /* When tiling is enabled, the heap should be a tight fit */
555 unsigned heap_offset = t->heap_start - heap->gpu_va;
556 if ((heap_offset + heap_size) != heap->length) {
557 pandecode_msg("XXX: heap size %d (expected %d)\n",
558 heap_size, heap->length - heap_offset);
559 }
560
561 /* We should also have no other flags */
562 if (tiler_flags)
563 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
564 } else {
565 /* When tiling is disabled, we should have that flag and no others */
566
567 if (tiler_flags != MALI_TILER_DISABLED) {
568 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
569 tiler_flags);
570 }
571
572 /* We should also have an empty heap */
573 if (heap_size) {
574 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
575 heap_size);
576 }
577
578 /* Disabled tiling is used only for clear-only jobs, which are
579 * purely FRAGMENT, so we should never see this for
580 * non-FRAGMENT descriptors. */
581
582 if (!is_fragment)
583 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
584 }
585
586 /* We've never seen weights used in practice, but we know from the
587 * kernel these fields is there */
588
589 bool nonzero_weights = false;
590
591 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
592 nonzero_weights |= t->weights[w] != 0x0;
593 }
594
595 if (nonzero_weights) {
596 pandecode_log(".weights = {");
597
598 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
599 pandecode_log("%d, ", t->weights[w]);
600 }
601
602 pandecode_log("},");
603 }
604
605 pandecode_indent--;
606 pandecode_log("}\n");
607 }
608
609 static void
610 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment)
611 {
612 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
613 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
614
615 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
616 pandecode_indent++;
617
618 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
619 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
620
621 pandecode_log(".format = ");
622 pandecode_log_decoded_flags(fb_fmt_flag_info, s->format);
623 pandecode_log_cont(",\n");
624
625 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", s->width + 1);
626 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", s->height + 1);
627
628 MEMORY_PROP(s, framebuffer);
629 pandecode_prop("stride = %d", s->stride);
630
631 /* Earlier in the actual commandstream -- right before width -- but we
632 * delay to flow nicer */
633
634 pandecode_log(".clear_flags = ");
635 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
636 pandecode_log_cont(",\n");
637
638 if (s->depth_buffer | s->depth_buffer_enable) {
639 MEMORY_PROP(s, depth_buffer);
640 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s->depth_buffer_enable));
641 }
642
643 if (s->stencil_buffer | s->stencil_buffer_enable) {
644 MEMORY_PROP(s, stencil_buffer);
645 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s->stencil_buffer_enable));
646 }
647
648 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
649 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
650 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
651 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
652 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
653 }
654
655 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
656 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
657 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
658 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
659 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
660 }
661
662 if (s->clear_stencil) {
663 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
664 }
665
666 MEMORY_PROP(s, unknown_address_0);
667 const struct midgard_tiler_descriptor t = s->tiler;
668 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment);
669
670 pandecode_indent--;
671 pandecode_log("};\n");
672
673 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
674 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
675 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
676 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
677
678 printf(".zero3 = {");
679
680 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
681 printf("%X, ", s->zero3[i]);
682
683 printf("},\n");
684
685 printf(".zero6 = {");
686
687 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
688 printf("%X, ", s->zero6[i]);
689
690 printf("},\n");
691 }
692
693 static void
694 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
695 {
696 pandecode_log(".unknown%d = {", name);
697
698 for (int i = 0; i < count; ++i)
699 printf("%X, ", slide[i]);
700
701 pandecode_log("},\n");
702 }
703
704 #define SHORT_SLIDE(num) \
705 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
706
707 static void
708 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
709 {
710 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
711 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
712
713 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
714 pandecode_indent++;
715
716 SHORT_SLIDE(1);
717
718 pandecode_indent--;
719 printf("},\n");
720 }
721
722 /* Extracts the number of components associated with a Mali format */
723
724 static unsigned
725 pandecode_format_component_count(enum mali_format fmt)
726 {
727 /* Mask out the format class */
728 unsigned top = fmt & 0b11100000;
729
730 switch (top) {
731 case MALI_FORMAT_SNORM:
732 case MALI_FORMAT_UINT:
733 case MALI_FORMAT_UNORM:
734 case MALI_FORMAT_SINT:
735 return ((fmt >> 3) & 3) + 1;
736 default:
737 /* TODO: Validate */
738 return 4;
739 }
740 }
741
742 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
743
744 static unsigned
745 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
746 {
747 unsigned mask = 0;
748 assert(MALI_CHANNEL_RED == 0);
749
750 for (unsigned c = 0; c < 4; ++c) {
751 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
752
753 if (chan <= MALI_CHANNEL_ALPHA)
754 mask |= (1 << chan);
755 }
756
757 return mask;
758 }
759
760 /* Validates that a (format, swizzle) pair is valid, in the sense that the
761 * swizzle doesn't access any components that are undefined in the format.
762 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
763 * omitted */
764
765 static bool
766 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
767 {
768 unsigned nr_comp = pandecode_format_component_count(fmt);
769 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
770 unsigned valid_mask = (1 << nr_comp) - 1;
771 unsigned invalid_mask = ~valid_mask;
772
773 if (access_mask & invalid_mask) {
774 pandecode_msg("XXX: invalid components accessed\n");
775 return false;
776 }
777
778 /* Check for the default non-swizzling swizzle so we can suppress
779 * useless printing for the defaults */
780
781 unsigned default_swizzles[4] = {
782 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
783 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
784 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
785 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
786 };
787
788 return (swizzle == default_swizzles[nr_comp - 1]);
789 }
790
791 /* Maps MALI_RGBA32F to rgba32f, etc */
792
793 static void
794 pandecode_format_short(enum mali_format fmt, bool srgb)
795 {
796 /* We want a type-like format, so cut off the initial MALI_ */
797 char *format = pandecode_format(fmt);
798 format += strlen("MALI_");
799
800 unsigned len = strlen(format);
801 char *lower_format = calloc(1, len + 1);
802
803 for (unsigned i = 0; i < len; ++i)
804 lower_format[i] = tolower(format[i]);
805
806 /* Sanity check sRGB flag is applied to RGB, per the name */
807 if (srgb && lower_format[0] != 'r')
808 pandecode_msg("XXX: sRGB applied to non-colour format\n");
809
810 /* Just prefix with an s, so you get formats like srgba8_unorm */
811 if (srgb)
812 pandecode_log_cont("s");
813
814 pandecode_log_cont("%s", lower_format);
815 free(lower_format);
816 }
817
818 static void
819 pandecode_swizzle(unsigned swizzle, enum mali_format format)
820 {
821 /* First, do some validation */
822 bool trivial_swizzle = pandecode_validate_format_swizzle(
823 format, swizzle);
824
825 if (trivial_swizzle)
826 return;
827
828 /* Next, print the swizzle */
829 pandecode_log_cont(".");
830
831 static const char components[] = "rgba01";
832
833 for (unsigned c = 0; c < 4; ++c) {
834 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
835
836 if (chan >= MALI_CHANNEL_RESERVED_0) {
837 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
838 continue;
839 }
840 pandecode_log_cont("%c", components[chan]);
841 }
842 }
843
844 static void
845 pandecode_rt_format(struct mali_rt_format format)
846 {
847 pandecode_log(".format = {\n");
848 pandecode_indent++;
849
850 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
851 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
852 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
853
854 pandecode_prop("block = %s",
855 pandecode_mfbd_block_format(format.block));
856
857 /* TODO: Map formats so we can check swizzles and print nicely */
858 pandecode_log("swizzle");
859 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
860 pandecode_log_cont(",\n");
861
862 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
863 MALI_NEGATIVE(format.nr_channels));
864
865 pandecode_log(".flags = ");
866 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
867 pandecode_log_cont(",\n");
868
869 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
870 * which is a faster hardware-based alternative to the wallpaper method
871 * to preserve framebuffer contents across frames. In practice, MFBD
872 * preload is buggy on Midgard, and so this is a chicken bit. If this
873 * bit isn't set, most likely something broke unrelated to preload */
874
875 if (!format.no_preload) {
876 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
877 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
878 }
879
880 if (format.zero)
881 pandecode_prop("zero = 0x%" PRIx32, format.zero);
882
883 pandecode_indent--;
884 pandecode_log("},\n");
885 }
886
887 static void
888 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
889 {
890 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
891 pandecode_indent++;
892
893 for (int i = 0; i < MALI_NEGATIVE(fb->rt_count_1); i++) {
894 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
895 struct pandecode_mapped_memory *mem =
896 pandecode_find_mapped_gpu_mem_containing(rt_va);
897 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
898
899 pandecode_log("{\n");
900 pandecode_indent++;
901
902 pandecode_rt_format(rt->format);
903
904 if (rt->format.block == MALI_MFBD_BLOCK_AFBC) {
905 pandecode_log(".afbc = {\n");
906 pandecode_indent++;
907
908 char *a = pointer_as_memory_reference(rt->afbc.metadata);
909 pandecode_prop("metadata = %s", a);
910 free(a);
911
912 pandecode_prop("stride = %d", rt->afbc.stride);
913 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
914
915 pandecode_indent--;
916 pandecode_log("},\n");
917 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
918 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
919 rt->afbc.metadata,
920 rt->afbc.stride,
921 rt->afbc.unk);
922 }
923
924 MEMORY_PROP(rt, framebuffer);
925 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
926
927 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
928 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
929 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
930 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
931 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
932 }
933
934 if (rt->zero1 || rt->zero2 || rt->zero3) {
935 pandecode_msg("XXX: render target zeros tripped\n");
936 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
937 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
938 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
939 }
940
941 pandecode_indent--;
942 pandecode_log("},\n");
943 }
944
945 pandecode_indent--;
946 pandecode_log("};\n");
947 }
948
949 static unsigned
950 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment)
951 {
952 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
953 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
954
955 if (fb->sample_locations) {
956 /* The blob stores all possible sample locations in a single buffer
957 * allocated on startup, and just switches the pointer when switching
958 * MSAA state. For now, we just put the data into the cmdstream, but we
959 * should do something like what the blob does with a real driver.
960 *
961 * There seem to be 32 slots for sample locations, followed by another
962 * 16. The second 16 is just the center location followed by 15 zeros
963 * in all the cases I've identified (maybe shader vs. depth/color
964 * samples?).
965 */
966
967 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
968
969 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
970
971 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
972 pandecode_indent++;
973
974 for (int i = 0; i < 32 + 16; i++) {
975 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
976 }
977
978 pandecode_indent--;
979 pandecode_log("};\n");
980 }
981
982 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
983 pandecode_indent++;
984
985 pandecode_prop("unk0 = 0x%x", fb->unk0);
986
987 if (fb->sample_locations)
988 pandecode_prop("sample_locations = sample_locations_%d", job_no);
989
990 /* Assume that unknown1 was emitted in the last job for
991 * now */
992 MEMORY_PROP(fb, unknown1);
993
994 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
995 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
996 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
997 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
998
999 pandecode_prop("unk1 = 0x%x", fb->unk1);
1000 pandecode_prop("unk2 = 0x%x", fb->unk2);
1001 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1002 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1003
1004 pandecode_log(".mfbd_flags = ");
1005 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1006 pandecode_log_cont(",\n");
1007
1008 if (fb->clear_stencil)
1009 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1010
1011 if (fb->clear_depth)
1012 pandecode_prop("clear_depth = %f", fb->clear_depth);
1013
1014 /* TODO: What is this? Let's not blow up.. */
1015 if (fb->unknown2 != 0x1F)
1016 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1017
1018 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1019 MEMORY_PROP(fb, scratchpad);
1020 const struct midgard_tiler_descriptor t = fb->tiler;
1021 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment);
1022
1023 if (fb->zero3 || fb->zero4) {
1024 pandecode_msg("XXX: framebuffer zeros tripped\n");
1025 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1026 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1027 }
1028
1029 pandecode_indent--;
1030 pandecode_log("};\n");
1031
1032 gpu_va += sizeof(struct bifrost_framebuffer);
1033
1034 if ((fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment) {
1035 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1036 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1037
1038 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1039 pandecode_indent++;
1040
1041 MEMORY_PROP(fbx, checksum);
1042
1043 if (fbx->checksum_stride)
1044 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1045
1046 pandecode_log(".flags = ");
1047 pandecode_log_decoded_flags(mfbd_extra_flag_info, fbx->flags);
1048 pandecode_log_cont(",\n");
1049
1050 if (fbx->flags & MALI_EXTRA_AFBC_ZS) {
1051 pandecode_log(".ds_afbc = {\n");
1052 pandecode_indent++;
1053
1054 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1055 pandecode_prop("depth_stencil_afbc_stride = %d",
1056 fbx->ds_afbc.depth_stencil_afbc_stride);
1057 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1058
1059 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
1060 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1061 pandecode_prop("zero1 = 0x%" PRIx32,
1062 fbx->ds_afbc.zero1);
1063 pandecode_prop("padding = 0x%" PRIx64,
1064 fbx->ds_afbc.padding);
1065 }
1066
1067 pandecode_indent--;
1068 pandecode_log("},\n");
1069 } else {
1070 pandecode_log(".ds_linear = {\n");
1071 pandecode_indent++;
1072
1073 if (fbx->ds_linear.depth) {
1074 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1075 pandecode_prop("depth_stride = %d",
1076 fbx->ds_linear.depth_stride);
1077 }
1078
1079 if (fbx->ds_linear.stencil) {
1080 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1081 pandecode_prop("stencil_stride = %d",
1082 fbx->ds_linear.stencil_stride);
1083 }
1084
1085 if (fbx->ds_linear.depth_stride_zero ||
1086 fbx->ds_linear.stencil_stride_zero ||
1087 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1088 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1089 pandecode_prop("depth_stride_zero = 0x%x",
1090 fbx->ds_linear.depth_stride_zero);
1091 pandecode_prop("stencil_stride_zero = 0x%x",
1092 fbx->ds_linear.stencil_stride_zero);
1093 pandecode_prop("zero1 = 0x%" PRIx32,
1094 fbx->ds_linear.zero1);
1095 pandecode_prop("zero2 = 0x%" PRIx32,
1096 fbx->ds_linear.zero2);
1097 }
1098
1099 pandecode_indent--;
1100 pandecode_log("},\n");
1101 }
1102
1103 if (fbx->zero3 || fbx->zero4) {
1104 pandecode_msg("XXX: fb_extra zeros tripped\n");
1105 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1106 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1107 }
1108
1109 pandecode_indent--;
1110 pandecode_log("};\n");
1111
1112 gpu_va += sizeof(struct bifrost_fb_extra);
1113 }
1114
1115 if (is_fragment)
1116 pandecode_render_target(gpu_va, job_no, fb);
1117
1118 /* Passback the render target count */
1119 return MALI_NEGATIVE(fb->rt_count_1);
1120 }
1121
1122 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1123 * count */
1124
1125 static void
1126 pandecode_padded_vertices(unsigned shift, unsigned k)
1127 {
1128 unsigned odd = 2*k + 1;
1129 unsigned pot = 1 << shift;
1130 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1131 }
1132
1133 /* Given a magic divisor, recover what we were trying to divide by.
1134 *
1135 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1136 * 0 <= m < 2^N, for N bits in m.
1137 *
1138 * Let q represent the number we would like to divide by.
1139 *
1140 * By definition of a magic divisor for N-bit unsigned integers (a number you
1141 * multiply by to magically get division), m is a number such that:
1142 *
1143 * (m * x) & (2^N - 1) = floor(x/q).
1144 * for all x on Z where 0 <= x < 2^N
1145 *
1146 * Ignore the case where any of the above values equals zero; it is irrelevant
1147 * for our purposes (instanced arrays).
1148 *
1149 * Choose x = q. Then:
1150 *
1151 * (m * x) & (2^N - 1) = floor(x/q).
1152 * (m * q) & (2^N - 1) = floor(q/q).
1153 *
1154 * floor(q/q) = floor(1) = 1, therefore:
1155 *
1156 * (m * q) & (2^N - 1) = 1
1157 *
1158 * Recall the identity that the bitwise AND of one less than a power-of-two
1159 * equals the modulo with that power of two, i.e. for all x:
1160 *
1161 * x & (2^N - 1) = x % N
1162 *
1163 * Therefore:
1164 *
1165 * mq % (2^N) = 1
1166 *
1167 * By definition, a modular multiplicative inverse of a number m is the number
1168 * q such that with respect to a modulos M:
1169 *
1170 * mq % M = 1
1171 *
1172 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1173 *
1174 */
1175
1176 static void
1177 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1178 {
1179 #if 0
1180 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1181 * shift) the most lame way possible... just repeatedly add.
1182 * Asymptoptically slow but nobody cares in practice, unless you have
1183 * massive numbers of vertices or high divisors. */
1184
1185 unsigned inverse = 0;
1186
1187 /* Magic implicitly has the highest bit set */
1188 magic |= (1 << 31);
1189
1190 /* Depending on rounding direction */
1191 if (extra)
1192 magic++;
1193
1194 for (;;) {
1195 uint32_t product = magic * inverse;
1196
1197 if (shift) {
1198 product >>= shift;
1199 }
1200
1201 if (product == 1)
1202 break;
1203
1204 ++inverse;
1205 }
1206
1207 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1208
1209 /* Recall we're supposed to divide by (gl_level_divisor *
1210 * padded_num_vertices) */
1211
1212 unsigned padded_num_vertices = inverse / orig_divisor;
1213
1214 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1215 #endif
1216 }
1217
1218 static void
1219 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1220 mali_ptr addr, int job_no, char *suffix,
1221 int count, bool varying)
1222 {
1223 char *prefix = varying ? "varyings" : "attributes";
1224
1225 if (!addr) {
1226 pandecode_msg("no %s\n", prefix);
1227 return;
1228 }
1229
1230 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1231
1232 pandecode_log("union mali_attr %s_%d[] = {\n", prefix, job_no);
1233 pandecode_indent++;
1234
1235 for (int i = 0; i < count; ++i) {
1236 pandecode_log("{\n");
1237 pandecode_indent++;
1238
1239 enum mali_attr_mode mode = attr[i].elements & 7;
1240
1241 if (mode == MALI_ATTR_UNUSED)
1242 pandecode_msg("XXX: unused attribute record\n");
1243
1244 mali_ptr raw_elements = attr[i].elements & ~7;
1245 char *a = pointer_as_memory_reference(raw_elements);
1246 pandecode_prop("elements = (%s) | %s", a, pandecode_attr_mode(mode));
1247 free(a);
1248
1249 pandecode_prop("shift = %d", attr[i].shift);
1250 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1251 pandecode_prop("stride = 0x%" PRIx32, attr[i].stride);
1252 pandecode_prop("size = 0x%" PRIx32, attr[i].size);
1253
1254 /* Decode further where possible */
1255
1256 if (mode == MALI_ATTR_MODULO) {
1257 pandecode_padded_vertices(
1258 attr[i].shift,
1259 attr[i].extra_flags);
1260 }
1261
1262 pandecode_indent--;
1263 pandecode_log("}, \n");
1264
1265 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1266 i++;
1267 pandecode_log("{\n");
1268 pandecode_indent++;
1269 pandecode_prop("unk = 0x%x", attr[i].unk);
1270 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1271 if (attr[i].zero != 0)
1272 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1273 pandecode_prop("divisor = %d", attr[i].divisor);
1274 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1275 pandecode_indent--;
1276 pandecode_log("}, \n");
1277 }
1278
1279 }
1280
1281 pandecode_indent--;
1282 pandecode_log("};\n");
1283 }
1284
1285 static mali_ptr
1286 pandecode_shader_address(const char *name, mali_ptr ptr)
1287 {
1288 /* TODO: Decode flags */
1289 mali_ptr shader_ptr = ptr & ~15;
1290
1291 char *a = pointer_as_memory_reference(shader_ptr);
1292 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1293 free(a);
1294
1295 return shader_ptr;
1296 }
1297
1298 static bool
1299 all_zero(unsigned *buffer, unsigned count)
1300 {
1301 for (unsigned i = 0; i < count; ++i) {
1302 if (buffer[i])
1303 return false;
1304 }
1305
1306 return true;
1307 }
1308
1309 static void
1310 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1311 {
1312 if (all_zero((unsigned *) stencil, sizeof(stencil) / sizeof(unsigned)))
1313 return;
1314
1315 const char *func = pandecode_func(stencil->func);
1316 const char *sfail = pandecode_stencil_op(stencil->sfail);
1317 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1318 const char *dppass = pandecode_stencil_op(stencil->dppass);
1319
1320 if (stencil->zero)
1321 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1322
1323 pandecode_log(".stencil_%s = {\n", name);
1324 pandecode_indent++;
1325 pandecode_prop("ref = %d", stencil->ref);
1326 pandecode_prop("mask = 0x%02X", stencil->mask);
1327 pandecode_prop("func = %s", func);
1328 pandecode_prop("sfail = %s", sfail);
1329 pandecode_prop("dpfail = %s", dpfail);
1330 pandecode_prop("dppass = %s", dppass);
1331 pandecode_indent--;
1332 pandecode_log("},\n");
1333 }
1334
1335 static void
1336 pandecode_blend_equation(const struct mali_blend_equation *blend)
1337 {
1338 if (blend->zero1)
1339 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1340
1341 pandecode_log(".equation = {\n");
1342 pandecode_indent++;
1343
1344 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1345 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1346
1347 pandecode_log(".color_mask = ");
1348 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1349 pandecode_log_cont(",\n");
1350
1351 pandecode_indent--;
1352 pandecode_log("},\n");
1353 }
1354
1355 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1356
1357 static unsigned
1358 decode_bifrost_constant(u16 constant)
1359 {
1360 float lo = (float) (constant & 0xFF);
1361 float hi = (float) (constant >> 8);
1362
1363 return (hi / 255.0) + (lo / 65535.0);
1364 }
1365
1366 static mali_ptr
1367 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1368 {
1369 struct bifrost_blend_rt *b =
1370 ((struct bifrost_blend_rt *) descs) + rt_no;
1371
1372 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1373 pandecode_indent++;
1374
1375 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1376 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1377 b->constant, decode_bifrost_constant(b->constant));
1378
1379 /* TODO figure out blend shader enable bit */
1380 pandecode_blend_equation(&b->equation);
1381 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1382 pandecode_prop("index = 0x%" PRIx16, b->index);
1383 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1384
1385 pandecode_indent--;
1386 pandecode_log("},\n");
1387
1388 return 0;
1389 }
1390
1391 static mali_ptr
1392 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1393 {
1394 if (all_zero((unsigned *) blend, sizeof(blend) / sizeof(unsigned)))
1395 return 0;
1396
1397 pandecode_log(".blend = {\n");
1398 pandecode_indent++;
1399
1400 if (is_shader) {
1401 pandecode_shader_address("shader", blend->shader);
1402 } else {
1403 pandecode_blend_equation(&blend->equation);
1404 pandecode_prop("constant = %f", blend->constant);
1405 }
1406
1407 pandecode_indent--;
1408 pandecode_log("},\n");
1409
1410 /* Return blend shader to disassemble if present */
1411 return is_shader ? (blend->shader & ~0xF) : 0;
1412 }
1413
1414 static mali_ptr
1415 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1416 {
1417 struct midgard_blend_rt *b =
1418 ((struct midgard_blend_rt *) descs) + rt_no;
1419
1420 /* Flags determine presence of blend shader */
1421 bool is_shader = (b->flags & 0xF) >= 0x2;
1422
1423 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1424 pandecode_indent++;
1425
1426 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1427
1428 union midgard_blend blend = b->blend;
1429 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1430
1431 pandecode_indent--;
1432 pandecode_log("};\n");
1433
1434 return shader;
1435 }
1436
1437 /* Attributes and varyings have descriptor records, which contain information
1438 * about their format and ordering with the attribute/varying buffers. We'll
1439 * want to validate that the combinations specified are self-consistent.
1440 */
1441
1442 static int
1443 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1444 {
1445 char base[128];
1446 char *prefix = varying ? "varying" : "attribute";
1447 unsigned max_index = 0;
1448 snprintf(base, sizeof(base), "%s_meta", prefix);
1449
1450 pandecode_log("struct mali_attr_meta %s_%d%s[] = {\n", base, job_no, suffix);
1451 pandecode_indent++;
1452
1453 struct mali_attr_meta *attr_meta;
1454 mali_ptr p = varying ? (v->varying_meta & ~0xF) : v->attribute_meta;
1455
1456 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1457
1458 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1459 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1460 sizeof(*attr_mem));
1461
1462 /* If the record is discard, it should be zero for everything else */
1463
1464 if (attr_meta->format == MALI_VARYING_DISCARD) {
1465 uint64_t zero =
1466 attr_meta->index |
1467 attr_meta->unknown1 |
1468 attr_meta->unknown3 |
1469 attr_meta->src_offset;
1470
1471 if (zero)
1472 pandecode_msg("XXX: expected empty record for varying discard\n");
1473
1474 /* We want to look for a literal 0000 swizzle -- this
1475 * is not encoded with all zeroes, however */
1476
1477 enum mali_channel z = MALI_CHANNEL_ZERO;
1478 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1479 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1480
1481 if (!good_swizzle)
1482 pandecode_msg("XXX: expected zero swizzle for discard\n");
1483
1484 if (!varying)
1485 pandecode_msg("XXX: cannot discard attribute\n");
1486
1487 /* If we're all good, omit the record */
1488 if (!zero && varying && good_swizzle) {
1489 pandecode_log("/* discarded varying */\n");
1490 continue;
1491 }
1492 }
1493
1494 if (attr_meta->index > max_index)
1495 max_index = attr_meta->index;
1496
1497 if (attr_meta->unknown1 != 0x2) {
1498 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1499 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1500 }
1501
1502 if (attr_meta->unknown3) {
1503 pandecode_msg("XXX: unexpected unknown3 set\n");
1504 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1505 }
1506
1507 pandecode_make_indent();
1508 pandecode_format_short(attr_meta->format, false);
1509 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1510
1511 if (attr_meta->src_offset)
1512 pandecode_log_cont("[%u]", attr_meta->src_offset);
1513
1514 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1515
1516 pandecode_log_cont(";\n");
1517 }
1518
1519 pandecode_indent--;
1520 pandecode_log("};\n");
1521
1522 return count ? (max_index + 1) : 0;
1523 }
1524
1525 static void
1526 pandecode_indices(uintptr_t pindices, uint32_t index_count, int job_no)
1527 {
1528 struct pandecode_mapped_memory *imem = pandecode_find_mapped_gpu_mem_containing(pindices);
1529
1530 if (imem) {
1531 /* Indices are literally just a u32 array :) */
1532
1533 uint32_t *PANDECODE_PTR_VAR(indices, imem, pindices);
1534
1535 pandecode_log("uint32_t indices_%d[] = {\n", job_no);
1536 pandecode_indent++;
1537
1538 for (unsigned i = 0; i < (index_count + 1); i += 3)
1539 pandecode_log("%d, %d, %d,\n",
1540 indices[i],
1541 indices[i + 1],
1542 indices[i + 2]);
1543
1544 pandecode_indent--;
1545 pandecode_log("};\n");
1546 }
1547 }
1548
1549 /* return bits [lo, hi) of word */
1550 static u32
1551 bits(u32 word, u32 lo, u32 hi)
1552 {
1553 if (hi - lo >= 32)
1554 return word; // avoid undefined behavior with the shift
1555
1556 return (word >> lo) & ((1 << (hi - lo)) - 1);
1557 }
1558
1559 static void
1560 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool noninstanced)
1561 {
1562 pandecode_log_cont("{\n");
1563 pandecode_indent++;
1564
1565 /* Decode invocation_count. See the comment before the definition of
1566 * invocation_count for an explanation.
1567 */
1568
1569 unsigned size_x = bits(p->invocation_count, 0, p->size_y_shift) + 1;
1570 unsigned size_y = bits(p->invocation_count, p->size_y_shift, p->size_z_shift) + 1;
1571 unsigned size_z = bits(p->invocation_count, p->size_z_shift, p->workgroups_x_shift) + 1;
1572
1573 unsigned groups_x = bits(p->invocation_count, p->workgroups_x_shift, p->workgroups_y_shift) + 1;
1574 unsigned groups_y = bits(p->invocation_count, p->workgroups_y_shift, p->workgroups_z_shift) + 1;
1575 unsigned groups_z = bits(p->invocation_count, p->workgroups_z_shift, 32) + 1;
1576
1577 /* Even though we have this decoded, we want to ensure that the
1578 * representation is "unique" so we don't lose anything by printing only
1579 * the final result. More specifically, we need to check that we were
1580 * passed something in canonical form, since the definition per the
1581 * hardware is inherently not unique. How? Well, take the resulting
1582 * decode and pack it ourselves! If it is bit exact with what we
1583 * decoded, we're good to go. */
1584
1585 struct mali_vertex_tiler_prefix ref;
1586 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, noninstanced);
1587
1588 bool canonical =
1589 (p->invocation_count == ref.invocation_count) &&
1590 (p->size_y_shift == ref.size_y_shift) &&
1591 (p->size_z_shift == ref.size_z_shift) &&
1592 (p->workgroups_x_shift == ref.workgroups_x_shift) &&
1593 (p->workgroups_y_shift == ref.workgroups_y_shift) &&
1594 (p->workgroups_z_shift == ref.workgroups_z_shift) &&
1595 (p->workgroups_x_shift_2 == ref.workgroups_x_shift_2);
1596
1597 if (!canonical) {
1598 pandecode_msg("XXX: non-canonical workgroups packing\n");
1599 pandecode_msg("expected: %X, %d, %d, %d, %d, %d\n",
1600 ref.invocation_count,
1601 ref.size_y_shift,
1602 ref.size_z_shift,
1603 ref.workgroups_x_shift,
1604 ref.workgroups_y_shift,
1605 ref.workgroups_z_shift,
1606 ref.workgroups_x_shift_2);
1607
1608 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1609 pandecode_prop("size_y_shift = %d", p->size_y_shift);
1610 pandecode_prop("size_z_shift = %d", p->size_z_shift);
1611 pandecode_prop("workgroups_x_shift = %d", p->workgroups_x_shift);
1612 pandecode_prop("workgroups_y_shift = %d", p->workgroups_y_shift);
1613 pandecode_prop("workgroups_z_shift = %d", p->workgroups_z_shift);
1614 pandecode_prop("workgroups_x_shift_2 = %d", p->workgroups_x_shift_2);
1615 }
1616
1617 /* Regardless, print the decode */
1618 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1619 size_x, size_y, size_z,
1620 groups_x, groups_y, groups_z);
1621
1622 /* TODO: Decode */
1623 if (p->unknown_draw)
1624 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1625
1626 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1627
1628 if (p->draw_mode != MALI_DRAW_NONE)
1629 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1630
1631 /* Index count only exists for tiler jobs anyway */
1632
1633 if (p->index_count)
1634 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1635
1636 if (p->offset_bias_correction)
1637 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1638
1639 /* TODO: Figure out what this is. It's not zero */
1640 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1641
1642 pandecode_indent--;
1643 pandecode_log("},\n");
1644 }
1645
1646 static void
1647 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1648 {
1649 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1650 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1651
1652 pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%"PRIx64"_%d[] = {\n",
1653 pubufs, job_no);
1654 pandecode_indent++;
1655
1656 for (int i = 0; i < ubufs_count; i++) {
1657 pandecode_log("{\n");
1658 pandecode_indent++;
1659
1660 unsigned size = (ubufs[i].size + 1) * 16;
1661 mali_ptr addr = ubufs[i].ptr << 2;
1662
1663 pandecode_validate_buffer(addr, size);
1664
1665 char *ptr = pointer_as_memory_reference(ubufs[i].ptr << 2);
1666 pandecode_prop("size = %u", size);
1667 pandecode_prop("ptr = (%s) >> 2", ptr);
1668 pandecode_indent--;
1669 pandecode_log("},\n");
1670 free(ptr);
1671 }
1672
1673 pandecode_indent--;
1674 pandecode_log("};\n");
1675 }
1676
1677 static void
1678 pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1679 {
1680
1681 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1682
1683 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1684
1685 if (scratchpad->zero) {
1686 pandecode_msg("XXX: scratchpad zero tripped");
1687 pandecode_prop("zero = 0x%x\n", scratchpad->zero);
1688 }
1689
1690 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1691 pandecode_indent++;
1692
1693 pandecode_prop("flags = 0x%x", scratchpad->flags);
1694 MEMORY_PROP(scratchpad, gpu_scratchpad);
1695
1696 pandecode_indent--;
1697 pandecode_log("};\n");
1698 }
1699
1700 static unsigned shader_id = 0;
1701
1702 static void
1703 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1704 bool is_bifrost, unsigned nr_regs)
1705 {
1706 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1707 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1708
1709 /* Compute maximum possible size */
1710 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1711
1712 /* Print some boilerplate to clearly denote the assembly (which doesn't
1713 * obey indentation rules), and actually do the disassembly! */
1714
1715 printf("\n\n");
1716
1717 char prefix[512];
1718
1719 snprintf(prefix, sizeof(prefix) - 1, "shader%d - %s shader: ",
1720 shader_id++,
1721 (type == JOB_TYPE_TILER) ? "FRAGMENT" : "VERTEX");
1722
1723 if (is_bifrost) {
1724 disassemble_bifrost(code, sz, false);
1725 } else {
1726 disassemble_midgard(code, sz, true, nr_regs, prefix);
1727 }
1728
1729 printf("\n\n");
1730 }
1731
1732 static void
1733 pandecode_texture(mali_ptr u,
1734 struct pandecode_mapped_memory *tmem,
1735 unsigned job_no, unsigned tex)
1736 {
1737 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
1738
1739 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
1740 pandecode_indent++;
1741
1742 struct mali_texture_format f = t->format;
1743
1744 /* See the definiton of enum mali_texture_type */
1745
1746 bool is_cube = f.type == MALI_TEX_CUBE;
1747 unsigned dimension = is_cube ? 2 : f.type;
1748
1749 pandecode_make_indent();
1750
1751 /* TODO: Are there others? */
1752 bool is_zs = f.format == MALI_Z32_UNORM;
1753
1754 /* Recall Z/S switched the meaning of linear/tiled .. */
1755 if (is_zs && f.layout == MALI_TEXTURE_LINEAR)
1756 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1757
1758 /* Print the layout. Default is linear; a modifier can denote AFBC or
1759 * u-interleaved/tiled modes */
1760
1761 if (f.layout == MALI_TEXTURE_AFBC)
1762 pandecode_log_cont("afbc");
1763 else if (f.layout == MALI_TEXTURE_TILED)
1764 pandecode_log_cont(is_zs ? "linear" : "tiled");
1765 else if (f.layout == MALI_TEXTURE_LINEAR)
1766 pandecode_log_cont("linear");
1767 else
1768 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
1769
1770 pandecode_swizzle(t->swizzle, f.format);
1771 pandecode_log_cont(" ");
1772
1773 /* Distinguish cube/2D with modifier */
1774
1775 if (is_cube)
1776 pandecode_log_cont("cube ");
1777
1778 pandecode_format_short(f.format, f.srgb);
1779 pandecode_swizzle(f.swizzle, f.format);
1780
1781 /* All four width/height/depth/array_size dimensions are present
1782 * regardless of the type of texture, but it is an error to have
1783 * non-zero dimensions for unused dimensions. Verify this. array_size
1784 * can always be set, as can width. */
1785
1786 if (t->height && dimension < 2)
1787 pandecode_msg("XXX: nonzero height for <2D texture\n");
1788
1789 if (t->depth && dimension < 3)
1790 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1791
1792 /* Print only the dimensions that are actually there */
1793
1794 pandecode_log_cont(": %d", t->width + 1);
1795
1796 if (dimension >= 2)
1797 pandecode_log_cont("x%u", t->height + 1);
1798
1799 if (dimension >= 3)
1800 pandecode_log_cont("x%u", t->depth + 1);
1801
1802 if (t->array_size)
1803 pandecode_log_cont("[%u]", t->array_size + 1);
1804
1805 if (t->levels)
1806 pandecode_log_cont(" mip %u", t->levels);
1807
1808 pandecode_log_cont("\n");
1809
1810 if (f.unknown1 | f.zero) {
1811 pandecode_msg("XXX: texture format zero tripped\n");
1812 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
1813 pandecode_prop("zero = %" PRId32, f.zero);
1814 }
1815
1816 if (!f.unknown2) {
1817 pandecode_msg("XXX: expected unknown texture bit set\n");
1818 pandecode_prop("unknown2 = %" PRId32, f.unknown1);
1819 }
1820
1821 if (t->swizzle_zero) {
1822 pandecode_msg("XXX: swizzle zero tripped\n");
1823 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
1824 }
1825
1826 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
1827 pandecode_msg("XXX: texture zero tripped\n");
1828 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
1829 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
1830 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
1831 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
1832 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
1833 }
1834
1835 pandecode_log(".payload = {\n");
1836 pandecode_indent++;
1837
1838 /* A bunch of bitmap pointers follow.
1839 * We work out the correct number,
1840 * based on the mipmap/cubemap
1841 * properties, but dump extra
1842 * possibilities to futureproof */
1843
1844 int bitmap_count = MALI_NEGATIVE(t->levels);
1845
1846 /* Miptree for each face */
1847 if (f.type == MALI_TEX_CUBE)
1848 bitmap_count *= 6;
1849
1850 /* Array of textures */
1851 bitmap_count *= MALI_NEGATIVE(t->array_size);
1852
1853 /* Stride for each element */
1854 if (f.manual_stride)
1855 bitmap_count *= 2;
1856
1857 /* Sanity check the size */
1858 int max_count = sizeof(t->payload) / sizeof(t->payload[0]);
1859 assert (bitmap_count <= max_count);
1860
1861 for (int i = 0; i < bitmap_count; ++i) {
1862 /* How we dump depends if this is a stride or a pointer */
1863
1864 if (f.manual_stride && (i & 1)) {
1865 /* signed 32-bit snuck in as a 64-bit pointer */
1866 uint64_t stride_set = t->payload[i];
1867 uint32_t clamped_stride = stride_set;
1868 int32_t stride = clamped_stride;
1869 assert(stride_set == clamped_stride);
1870 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
1871 } else {
1872 char *a = pointer_as_memory_reference(t->payload[i]);
1873 pandecode_log("%s, \n", a);
1874 free(a);
1875 }
1876 }
1877
1878 pandecode_indent--;
1879 pandecode_log("},\n");
1880
1881 pandecode_indent--;
1882 pandecode_log("};\n");
1883 }
1884
1885 static void
1886 pandecode_vertex_tiler_postfix_pre(
1887 const struct mali_vertex_tiler_postfix *p,
1888 int job_no, enum mali_job_type job_type,
1889 char *suffix, bool is_bifrost)
1890 {
1891 mali_ptr shader_meta_ptr = (u64) (uintptr_t) (p->_shader_upper << 4);
1892 struct pandecode_mapped_memory *attr_mem;
1893
1894 unsigned rt_count = 1;
1895
1896 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1897 * are the only things actually needed from the FBD, vertex/tiler jobs
1898 * no longer reference the FBD -- instead, this field points to some
1899 * info about the scratchpad.
1900 */
1901 if (is_bifrost)
1902 pandecode_scratchpad(p->framebuffer & ~FBD_TYPE, job_no, suffix);
1903 else if (p->framebuffer & MALI_MFBD)
1904 rt_count = pandecode_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
1905 else if (job_type == JOB_TYPE_COMPUTE)
1906 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
1907 else
1908 pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no, false);
1909
1910 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
1911 int texture_count = 0, sampler_count = 0;
1912
1913 if (shader_meta_ptr) {
1914 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr);
1915 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, shader_meta_ptr);
1916
1917 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", shader_meta_ptr, job_no, suffix);
1918 pandecode_indent++;
1919
1920 /* Save for dumps */
1921 attribute_count = s->attribute_count;
1922 varying_count = s->varying_count;
1923 texture_count = s->texture_count;
1924 sampler_count = s->sampler_count;
1925
1926 if (is_bifrost) {
1927 uniform_count = s->bifrost2.uniform_count;
1928 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
1929 } else {
1930 uniform_count = s->midgard1.uniform_buffer_count;
1931 uniform_buffer_count = s->midgard1.uniform_buffer_count;
1932 }
1933
1934 mali_ptr shader_ptr = pandecode_shader_address("shader", s->shader);
1935
1936 pandecode_prop("texture_count = %" PRId16, s->texture_count);
1937 pandecode_prop("sampler_count = %" PRId16, s->sampler_count);
1938 pandecode_prop("attribute_count = %" PRId16, s->attribute_count);
1939 pandecode_prop("varying_count = %" PRId16, s->varying_count);
1940
1941 unsigned nr_registers = 0;
1942
1943 if (is_bifrost) {
1944 pandecode_log(".bifrost1 = {\n");
1945 pandecode_indent++;
1946
1947 pandecode_prop("uniform_buffer_count = %" PRId32, s->bifrost1.uniform_buffer_count);
1948 pandecode_prop("unk1 = 0x%" PRIx32, s->bifrost1.unk1);
1949
1950 pandecode_indent--;
1951 pandecode_log("},\n");
1952 } else {
1953 pandecode_log(".midgard1 = {\n");
1954 pandecode_indent++;
1955
1956 pandecode_prop("uniform_count = %" PRId16, s->midgard1.uniform_count);
1957 pandecode_prop("uniform_buffer_count = %" PRId16, s->midgard1.uniform_buffer_count);
1958 pandecode_prop("work_count = %" PRId16, s->midgard1.work_count);
1959 nr_registers = s->midgard1.work_count;
1960
1961 pandecode_log(".flags = ");
1962 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
1963 pandecode_log_cont(",\n");
1964
1965 pandecode_prop("unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
1966
1967 pandecode_indent--;
1968 pandecode_log("},\n");
1969 }
1970
1971 if (s->depth_units || s->depth_factor) {
1972 pandecode_prop("depth_factor = %f", s->depth_factor);
1973 pandecode_prop("depth_units = %f", s->depth_units);
1974 }
1975
1976 if (s->alpha_coverage) {
1977 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
1978 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
1979
1980 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
1981 invert_alpha_coverage ? "~" : "",
1982 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
1983 }
1984
1985 if (s->unknown2_3 || s->unknown2_4) {
1986 pandecode_log(".unknown2_3 = ");
1987
1988 int unknown2_3 = s->unknown2_3;
1989 int unknown2_4 = s->unknown2_4;
1990
1991 /* We're not quite sure what these flags mean without the depth test, if anything */
1992
1993 if (unknown2_3 & (MALI_DEPTH_TEST | MALI_DEPTH_FUNC_MASK)) {
1994 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
1995 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
1996
1997 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
1998 }
1999
2000 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2001 pandecode_log_cont(",\n");
2002
2003 pandecode_log(".unknown2_4 = ");
2004 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2005 pandecode_log_cont(",\n");
2006 }
2007
2008 if (s->stencil_mask_front || s->stencil_mask_back) {
2009 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2010 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2011 }
2012
2013 pandecode_stencil("front", &s->stencil_front);
2014 pandecode_stencil("back", &s->stencil_back);
2015
2016 if (is_bifrost) {
2017 pandecode_log(".bifrost2 = {\n");
2018 pandecode_indent++;
2019
2020 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2021 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2022 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2023 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2024
2025 pandecode_indent--;
2026 pandecode_log("},\n");
2027 } else if (s->midgard2.unknown2_7) {
2028 pandecode_log(".midgard2 = {\n");
2029 pandecode_indent++;
2030
2031 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2032 pandecode_indent--;
2033 pandecode_log("},\n");
2034 }
2035
2036 if (s->unknown2_8)
2037 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
2038
2039 if (!is_bifrost) {
2040 /* TODO: Blend shaders routing/disasm */
2041
2042 union midgard_blend blend = s->blend;
2043 pandecode_midgard_blend(&blend, false);
2044 }
2045
2046 pandecode_indent--;
2047 pandecode_log("};\n");
2048
2049 /* MRT blend fields are used whenever MFBD is used, with
2050 * per-RT descriptors */
2051
2052 if (job_type == JOB_TYPE_TILER) {
2053 void* blend_base = (void *) (s + 1);
2054
2055 for (unsigned i = 0; i < rt_count; i++) {
2056 mali_ptr shader = 0;
2057
2058 if (is_bifrost)
2059 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2060 else
2061 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2062
2063 if (shader & ~0xF)
2064 pandecode_shader_disassemble(shader, job_no, job_type, false, 0);
2065 }
2066 }
2067
2068 if (shader_ptr & ~0xF)
2069 pandecode_shader_disassemble(shader_ptr, job_no, job_type, is_bifrost, nr_registers);
2070 } else
2071 pandecode_msg("<no shader>\n");
2072
2073 if (p->viewport) {
2074 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2075 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2076
2077 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2078 pandecode_indent++;
2079
2080 pandecode_prop("clip_minx = %f", f->clip_minx);
2081 pandecode_prop("clip_miny = %f", f->clip_miny);
2082 pandecode_prop("clip_minz = %f", f->clip_minz);
2083 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2084 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2085 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2086
2087 /* Only the higher coordinates are MALI_POSITIVE scaled */
2088
2089 pandecode_prop("viewport0 = { %d, %d }",
2090 f->viewport0[0], f->viewport0[1]);
2091
2092 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2093 f->viewport1[0] + 1, f->viewport1[1] + 1);
2094
2095 pandecode_indent--;
2096 pandecode_log("};\n");
2097 }
2098
2099 if (p->attribute_meta) {
2100 unsigned max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2101
2102 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2103 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false);
2104 }
2105
2106 /* Varyings are encoded like attributes but not actually sent; we just
2107 * pass a zero buffer with the right stride/size set, (or whatever)
2108 * since the GPU will write to it itself */
2109
2110 if (p->varying_meta) {
2111 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2112 }
2113
2114 if (p->varyings) {
2115 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2116
2117 /* Number of descriptors depends on whether there are
2118 * non-internal varyings */
2119
2120 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true);
2121 }
2122
2123 if (p->uniform_buffers) {
2124 if (uniform_buffer_count)
2125 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2126 else
2127 pandecode_msg("XXX: UBOs specified but not referenced\n");
2128 } else if (uniform_buffer_count)
2129 pandecode_msg("XXX: UBOs referenced but not specified\n");
2130
2131 /* We don't want to actually dump uniforms, but we do need to validate
2132 * that the counts we were given are sane */
2133
2134 if (p->uniforms) {
2135 if (uniform_count)
2136 pandecode_validate_buffer(p->uniforms, uniform_count * 16);
2137 else
2138 pandecode_msg("XXX: Uniforms specified but not referenced");
2139 } else if (uniform_count)
2140 pandecode_msg("XXX: UBOs referenced but not specified\n");
2141
2142 if (p->texture_trampoline) {
2143 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
2144
2145 if (mmem) {
2146 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
2147
2148 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
2149 pandecode_indent++;
2150
2151 for (int tex = 0; tex < texture_count; ++tex) {
2152 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2153 char *a = pointer_as_memory_reference(*u);
2154 pandecode_log("%s,\n", a);
2155 free(a);
2156 }
2157
2158 pandecode_indent--;
2159 pandecode_log("};\n");
2160
2161 /* Now, finally, descend down into the texture descriptor */
2162 for (unsigned tex = 0; tex < texture_count; ++tex) {
2163 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2164 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2165 if (tmem)
2166 pandecode_texture(*u, tmem, job_no, tex);
2167 }
2168 }
2169 }
2170
2171 if (p->sampler_descriptor) {
2172 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2173
2174 if (smem) {
2175 struct mali_sampler_descriptor *s;
2176
2177 mali_ptr d = p->sampler_descriptor;
2178
2179 for (int i = 0; i < sampler_count; ++i) {
2180 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2181
2182 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2183 pandecode_indent++;
2184
2185 pandecode_log(".filter_mode = ");
2186 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2187 pandecode_log_cont(",\n");
2188
2189 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2190 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2191
2192 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2193 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2194 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2195
2196 pandecode_prop("compare_func = %s", pandecode_alt_func(s->compare_func));
2197
2198 if (s->zero || s->zero2) {
2199 pandecode_msg("XXX: sampler zero tripped\n");
2200 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2201 }
2202
2203 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2204
2205 pandecode_prop("border_color = { %f, %f, %f, %f }",
2206 s->border_color[0],
2207 s->border_color[1],
2208 s->border_color[2],
2209 s->border_color[3]);
2210
2211 pandecode_indent--;
2212 pandecode_log("};\n");
2213 }
2214 }
2215 }
2216 }
2217
2218 static void
2219 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2220 {
2221 if (!(p->position_varying || p->occlusion_counter || p->flags))
2222 return;
2223
2224 pandecode_log(".postfix = {\n");
2225 pandecode_indent++;
2226
2227 MEMORY_PROP(p, position_varying);
2228 MEMORY_PROP(p, occlusion_counter);
2229
2230 if (p->flags)
2231 pandecode_prop("flags = %d", p->flags);
2232
2233 pandecode_indent--;
2234 pandecode_log("},\n");
2235 }
2236
2237 static void
2238 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2239 {
2240 pandecode_log_cont("{\n");
2241 pandecode_indent++;
2242
2243 pandecode_prop("unk2 = 0x%x", v->unk2);
2244
2245 if (v->zero0 || v->zero1) {
2246 pandecode_msg("XXX: vertex only zero tripped");
2247 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2248 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2249 }
2250
2251 pandecode_indent--;
2252 pandecode_log("}\n");
2253 }
2254
2255 static void
2256 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2257 {
2258
2259 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2260 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2261
2262 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2263 pandecode_indent++;
2264
2265 if (h->zero) {
2266 pandecode_msg("XXX: tiler heap zero tripped\n");
2267 pandecode_prop("zero = 0x%x", h->zero);
2268 }
2269
2270 for (int i = 0; i < 12; i++) {
2271 if (h->zeros[i] != 0) {
2272 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2273 i, h->zeros[i]);
2274 }
2275 }
2276
2277 pandecode_prop("heap_size = 0x%x", h->heap_size);
2278 MEMORY_PROP(h, tiler_heap_start);
2279 MEMORY_PROP(h, tiler_heap_free);
2280
2281 /* this might point to the beginning of another buffer, when it's
2282 * really the end of the tiler heap buffer, so we have to be careful
2283 * here. but for zero length, we need the same pointer.
2284 */
2285
2286 if (h->tiler_heap_end == h->tiler_heap_start) {
2287 MEMORY_PROP(h, tiler_heap_start);
2288 } else {
2289 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2290 pandecode_prop("tiler_heap_end = %s + 1", a);
2291 free(a);
2292 }
2293
2294 pandecode_indent--;
2295 pandecode_log("};\n");
2296 }
2297
2298 static void
2299 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2300 {
2301 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2302 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2303
2304 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2305
2306 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2307 pandecode_indent++;
2308
2309 if (t->zero0 || t->zero1) {
2310 pandecode_msg("XXX: tiler meta zero tripped\n");
2311 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2312 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2313 }
2314
2315 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2316 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2317
2318 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2319 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2320
2321 for (int i = 0; i < 12; i++) {
2322 if (t->zeros[i] != 0) {
2323 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2324 i, t->zeros[i]);
2325 }
2326 }
2327
2328 pandecode_indent--;
2329 pandecode_log("};\n");
2330 }
2331
2332 static void
2333 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2334 {
2335 pandecode_log(".gl_enables = ");
2336
2337 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2338
2339 pandecode_log_cont(",\n");
2340 }
2341
2342 static void
2343 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2344 {
2345 if (u.pointer == 0x0)
2346 return;
2347
2348 pandecode_log(".primitive_size = {\n");
2349 pandecode_indent++;
2350
2351 if (constant) {
2352 pandecode_prop("constant = %f", u.constant);
2353 } else {
2354 MEMORY_PROP((&u), pointer);
2355 }
2356
2357 pandecode_indent--;
2358 pandecode_log("},\n");
2359 }
2360
2361 static void
2362 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2363 {
2364 pandecode_log_cont("{\n");
2365 pandecode_indent++;
2366
2367 /* TODO: gl_PointSize on Bifrost */
2368 pandecode_primitive_size(t->primitive_size, true);
2369
2370 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2371
2372 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2373 || t->zero6 || t->zero7 || t->zero8) {
2374 pandecode_msg("XXX: tiler only zero tripped\n");
2375 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2376 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2377 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2378 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2379 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2380 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2381 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2382 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2383 }
2384
2385 pandecode_indent--;
2386 pandecode_log("},\n");
2387 }
2388
2389 static int
2390 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2391 const struct pandecode_mapped_memory *mem,
2392 mali_ptr payload, int job_no)
2393 {
2394 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2395
2396 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true);
2397
2398 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2399 pandecode_indent++;
2400
2401 pandecode_log(".prefix = ");
2402 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2403
2404 pandecode_log(".vertex = ");
2405 pandecode_vertex_only_bfr(&v->vertex);
2406
2407 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2408
2409 pandecode_indent--;
2410 pandecode_log("};\n");
2411
2412 return sizeof(*v);
2413 }
2414
2415 static int
2416 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2417 const struct pandecode_mapped_memory *mem,
2418 mali_ptr payload, int job_no)
2419 {
2420 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2421
2422 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true);
2423
2424 pandecode_indices(t->prefix.indices, t->prefix.index_count, job_no);
2425 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2426
2427 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2428 pandecode_indent++;
2429
2430 pandecode_log(".prefix = ");
2431 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2432
2433 pandecode_log(".tiler = ");
2434 pandecode_tiler_only_bfr(&t->tiler, job_no);
2435
2436 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2437
2438 pandecode_indent--;
2439 pandecode_log("};\n");
2440
2441 return sizeof(*t);
2442 }
2443
2444 static int
2445 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2446 const struct pandecode_mapped_memory *mem,
2447 mali_ptr payload, int job_no)
2448 {
2449 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2450
2451 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false);
2452
2453 pandecode_indices(v->prefix.indices, v->prefix.index_count, job_no);
2454
2455 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2456 pandecode_indent++;
2457
2458 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2459 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2460
2461 bool instanced = v->instance_shift || v->instance_odd;
2462 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2463
2464 pandecode_log(".prefix = ");
2465 pandecode_vertex_tiler_prefix(&v->prefix, job_no, !instanced && is_graphics);
2466
2467 pandecode_gl_enables(v->gl_enables, h->job_type);
2468
2469 if (v->instance_shift || v->instance_odd) {
2470 pandecode_prop("instance_shift = 0x%d /* %d */",
2471 v->instance_shift, 1 << v->instance_shift);
2472 pandecode_prop("instance_odd = 0x%X /* %d */",
2473 v->instance_odd, (2 * v->instance_odd) + 1);
2474
2475 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2476 }
2477
2478 if (v->offset_start)
2479 pandecode_prop("offset_start = %d", v->offset_start);
2480
2481 if (v->zero5) {
2482 pandecode_msg("XXX: midgard payload zero tripped\n");
2483 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2484 }
2485
2486 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2487
2488 pandecode_indent--;
2489 pandecode_log("};\n");
2490
2491 return sizeof(*v);
2492 }
2493
2494 static int
2495 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2496 mali_ptr payload, int job_no,
2497 bool is_bifrost)
2498 {
2499 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2500
2501 bool fbd_dumped = false;
2502
2503 if (!is_bifrost && (s->framebuffer & FBD_TYPE) == MALI_SFBD) {
2504 /* Only SFBDs are understood, not MFBDs. We're speculating,
2505 * based on the versioning, kernel code, etc, that the
2506 * difference is between Single FrameBuffer Descriptor and
2507 * Multiple FrmaeBuffer Descriptor; the change apparently lines
2508 * up with multi-framebuffer support being added (T7xx onwards,
2509 * including Gxx). In any event, there's some field shuffling
2510 * that we haven't looked into yet. */
2511
2512 pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true);
2513 fbd_dumped = true;
2514 } else if ((s->framebuffer & FBD_TYPE) == MALI_MFBD) {
2515 /* We don't know if Bifrost supports SFBD's at all, since the
2516 * driver never uses them. And the format is different from
2517 * Midgard anyways, due to the tiler heap and scratchpad being
2518 * moved out into separate structures, so it's not clear what a
2519 * Bifrost SFBD would even look like without getting an actual
2520 * trace, which appears impossible.
2521 */
2522
2523 pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2524 fbd_dumped = true;
2525 }
2526
2527 uintptr_t p = (uintptr_t) s->framebuffer & FBD_MASK;
2528 pandecode_log("struct mali_payload_fragment payload_%"PRIx64"_%d = {\n", payload, job_no);
2529 pandecode_indent++;
2530
2531 /* See the comments by the macro definitions for mathematical context
2532 * on why this is so weird */
2533
2534 if (MALI_TILE_COORD_FLAGS(s->max_tile_coord) || MALI_TILE_COORD_FLAGS(s->min_tile_coord))
2535 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
2536
2537 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
2538 MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT,
2539 MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT);
2540
2541 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
2542 (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT,
2543 (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT);
2544
2545 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2546 * we have to fallback on offsets. */
2547
2548 const char *fbd_type = s->framebuffer & MALI_MFBD ? "MALI_MFBD" : "MALI_SFBD";
2549
2550 /* TODO: Decode */
2551 unsigned extra_flags = (s->framebuffer & ~FBD_MASK) & ~MALI_MFBD;
2552
2553 if (fbd_dumped)
2554 pandecode_prop("framebuffer = framebuffer_%d_p | %s | 0x%X", job_no,
2555 fbd_type, extra_flags);
2556 else
2557 pandecode_prop("framebuffer = %s | %s | 0x%X", pointer_as_memory_reference(p),
2558 fbd_type, extra_flags);
2559
2560 pandecode_indent--;
2561 pandecode_log("};\n");
2562
2563 return sizeof(*s);
2564 }
2565
2566 static int job_descriptor_number = 0;
2567
2568 int
2569 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost)
2570 {
2571 struct mali_job_descriptor_header *h;
2572
2573 int start_number = 0;
2574
2575 bool first = true;
2576 bool last_size;
2577
2578 do {
2579 struct pandecode_mapped_memory *mem =
2580 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2581
2582 void *payload;
2583
2584 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2585
2586 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2587 * high 32-bits of the 64-bit pointer are reused to store
2588 * something else.
2589 */
2590 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2591 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2592 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2593
2594 payload = pandecode_fetch_gpu_mem(mem, payload_ptr,
2595 MALI_PAYLOAD_SIZE);
2596
2597 int job_no = job_descriptor_number++;
2598
2599 if (first)
2600 start_number = job_no;
2601
2602 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2603 pandecode_indent++;
2604
2605 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2606
2607 /* Save for next job fixing */
2608 last_size = h->job_descriptor_size;
2609
2610 if (h->job_descriptor_size)
2611 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2612
2613 if (h->exception_status && h->exception_status != 0x1)
2614 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2615 h->exception_status,
2616 (h->exception_status >> 16) & 0xFFFF,
2617 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2618 h->exception_status & 0xFF);
2619
2620 if (h->first_incomplete_task)
2621 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2622
2623 if (h->fault_pointer)
2624 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2625
2626 if (h->job_barrier)
2627 pandecode_prop("job_barrier = %d", h->job_barrier);
2628
2629 pandecode_prop("job_index = %d", h->job_index);
2630
2631 if (h->unknown_flags)
2632 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2633
2634 if (h->job_dependency_index_1)
2635 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2636
2637 if (h->job_dependency_index_2)
2638 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2639
2640 pandecode_indent--;
2641 pandecode_log("};\n");
2642
2643 /* Do not touch the field yet -- decode the payload first, and
2644 * don't touch that either. This is essential for the uploads
2645 * to occur in sequence and therefore be dynamically allocated
2646 * correctly. Do note the size, however, for that related
2647 * reason. */
2648
2649 switch (h->job_type) {
2650 case JOB_TYPE_SET_VALUE: {
2651 struct mali_payload_set_value *s = payload;
2652 pandecode_log("struct mali_payload_set_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2653 pandecode_indent++;
2654 MEMORY_PROP(s, out);
2655 pandecode_prop("unknown = 0x%" PRIX64, s->unknown);
2656 pandecode_indent--;
2657 pandecode_log("};\n");
2658
2659 break;
2660 }
2661
2662 case JOB_TYPE_TILER:
2663 case JOB_TYPE_VERTEX:
2664 case JOB_TYPE_COMPUTE:
2665 if (bifrost) {
2666 if (h->job_type == JOB_TYPE_TILER)
2667 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no);
2668 else
2669 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no);
2670 } else
2671 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no);
2672
2673 break;
2674
2675 case JOB_TYPE_FRAGMENT:
2676 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost);
2677 break;
2678
2679 default:
2680 break;
2681 }
2682
2683 /* Handle linkage */
2684
2685 if (!first) {
2686 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2687
2688 if (last_size)
2689 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no);
2690 else
2691 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no);
2692 }
2693
2694 first = false;
2695
2696 } while ((jc_gpu_va = h->job_descriptor_size ? h->next_job_64 : h->next_job_32));
2697
2698 return start_number;
2699 }